The Voyager 1 spacecraft, which was designed for a five-year mission to observe Jupiter and Saturn, is about to reach a distance of one light-day from Earth after nearly 50 years in flight. According to NASA, the space probe will get to that marker on Nov. 18, 2026, at 5:16:07 a.m. EST, making it the most distant artificial object in history. The agency has now posted a countdown clock to mark the historic moment it becomes the first human-made object to reach that distance.

One light-day isn’t a span of time, but the distance it takes for light to travel in 24 hours. Since light travels at around 186,282 miles per second or 300,000 km per second, that’s a distance of about 16 billion miles or 26 billion kilometers. By comparison, the Moon is 238,855 miles (384,400 kilometers) from Earth, while our planet is 93 million miles (149.6 million kilometers) from the Sun.

Voyager 1’s estimated speed in relation to the Sun is 37,844.49 mph, which makes it pretty fast (but still not the fastest human-made object ever built — the Parker Solar Probe’s holds the record at 394,736 mph). Despite that, it will still take the spacecraft 49 years, 2 months, and 13 days to get to one light-day from Earth.

This is actually one of Voyager 1’s achievements, as NASA scientists have extended its lifespan by nearly ten times the length of its original mission timeline. It also hasn’t been easy for the space probe — in 2024, it suffered from a critical memory error, which engineers were able to fix after weeks of diagnosis. The distance has made repairing the spacecraft quite a challenge, as the latency at that time was already more than 22 hours, with round-trip communications taking at least 45 hours. Earlier this year, NASA had to use an emergency instrument shutdown procedure after an unexpected drop in power levels.

Experts estimate that the three radioisotope thermoelectric generators (RTGs) that power the spacecraft have enough juice in them to allow Voyager 1 to continue operating until the 2030s. So, unless it encounters a major issue, we can expect it to continue sending data back to Earth as it continues to travel through interstellar space. As the space probe continues its journey, its heading is set towards a star called AC +79 3888 or Gliese 445. This heavenly body is located within the Ursa Minor constellation, with the ship expected to pass within 1.7 light-years of the star in 17 light-years. But given that Voyager 1 only travels at 0.00564325% of the speed of light, we estimate that it will take the ship more than 300,000 years to reach its next destination.

The Air Force Research Lab (AFRL), the research and development arm of the US Air Force, has successfully tested Rocket Cargo capabilities, which would allow military units to resupply via cargo pods dropped from space. According to the AFRL’s X post, the cargo pods, which were onboard a SpaceX Starfall capsule, launched from the Cape Canaveral Space Force Station as part of the Bandwagon-5 “rideshare” mission carried by a Falcon 9 rocket. The capsule splashed down in the Pacific Ocean for this test, with the unit containing mission equipment like MREs, bottled water, a SATCOM-enabled tracker, spare batteries, and other equipment. The AFRL also fit-checked a UNI-PAC III life raft to see if it would fit inside the Starfall capsule for a potential test in the future.

The mission is a test of Rocket Cargo’s point-to-point delivery (P2D) capabilities, which should allow forces to resupply virtually anywhere in the world without needing to risk other assets like aircraft and aircrews, especially in areas where there is active combat.

“This was a critical achievement to demonstrate the potential of Rocket Cargo delivery for point-to-point rapid global mobility for the Future Force," AFRL Technology Transition Office director Adam Remaly said in a statement. "The cross-discipline collaborations and ability to unleash the speed of our defense industrial base are key to maximizing capability and winning the future.”

The Starfall capsule, which looks like a hockey puck, measures about 10 feet across and has a capacity of up to 2,200 pounds. It uses a carbon fiber heat shield to protect its contents during reentry and uses compressed gas thrusters for attitude control in flight. Finally, it deploys a parachute system to slow down the entire system for landing. While 2,200 pounds might already seem like a substantial capacity, the Air Force is reportedly hoping that it could increase the limit in the order of tens of tons, potentially allowing it to resupply ammunition, weapon systems, and even deploy vehicles from space. This would have to wait for the larger SpaceX Starship rockets, which are still in development, though.

Even though the latest test seemingly relies mostly on SpaceX technologies, it’s not the only company developing that capability for the U.S. military. Sierra Space, a subsidiary of the Sierra Nevada Corporation, is also working on the Ghost reentry vehicle and has already conducted test drops of the system from heights of 35,000 and 68,500 feet.

The capability to resupply troops across long distances without needing to rely on a long logistical tail is crucial for the U.S., especially as it faces a potential adversary in the Pacific. It’s also for this reason that the U.S. Navy has been experimenting with 3D-printing technologies, allowing it to print parts onboard carriers and at forward bases instead of waiting for shipments from across the ocean.

A new 'Reaper' jet-powered drone project just unofficially broke the previous drone speed world record of 465 mph by hitting 491 mph. Tomas Salvo is an aerospace engineering student at Georgia Tech, and his project was assembled in his dorm room. The drone is comprised of a custom-molded carbon fiber shell, laser-cut internal structure, and a 250N jet engine with a titanium 3D-printed nozzle, costing $13,000 to build. Tomas Salvo aims to win the $800,000 Boom Prize posted by Boom Supersonic CEO Blake Scholl.

Tomas Salvo’s Reaper project, which aimed to beat Niels Herbrich’s world record for the fastest jet-powered RC aircraft, went viral and caught the attention of Boom Supersonic CEO Blake Scholl, who sponsors the $800,000 Boom Prize ($750,000 cash, $50,000 Boom stock). The prize’s conditions are: First, pass Mach 1 for at least 5 seconds while climbing or level, not descending. Second, it must be an air-breathing aircraft with no on-board oxidizers or rockets for propulsion. Third, it must land intact, and lastly, the same aircraft (wings, fuselage, engine, and flight computer) must fly twice in one day. The participants must be individuals or teams, not corporations or people backed by venture capitalists or government funding; they want garage hobbyists.

We recently covered Mach Initiative’s Kingchaser, a 3D-printed jet-powered RC aircraft also aimed at beating Herbrich’s world record. They were beaten by Tomas Salvo, but Tomas’ run was not officially recorded by Guinness. The Kingchaser project was inspired by the Boom Prize; however, since they are not US citizens, they are not eligible to win it, but are still aiming to have an officially recorded Guinness World Record. The Reaper has a smaller engine than the Kingchaser at only 250N, compared to the Kingchaser’s 300N, but achieved a flight speed of 491 mph, beating the Kingchaser’s test flight of only 273 mph, as the latter did not have enough space to accelerate. However, Mach Initiative estimates a top speed of 550 mph if they had the space.

While the Reaper isn’t made out of PETG like the Kingchaser, it still utilizes some 3D printing, albeit in a much more expensive fashion — it has a titanium 3D-printed nozzle for its engine that replaces the stock nozzle, allowing the engine to sit further forward, improving the center of gravity without adding ballast on the front, weighing in at a total of 11 pounds. It’s made out of custom-molded carbon fiber on the outside with a laser-cut internal structure, all designed by Tomas Salvo and assembled in his dorm.

The Reaper was able to achieve its 491 mph speed while carrying a 1.2kg payload, with Tomas’ post further explaining that it could hold the speed for four minutes on a full tank, climb to 6,500 feet in under 20 seconds, and fly for a distance of 31 miles at full speed.

Before the Boom Prize, Tomas was able to land a few sponsorships that got him a JetCat 250N engine; however, we’re not sure how this fits with the Boom Prize’s rules barring corporate sponsorship. Since the Boom Prize's announcement, however, Tomas has resorted to crowdfunding, asking for $50,000 in August after raising $20,000, the source of which he did not disclose.

Both the Kingchaser and the Reaper are aiming at Herbrich’s world record, but they are still far from the 1,234.8 km/h or 767.3 mph Mach 1 speed that they need to surpass in order to win the Boom Prize. Mach Initiative responded to Blake Scholl’s X post about the Boom Prize, saying “challenge accepted,” but the rules may make them ineligible for the prize, making the Reaper a real contender for the $800,000, as it currently sits unrivaled by any other US project.

The U.S. Department of Justice (DOJ) just announced indictments resulting from Operation Ghostbusters, which investigated widespread student loan fraud. According to the agency, bad actors used stolen identities to sign up for student loans and used AI to complete the application, making it appear that the fake student was attending online classes.

While the schools receive the majority of the proceeds of the fraudulent student loan, the remaining amount is funneled into the bank accounts of the fraudsters, who then use mules to move the money from U.S. bank accounts to overseas conspirators. On the other hand, unwitting victims who have skipped college and instead directly joined the workforce after high school are saddled with student debt for a degree they never even signed up for.

Multiple community colleges and universities across the United States started suspecting fraud in the Federal Student Aid (FAFSA) program in 2021. Estimates suggest the U.S. government lost over $200 million because of this activity, and the U.S. Department of Education Office of the Inspector General referred the case to the FBI Richmond Field Office. Currently, more than 40 open fraud cases involve 26 FBI field offices and 20 U.S. Attorney’s offices. The DOJ called it Operation Ghostbusters, which has already resulted in the indictment of five individuals.

How much money are you willing to spend during the Black Friday sales?

“Today, we are announcing the first results of Operation Ghostbusters — a nationwide effort to stop a major fraud scheme targeting federal student aid, low-income Americans, and the American taxpayer,” Assistant Attorney General Colin M. McDonald of the Justice Department’s National Fraud Enforcement Division said in the statement.

“The Department is charging five defendants and announcing the seizure of more than $750,000 in cryptocurrency tied to this scheme through a network associated with Kenyan nationals. This is the first international seizure of its kind in Kenya. If you steal American identities, if you steal taxpayer dollars, if you exploit programs meant to help students build a future — we will find you. Whether you are in Nebraska, New Jersey, Virginia, or Nairobi, we will pursue you.”

While online fraud has existed since the birth of the internet, AI has let scammers become more sophisticated and target bigger payouts. One example is a scammer who used an AI-generated image of Brad Pitt to scam a 53-year-old woman out of $850,000. However, larger groups also use AI tools for their money-making schemes, such as North Korea, which has reportedly used AI to help its workers get jobs in Western companies and earn money to fund the nation’s coffers. The U.S. is taking steps to shut down student loan fraud by deploying automated fraud detection systems, but the added burden of identification and false positives is negatively affecting the country's most vulnerable students.

Darkbloom, a decentralized AI inference network that aggregates and sells spare compute capacity from Macs, told its members to expect revenues to fall as token prices continue to drop. According to The Register, the price per million tokens for Anthropic dropped by 77% in under five months, falling to $0.62 in October 2026 from $2.65 in May.

“Data centers have become cheaper. Output per NVIDIA cluster is up ~20x this year, and their margins are wide enough to cut prices 90 percent and still profit at 50 percent,” said Kydo, a researcher from Eigen Labs, the developer of the compute platform. “The market caught up to our price. We could go lower, but then most of you would earn even less [than] nothing.”

It’s estimated that someone running an M5 Pro Mac Mini with 64GB of Unified Memory could earn as much as $192 per month. This is the best-case scenario, though; the median earnings of the top 50 operators on Darkbloom’s leaderboard are about $159.86 per month. This translates to about $5.26 daily or $1,918.50 annually. However, if we account for Anthropic’s 77% token price drop, earnings will fall to a measly $36.77 per month, or about $1.23 per day.

AI token prices are falling so fast that they've even outpaced Moore’s Law, with some reports saying AI is becoming 50% cheaper every quarter and up to 13 times more affordable every year. News of companies blowing through their annual AI budgets in months has less to do with per-token price and more with the number of tokens, as agentic AI use outpaces token price drops. Aside from that, OpenRouter, Darkbloom’s primary client, has started prioritizing decode speed, quantization, and prefill speed, putting users without a top-of-the-line Max or Ultra Apple chip at a disadvantage.

Despite that, Darkbloom says it will not sit idly by and let some users lose out because of these changes. The network is analyzing model demand to advise members on which models are in higher demand right now. It’s also looking at other kinds of workload beyond inference, including distributed CI/CD jobs, improving network performance to better compete against other computer providers/aggregators, and finding other clients outside of OpenRouter.

In case you're not yet subscribed to our Tom's Hardware Premium subscription service, here's what you've been missing out on this week. We detail our thoughts on Anthropic's filings pre-IPO, Cadence's new suite of LLMs to power its EDA tools, how Nvidia is securing its entire supply chain to serve AI data center demand, how Vera CPUs are landing in the cloud, an update to the ongoing TP-Link saga, and more. Remember, to access any of the articles listed here, you'll need to arm yourself with a subscription.

We've been hard at work behind the scenes to determine exactly how locally-run AI models and workloads stress consumer-level components. While throughput and tokens per second are indeed important factors, our resident CPU expert Jake Roach has begun laying out his vision for how Agentic workloads tax CPUs, and how we're building benchmarks in this newfangled AI world that the PC hardware space finds itself increasingly entrenched in. It's not often that readers get the chance to look at our behind-the-scenes work, and this piece offers a look at some of the foundational exploration we're doing around how to define what an agentic AI CPU benchmark might look like in the future.

• Our first step at defining a CPU benchmark for agentic AI — how we can measure throughput of agentic AI workloads without any GPU involved

As the rise of Artificial Intelligence increasingly becomes attached to powerful, locally-run hardware, you might encounter terminology and nomenclature that you may otherwise not be familiar with. So, we've published a primer on all of the basic terminology that you need to know moving forward. Currently, you get breakdowns of how inference works, details on quantization, context, KV cache, and more. We're striving to update this page as our exploration with local AI grows, to serve as a handy bookmark. It's not a mere dictionary either, as we contextualize how each of these things relates to the larger overall model, how it might be served, or where you might encounter it when trying to run an AI model of your own.

• A primer on key AI LLM concepts — KV cache, weights, quantization, and how inference actually works

Nvidia's Vera CPUs were a salient part of a swath of Oracle announcements this week. The chip, usually found alongside Nvidia's NVL72 rack, is now generally available, but curiously, standalone Vera instances are also available on OCI superclusters. The CPU is specifically designed with Agentic AI workloads in mind, and it just so happens that we've benchmarked an Nvidia Vera server just this week to highlight the efficacy of how well Vera might perform under certain workloads.

• Standalone Nvidia Vera CPUs land in Oracle Cloud's multi-gigawatt deployment — our independent benchmarking of a local server reveals Vera's massive memory bandwidth lead over dual Xeon 6 server

Last week, we covered TP-Link's proposal to allow the router manufacturer to sell its Wi-Fi 8 products in the U.S. Now, according to a new report, it's alleged by the FCC that TP-Link's own reports contain "false or incorrect information in its applications" related to equipment authorizations for four different models of Wi-Fi 8 routers that the company hopes to sell stateside. As the saga continues to develop, we attained a statement from TP-Link on the subject, as the situation continues to develop.

• FCC rebuffs TP-Link again as it attempts to sell Wi-Fi 8 gear in the US — FCC alleges inaccurate documentation and out-of-spec power levels for routers

Anthropic's large and sprawling S1 filing made headlines this week as the company prepares for an IPO. Within the document, it details many (but not all) of the numbers driving one of the world's foremost AI labs. Once you get past the mind-bogglingly high $800 billion in GPU and TPU contracts, the company has shown that most of its compute capacity isn't actually served to customers, and is instead dedicated to R&D for new models, features, and testing. There's a huge amount of information to sink your teeth into here, and despite what many might believe, the company's revenue per compute dollar spent is actually in Anthropic's favor, shattering the myth that tokens cost more in compute than they do for a user.

• Anthropic’s IPO filing shows most of its compute goes to R&D, not serving customers — company details massive $800 billion in TPU and GPU contracts

As neoclouds and hyperscalers (who rent compute out to companies like Anthropic) scale up and scale out, domains become exceedingly important. With copper materials coming up to their natural limits, eyes have turned to the photonic domain to alleviate bandwidth bottlenecks. Now, PsiQuantum has entered the domain, showcasing a prototype photonic switch that loses less light than other devices on the market, and the company expects to develop and mature its solution and is likely to enter the photonic switch market.

• PsiQuantum turns its photonics know-how into an optical switch for supercharging AI data center GPU networks — 8x8 lab prototype loses less light than rivals, with plans to develop larger devices

Cadence has revealed Maestro, a suite of tuned LLM's, designed specifically to work in EDA flows using the company's in-house suite of agentic tools. We've covered the rise of agentic EDA tools closely, and Cadence's Maestro suite appears to be the company's own shot across the bow in helping its customers automate elements of their own chip designs. In initial results compared with general-purpose AI models, Cadence claims up to a 30% improvement in the quality of results and that Maestro effectively uses fewer tokens for the same task, with up to 5x the token efficiency compared to off-the-shelf AI models.

• Cadence reveals Maestro, fine-tuned AI models built to run specialized chip design agents, with early access coming for 'select customers' in Q4 2026 — claims up to 30% higher quality results and up to 5x token efficiency over a general-purpose AI model

But all of those new AI chips have to come from somewhere. With the ongoing, ceaseless AI buildout, firms are looking to secure their supply chains, with Nvidia one of the chief companies among them. As demand continues to soar, the very real limitation of fabs is particularly highlighted in the realm of the substrates used to create PCBs. Ajinomoto Build-up Film, or ABF, is one of the hottest commodities in the chipmaking industry as a result, and Nvidia has announced plans to secure the material from Ibiden, in addition to other critical materials used in chip packaging, testing, and more. Of course, Nvidia isn't doing this out of the goodness of its own heart; it's doing it in exchange for reserved production capacity.

• Nvidia is bankrolling AI accelerator supply chain buildout to reserve production capacity — announces billion-dollar deals with six major suppliers across substrates, advanced packaging, semiconductor testing, and optical engines

The ramifications of AI are still being felt across the globe, but if you've been on social media over the past several weeks, you may have found that there's one particular trend: Decompilation and recompilation of video games, recompiled to work natively on PCs. We detail how additional software tools like Reverse Engineer Anything and Ghidra are powering the boom.

• Why native PC conversions of console games are suddenly everywhere — Reverse Engineer Anything gives AI agents a unified toolkit, and byte-for-byte checks keep hallucinations in line

To read any of the articles we mention above, as well as unlock Bench, you'll need to purchase a Tom's Hardware Premium subscription. Your support goes a long way in helping us to publish deeper, more insightful articles that might not otherwise be covered.

Reverse-engineering games for converting them between systems is currently a hot topic, and it's getting messier by the day. According to multiple user claims, Anthropic's Claude Code and/or Opus 5.5 model are blocking attempts at innocuous reverse-engineering, including but not limited to those using the popular Reverse Engineer Anything (REA) tool. That issue may be linked to a text rule that's now in force, which allegedly instructs the model to avoid "[recreating] a specific existing game." The updated rule appears to be a move to limit a flood of AI-cracked and cloned titles, with some game devs even recently being hit by extortion scams that threaten to clone their game demos into complete games before they have a chance to finish and release the original.

According to Twitter user Chasm, the game-related textual rule is "Do build original games, games in a genre, and games from the user’s own designs and assets. Don’t build a recreation of a specific existing game, even if it is renamed or called original." There is no official or additional confirmation if this is the actual rule text, but it's plausible enough given the very clear refusals that devs have been encountering over the past week.

This situation is tricky to analyze due to the varying types of reports. First, the existence of a game-related rule is real, as Claude Code's GitHub is seeing an influx of issues complaining about refusals to do standard game work halfway through development, when replicating a game mechanic, and patching a tabletop-like game. It's possible that some of those reports have been made by users with nefarious purposes, but at least one is a bona fide developer with a real track record.

As for reverse engineering, there had already been a smattering of user reports across Reddit and other platforms, wherein Claude refused to perform tasks like (allegedly) analyzing a simple Windows crash dump, clearing up junkware on a Samsung TV, or just using REA altogether.

The simultaneous timing of both game- and cyber-security restrictions may be a coincidence, and yet there are complaints from purported game developers saying that Claude blocked their own games' anti-cheat development and security reviews, under the justification that its safeguards were triggered.

Anthropic introduced its Cyber Verification Program on October 6, essentially saying that access to its frontier models' advanced cybersecurity capabilities would be limited to entities covered by Project Glasswing, taking effect immediately.

It's possible that whichever ruleset was applied is the same one that's causing the cybersecurity-related refusals. Still, it's odd to see that 24 days have passed since then, and yet reports of bot refusal are only coming in since the last week or so.

As for possible motivations behind these restrictions, they are twofold. The first and most obvious one is that Anthropic — and, by extension, other AI players — are preemptively protecting themselves from potential lawsuits from software companies, including game studios, for offering the means to reverse-engineer their software.

There is no official statement on these matters from Anthropic or other AI companies so far. Anthropic isn't the only one playing this game, too, as recently OpenAI deactivated the account of another person trying to clean up their TV's software.

The recent rash of AI-powered clones of application suites like Adobe's suite and Microsoft Office 365, as well as the countless recomps and decomps of console games, likely has most every software company on earth rethinking their life choices. And probably also in frequent contact with their legal teams.

The online reaction was swift, suggesting that users looking to perform reverse-engineering work simply use Chinese, open-weight, or local LLMs instead of Claude or GPT. This, at the same time, highlights the irony of frontier AI companies that copied the world's information and art for their training sets and are now concerned with copyright infringement.

And yet there are situations that merit some discussion, like the recent controversy about the Steam Next Fest being chock-full of AI slop copied from existing titles. That particular situation is getting particularly nasty, as game developers are reportedly being extorted under threat that their game demos will be cloned into complete games before they have a chance to finish and release the original.

Adding insult to injury, this is seemingly being done to small studios whose main priority is usually "having food," rather than big publishers looking to microtransaction everything into a new yacht for shareholders.

Exin Therapeutics, a biotech startup focusing on next-generation neurotherapeutics, claims it has successfully conducted brain surgery on a mouse using an AI agent. Its co-founder and CEO, Gabriel Ocana-Santero, shared the entire procedure on X, showing how the AI agent chose the gene therapy vector and dosage and the brain region where it would be applied. The AI agent also drove the robotic arms that injected the medication into the mouse’s brain. Thankfully, the operation seems to have been a success, with the mouse recovering well and recorded moving around in its enclosure.

“Autonomous surgical robots exist and they are better at the mechanics than our bench rig, [to be honest]. But in every one of them, a human decided what operation to perform, and the robot executed it,” Ocana-Santero said. “Nobody told this agent what to give or where to put it. It got a phenotype and a shelf of candidates and had to work out the therapy and the coordinates itself. I am not aware of an agent previously choosing a treatment and a brain target and then delivering it into a living mammal.”

The team used Anthropic Sonnet, which received a mouse model of epilepsy, control of a motorized surgery rig, and 19 gene therapy vectors in a spreadsheet. Aside from that, it also had access to three research papers, allowing it to choose the brain region where it should apply the treatment. Nevertheless, a licensed veterinarian was purportedly overseeing the entire operation, with every irreversible step requiring approval before the AI agent could proceed. Aside from that, the entire system also had an emergency stop button that would cut the power to every motor in the surgery rig.

The only thing that the vet surgeon on duty did was to do the craniotomy procedure to expose the brain, as the AI agent wasn’t allowed to do that yet for safety reasons. There were also several safeguards that the AI agents were banned from overriding, which even the human operators couldn’t bypass, necessitating a manual fix to the agent’s code during the operation.

What’s interesting, though, is that the startup only used Anthropic’s mid-range Sonnet model for the operation. It initially tried using Opus 5.5, but it immediately refused to do any work after it detected that it was working on a biological experiment. The older Opus 5 model was able to run the entire procedure on paper but refused to do live surgery. “The block isn't the rig and isn't the inputs. It's me. My own output was stopped when I started writing the plan for this surgery... if I try, it gets cut off the same way, and the worst version of that is me cut off partway through a live procedure with a needle somewhere in the brain,” the LLM reportedly told the Exin Therapeutics team. “Starting something I can't finish on an anesthetized animal is worse than not starting.”

Despite the seeming success of the operation, some people are concerned about the ethics of such a procedure. While AI LLMs have been useful in medical research, adding them to sinus surgery has seen malfunctions skyrocket by more than 12 times. Aside from that, relying on an AI to make every decision in such a risky procedure puts accountability in question, especially if something goes wrong. We also haven’t even touched the long-standing ethical question of animal experimentation in the medical setting. Still, some users applauded the experiment, with one comment saying that they’re excited about the possibilities of these integrated technologies, as their son has been suffering from drug-resistant epilepsy for 19 years.

OpenAI CEO Sam Altman has data centers full of memory, but seemingly it's the 32GB of RAM in your desktop that's the problem. In an interview over beers with the YouTube channel Cheeky Pint, Altman said that local inference, and, by extension, people having DRAM in computers at home, is "quite terrible."

When discussing local inference on home computers with open-source models, Altman discussed the assumption that people will be able to run local AI on their own computers, the ones they may already own or want to build. Then he explained why that's a bad idea (at the 29-minute mark):

"From a global allocation perspective, actually, it's quite terrible to have everybody have a bunch of the world's limited supply of DRAM sitting in their own computer and being used 5% of the time," Altman said. "You want maximum utilization in the cloud if you're trying to deal with this very supply-constrained environment. And somehow many people's intuition is the opposite of that."

While OpenAI does have some open-source models, its primary business is in closed-source models that run in data centers, so he's really defending the company's business model.

OpenAI isn't exactly threatened by some laptops and gaming rigs just yet, but local AI could mean people one day paying less for subscriptions to models like GPT-6 Luna if they have the power to run them at home. Altman's talk suggests that the company would prefer that more RAM go to OpenAI's data centers rather than home computers, which also sparks enthusiast fears of not being able to own your own hardware or software, but rather rent an agent and a computer in the cloud.

Some of what Altman is saying is based on an idea that data center buildouts won't slow down. It also assumes that RAM being used in AI data centers 24/7 is more worthwhile than RAM that is used for spreadsheets, web browsing, gaming, and maybe even local AI occasionally.

The rest of the wide-ranging interview includes talk of this week's PHP vulnerabilities, sandboxing models in training, and OpenAI's research in mathematics.

PsiQuantum, a quantum computing company, announced in a release earlier this week that it has turned its low-loss silicon photonics into an optical switch for use in AI data centers. Every optical switch typically comes with loss of light: the more light a switch receives, the less reaches the other side, with the amount of light lost measured in decibels. Typical data center optics can tolerate half the light (3 dB) being lost. PsiQuantum's switch debuted with a demo that typically loses only 1.2 dB of signal, or about a quarter of the light in lab tests, and the all-optical switch can achieve it without an amplifier. The prototype is, however, comparatively small, with eight inputs and outputs (8x8) against competitors’ 32 to 300, but the company says larger devices are in development.

The optical circuit switch (OCS) passes light straight through, from fiber to fiber, without converting it to or from electrical signals. Regular data center switches make that conversion at every hop, which takes power and adds latency. An OCS, instead, uses a fixed light path that it maintains until a change is ordered. Its main penalty comes down to how much light is lost when passing through the switch. PsiQuantum’s switch, which uses circuits etched on a silicon chip with no moving parts, unlike Google’s and Lumentum’s tiny mirrors, or microelectromechanical systems (MEMS), would be used alongside regular packet switches.

Skipping those conversions matters more as AI clusters grow, because moving data between GPUs is quickly becoming a bottleneck. Optics are a high-bandwidth improvement over copper, which has limitations in reach, although a shortage of lasers is a problem for the alternative. Improving communication bandwidth and distance is the goal, so that different racks of GPUs can talk at speed and act as one.

Switches similar to PsiQuantum’s made by Salience Labs and iPronics use built-in amplification and, at least in the former case, are also rated at under 2 dB, or under 37% of the light lost. Google’s mirror-based switch, Palomar, can also typically stay under 2 dB and has 136 ports. Salience’s switch sits at 32 and Lumentum’s at 300, which leaves the eight-port prototype a ways behind on how many fibers it can handle.

PsiQuantum's switch is an 8x8, meaning eight in and eight out, and “every input can reach every output.” It’s also “strictly non-blocking,” so a new connection never disturbs ones already in place. In the lab, it measured a typical 1.2 dB insertion loss, 40 dB of crosstalk (about 0.01% of the light leaking to other ports), and under 3 ms of reconfiguration time, all of which exceed “initial targets set for these devices.” The loss figure is measured fiber to fiber, which includes getting light onto and off the chip but not connector losses, PsiQuantum told Tom’s Hardware Premium.

The platform loses very little light at each component, which matters because a quantum computer can tolerate only so much lost light. PsiQuantum’s 2025 Nature paper reports silicon nitride waveguides losing as little as 0.5 dB per meter (about 11% of the light per meter), splitters 0.5 millidecibels, and crossings 1.2 millidecibels, the last two a few hundredths of a percent of the light or less each.

The company also says chip-to-fiber coupling has reached about 60 millidecibels, roughly 1.4% of the light lost, in its best published lab results, although without a full fiber array attached. Those figures set a best case, not a worst case, for a 64-port switch, since light crosses more components as ports are added and the losses add up.

How much loss is acceptable varies with the optics at the end of each link. Coherent, an optics maker, says that using an OCS can add “up to 3 dB additional insertion loss,” but its transceivers are able to work with that. Lumentum told Gazettabyte that 3 dB is good enough to work with off-the-shelf FR4, DR4, and DR8 optical modules using the switch; those are the standard plug-in options used to link GPU servers and switches, at distances of 500 meters to 2 kilometers. PsiQuantum’s 8x8 is built for the O-band, the wavelength range those modules use. PsiQuantum told Semafor that, to be workable, its switch needed insertion loss under 2 to 3 dB.

The 8x8 prototype was designed for conventional data center traffic and comes in well below the 3 dB loss target in a request from Nvidia for proposals, the company says. The port count marks it as a demonstration vehicle, but PsiQuantum has a 64x64 version taped out with a 72x72 underway and is “actively developing larger devices.” With crossings that lose so little light, the scale-up might hold. PsiQuantum declined to share details on future products, though it says the loss, which may be slightly higher, will stay very low.

Speed remains less important than insertion loss for how OCSes are currently used. Google, which has the most deployments, numbering in the “tens of thousands,” uses its switches to resize TPU pods or to route around failed chips, a job its “millisecond-scale” MEMS switches handle. PsiQuantum’s under 3 ms is within that class, but Salience and iPronics claim less than 300 microseconds for their amplified chip switches, which iPronics highlights for mid-workload network rewiring to hide the switch time while chips compute.

There is quite a lot of room for speed to improve, and PsiQuantum already has a material for that, used in an existing 64-port switch for its quantum systems, which need to switch light at gigahertz rates. The material, barium titanate (BTO), changes how it bends light when a voltage is applied, and the company chose it over lithium niobate, a more common selection, because BTO had the strongest effect it could find. This keeps switching elements short and losses low at speeds roughly a million times faster than the demo. BTO is “not required for the demonstration vehicle or initial products,” which use a thermally switched silicon nitride platform, but it is “an additional path to faster, more advanced future products.”

An argument for starting now is that building thousands of such switches for conventional applications would reduce the risk for the packaging, assembly, and supply chain of parts PsiQuantum builds for quantum computing. AI and quantum computing are “complementary technologies,” it says, and this is the chance to put its technology into action. When PsiQuantum hired Rob Soderbery as an executive vice president in July, it said he would advance its silicon photonics platform, “accelerating its application across AI networking and other next-generation computing infrastructure,” hinting at the roadmap. The company’s CHIPS award was finalized in September for “high-performance optical switches,” among other components.

Co-founder Pete Shadbolt praised Google’s alternative MEMS switch but would prefer to have similar-quality switching on a monolithic chip that needs no moving parts. Nvidia has a stake through its venture arm, NVentures, and collaborates with PsiQuantum on “GPU-QPU integration,” that is, graphics and quantum processing units, as well as silicon photonics. Nvidia also joined rival iPronics’ $125 million round in September to develop its second-generation optical switch.

Right now, inside Google, OCSes rewire AI clusters, and research firm Cignal AI expects most deployments to stay there through the end of the decade; outside of Google, OCS is still at “proof-of-concept and early trial.” The firm raised its market estimates earlier this year, though, roughly threefold for 2026 and by more than 40% for 2029. An opening exists for PsiQuantum in GPU networks, which ties into Nvidia’s interest, though adoption there is expected to be limited until then. The company says samples and testing kits have already gone to several computing companies, including GPU vendors and hyperscalers.

PsiQuantum should have commercial products “announced in the coming months,” along with updates on larger devices, while competitor Salience has already demonstrated 64 ports at this year’s European Conference on Optical Communication. Salience’s roadmap already has 128- and 256-port products. Scaling is always an issue, but if PsiQuantum’s figures, which haven’t been independently verified, hold up at 64 ports without amplifiers, then larger monolithic OCSes are on the way.

As hardware for running local AI models continues to move to the forefront, such as Nvidia's RTX Spark, a new set of acronyms, definitions, and terminology is popping up. Given the pace of releases, we've put together a list of AI's new nomenclature to familiarize you with new concepts and pieces of terminology that you'll likely see in Tom's Hardware Premium articles.

This primer should get you started on the world of AI and be handy the next time news on the topic comes around, and possibly when you're downloading some models to try out and tweak for yourself. For the context of this article, we're going to stick with notions related to text LLMs, or in common parlance, chat, programming, and most agentic bots.

Model

This is probably the easiest one: a model is the LLM itself (a Large Language Model), computed from training sets of mostly text data on the scale of terabytes to petabytes. That training information comes from the internet, books, video and audio transcripts, and other sources.

The model is roughly analogous to a set of highly advanced layers of statistical analysis of relationships between tokens — words or parts of words, computer code, punctuation, and other symbols. The tokens themselves aren't part of the model, but they are packaged along with it. An AI model generates tokens in its outputs.

Weights

Those computing layers within an AI model would be useless without corresponding weights, which represent how strong the connection is between sections of tokens, sort of like how a pathway between neurons is more traveled than others. When the model runs, these values end up establishing how tokens end up relating to each other.

Example: the weights of the connections between "na" and each next part for "naval", "nachos", or "naruto" will all be different, and the higher the weight, the more often that word appears in training data. Importantly, weights cannot change without re-training the model, or parts of it.

When you see LLM specs online, you'll see something like "4B", "8B+", "20B", and so on and so forth. That shows how many parameters, or weights, the model uses, and it's generally measured in billions for consumer-grade models, and trillions for frontier models like GPT, Claude, and Kimi.

Note that a model having more weights than another doesn't automatically mean it's smarter. Just like with CPU clock speeds, it's not just about getting a higher number — bigger is usually better, yes, but you want those weights to be meaningful for the output.

You may also have heard about "open-weight" models versus "closed" models. Those refer to models like Llama, DeepSeek, and Kimi, which are effectively free-to-use, since the entities behind them freely publish the weights online. You can tweak the weights to your heart's content, if you wish, and come up with your own spin on things, as long as you have the hardware to run it. Note that downloading a free model doesn't reveal anything about how it was trained or how the weights were calculated — think downloading a free customizable application, not open-source software.

Data types and quantization

FP16, INT8, FP4, all sound like acronyms in some military handbook, but they're simple to interpret. All programs use numbers of varying width and precision, depending on their purpose: a simple money counter in your favorite game might be fine with a 32-bit integer that can represent 0 to 2^32 (INT32), while a value held in memory for a highly precise scientific measurement may need floating-point 64-bit arithmetic (FP64).

As memory is an exceedingly finite resource, good developers try to use the smallest units they can get away with, a notion that's particularly important in the VRAM-limited world of GPUs. Predictably, these chips became particularly adept at handling lower-precision values as a tradeoff between processing speed and numerical accuracy. A particular datatype isn't better or worse, as that assessment will depend on what it's used for.

When training an AI model, the labs will want to use as high a precision as they can for the weights, but keeping that in the final version would make for a prohibitively large and slow LLM. That's where the concept of quantization comes in, which is a fancy word for "clever rounding." As one of the steps after training, several passes are made to transform highly precise weights. Each AI lab has its own secret sauce about how to do this, as it's a critical step.

Frontier models predominantly use relatively high precision for their weights (BF16), in order to maintain response quality, nuance, and reasoning ability (an analogy would be deciding how accurate you want a GPS location to be — miles or inches). Open-weight models often have several versions with different quantization levels depending on what your hardware can bear — for example, Llama 3.3 70B can be run in full BF16 or quantized INT4 flavors.

The lower the precision, the more likely it can be run on consumer hardware. AI models listed on Hugging Face may have several quantization levels to fit certain levels of hardware, trading response accuracy for size. Consumer-grade models use smaller data types so they can fit on your Mac Studio, RTX Spark, or graphics card.

Context window

"Context window" is another term you've heard a lot and probably seen firsthand, likely in Claude Code's UI as a broad indicator of how long your session can go before the bot starts acting a little dumb. It's measured in tokens, with one million being a common number for frontier models, and it's almost directly analogous to RAM in a computer.

The context window is the LLM's thinking space, and it needs to contain everything associated with your session, not just the current question. Along with all your questions and its replies, it contains system-level prompts, all files you uploaded, the content of links it looked up, plus any associated data.

Crucially, while for you all these things are distinct concepts, for the LLM, they get reduced to tokens in one large pile: the context window. This is also why separating "data" from "code" is technically impossible — the best that modern models do is an approximation of that distinction.

LLMs don't think — when answering a new question, they re-run a statistics program over all the tokens and predict the next handful, so having a larger context window is beneficial (at the cost of VRAM), hence the one-million figure that's currently popular. But even statistical algorithms have their limits.

Even if you increase the window to, say, three million tokens, you may find that your LLM's prediction starts breaking down even though it hasn't run out. This is amusingly called the "lost in the middle" effect, where the bot will pay attention to the start and end of the conversation but will forget things down the middle. Thus, each model is finely tuned to the size of its context window.

KV Cache

If you've looked into using LLMs locally, or by way of running a remote GPU, you may have heard about the "KV cache" time and again, and that you need to reserve VRAM for it, or store it on a fast SSD if you absolutely must. This is a set of data that's created when the model runs, but that's actually reusable between queries in the same session.

The data that's saved and reused can be described as "the relations between tokens for this session." For example, the ever-popular sentence "you're right to push back" can be tokenized as "you-'re-ri-ght-to-pu-sh-back". The dashes between each token represent a vector (or array), called KV for key-value, that is used for determining which token gets picked next.

Since the text generation is linear, the math for calculating each new token needs every KV preceding it. Thus, if you're asking repeated questions in the same session, you'll want the KVs reused to avoid the expense of recalculating them over and over again not just with each round of conversation, but with each additional token.

Remember, an LLM doesn't simply answer the question you just posted — it always analyzes the entire conversation from the top to eventually generate tokens for your response. In a long conversation with 300,000 tokens, not having the cache would mean recalculating the KVs for all of them, a slow operation that gets disproportionately slower still as the interactions grow. This excellent blog post has a more detailed example.

Remember quantization? That's also applied here — modern models reduce the precision of the KV cache's content compared to their initial values, trading a minor accuracy loss for all new tokens in the session, in exchange for keeping VRAM usage down. In fact, with large context windows and conversations, the KV cache may well eat more space than the model and weights themselves.

Mixture-of-Experts (MoE)

Previously, AI firms touted how many more parameters their newest model had. Now, the race seems to have shifted to who can include more experts in any given model. That description may conjure a mental image of many tiny scientists living inside a knowledge laboratory, each handling a specific topic, but that's not quite how it works.

Mixture-of-experts, or MoE, is essentially a set of sub-models within the main model. When training starts, all sub-models get the same dataset, but the choice of weights is randomized ever so slightly for each one. Each sub-model will value tokens a bit differently from the rest, and due to how the training process reinforces patterns (called the "rich-get-richer" effect), related tokens will tend to converge together, so you end up with "experts" on sets of tokens.

A common misconception is to think that these experts are related to topics (like programming), but they're only establishing relations between tokens, meaning they'll be more accurate for certain words and text patterns, but devoid of any concept of what they're about. The fact that an expert might offer pretty good responses in regard to landscape painting is a happy accident, not a feature.

You may be thinking that this makes the models larger, and that maybe just the one generalist expert may prove a wiser choice. That's correct, but the trick is that the models only activate certain experts, meaning that only a portion of the model is actually running, thus at higher speed than a generic LLM.

This model selection happens on a per-token basis, too, as a router re-selects experts for each incoming token. For visualizing this, think of brain scan patterns where different areas light up depending on words heard. In practical terms, it means the model can answer more accurately in certain sections, while answering the question faster — albeit at the cost of higher VRAM usage.

Retrieval-Augmented Generation (RAG)

RAG is a concept that's pretty easy to describe, and yet one of the most powerful ones when used with any given model. In short, it can be described as: your chat application runs a few Google searches, feeds them to the model, and voila, there's now much more and fresher information to answer your question. This additional information can come in most any form: company documents, reports, emails, or spreadsheets — anything that you might supply to the model yourself, directly tokenized into its context window.

While broadly speaking most contemporary LLM chat applications run searches, activating RAG, usually triggered by selecting a "web search" switch or sometimes enabling "thinking" or "high effort" mode, will explicitly instruct the application to find as much data as it believes necessary to answer a query. Given that pure context is often key to answering many questions, and do so in an informational manner, RAG is among the most powerful tools available.

Fun fact: the search queries the application will search for are usually generated by a smaller, separate model that will, for example, say what web searches are bound to give the best results, or what to look for in a linked Office 365 account.

The main drawback is that it can be extremely computing-intensive, as the text for your questions and responses may consume, say, 10,000 tokens, yet the information in the retrieved webpages can take hundreds of thousands of tokens all on their own. For example, the Perplexity service was built with an emphasis on RAG usage, and judging from personal experience and user reports, it's usually much more competent and context-aware when answering questions, even though it's underpinned by the same models as everyone else.

Inference, Prefill and Decode

Inference is the act of using a trained model to generate outputs or tokens. It happens every time you send a prompt, and it's where all of the elements that we've discussed above come into play. Inference can be split into two primary parts: Prefill and Decode.

Prefill is the first phase of inference. Essentially, all of the input history (or your prompt) is tokenized and then processed by an AI model. This can include documents, code, and more. These inputs also dictate what gets put into your KV cache. Prefill is processed all at once in parallel, which often means that prefill times are often faster than overall outputs.

Prefill is a large part of the overall workload when measuring how long it takes for a specific chip or piece of hardware can take to output its first token, also known as "time to first token" (TTFT), a common AI benchmark, usually reported alongside prefill tokens per second. Longer prompts take longer to prefill, and context, KV cache, and everything else we've discussed are taken into account.

Decode is the second phase of inference. After a prefill has processed your input and has built the KV cache, the AI model then generates its response one token at a time, which must be done sequentially in order to determine the ones that come after, which means that the step cannot be performed in parallel, unlike prefill. Decode is usually also dictated by context, and reads from a model's KV cache during this step. Decode is usually limited by memory bandwidth, instead of raw compute, as streaming single tokens from weights and reading the KV cache is a bandwidth-intensive task, meaning memory speeds matter more. You'll usually see Decode speeds referred to as tokens-per-second (TPS).

Speculative Decode is a technique that makes those single tokens in the decode step stream faster, without changing the overall output. A drafting mechanism guesses several upcoming tokens, and then the model attempts to verify them in a parallel pass (almost like an additional prefill step). The model then verifies the pass itself; if the guesses are good, then the model's outputs are indeed correct, you get tokens for less of a computational cost during a given decode pass, and when they're bad, they are rejected by the main model during verification, since the main model's output is always what's kept.

Multi-token prediction (MTP) attempts to predict the next set of tokens at once, rather than only serving the next token. This is often built into AI models when they are trained with multiple steps. This can serve as the built-in drafting mechanism required for speculative decode.

This is a non-exhaustive list of concepts critical to understanding how AI models and LLMs work, but it is by no means complete. We will continue to update and refine this article as we add more.

Some 700 Micron employees in Taiwan will receive five years’ worth of salary in bonuses as the company announced three bonuses in a single payout. According to Digitimes, this accounts for about 17% of the over 4,300 production-line employees working at Micron Taiwan. The remaining staff will not come out empty-handed, though, as they will receive three bonuses in the month of October. The news comes as Micron continues to battle worker unrest and unions fighting for increased pay.

The lump-sum payout consists of three incentives — an annual performance bonus that’s estimated to be between 7.5 and 12.5 months of salary for the majority of workers, a mass-production bonus amounting to 20% of their annual salary, plus a one-time $31,000 (NT$1,000,000) appreciation bonus which will be prorated for employees who haven’t reached a full year of service. Micron workers across the globe also expect a new quarterly profit bonus which will be paid 50% in cash and 50% in stock. This bonus will start in January 2027, with the first one to be paid 100% in cash.

Micron is following in the footsteps of Samsung and SK hynix, which have recently paid out massive bonuses to their workers of up to $400,000. This is a direct effect of the AI-driven memory chip boom, which saw the big three memory chip manufacturers — Micron, Samsung, and SK hynix — raking in massive orders from AI hyperscalers. The record profits that these companies are drawing from the appetite for AI data centers have led workers on production lines to demand a slice of the pie. Samsung workers voted to accept a $340,000 average bonus in May 2026, while SK hynix workers are set to receive $477,000 in bonuses this year, which could potentially grow to almost $900,000 in 2027.

Despite this seemingly massive payout, the workers’ union at Micron Taoyuan is still in the midst of a dispute with management regarding their pay. Some officials of the group have already launched a one-day strike as they call for clearer profit-sharing arrangements and terms for bonuses, with nearly 70% of members at the Taoyuan site turning down management’s latest offer. The two sides are tentatively set to meet on November 6 again for further negotiations. On the other hand, the union at Micron Taichung has already agreed to a 10% employer’s contribution towards worker pensions and a tripling of their day-shift allowance.

These bonuses allow workers to get a share of the massive profits that the chipmakers are raking in at the moment, allowing them to stay competitive for chip talent. However, their reliance on performance and profit bonuses also means that these incentives could fall precipitously if AI demand falls and the demand cycle returns to normal.

In a world where youth unemployment is growing as AI replaces more and more entry-level jobs, Cisco president and chief product officer Jeetu Patel is going for a different approach. According to a report by Fortune, he's given a 27-year-old ex-startup founder — brought into the company by an earlier acquisition, and previously languishing in a role five levels below — a new job to refactor 98% of the company's legacy code within 24 months.

The employee, Padmanabhan Krishnamurthy, joined Cisco after it acquired EzDubs, a translation startup that Krishnamurthy helped build in 2023. While having executive experience, Patel found that "Paddy" was left "buried away in the organization doing something unstrategic."

According to the report, Patel spoke to Krishnamurthy in Cisco's office in San Francisco, where Krishnamurthy told him that he could rewrite Cisco's legacy code in a team of just two people, and that "even the most AI-forward people on [Patel's] own team weren't thinking that way."

As a result, he's now been tasked with rewriting 98% of Cisco's legacy code within 24 months, with four weeks as a proof of concept. "I've given him unlimited tokens because he's going to refactor code," according to Patel, telling him "you can go nuts."

AI coding has skyrocketed in recent months, with GitHub forced to deploy 4 million extra CPU cores and 160 petabytes of storage to cope. However, the practice is not without its pitfalls, with rogue AI agents notorious for taking unsolicited actions including deleting entire company databases, TV show edits, and production setups.

Patel hopes that there are more young hires in the Krishnamurthy mould that he can find, both in the company and externally to hire. He's looking for hires to "challenge the system and the status quo," to be "an energizer bunny that just goes and kind of revs everyone up."

Patel believes that current corporate structures work against this kind of thinking. "If you don't hire the early-in-career [workers], and you just let your workforce continue the way that they were, where do fresh ideas get injected? How do you challenge the status quo?"

He acknowledges that this will create friction, putting Gen Z employees among decades-old employees. "Tension is good.... I actually try to create tension in the teams because you want to have enough friction on the ideas. You don't want to have friction with the people and personalities."

Patel believes that a combination of experienced, scale operators at Cisco, alongside external hires and "entrepreneurial founders, CEOs and engineers who are AI-native" will bring results. "Either of these groups by themselves aren't nearly as potent as when all three work together. It's what creates magic."

A criminal investigation into the environmental damage caused by Google’s Finnish subsidiary has been launched. Tuike Finland Oy is being investigated after five complaints were received regarding the Muhos data center project, according to an official police statement published today [machine translation].

The police say that the investigation is backed by Finland’s Licensing and Supervision Authority (LVV). Google/Tuike Finland Oy is accused of suspected breaches of environmental law and environmental damage. Other lesser criminal offenses related to ambiguities and shortcomings in the drainage and water management are also being investigated.

Since Google representatives have previously publicly apologized for the unauthorized forest clearings near Muhos, which have turned 816 acres of forest into an expanse of mud, there seems little doubt it is guilty of some wrongdoing. Rather, it is the scale of the punishment and any proposed remedies that will probably be of most concern to execs from the tech giant now.

As well as the awful mess Google has made near Muhos, it seems to have also jumped the gun and plowed ahead with data center development by clearing 500 acres of forest near Otanmäki.

Though the horse has already bolted, Reuters reports that a Google rep said the company would cooperate with the ⁠police investigation and continue to focus on maintaining compliance at its sites. Maintaining?

Last week we reported that wolves had descended on Finnish villages after their habitats had been destroyed to make way for the Google AI data centers. Data center developments were already a thorny issue with locals. Potential construction and operating noise, pollution, utility competition, and environmental damage concerns were overshadowed – for some – by the presence of wild animals that had moved out of the forest.

Google sees Finland as an attractive location for data center development with its cool climate, affordable power, and government endorsement of AI projects. It has also benefitted from being able to develop in low population density areas and from a populace with a low propensity for protest. However, Google is testing the Finnish public with its trampling of the environment. The wolves, pollution, potentially 40% higher electricity bills, and other impacts weren’t helping its case, even before this criminal investigation.

Finnish government representatives have yet to comment on the case.

From hereon, this month shall be called SpookITober. In yet another high-profile cybersecurity incident this month, Binance.com, the world's largest cryptocurrency exchange, has been hacked, to the tune of $5.6 billion taken in various cryptocurrencies. As of this writing, withdrawals at the exchange are suspended until the company sorts out the situation, though it has stated repeatedly that "user funds remain safe," adding that it "will cover all losses." The U.S.-specific Binance.us website is so far reportedly unaffected. At least two security agencies, including TRM, have pinned the attack on an elite North Korean hacking unit.

"Our initial investigation suggested that the security incident may have involved a compromised laptop, however our subsequent investigation has uncovered additional points of breach," the company said in an update on Friday. "This was a complex attack, and we are doing all we can in coordination with law enforcement to resolve it. We are still gathering the facts, and will share more directly as we have them."

"TRM attributes this attack to TraderTraitor, North Korea’s premier hacking unit," the security firm said. "We base this assessment on robust on and off-chain intelligence sources," continuing to note that the group has been named responsible for a $1.5 billion Bybit hack last year. Per TRM, this latest instance bears many of the same hallmarks as earlier North Korean heists. According to CNN, an unnamed analyst at Pangolin Intelligence corroborated this story by "Analyzing the on-chain activity related to this hack and correlating with information gathered from previous known incidents [involving the North Korean cyber unit]."

Preliminary info breaks down the stolen amounts like so: $1.57b in Tether dollars (USDT), $1.45b in Bitcoin (BTC), $875m in Ethereum (ETH), $823m in Circle dollars (USDC), $234m in Tron (TRX), $88.5m in Ripple dollars (RLUSD), $75M in Wrapped-Beacon Ethereum (WBETH), $50m in Tether Gold (XAUT), and $447.5 million in other cryptocurrencies. The "dollar" currencies are all stablecoins, meaning they're pegged to $1 to represent dollars in the virtual space. Meanwhile, Tether Gold is pegged to the value of one troy ounce.

The stablecoin issuers all moved to blacklist the thief's main wallet (you can see it here on the Ethereum blockchain) in their smart contracts, a move that means the amounts in USDT, USDC, and RLUSD are frozen in place, totaling around $2.4 billion. It's possible, but not guaranteed, that the gold-pegged XAUT will also be frozen.

Note that this does not mean those funds have been recovered — the thief simply cannot do anything with those tokens, be it moving or exchanging. Normally in these situations, restitution from the frozen funds happens only after complainants provide proof of ownership, something that usually means copious time spent in courts.

The situation is a lot more grim for the amounts in pure cryptocurrencies like Bitcoin and Ethereum, among others. Since there is no governing entity for their blockchains, the most likely prospect is that the remaining $3.2 billion or so is forever lost in the ether — though some networks incorporate clawback mechanisms. Exchanges may block the thief when he eventually tries to cash out his virtual dough, but historically this hasn't been a reliable means of recovery.

Note that this is a developing story, so the tallies are subject to change. Should that amount be final, this heist would exceed the $1.5b stolen from Bybit in February 2025, of which only a paltry portion has so far been recovered. There's little info so far on the miscreant's means of larceny. Apparently, the Binance 14 , Binance 16, and Hot Wallet 34 hot wallets, which hold pooled customer funds, were drained. It's currently not known if this happened via stolen private keys, a compromised signing process, or another type of intrusion.

Binance has assured users that their funds are safe, and most importantly, that it will cover all losses. For what it's worth, the company has a means of insurance and a reasonable track record in these situations, though the amount of bullion pilfered today is exorbitant.

The company created its SAFU reserve fund in 2018, where it deposits 10% of its trading fees to cover any losses incurred from security issues. That fund currently sits at 15,000 BTC in it (verifiable here), or about $1.17b at the current BTC-USD price of $77,858. SAFU is a cold wallet, meaning its private keys are stored offline.

Back in May 2019, thieves made out with 7,000 BTC from Binance, amounting to around $40 million at the time. The company covered the entirety of the losses with SAFU funds.

However, while $1.17b is certainly a tidy sum, 'tis but a morsel of what currently looks like a multi-billion loss. Binance co-founder, shareholder, and former CEO Changpeng Zhao (CZ) stated in a tweet that he too expects that users will be made whole.

As a final note, Binance is alerting everyone to be on the lookout for contact attempts impersonating the company and/or referencing the current situation. Expect a flood of emails, SMS, and phone scams in the coming months. Be especially wary of any "compensation" or "withdrawal restoration" offers, and never click any e-mail links or otherwise offer any account information — always use the official website or applications yourself directly.

South Korean authorities have identified 80 Tesla drivers who allegedly forced Full Self-Driving (FSD) onto cars that were never cleared for it, Electrek reports. The Seoul Metropolitan Police Agency says it confiscated 21 of the devices used to switch the feature on. The unlock modules reportedly connect to the car’s CAN bus (the internal network that links its electronic control units) and get around the regional software lock that keeps FSD disabled. That works because Tesla builds the same computer into every car, whether or not the feature is approved locally.

Some suspects bought ready-made units from Europe and China, and one reportedly plugged a European-made module into his own Tesla. Others skipped the shopping. One suspect, an IT worker, built his device using information he found on GitHub, which suggests the barrier to entry for this kind of modification is lower than it sounds. The group also traded tips online, including how to stop the car from keeping any record that FSD had been activated. Concealing the activation history is likely an attempt to avoid detection by Tesla, which has been looking for these devices itself.

Then there is the legal risk, where prosecutors are using the Motor Vehicle Management Act, which bans unauthorized software changes that could affect vehicle safety. In Korea, the maximum penalty is two years in prison or a fine of up to 20 million won, or roughly $15,000, a steep price for a driver-assistance feature that many of these owners had already paid for.

South Korea’s transport ministry first flagged suspicious activations to police in April. “Safety risks were confirmed, including the possibility that FSD may fail to properly recognize driving conditions or operate differently from the driver’s intentions,” a police official said then.

Notably, the root of the problem is a certification gap. Tesla began offering FSD (Supervised) in South Korea in late 2025, but only on US-built Model S, Model X, and Cybertruck units, which skip local safety certification under the Korea-US free trade agreement. The Model 3 and Model Y arrive from Giga Shanghai and have yet to pass. Nobody agrees on how many Korean Teslas are eligible, although Carisyou Data Lab figures cited by MK last month indicate 47,458 of 229,311 cars (20.7%).

The South Korean investigation follows Tesla’s own efforts to stop unauthorized activation. In April, the company began remotely detecting unlock devices and disabling FSD access on affected vehicles. It has also warned owners that unauthorized modifications could void warranty coverage and leave them liable for crashes. Moreover, the restrictions are not the first time Tesla’s software protections have come under scrutiny. In August 2023, researchers exploited a hardware security flaw in the AMD processor powering Tesla’s media control unit to unlock certain paid features. The exploit targeted the car’s AMD-based system-on-chip, although researchers did not confirm that every premium feature could be unlocked.

Tesla has also faced regulatory scrutiny over software restrictions. In December 2021, the US National Highway Traffic Safety Administration investigated the company after an over-the-air update allowed drivers to play certain games while their vehicles were moving. The case highlighted how software changes can raise safety concerns when features are accessible under conditions regulators consider unsafe. Hardware reliability is another challenge for Tesla. Last year, the company detailed how it identifies defective processor cores in its Dojo AI supercomputers, where a single faulty component can compromise lengthy training runs.

SpaceXSI, Elon Musk’s data center business powering Grok, is reportedly buying 418 properties in Colonial Hills, Southaven, which sits about half a mile from the powerplant that the company will use to power its data centers in the area. According to The Guardian, real estate agents have been going door to door, offering owners deals on behalf of the company provided that they do not join a class-action lawsuit over the noise pollution coming from the gas turbines that power its GPUs.

MZX, the SpaceXSI subsidiary that operates the Colossus II site, circulated an open letter to nearby residents, saying that it “requires property for its business purposes,” with Southaven mayor Darren Musselwhite saying that they will be used to house SpaceXSI staff that the company will need as it expands operations. Musk himself is reportedly staying in the area in an Airstream trailer four nights per week to personally oversee the data center build-out needed to power Grok AI.

The bigger issue here, though, is the requirement that sellers must sign an agreement that prevents them from joining the ongoing case against the company for noise pollution. The Colossus 1 and 2 sites have been in hot water since they started operating, primarily for their use of unpermitted turbines and the air pollution that they bring. SpaceXSI has since promised that it will remove all its unpermitted turbines in a year, especially as its 1.2-gigawatt powerplant comes online. However, this apparently hasn’t resolved the noise pollution issue that the community has been complaining about.

“It sounds like an airplane engine but it’s continuous, it’s incredibly loud. I’ve been woken up at 2am by it, I’ve been woken up at 3am by it. I sometimes have to just get away from it because it makes me so angry,” piano teacher Elizabeth Rose, who lives a couple of miles from the site, told the publication. One of the plaintiffs in the class-action, Jason Haley, also said to The Guardian, “There’s a high-pitched hum and then another sound like an airplane over your house. It’s been so torturous I can’t sleep; I just leave the house or turn on some music until I fall asleep completely exhausted. It drives you insane. You also lose hope because you’re being ignored or dismissed. There’s no escape.”

The company has denied all the allegations against it, saying that it has adjusted its power output and installed sound-dampening walls to reduce the impact on the neighborhood. Musselwhite also said that city monitors have not detected the site breaching the local ordinance limit of 70 decibels, although some experts say that persistent noise as low as 50 decibels could have adverse health impacts.

Opposition against the noise pollution and Musk’s purchases aren’t unanimous, though. Some residents aren’t as affected by the noise pollution, while local businesses are experiencing increased activity. Other people have also agreed to sell just so they could get away from the noise. The mayor says that the buyout isn’t intended to turn Colonial Hills into a company town. “The buyout has multiple purposes which are to provide homes for some employees and a goodwill gesture from a good corporate citizen for those who perceive the industrial development near them as negative,” Musselwhite said.

Concerns have been mounting everywhere about the environmental impact of data centers, with alarms ringing at various environmental think tanks. The Sierra Club, one of the oldest environmental preservation organizations in the U.S., has now put forth an estimate that by 2030, U.S. data center buildouts will be outputting as much carbon dioxide as the entirety of Mexico, pushing back progress on greenhouse gas reduction by 15 years, Bloomberg reports.

The Club's principal advisor for climate and energy, Jeremy Fusher, puts it plainly: "the numbers continue to sort of boggle our minds." He's referring to an estimate that with the output from pending data center builds, the output from the U.S. power sector is set to put 2.1 billion metric tons of CO2 in the atmosphere by the end of the decade, an uptick of 42% compared to 2025 figures. The researchers' previous estimates were between a quarter and half of the current projections.

An increasing number of countries and states are enacting laws demanding that new data center builds have a large portion of their power generated by green sources. However, those laws rarely apply retroactively, and oftentimes have a phase-in period. Even the Irish law on the matter demanding that new data centers be fed by 80% of renewable power only requires that that happens six years after they're powered up.

In the current AI-driven market, speed of construction is king. Roughly 40% of U.S. data centers are fed by natural gas turbines, and the situation doesn't look like it'll be changing much. It'll likely get worse, as the addition of green sources almost assuredly won't keep up with demand. Much like RAM, GPUs, and storage, turbines were all slurped up by data centers. Their prices have risen by 50%, and production is backlogged for years, with one maker stating it's opening slots for production in 2032.

Having said all that, Brendan Pierpont from Energy Innovation, another climate-concerned organization, points out that although the current path looks dangerous for the environment, it's hard to make a hard prediction of how the situation will play out. He justifies his position with common data center roadblocks: the availability of turbines, delays in grid connections, and other delay factors.

Truth be told, at this point it's hard to predict how many data centers will actually reach completion, given that requests for permits far exceed available grid capacity across the U.S., and a sizable number of buildouts are encountering strong opposition. Plus, although the AI bubble is still alive and well, concerns started to arise about the economics of hyperscalers, indicating some may go the way of the dodo — and their projects along with them.

Multiple sources claim that a top executive at a Singapore AI firm suspected of steering Nvidia AI GPUs to China and who has close business ties with Jensen Huang was apparently passing information to Beijing as a Chinese intelligence asset. According to The New York Times, U.S. intelligence has been aware of her activities for over a year already and has even briefed President Donald Trump on the situation. However, Trump reportedly chose not to do anything about the information, as he apparently did not want to affect Nvidia sales or antagonize China.

The suspect, Alice Huang or Huang Le, was listed as the managing director of Singaporean AI firm Megaspeed, an Nvidia cloud partner that builds data centers and rents out the compute to Chinese firms like Alibaba. The investigation into the firm began when the U.S. Commerce Department visited the firm’s Malaysia data center and found the Nvidia equipment that it ordered still packed in boxes. This raised suspicion that the machines were either just shipped in or about to be shipped out, instead of being installed on the site. However, it remains unclear what made U.S. intelligence agencies investigate Ms. Huang. Megaspeed was accused of helping smuggle $2 billion worth of Nvidia chips last year.

There have been multiple instances where Ms. Huang was spotted in public with the Nvidia chief. She was photographed sitting beside Jensen Huang alongside other executives in an informal event around Computex 2024 in a LinkedIn post by NetThunder, identified as Megaspeed Chairwoman Alice Wong. About a year later, around the same time as Computex 2025, she was once again seen with Jensen Huang leaving a restaurant in Taipei alongside other business associates.

“Megaspeed has been no different than many other regional cloud customers. Jensen’s interactions with Ms. Huang are not unusual or unexpected. Meetings with customers are part of a CEO's job; Jensen talks with thousands of customer executives every year,” an Nvidia spokesperson told Tom's Hardware. “Our regional cloud customers receive ordinary business information — not sensitive NVIDIA trade secrets — and zero information relevant to national security." Still, the possibility that Chinese intelligence is targeting a senior figure close to President Trump is apparently troubling some U.S. officials.

The Chinese embassy in Washington denied the allegations, saying that it was merely “groundless smears and slanders” against China. Nevertheless, it would also make sense for Beijing to keep tabs on senior executives of the top American firms that arguably single-handedly fueled the AI boom. And given Jensen Huang’s status as a celebrity CEO, it’s not surprising that he is surrounded by all kinds of people from both the U.S. and China. Sources cited in the report claimed that Mr. Huang hasn’t yet been given a defensive briefing — a warning that U.S. counterintelligence agencies would sometimes give to American citizens who have been targeted by foreign intelligence services — even though the surveillance against Alice Huang continues. The report speculates it is possible this was done to avoid alerting Alice Huang that she was being investigated.

As for Jensen, U.S. officials do not believe that he is aware of Ms. Huang’s intelligence activities. "We're confident that if the U.S. government believes that any person is a threat to American industry, it would let us know," Nvidia told Tom’s Hardware. "We often receive updates from the U.S. government regarding parties and entities of concern and always act accordingly."

Nvidia is reportedly facing questions from the United States Department of Justice over whether its purchase of Hugging Face will have an impact on competition in the growing AI space. Nvidia agreed to buy Hugging Face in September, paying almost $13 billion. Hugging Face offers an open-source platform for the hosting and testing of new AI models, and that appears to be behind the DOJ's concerns. Reuters reports that the department has made inquiries about how the Hugging Face deal could change the landscape for Nvidia's competitors. The move could be seen as a way for Nvidia to lock in a key distribution channel for its own products. Hugging Face also sells AI infrastructure. Nvidia has declined to comment on the story.

This isn't the first time that Nvidia has found itself in the DOJ's crosshairs. As Nvidia becomes more vital to the tech industry as a whole and AI technologies specifically, the company has come under increasing scrutiny within the U.S. government, stretching back to the Biden administration. Things didn't improve when the Trump administration took over, with the U.S. president reportedly tempted to break the company up — despite being unfamiliar with what it was. More recently, Nvidia and other AI hardware makers found themselves under export controls. Nvidia's powerful AI-focused hardware can only be sold to approved entities, while it cannot ship to countries including China, Iran, North Korea, and Russia.

At home, the DOJ has also looked into Nvidia's purchase of SchedMD, a company whose software is used to run supercomputers that have ties to national security. Nvidia promised to continue to support the software as part of a larger three-year investigation into its conduct. Officials have also considered whether Nvidia dangled a larger allocation of its AI chips in front of customers who also purchased its networking hardware as well as Groq chips — Nvidia bought Groq for $20 billion in December of 2025. The deal included the acquihires of key members of the AI chip startup's team, including its CEO.

As for Hugging Face, Nvidia has similarly confirmed that it will continue to support a range of chip and AI providers once the buyout completes. It's expected that Hugging Face will officially become part of Nvidia in the first half of 2027.

Systems administrators are among the first to point out lax security policies in other companies' or people's setups. Sometimes, they are the ones who get bitten by not taking the same advice they will so readily dispense. Such is the tale of Reddit user NakamuraHwang, who had files on their Synology DS918+ NAS encrypted by a ransomware attack after making a series of unforced errors.

Hwang appears to be quite technically proficient and provided a detailed write-up of the situation. The key points, however, are that the unit in question (a) had its web interface exposed to the internet and (b) was roughly 18 months out-of-date with security patches. Just like the recent PHP snafu has adeptly shown, complacency is a far worse enemy than an army of AI bots.

The tech has a 3-2-1 backup setup, so they seemingly didn't lose any data. They attached the disks to a Proxmox machine for further inspection — Synology's DSM operating system uses mdraid underneath, so you can take the disks to another box and mount them there. Investigation of logs and the operating system files revealed that the attack was likely performed by abusing the CVE-2024-10441 remote code execution vulnerability, revealed a year and a half ago.

As most any other sysadmin will tell you, it's never advisable to expose devices directly to the internet without some layer of protection in between, chiefly but not only a VPN or an SSH tunnel. Tailscale, for instance, makes having a VPN fairly painless these days, and its free plan covers most home installations.

Synology had published its own advisory regarding the exploit, but that's where the NAS' behavior regarding updates complicates matters. According to multiple comments on the thread and in other forums, the UI will happily tell you that your device is up to date, all while there are critical security updates that need to be manually installed. The manual for DSM indicates that units from 2025 onwards have automatic updates, apparently implying that those of previous makes need to be configured so.

Furthermore, Synology apparently doesn't have a direct channel to warn its customers of critical security vulnerabilities —a simple e-mail notification would suffice. Although the firm is seemingly shifting away from consumer gear, it still sells small units that regular users might buy, and it's not reasonable to expect them to even be aware of security bulletins. Plus, package updates like the patch for the PHP2Shell vulnerability have to be installed separately, too.

Agentic AI has set off a firestorm in the world of CPUs over the past six months, yet it’s an area of performance that still doesn’t have solid, repeatable benchmarks to evaluate where the latest hardware falls. We’ve been working on an internal set of tests to evaluate at least one aspect of contentious agentic loops end-to-end, isolating CPU and memory performance from the accelerator to understand how these components contribute to the agentic loop.

This is still a work in progress, but we wanted to bring Tom’s Hardware Premium readers behind the scenes to understand how we’re going about tackling this problem, our methodology for testing, and detail what we’ve learned about agentic loops along the way. We've prototyped the test and settled the central design, but there's still work to do before the numbers it produces hold up to scrutiny.

Given that this area of computing is bleeding-edge and rapidly evolving, we undoubtedly have blind spots here. That’s OK. The testing we’re doing is built to look at how agentic performance scales across the CPU as sessions increase, providing an ideal throughput for each chip that provides as many agent sessions as possible while remaining interactive.

Toward the end of this article, we've included a section on some of the shortfalls of this approach, as well as some ideas we’re toying with to tackle things like multi-turn agentic flows. And keep in mind that these tests aren't meant to replace our other testing of local AI, but to supplement it. As this space continues to evolve and the hardware characteristics are better defined, we’ll be able to make more informed decisions about where this testing goes in the future.

Defining an agentic AI loops

"AI agents" imply some little workers that, strapped with an LLM, can go and handle whatever task you throw at them. It's a fine description, but we'll need a bit more a technical definition if we're going to benchmark the performance of agentic AI.

There are two parts here that are important. First are functions. With agentic AI, we give the LLM access to tools, which, in the case of popular agentic frameworks like LangGraph, are Python functions. The tool may be reading or writing a file, querying a database, running tests on code, or just about anything else you can define in a function. Here's a basic example.

@tool    def run_command(command: str) -> str:        """Run a build or test command in the repository root. Allowed: make [targets], python -m pytest [args]."""        return traced("run_command", {"command": command}, lambda: ws.run_command(command))

This is how you write a tool call with LangGraph inside of Python, and it's an example of a tool call that the we've given the LLM access to through our LangGraph persona. We identify the function as a tool the LLM can call with "@tool," and then we define the function like normal Python with "def run_command," passing through the command as an argument. The extra "-> str" on the end notes that the argument will be a string, and the function should return a string.

That's followed up by a description of the tool that the LLM can read (triple quotes just allow you to space this across multiple lines in Python). We're prompting the LLM that it can run build or test commands using make or pytest. For this example, we wrap the function up with the anonymous lambda function and pass it back to an overarching trace function. That's not important right now, however. The goal is to illustrate what a tool call looks like.

This is only one of a handful of tools available to our coding agent. It can also read and write files, edit files, and search code using regular expressions, all defined in the same way with Python functions. Agentic AI is often described as a "loop," however, and that's because the task often requires calling a tool multiple times and/or calling different tools.

If we're working a Python project, for instance, we may prompt the LLM to "debug this Python script." It may call the "list files" tool to evaluate what Python script we're talking about first. Then, it will call the "read file" tool to understand it, or maybe "search code" to find a spot to start troubleshooting. Then it might run "edit file" and "run command," and if the file isn't fixed, it may loop back over these tools a few times.

This is the agentic loop. We prompt the LLM, it evaluates and starts calling tools it needs to form an answer. Then, it gets back the information from the tool calls, it re-evaluates, and then it may call more.

Starting with concurrency

We’ll dig into the specific tests and what they’re doing later in this article, but it’s good to set the stage with why our benchmark is designed the way it is. We developed the test on an Nvidia DGX Spark (GB10) featuring 128 GB of unified memory, a Blackwell GPU, and the relatively weak Arm-based Grace CPU. Above, you can see the time spent running these agent loops on the GB10 with gemma4-26b-a4b. The model isn’t super important here as long as we're able to call tools, and we’re able to re-record the tests using different models as they continue to release.

The point here is the share of time spent, not the model or the specific work the agents are doing (though that does provide useful context). Each result here is a list of sequential tasks for the agent to execute against the model, and nearly all of the time is spent on the GPU. Even relatively demanding CPU tasks, like with our coding agent, represent just a small fraction of the computational time. Running one or even a few agents isn’t a problem for the CPU when we have access to a large, powerful GPU with a sufficiently large model running on it.

Looking closer at one specific set of tasks, you can see that there are plenty of tasks for the CPU to execute during the window, but those tasks happen so quickly that it barely registers in the overall time spent.

These tasks aren’t evenly spaced, either. Depending on when the agent needs to call tools, the shape of the workload can look very different. Above, for example, you can see our document agent, which is tasked with scanning and extracting information from a set of PDF files. Nearly all of the agent time is spent at the very beginning of the task, after which it’s almost pure GPU inference.

Obviously, a more powerful GPU will churn through the inference step faster, but we’d have to create a severe imbalance between the CPU and GPU for the CPU portion of a single agentic loop to create a bottleneck — at least with these common tools. Agents don’t just handle a sequential list of tasks, however. They spawn sub-agents, and if multiple sessions are running on the same system in sandboxed environments, contention for limited resources becomes a real issue.

Instead of designing the test platform in such a way that creates a severe CPU/GPU imbalance, we looked toward concurrency. The idea is to sweep across the chip with concurrent sessions running a list of tasks in a loop in their own processes. As the sweep scales up, contention across the CPU starts to appear, and minor differences in tool calls, code execution and compilation, web browsing, and the variety of other tasks our agents do pile on.

How the benchmark works

With concurrency as our north star, we can start breaking down the design of the benchmark and how it works internally. The overall design is simple. Custom agents built with LangGraph have a list of 30 tasks to execute in sequence against a llama.cpp server running gemma4-26b-a4b. Instead of going directly to the server, however, the agents go through a proxy server that records a trace.

During live recording, the agent communicates with the proxy server, which passes a live call to the llama.cpp server. The proxy server meanwhile writes every logged request to the LLM and the response from it to an overarching JSON. Tool calls bypass the server chain and are logged to their own JSON files. Finally, for each task, a correct response is generated, hashed, and stored as its own JSON.

Recording everything allows us to remove the live model entirely from the equation. The agents still run as normal, but instead of running against a live LLM, they run against that proxy, replaying the traces captured. The requests and tool calls are all happening live, but the agents are immediately handed back a response from the trace based on the live inference capture.

Starting with the request from the agent, it’s matched by session label and hash to ensure it’s accurate before the response is looked up in the master JSON. If there’s a mismatch, the run stops and is marked as invalid. If everything checks out, the response is passed back, and the agent session(s) call the tools they need. The tool calls are compared to the JSONs generated during recording, and the recorded result is passed back. This is critical.

The test compares the live tool call to the recorded result, but it always passes back the recorded result to keep the test from drifting, even if the tool call is correct. True tool mismatches are logged. The test proceeds, but it’s marked as invalid for each level with tool mismatches.

This setup makes redesigning and expanding the benchmark simple. We can write a new persona in Python using LangGraph, or expand to other frameworks like CrewAI. Additionally, we can rerecord traces against different models in the future, perhaps some with expanded capabilities. As long as we’re communicating over an OpenAI-compatible endpoint, we can mix and match pieces of our testing to represent a broader range of workloads.

The test starts at a single worker process and scales up to the maximum thread count of the CPU, running three repeats per level. Each worker process operates off a list of 30 tasks, with new worker processes starting at the next point in the list. So, worker session one would start at task one, worker session two would start at task two, and so on. Tasks do not take an equal amount of time, so this offset creates imbalanced contention as tasks overlap.

What this actually measures

The metric we get out of the test is sessions per second. A session is a single task completed from end-to-end, and we measure the number of sessions completed per second at each level. Naturally, at lower levels, we complete fewer sessions per second, often falling below 1, meaning the task takes longer than one second to complete. The tasks are completed sequentially. What we’re looking for is how chips diverge as sessions increase, measuring the throughput of agentic work on the CPU.

We derive two headline scores from the sweep. First is peak throughput, which is the highest sessions per second we recorded, as well as the concurrency level at which it was recorded. Next is ideal throughput, which is the point at which the slowest 5% of sessions take no longer than 1.5x the time it takes for a single session to complete at the base concurrency level. We expect that tasks will take longer as the number of sessions increases, but past that 1.5x mark, the length of the task gets significantly longer, and depending on the processor’s topology, throughput can start to degrade.

It’s worth highlighting again that there is no GPU delay in this concurrency sweep. We’ve removed the GPU entirely, throwing as much concurrent work at the CPU as possible to determine where that peak throughput point lies. However, the GPU still plays a significant role, so the results should be viewed in a realistic context.

As more agents work on the CPU, they’ll compete for resources on the chip, but this balance can easily shift. If there isn’t enough GPU horsepower, agents may queue for a slot, and that’s not effectively utilizing the compute on either processor. Our testing is more than theoretical throughput from napkin math, but it’s less than a true end-to-end system. There are too many variables to account for in that setup for us to produce any useful numbers that are broadly applicable.

However, we can run the sweep with live inference to evaluate system-level performance. For live runs, the responses handed back to agents are still served from the recording to keep runs equal. However, the request is recorded by the llama.cpp server, which generates the same number of tokens before being discarded. This accounts for the normal inference pipeline, including prefill, decode, and cache performance, as well as agents queuing for a slot.

Both numbers are important. Live sweeps are a realistic look at local AI across the entire system, while frozen replays measure the peak throughput a CPU is capable of with access to all the GPU compute (theoretically) that it could want.

Our starting agent personas

We’ve defined six agent personas. A “persona” is a Python script built with LangGraph that defines what the agent can do. The agents tool calls are Python functions, defined in that main script that connects our LangGraph nodes. Each persona is working off its own unique list of tasks from local documents. For example, the research persona is able to answer questions by searching in a Wikipedia corpus of about 50,000 articles.

Above, you can see an overview of each persona. For five out of the six, we start with some sort of planning stage, followed by the main agent loop, and finish with a report. For the analyst persona, for instance, the planner takes the questions and splits them into structured output with two to four steps. Analysts then run in parallel, searching the database, and the writer turns those findings into a response.

The one exception is the coder persona, which works a bit differently. It’s working off four projects, attempting to address a bug from an issue report (similar to a GitHub issue report). The coder finds the bug, runs tests to confirm the fix, and then passes it off to a reviewer step. The reviewer step (naturally) reviews the fix before applying it.

For each of the personas, the current test setup looks at single-turn tasks. Agents make multiple loops to the LLM depending on the task, and often use tools several times, but it’s all to resolve a single task with no previous context. For instance, when a customer support query comes up, the agent may take multiple passes at the knowledge base to answer that query. However, once the agent has answered the query and moved on to the next, it’s treated as a separate action without context from the first query.

This is not completely representative of agentic workflows, which are often multi-turn processes. We're able to script multi-turn tasks, but that creates greater divergence risk, as well as risk of a crash. Running this locally, it's not hard to spill over the context window after a few turns of complex interactions. It's possible to do, but it's not on the top of our list of priorities right now.

We won’t cover all of the tasks the current personas do because, frankly, it’s a lot and most of it isn’t that interesting. Still, we want to provide at least one example for each persona so you can get an idea of what’s going on under the hood.

All of the other tasks are very similar to what you can see in the examples above. With the documents persona, for example, the agent is always approving the correct invoices, but the mix of documents and types of errors change between sets. Similarly, the support persona is always answering support questions, but they focus in different areas. Some ask to troubleshoot error codes, others (like the example above) ask about policies, and more still include multiple questions within one query, tasking the agent with piecing together multiple areas of the knowledge base.

Shortcomings of the current approach

Although we’ve tried to create a broad range of agentic work here, there are some shortcomings of the current version of this test. Most notably, our testing and recordings only come from two types of systems: a 64-core Threadripper test bed with 128GB of DDR5-6000 RDIMMs, and a DGX Spark.

As we continue to test other platforms, we may need to adjust the tests, rerun recordings, or swap the model. We’d have to re-gather all of our results if we make those changes. By other platforms, we’re not just talking about an Intel system, either. We’ve seen a variety of different local AI setups, from Mac Studio clusters to GPU-ridden workstations. The load on the CPU completely changes depending on the GPU performance, model, quantization level, and server settings.

The other shortcoming is API performance. The test is designed in such a way that we can easily connect to any model from a frontier lab through their API. Although that model runs in the cloud, the tools themselves run locally unless you’re using an entirely cloud-based sandbox. If you’re working through an AI-assisted development workflow on a budget laptop, for instance, all of the compilation, testing, and agent code still runs on your machine.

Although we have plans to look into this hybrid model with a weaker host PC and a cloud API, we didn’t explore it too much during initial development. API usage is expensive, and we don’t want to be throwing away tokens trying to get the initial test suite set up correctly.

Finally, sandboxing. The current test is isolated through a Docker image, which introduces negligible overhead (<2%) compared to a bare-metal run. We’ll need to explore the overhead of sandboxing local agents via different means in the future, which is an area we really haven’t explored much in this test.

Continuing to learn

Agentic AI is rapidly evolving, so much so that you’ll get different definitions of what characterizes agentic workloads and hardware built for agentic workflows depending on who you ask. Add on top of that rapid hardware and model changes, and it’s obvious that we could never define a bulletproof benchmark that can stand the test of time. That’s a fool’s errand. What we have here is our starting point, which we will continue to build on top of.

From what we have now, the next steps are to get the first real numbers. We've done some basic tests to ensure the system is set up correctly (for instance, running a free-threaded version of Python to test GIL impact within a single session), but the focus here was getting the basic system in place to record traces and replay them.

We're also relying exclusively on LangGraph here. We need to explore other frameworks like CrewAI, as well as the "one-click" agents like Hermes and OpenClaw.

Fujifilm plans to boost production of magnetic tape by as much as 5x as it works to keep up with demand for the data storage medium that has seen a jump in popularity amid ongoing global AI data center buildouts. The Japanese firm will spend 10 billion yen ($63.3 million) to increase capacity at its Odawara plant throughout its fiscal year 2028. The move will enable it to produce more advanced tapes with capacities of up to 100TB each, Nikkei Asia reports. The expansion of production capacity will mark the company's largest investment in tape-based storage in around 20 years. Fujifilm first began working on magnetic tape technology back in 1954 when it was used for television broadcasting equipment.

Nowadays, magnetic tape is the go-to storage option for data centers everywhere as a form of long-term data backup. While tapes lag behind solid-state drives and traditional spinning hard disks in terms of performance, tapes are hard to beat when it comes to their durability and relatively low cost for storing huge amounts of data. Magnetic tape storage can cost as little as one-quarter that of hard disks, and even less when compared to faster SSDs. That makes them the perfect solution for so-called cold data, information that is rarely accessed but could be considered mission-critical and must be stored safely. Tape storage isn't just cheaper, either, with research showing that it's also less polluting than hard disks and SSDs.

While Fujifilm will produce 100TB tapes, the Linear Tape-Open (LTO) roadmap allows for considerably larger capacities, all the way to 365TB per cartridge. By comparison, SSDs top out at around 256TB and cost considerably more.

Fujifilm is far from the only storage company looking to take advantage of the AI data center boom. Toshiba announced earlier this month its plans to double hard disk production capacity amid shortages, and the company, alongside Seagate and WD, is working towards ever-increasing capacities. Kioxia and SanDisk are also getting in on the act after detailing 332-layer 3D NAND flash in August. The pair hope to bring higher storage capacities to SSDs for data center use.

The demand for AI data center storage is so high that it's at the heart of a global shortage that shows little sign of improving anytime soon. Scarcity has seen prices skyrocket, while manufacturers have chosen to divert resources away from consumer-focused products, instead focusing on the more lucrative enterprise market.

Anthropic spent less than a third of its $8.9 billion first-half compute bill on serving customers, with most of the rest going into training, research, and running Claude internally, its S-1 filing shows. It has meanwhile signed up for hundreds of billions of dollars more in TPU and GPU capacity. By our math, it has meanwhile signed up for roughly $800 billion more in TPU and GPU capacity, and revenue per compute dollar jumped from $0.40 in Q1 2025 to $2.08 in Q2 2026, which the company links to capacity purchased 12–18 months ahead.

Anthropic’s IPO is likely to beat that of rival OpenAI, which filed confidentially before CEO Sam Altman said its own won’t come this year, amid widespread doubts about an AI bubble. Whether AI labs can cover what it costs to serve their models is an open question.

Anthropic appears to do so easily: it spent $2.4 billion serving customers in the first half of the year, versus $16.3 billion in revenue. Paying for what it has on order is more difficult. Its five compute deals with the biggest stated amounts would cost an estimated $103 billion a year on average once delivered, more than twice its revenue at Q2’s pace of about $46 billion a year. The company also warns that margins may compress as new hardware lands ahead of the revenue it will support.

The company has three main costs: the cost of revenue, which is the inference cost of running Claude for paying customers; R&D, which is mostly compute for training and research into new models plus any internal use; and sales and marketing, which covers the cut taken by channel partners and cloud platform resellers.

Comparing the first half of last year to this year’s, cost of revenue is up about 2.3x, but with revenue up 14x, that means Anthropic keeps an estimated 85 cents for every dollar after serving costs, up from just 12 cents. R&D is up about 5 times, with 66% or more of the compute going to it, and that’s the main source of its operating loss. Sales and marketing is up 16x, and doesn’t show up in cost of revenue, but its partner fees and marketing would lower what it keeps to 66 cents a dollar. The trend here is that serving Claude now more than covers its costs.

Serving has kept getting cheaper for customers since June: in September, Anthropic’s Opus 5.5 model cut costs per task by about 40% versus Opus 5, says the company, and Fable 5.1 cut the cost of typical workloads by about 25% against Fable 5; each came from “a combination of a lower price and improved efficiency,” made “while preserving our margins.”

As for compute, which runs through all three costs, Anthropic runs it as “a single, shared resource across all workloads” and doesn’t disclose how it decides the split between training, research, and serving. Serving was limited, as the company has “consistently operated in a compute-constrained environment where customer demand exceeds available compute,” which suggests serving costs were muted by a lack of supply, not just efficiency.

Anthropic runs a “multi-chip compute architecture across AWS Trainium, Google TPUs, Nvidia GPUs, and AMD GPUs,” without giving any percentage splits or GW measurements. For TPUs, it has five-year equipment leases with Broadcom for “TPU-based compute capacity,” with rent of “approximately $161.2 billion excluding interest”; some of that capacity came online this year, with more from 2027. Another $114.5 billion or more is with Google “between April 2026 and July 2033,” with more from Google and other TPU providers from 2028 to 2031, estimated to be in the $125–$150 billion range.

Trainium comes through Amazon, with a deal for “at least $110.0 billion between May 2026 and April 2036,” though the filing doesn’t split AWS spend by chip. Nvidia GPUs come through a series of deals, including “up to $120.5 billion … extending into 2029” with SpaceXAI, “at least $44.9 billion” extending into 2033 with Microsoft, “up to approximately $44.6 billion” also into 2033 with Nscale, and about $41.3 billion into 2034 with Lambda. There is also an AMD hardware supply agreement which, in aggregate, may “exceed $20.0 billion” if Anthropic takes the full initial capacity. These deals each have their own conditions for termination.

Anthropic’s minimum purchase commitments, which exclude equipment leases such as Broadcom’s, have gone from about $55 billion at the end of last year to about $265.6 billion by the end of Q2 this year. Since then, it has signed about $245 billion in new minimums, leases included, and TPUs are the biggest line overall. The Nvidia capacity, Microsoft’s aside, is also the easiest for Anthropic to walk away from.

With all these multi-year commitments, bought well ahead of delivery, expenses are paid and recognized “as compute becomes available to us,” the company says. That compute base “did not fluctuate with in-period demand,” so revenue per compute dollar increased as demand surged. When demand outruns that capacity, Anthropic buys short-term compute, which “typically carries a higher cost per unit.”

Our $103 billion estimate is straight-line, spreading each deal evenly over its term except SpaceXAI, which is at its full monthly rate, while payments ramp up as capacity is delivered. Also not included in the estimate is interest on the Broadcom leases. Compute also does not necessarily have to earn revenue, especially in R&D.

The filing has many blanks, such as the total commitment figure, the contractual obligations table, Q3 revenue, and Run-Rate Revenue, which an amended S-1 would normally fill in before pricing. No date for that has been given. Microsoft’s minimum starts in November, and SpaceXAI’s full delivery is expected in December, with non-cancelable periods ending in January and March 2027, after which either side can cancel on 90 days’ notice. More Broadcom TPU capacity and AMD’s first tranche follow next year, by which point most of these deals will be billing.

Malicious GitHub repositories masquerading as two well-known and legitimate PS5 game emulation tools have been identified as spreading malware that steals gamers' passwords. The malware uses self-signed certificates, created using fake company names and on the same day the malware was built — suggesting that the malware campaign may be automated. Just two minutes after infection, a cleanup script scrubs the affected computer of the malware's files, leaving no trace that it was ever present, according to a report shared with Tom's Hardware by NordVPN's Threat Intelligence research team.

The campaign trades on the back of the popularity of two projects that have become popular within the PS5 emulation community. AnyPS5 is a portability layer that allows PS5 games to run as native applications on both Windows and Linux. Kyttyps5 is an emulator for running PS5 games on PCs, and both projects are freely available for download via their respective GitHub repositories. However, users who visit one of the malicious repositories posing as these tools are directed to download a zip file containing a Windows executable. Notably, the GitHub repositories used as part of the malware campaign rapidly rotate, and it's possible that GitHub is not the only source.

The malicious executable unpacks a hidden application before launching its own multi-stage loader. It first checks whether the PC is a legitimate machine or a security researcher's sandbox by checking if less than 2GB of RAM is available, if it's a virtual machine, if it's running on a cloud-based server, or if it doesn't have a physical GPU installed. If any of those criteria are met, the loader exits silently. If the machine is deemed to be a legitimate target, the malware adds exclusion rules to Microsoft Defender to ensure that it will never be scanned and identified during an antivirus scan. Finally, the payload is decrypted and executed.

That payload, we're told, is an executable protected by a commercial packer that is believed to be an infostealer — a type of malware whose job it is to steal passwords as well as any other sensitive data that it comes across.

While this appears to be a relatively new malware attack, there are some steps that you can take to avoid falling foul of it. If you're trying to download tools like AnyPS5 and Kyttyps5, it's vital to ensure that the repository belongs to the official project team. Double-check the owner of the project, look for download links that use third-party destinations, and check metadata like the project and files' creation dates.

If you do find yourself in a position where you have accessed a suspected malicious download, disconnect your computer from the internet and run a full antivirus scan. It's also recommended that you change your passwords from a separate device where possible.

PS5 emulation is likely to become more popular than ever once GTA 6 is released on November 19, 2026. The game will be a console exclusive at launch, with a PC version expected to arrive in early 2027.

David Durman, a research administrator at the University of Kentucky (UK), has successfully argued that using AI violates his beliefs as a humanist and has received Title VII religious accommodation from university officials not to use the technology. According to the Religion News Service (RNS), Durman believes in "the human part of life," which includes making his own decisions. However, he says that when the UK asked its employees to use AI, he felt that doing so would mean outsourcing his thinking to tools like ChatGPT and Microsoft Copilot, thus violating his deeply held beliefs.

“The process and work involved in thinking through situations, through my work, through my ethics, through the way I handle other people, is a key component in being able to develop myself as a moral person,” Durman told RNS in an interview. “And offloading that labor, regardless of whether it’s taxing or not, is undercutting my ability to be a moral actor.” Because of this, the research administrator was granted a Title VII exemption under the Civil Rights Act of 1964, which prohibited employment discrimination based on race, color, religion, sex, and national origin. While humanism is often a non-religious philosophy or a worldview, federal courts have ruled that it qualifies as a religion.

The American Humanist Association (AHA) assisted Durman with his request for the accommodation. “This is the first of its kind at the AHA and for humanism generally to exempt those who see AI use as flying in the face of their freedom of conscience,” AHA staff attorney Alexandria Huiskens said to RNS. She also added, “For us, humanism is a comprehensive way of life. It’s built on inherent worth and dignity, and everyone should have room to exercise this comprehensive way of life at work the same way Christians do, or Muslims do.”

It should be noted that while anyone can request a Title VII exemption, it will only be granted if the applicant has a deep moral conviction that could count as an article of faith. Aside from that, the law also states that the request should not place an “undue hardship on the employer’s business.” So, even though Durman is no longer personally required to use AI tools for his tasks, he is still expected to attend meetings and cooperate with colleagues who use AI or to submit materials to AI tools, as needed.

AI use in corporate settings has been one of the biggest questions ever since LLMs exploded in popularity with ChatGPT. While AI proponents have said that the technology would increase productivity multiple times, reports say that corporate America, except for Big Tech, is not seeing any financial returns. Even so, about half of all U.S. employees now use AI one way or another at work, even as many executives remain uncertain about AI’s return on investment. But as for Durman, he says that the fact that Americans are already thinking about the impact of AI is a good thing, and that he hopes that his story would help others who’d rather not be forced to use the technology.

Flock's AI surveillance cameras are continuing to cause controversy, this time in Dayton, Minnesota, where it is alleged that the company reinstalled two cameras despite express orders from the city not to do so. The city's local government posted on Facebook to explain, claiming that the Police Department was made aware of a maintenance work order on September 17, to which Dayton responded the next day, stating the program had been put on hold and no work orders were to be completed. However, Flock went ahead and reinstalled them anyway without authorization from the city.

Dayton’s city council had already voted to suspend the automatic license plate reader program and Flock cameras on August 25, following community concerns over privacy and data misuse, The Drive reports. Flock is infamous nationwide, with instances of its cameras being used for unjust purposes, such as a car reviewer mistakenly being pulled over because one of its cameras misread the tags. In another case, Flock’s chief information security officer also falsely assured the Oshkosh City Council that Flock did not create a heat map of individual license plates, when it in fact does have a heat map feature in the Flock system.

Flock cameras in the city had already been broken and vandalized, leading police to disable and remove the cameras to prevent further damage on August 20, before the program’s suspension became official, local media reports. Despite the suspension, almost a full month later, on September 17, Flock submitted the offending work order, which led to this latest controversy.

We don’t know when Flock reinstalled the cameras, but Dayton citizens had already complained and reported spotting the two Flock cameras, leading to the September 29 Facebook post. The local government further stated that the Dayton Police Department had been informed and ordered to remove the cameras.

It is unclear what happened in the chain of communication from when the license plate reader program was suspended and canceled its contract to notify Flock to cease work orders; either Flock defiantly ignored the orders or instructions had not reached the people responsible for installing the cameras; either way, this could be another blow to Flock's reputation.

Malaysian customs have seized over $12 million (49.5 million Malaysian ringgit) worth of advanced AI chips in a single bust, as the country clamps down on a suspected semiconductor smuggling syndicate. According to a Reuters report on October 28, officials found 30 servers containing the chips at Kuala Lumpur International Airport’s free trade zone on October 15th. Initial investigations revealed that the shipment originated in an Asian country and was bound for another undisclosed Asian country.

The airport’s customs director, Aspalila Ag Tuah, confirmed that the servers were falsely declared as ordinary electronic goods, saying the suspected scheme used Malaysia as a transit point to avoid restrictions on the goods’ onward movement. The Malaysian company involved in the shipment has been summoned to assist in the investigation.

Authorities suspect that the foiled smuggling attempt was carried out by the same group responsible for a similar attempt last June. On June 5, customs officers discovered 72 server units containing advanced AI chips at the Kuala Lumpur International Airport, valued at approximately $13 million. This shipment had also been falsely declared as computer components and was meant to be reshipped without the required permit. Aspalila said the network has been linked to seizures worth over $25 million so far across at least five cases. She declined to provide further information on the network or the current case, citing ongoing investigations.

Several key facts remain unknown, including the chip type, country of manufacture, final destination, and intended recipient — whether a company or a military organization. The U.S. had placed export controls on the movement of advanced Nvidia and other U.S.-origin AI chips to China, the most likely destination for the seized chips. Smuggling networks began using Malaysia, a U.S. ally not affected by the controls, as a hub, exporting the banned chips to the country before re-exporting them to China.

While the country’s geographic location and mature logistics infrastructure make it important for legitimate international trade, these characteristics also make it ideal for smuggling sensitive technology. To restrict this avenue, Malaysia subsequently imposed export controls on the ​movement of advanced AI chips of U.S. origin, at Washington’s request. Other Asian U.S. allies, such as Singapore and Taiwan, have implemented similar controls. Demand for powerful Nvidia GPUs remains high in China, fueling a booming market for smuggled hardware.

HP, Asus, and Acer have reportedly moved some of their manufacturing capacity back into China as increasing component costs — driven by AI — bite into their margins. According to Nikkei Asia, increasing bill-of-materials costs have made China’s experienced workforce, relatively lower prices, and established supply chain ecosystem a more attractive option. However, they are not going back to traditional suppliers like Foxconn, who have focused their production capacity on AI data centers. Instead, they’re signing contracts with smaller Chinese contract manufacturers, like Luxshare, Huaqin Technology, and Longcheer Technology, that are trying to establish themselves in PC manufacturing.

Many PC brands slowly moved out of China in 2024 and 2025 as the U.S.’s sanctions and tariffs on China have increased prices and forced them to find friendlier nations. While these nations do not have established supply chains and expertise like China, it at least spreads the risk as manufacturers avoid getting caught up in geopolitical tensions.

However, it seems that rising component prices have forced them to move back some of their capacity into China. The biggest culprit is the AI-induced memory and storage chip shortage, which has caused memory prices to climb an unprecedented 500% in 12 months. More than that, other parts like PCBs and fiberglass cloth used in motherboard manufacturing are seeing increasing prices, which Acer chairperson and CEO Jason Chen bemoaned as suppliers hyping shortage fears to protect margins.

While tech companies aren’t abandoning their “China+1” diversification strategy, the mature industry in China is helping offset increased component costs. Aside from that, the Chinese contract manufacturers often have better access to memory chips from domestic manufacturers like CXMT. "These Chinese contract manufacturers have been telling potential clients and computer makers that working with them could be a good option for securing better access to Chinese-made memory chips," a supply chain manager in the computer ecosystem told Nikkei Asia. "They are marketing themselves as a good alternative to even help sort out the supply chain."

Moving back some manufacturing capacity into China is easing some of PC manufacturers’ cost concerns, but it’s also negatively impacting the newly established supply chains in Southeast Asia. "We built two new production lines in Thailand for the PC clients, and now they are moving substantial orders back to China,” one manager at a PC component supplier told the publication. “Those two lines have not yet turned profitable, and I don't know when they will add the orders back."

Unfortunately, building an entire supply chain for PC manufacturing is slow and expensive. The U.S. has been trying to accomplish this for years, and, so far, only Anduril has attempted to build laptops with 90% of its component costs coming from the U.S. and other allied countries. Even then, it expects its devices to be around 20% more expensive than the competition, which it could afford by offering them mostly to critical industries that demand supply chain security, like the government and the military. But if you’re a PC manufacturer vying for customers in the consumer market, you’re likely incentivized to find the lowest possible costs to be able to compete in pricing.

Amazon wants its engineers to quit chasing token counts, warning that heavy AI usage does not automatically mean better work. In updated engineering guidelines published on October 27, the company laid out how developers and managers should gauge AI’s real impact — and the takeaway is blunt: consuming more tokens is not the same as delivering more value. Amazon is not backing away from the technology, though. “We are certain that whatever future best practices are, AI will be at the center of software engineering practice,” the company wrote.

The guidance targets ‘tokenmaxxing,’ a practice where employees deliberately inflate their AI usage to look productive on internal dashboards. According to the memo, metrics such as AI adoption rates, token consumption, code changes, and code reviews become ‘deeply flawed’ when they are disconnected from actual results. Turning them into targets, Amazon says, encourages wasteful behavior that clashes with its emphasis on ‘frugality and simplicity’ and can raise both compute bills and software complexity.

The company’s preferred yardstick is long-term results. “We believe that it is best to measure, and be measured by, long-term outcomes,” Amazon wrote. That does not mean managers should ignore usage data altogether, as input metrics remain useful when examined ‘with care and judgment’ alongside other indicators. Amazon suggests looking at both the teams using AI the least and those using it the most to determine whether adoption is translating into better delivery.

The push comes as Amazon, under CEO Andy Jassy, has been weaving AI deeply into its sprawling operations. Jassy said earlier this year that the technology will ‘change the scope and speed’ of everyday work, and he has suggested that automation could shrink the company’s corporate workforce over the next few years — though he also expects it to create ‘other jobs.’

The memo also pushes back on the idea that every task needs an AI tool. If current tooling does not benefit someone’s work, Amazon says, that is ‘a finding, not a failing.’ Engineers who can explain where today’s tools fall short are demonstrating judgment rather than reluctance to adopt AI.

In May, it emerged that Amazon employees were using AI unnecessarily to boost internal usage scores, amid pressure to use the company's tools. Workers reportedly used MeshClaw, an internal AI agent capable of handling tasks such as deploying code and interacting with Slack, to inflate their token counts.

Notably, Amazon is not the only company rethinking how it spends on AI. Uber president and COO Andrew Macdonald said in May that higher token usage has yet to translate into useful consumer features. “That link is not there yet,” he said. The remark stands out because Uber’s CTO previously revealed that the company had burned through its entire 2026 Claude Code budget by April.

Accenture appears to also be wrestling with the same problem on behalf of its clients. Leaked audio obtained by 404 Media in June captures executives acknowledging that token spending is climbing fast as companies move from simple chatbots to agentic workflows, and that nobody can reliably predict what a given task will cost or whether it was worth the money. Accenture’s agentic AI strategy lead described it as ‘not a niche problem,’ and the firm is reportedly building an advisory tool called ‘Token IQ’ to help customers manage the bill. OpenAI CEO Sam Altman has previously acknowledged that AI token costs had become 'a huge issue', as companies faced mounting bills from heavy AI use.

Amazon, however, says it has not imposed token caps on its employees. According to CNBC, however, it shut down an internal leaderboard in May after workers were found gaming it to climb the rankings — an early example of the behavior the new guidelines are meant to discourage.

A team of around four dozen U.S. Marines is joining a NATO-led operation in the Baltic Sea to help protect undersea internet cables and gas pipelines in the area. In a press release, the U.S. Marine Corps shared its plans to deploy a "specialized team" to support NATO's Enhanced Vigilance Activity Baltic Sentry operation. The move comes amid a backdrop of growing undersea cable interest, movements of Russia's shadow fleet, and fears of sabotage.

The U.S. Marines are being sent from Marine Corps Base Camp Lejeune in North Carolina to provide "advanced expeditionary surveillance and tactical reconnaissance capabilities" to the region in support of Finnish and other NATO allies. Marines from 2nd Reconnaissance Battalion, 2nd Marine Division, and the II Marine Expeditionary Force Information Group are joining the mission.

Capt. Johan Tillander, Chief of Operations for the Finnish Navy, explained that the "experience, skills, and systems of the U.S. Marine Corps are perfectly suited" for the operation, including "monitoring of the Baltic Sea."

While the press release stresses that the Marines are serving in a "supportive, enabling capacity," with European nations taking the lead, the Marines are there to provide "highly specialized, agile force multipliers."

Maj. Gen. Daniel Shipley, Commander of U.S. Marine Corps Forces, Europe and Africa, shared that its participation shows its ability to "rapidly project specialized capabilities where they are most effective," and as a result, "ensuring we can operate seamlessly as a cohesive, lethal team to deter hybrid threats" in the area.

The U.S. deployment means that the operation will be able to "leverage... unique capabilities of Reconnaissance Marines, including small Unmanned Aerial Systems and littoral sensors, to feed vital data directly" to NATO.

The Baltic Sea is the frontline in potential assaults against European undersea infrastructure. Several attempts have been made to sabotage cables in the area, and two cables were cut between Finland, Sweden and Europe back in 2024. Meanwhile, Finland's own concerns about shadow fleet operations led it to set up a dedicated maritime surveillance center earlier this year to monitor activity in the Baltic Sea.

Russia isn't mentioned in the U.S. press release, but the authoritarian nation is one of Europe's biggest threats in the region, and has had worsening relations with European nations for several years, particularly since it went to war with Ukraine in 2022. Over 500 incidents of Russian fishing boats loitering near undersea cables in Europe have been seen in the last four and a half years, showing a persistence in mapping and identifying, if not directly interfering, with global data links in the area.

Solidigm, an American subsidiary of SK hynix, one of the biggest flash chip makers on the planet, is reportedly in talks with the Commerce Department over plans to invest $45 billion in a new plant to be built in the United States. But, Bloomberg reports, there's a catch — Solidigm wants to be given the go-ahead to export equipment to upgrade its existing plant in Dalian, China. The company has purportedly told officials that the upgrades would allow the Chinese plant to then help finance the cost of its U.S.-based counterpart, according to unnamed sources. It's as-yet unclear how the proposal was received, or whether the Commerce Department is inclined to agree to the deal.

The U.S. government has been keen to drive at-home manufacturing wherever possible, including for chips used in its computers and data centers. National security is regularly cited as a key reason behind the move. It's also the reason for the government's stance on exporting technologies to China, including the equipment that Solidigm says will help upgrade its Dalian plant. Similar export restrictions are in place for high-end AI data center components, with a black market having sprung up to take the place of sanctioned sales. Earlier this month, it was reported that some of Nvidia's powerful Blackwell Ultra B300 GPUs are being used in China Telecom data centers, with no explanation of how they got there. A 2025 report claimed that a cool $1 billion worth of Nvidia AI chips had been smuggled into China, a figure that is sure to have increased significantly as 2026 draws to a close.

In this case, Solidigm's planned Chinese factory upgrades could be mutually beneficial, allowing funds to be transferred to the new U.S. plant. But that requires the U.S. government to give it permission to ship the machinery in the first place, something that is far from guaranteed. The Dalian factory produces NAND flash memory, the likes of which are used in everything from relatively simple devices to full-blown data centers. It also comes at a time when global NAND production is a hot-button topic. Samsung and SK hynix are upgrading their Chinese factories to produce more of the chips used in AI data centers in an attempt to meet growing demand. Manufacturers have already slashed production of consumer-focused memory in favor of the more lucrative data center-grade components, causing prices to spike globally. This volatility in the market as a whole was given as one of the reasons SK hynix chose to cancel a planned Solidigm IPO earlier this week.

Beyond dangling a $45 billion plant, Solidigm has also allegedly told the Commerce Department that its total investment in U.S. manufacturing could reach $100 billion eventually, with construction in New York penciled in for 2027. The company is fresh from its latest earnings call, where it announced revenues of $75 billion with $60 billion in operating profit, and $28 billion in net. The company made no mention of its Dalian plant or a planned US investment, but did warn that memory shortages are set to last through 2030.

Taiwan’s Ocean Affairs Council is working on an autonomous system that would allow drone fleets to coordinate and execute missions without human intervention in response to "grey-zone" activities carried out by China. According to the South China Morning Post, the council is working with the National Cheng Kung University, which has already completed simultaneous flight simulations of two drone fleets, with each fleet comprising 10 units. This is already the second step of the process, which has shifted from individual aircraft control to the coordination of two drone detachments.

This project is designed to help Taiwan’s Coast Guard Administration (CGA) to counter Beijing’s grey-zone tactics in the waters surrounding the island nation. There have already been instances of civilian Chinese ships damaging undersea cables that Taipei relies on, with the CGA increasing its defensive patrols around 24 undersea cables and closely monitoring nearly a hundred China-linked boats. Despite that, Chinese “research vessels” continue to loiter over Taiwan-U.S. undersea cables, mimicking the activities of Russian trawlers in the Atlantic.

“The next stage will involve hardware-in-the-loop testing, prototype integration and flight trials over actual maritime areas to assess reliability under different sea and weather conditions,” the council said. The researchers programmed the drones to have a minimum separation of 100 feet in 3D space, and they’re also incorporating leader-follower formation flying and dynamic obstacle avoidance. This should allow them to respond to quickly changing conditions out in the open sea and even deal with unexpected hazards.

More than that, the autonomous system Taipei is working on would let different types of drones coordinate their movements, divide tasks, and plan routes, primarily for maritime surveillance. This is especially important as it would allow them to use different types of drones to accomplish various missions, much like how different air forces use different aircraft types.

The project is currently focusing on three Western-made drones — the Jump 20, V-BAT, and Flexrotor. In fact, the Ocean Affairs Council just received the V-BAT, a VTOL drone built by U.S. defense firm Shield AI. This unit can take off vertically from the relatively constrained deck space of non-aircraft carrier ships and offers an endurance of over 12 hours for extended missions, making it ideal for the CGA, as it would not need to make major renovations to its ships to accommodate these drones while extending their reach.

Nvidia and Oracle have announced general availability of Nvidia’s Vera Rubin NVL72 rack and Vera CPUs on Oracle Cloud Infrastructure (OCI) Superclusters at Oracle AI World on Tuesday. The announcement comes hot on the heels of Google Cloud announcing general availability for Vera Rubin last week, putting two major hyperscaler feathers in Nvidia’s cap. However, Oracle’s announcement is interesting, specifically calling out standalone Vera instances on OCI Superclusters. We've included a few benchmarks we've performed on a Vera server for a forthcoming review to highlight some of the chip's performance characteristics, but do note that these tests were not conducted on OCI infrastructure.

Neoclouds have been banging the drum of Vera for the past several weeks, with instances under way at Nscale, CoreWeave, and Nebius, among others, so it makes sense to see Vera in the GPU-dominant and AI-focused neoclouds. A standalone deployment at a hyperscaler like Oracle sends a different signal, however.

Although hyperscalers maintain large fleets of hardware, sometimes even renting from the neoclouds, the CPU focus is primarily on x86, with OCI Superclusters otherwise offering the Ampere Altra for an Arm-based instance. It’s Nvidia’s first standalone CPU push into the hyperscale data center, placing Vera alongside the x86 competition regardless of the GPU compute.

Vera is not meant to upend the x86 guard, at least in isolation. Ian Buck, Nvidia’s mind behind CUDA and the face of Nvidia’s AI infrastructure buildout, said in March — when the chip was revealed — that the world couldn’t be served by a single, 88-core SKU. However, Vera is opening the door at hyperscalers like Oracle to expand in future CPU generations.

AMD and Intel maintain the lead when it comes to single-socket density, which is only set to expand in the next generation, as AMD’s upcoming EPYC Venice CPUs can pack up to 512 threads in a single socket. However, Vera offers some distinct advantages due to its topology depending on the workload, many of which are becoming increasingly relevant in the agentic AI era.

Memory bandwidth, for instance, is one of Vera’s strong suits, and that’s not just a marketing claim. Nvidia has said the memory system is capable of 1.2 TB/s of bandwidth, but we can see Nvidia’s lead outside of theoretical numbers in our testing below.

Take the industry-standard Stream benchmark as an example. Nvidia’s use of LPDDR5X-9600, along with a monolithic die and distributed mesh for caching, allows it to more than double the bandwidth available on a dual Intel Xeon 6980P system, even when it’s paired with MRDIMMs sporting speeds of 8,800 MT/s. We’re showing triad here, as it’s the most “demanding” of Stream’s operations, but the leads for Vera are similar across add, copy, and scale as well.

This bandwidth is critical for some types of agentic loops, particularly those dealing with large database queries or extensive documentation. You can see the OpenMP “spread” policy offers a massive boost to Vera’s memory bandwidth, as well. Both AMD and Intel maintain centralized memory control and system-level cache, while Vera distributes it in a mesh across the die. The result, at least in OpenMP set properly, is much higher memory bandwidth by binding processes as far away from each other as possible.

Another critical performance vector for agentic AI is Python. The agentic world is expanding, but much of the foundation of the AI software ecosystem is built on Python. Common agentic frameworks like CrewAI and LangGraph are written in Python, and although we’re seeing more frameworks built in languages like Rust and TypeScript, Python performance is still critical to agentic loops.

And, unsurprisingly, it’s an area where Vera is also strong, beating out CPUs that are much more powerful on paper. Popular agentic harnesses are built on Python, but so are many of the applications that also run on Python, including foundational AI libraries like PyTorch and TensorFlow.

Nvidia’s focus on agentic AI makes sense given the times we’re living in, but it can miss the forest for the trees. Oracle bringing Vera standalone instances online shifts the focus onto the broader implications of Vera’s entrance into the data center. There are some aspects of performance that characterize agentic AI workloads, but many agent workloads are regular data center workloads in the first place, just under high concurrency.

For instance, SQLite performance is important for any server querying a large database, regardless of its written by logic, a person, or an AI agent. Nvidia’s second-gen Scalable Coherency Fabric (SCF), which we saw at play in Stream, shows up in a real workload here, with these database queries largely bound by data movement across the system.

And that is what makes Vera so interesting in the context of Oracle’s announcement. It’s a CPU built for agentic AI, says Nvidia, but the crosstalk with general data center workloads is strong. It’s not just a CPU to use because you need access to Nvidia’s accelerators. Depending on the workload, it may even be a better choice than the x86 guard.

Ultimately, we have to see how customers of these hyperscalers use (or don’t use) Vera in production. It’s demonstrated its usefulness in agentic AI loops, but it’s also showing early promise in certain general-purpose workloads. Oracle’s standalone deployment is a sign that there’s some demand for Vera that isn't exclusively tied at the hip of a Rubin GPU.

Oracle is approaching cautiously, however, as any hyperscaler should be when bringing up a multi-gigawatt deployment. The company said it will also offer Nvidia’s HGX NVL8, bringing eight Rubin GPUs together in a cluster, connected to a dual, 64-core Intel Xeon 6 host with 4TB of DDR5 memory.

Our testing of Vera shared in this article is part of a larger benchmarking session we had with the chip, and we'll share the technical details of the test setup in our forthcoming review. We’ll be publishing our full results shortly on Tom’s Hardware.

A scary-looking humanoid robot with a chainsaw arm was shown in action on Japan's ANN News recently. The rail truck-mounted mech was being used by Japanese railway firm JR East to clear tree branches that had begun to impede rail tracks. The unnamed heavy machinery robot’s chainsaw and grabber arms are remotely controlled by a human in this instance. However, JR East is broadly integrating AI and robotic technologies across its operations to enhance safety, efficiency, and round-the-clock maintenance.

In the future, this robot could become autonomous with AI-based image analysis for recognizing rail obstructions that need to be addressed. While they’re at it, why not dispose of the wheeled vehicle and add some legs and a jet pack?

JR East is the largest passenger railway operator in Japan. It is famous for operating the northern Shinkansen (bullet train) routes, but also has business interests/revenue from shopping centers, hotels, smart cards, and more. The firm has a reputation for quite aggressive AI adoption, including company-wide AI tools, designing AI programs for disaster recovery, and it is even using drones equipped with AI image analysis for rail infrastructure inspections (PDF).

The source video and report (machine translation) indicate that the chainsaw mech you see has been available to JR East engineers since March, but the TV crew took the opportunity to catch a rare glimpse of this heavy machinery in action. JR East is turning to robots like this due to labor shortages. Apparently, operating a robot like this requires some skill, but so does chainsaw work at height.

There’s also been a recent shift in rail building and maintenance work to the daytime – despite some inconvenience to customers – to the benefit of workers and safety. Perhaps that’s why we now have a video of the chainsaw-wielding truck-mounted robot for the first time.

The combination of labor shortages and advances in robotics and AI makes development of this kind of rail maintenance robot seem very timely. If the metallic humanoid can be made autonomous, it could save a lot of labor and promote safety. What could go wrong?

A University of California, Berkeley (UC Berkeley) professor has stopped giving his students graded homework, saying that rampant AI use has made it pointless. According to the New York Post, renowned mathematician Richard Borcherds said that it’s now nearly impossible to tell the difference between a student-written submission and one that’s been created using artificial intelligence. Because of this, he now primarily uses in-person exams with no electronic devices to grade his students and see how well they are doing in his class.

“Online exams are now a waste of time as several students just use AI to answer the questions. I cannot tell which students are using AI, but I notice the average score increases a lot in online exams,” Borcherds told the publication. He also added, “Grading for correctness just adds a huge incentive for them to use AI for homework instead of doing it themselves and penalizes students for honesty. And paying to grade AI generated answers is a waste of time and money.”

This problem appeared when OpenAI released the first popular LLM, GPT-3.5, in November 2022. But while LLMs were prone to mistakes back then, making their inordinate use easy to spot, the technology has improved to the point that it’s now very hard to tell if it's been used for written work.

Various researchers have proven that students routinely get better scores when taking exams online. While they haven’t definitively proven that the students used AI to answer the questions, the difference between the scores was significant enough that the researchers concluded unmonitored online exams are no longer reflective of what the students have learned in the course.

A Brown University professor determined this earlier this year when his students scored an average of 96% for his take-home midterms — a huge jump from the historical average of 65% to 80%. Because of this, he made his students take their final exams in person, which resulted in only two students getting a score that’s within 10% of their midterm grades, with only one performing better.

However, not all AI tools are as keen to help students cheat, with Cursor AI telling one user to learn coding instead of asking it to generate code. Still, this is just one tool in a sea of many, making it next to impossible to stop students from using AI.

As for UC Berkeley, it has already looked at AI-detection software to try and deal with the situation, but determined that it wasn't reliable. Because of this, the University decided to give professors free rein on their own AI policies, as long as they spell out their AI rules in the course syllabi. As for Borcherds, he says that if the worst thing that AI can do is to make it harder for teachers to grade student homework, then he will be “very relieved."

Discover-Tech, an Israeli advertising technology company, has reportedly been linked to a suspected surveillance effort aimed at Starlink users, according to Reuters. The report, which cites six people familiar with the matter, claims the firm was the top bidder for ad space tied to Starlink traffic in more than 100 countries. Researchers suspect the aggressive bidding was meant to reveal where users are located or to enable further monitoring — though the findings do not show that anyone was actually tracked.

The data comes from TrendAI, the research arm of Trend Micro. Its October 15 report referred to the activity as ‘Shadow-Earth-79’ without naming a company, but the report claims the codename belongs to Discover-Tech. TrendAI says its attribution process is designed to avoid ‘hype-driven labels’ that could mislead security teams. “Our approach reduces the risk of misattribution, hype-driven labels, and unstable actor names that can mislead defenders, distort risk prioritization, and trigger inappropriate response actions.” The report says.

The method exploits how online ads are sold with advertisers bidding in real-time auctions for space on websites and in apps, and each bid request can carry details about the person on the other end. Collected over time and combined with other data, that information could be used to identify people, follow their activity across the web, or select targets for spyware.

The geography is what raised eyebrows. Shadow-Earth-79 was reportedly the largest ad buyer in Syria, Libya, Yemen, and South Sudan, and the second largest in Afghanistan, while also claiming unusually large shares of bids in Hungary, Israel, Egypt, Azerbaijan, and Venezuela. TrendAI says the pattern points to ‘targeting and tracking infrastructure rather than a commercial advertising operation.’ That is plausible, given how little legitimate Western advertising demand exists in sanctioned or war-torn markets.

Notably, Discover-Tech was already under scrutiny. On the same day as TrendAI’s report, Citizen Lab and ad security firm Confiant separately named the company, alleging that it used ad auctions to collect data on users across sensitive networks, including those tied to media organizations, governments, and tech firms. “The malicious advertisements were purchased by Discover-Tech, an Israeli company that presents itself as a legitimate advertising company, but which we found operates as a facilitator for multiple mercenary spyware partners.” Citizen Lab wrote.

Google and Comcast have since said they removed Discover-Tech from their advertising platforms, while the report says that a major U.S. technology company and an international organization have opened internal inquiries into whether the firm’s ads targeted their employees.

For Starlink, the stakes are obvious — the service works outside conventional telecom networks, which makes it a lifeline where governments restrict internet access, and makes its users an attractive target for those same governments. In April, Iran’s nationwide internet blackout passed 1,000 hours, with authorities reportedly going after Starlink terminals and imposing severe penalties for possessing one. Earlier this month, SpaceX also limited Starlink service for Ukraine’s high-speed military drones, which Ukrainian officials say complicates logistics and operations.

Advertising and consumer tech have been abused for tracking before. A backdoor was previously found in a popular iOS ad library that could collect device-location data and capture screenshots, while Chrysaor, Android spyware linked to the Pegasus family, was discovered by Lookout and Citizen Lab on a small number of devices. More recently, a researcher claimed to have found extensive data collection, including ad-related telemetry, inside a Samsung smart TV’s Tizen operating system, though those findings are still being investigated.

That said, it is still unclear if Discover-Tech actually used Starlink-related data to single out specific subscribers remains unverified, and neither the company nor SpaceX has commented.

TSMC is weighing raising prices for its Chip on Wafer on Substrate (CoWoS) advanced packaging technology by as much as 20%, reports Trendforce, citing supply chain sources and local media. The increase reflects strained capacity amid AI demand and longer delivery times for chipmaking tools. Such a hike would be significant, as CoWoS is already one of the tightest links in the AI supply chain. TSMC has yet to confirm the report.

CoWoS combines logic dies and high-bandwidth memory (HBM) stacks within a single package, which is how AI accelerators get the data bandwidth that demanding workloads require. Building these packages is only getting harder as chip designers cram in more compute and memory. Nvidia’s Blackwell already pairs two large GPU dies with eight HBM3E stacks, whereas the upcoming Rubin and Rubin Ultra are expected to push integration further, which will increase pressure on packaging and the supply chain that supports it.

Testing is turning into another bottleneck. Blackwell reportedly takes three to four times longer to test than Nvidia’s previous-generation Hopper, and Rubin could need as much as seven times longer. Extended test cycles keep equipment occupied for longer, which makes it harder to increase output as demand grows. GPUs are not the only products competing for capacity either, as server CPUs and custom ASICs also need advanced packaging, adding more strain to an already crowded supply chain.

Tools are also taking longer to arrive, as equipment lead times have reportedly stretched from three to six months to as long as eight months, making timely access to fab equipment increasingly important. As a result, TSMC has reportedly placed this year’s equipment orders at existing prices instead of negotiating discounts or extra features — a departure from its customary approach of using its purchasing power to secure better terms. Essentially, getting the tools on time now appears to matter more to the company than getting them cheaply.

The squeeze echoes a report from September, which said that TSMC’s projected semiconductor equipment requirements had almost doubled in six months as the company expanded production to meet AI demand. Cliff Hou, deputy co-COO at TSMC, said that by July the company’s expected equipment needs had reached 1.9 times its original estimate as AI demand accelerated. “Compared with the pattern 30 years ago, we have never seen the demand increase that much, that frequency of changes,” Hou said at the time.

TSMC is also trying to broaden its packaging supply chain. Earlier this month, GlobalFoundries agreed to produce silicon interposers for TSMC’s CoWoS in the U.S. under a five-year contract worth $2 billion, with production expected to ramp up in the first half of 2028. However, interposers are only one part of the packaging flow, so the agreement alone will not eliminate shortages.

Higher costs are not limited to packaging. TSMC is reportedly considering broader price increases for 2027, with wafer prices on advanced process technologies expected to rise by 5% to 10%. Customers using 2nm and 3nm-class nodes could face the steepest hikes, whereas those seeking capacity beyond their original allocations may have to pay additional premiums, which could push some costs up by as much as 25%.

Industry estimates cited by Trendforce put TSMC’s monthly capacity at over 106,000 wafers in 2026, with a target of 227,000 wafers in 2027 as the company works to keep pace with demand for increasingly complex AI processors. If both the packaging and wafer price increases materialize, the cost of AI hardware will likely climb further, and some of that will inevitably be passed down the chain.

Police are increasingly using AI-derived pretext to justify pulling drivers over, claims a report by the American Civil Liberties Union (ACLU) on October 26. According to Jay Stanley, a senior policy analyst with the ACLU Speech, Privacy, and Technology Project, police are using AI algorithms and automated license plate reader (ALPR) surveillance data to flag motorists whose travel histories appear “suspicious,” even without evidence of wrongdoing.

Condemning the approach and the use of mass surveillance in general, Stanley expressed concerns that this approach to law enforcement would lead to an uptick in the controversial “pretextual stops” and that “More and more people will be subject to police encounters as these AI algorithms inevitably sweep in far more innocent people than actual wrongdoers.” The report details several incidents, including one in which a woman claims an officer stopped her during a cross-country road trip because a surveillance system he was using had flagged her movements as suspicious. The system has accessed ALPR data on where she had been driving.

ALPRs installed along roads and highways photograph passing vehicles and record their registration numbers, locations, and timestamps. By combining sightings from different cameras, law enforcement can determine where a vehicle has traveled, which routes it used, and how frequently it visited particular locations. An AI system with access to the surveillance data can quickly analyze huge amounts of data and identify patterns it has been trained to deem as “suspicious,” triggering an alert to law enforcement. Stanley implies officers may receive the alert and look for a minor traffic violation to justify pulling the driver over. In some documented cases, officers have concealed the actual reason for the stop.

The report comes amid growing concerns about the impact of massive mass surveillance camera networks, such as privacy, abuse, and limited regulation. Several officers have been fired for abusing the Flock surveillance system. License plate readers have existed for years and have traditionally helped police track known suspects or locate a vehicle already associated with a crime, such as a stolen car. However, the current scale and blanket surveillance are unprecedented. Security researchers disclosed a network of over 335,000 Flock ALPR cameras across the United States.

Using AI, systems can also now analyze large collections of vehicle movements and generate new suspects based on behavioral patterns. Companies such as Flock Safety have been developing AI-enhanced surveillance systems that can identify “unusual” travel behavior and associations between vehicles, marketing them as effective crime-prevention tools.

Independent 3D modeling community Open3DLab has publicly warned that a web crawler, linked to the gaming platform Roblox, mapped "almost every project" on its website, along with sister sites SFMLab and SmutBase, downloading 145.5GB of data in the process. In a blog post, Open3DLab reports that "complete copies" of 1,027 files from 712 projects were sourced in the process, which it believes were taken for "AI training purposes."

The blog post goes into the specifics of the incident, along with some changes that Open3DLab has made since to try and avoid a repeat. According to the post, Open3DLab's staff noticed a crawler "making a large number of requests to our sites and download infrastructure" on October 18. The crawler used trellis2-data-sourcing/1.0 (3D foundation model research; [redacted]@roblox.com) as its user agent, with the data coming from a single Amazon AWS address. The crawler didn't respect the site's robots.txt file or Terms of Service, which explicitly forbid scraping. It also, according to the report, ignored creator licenses for the projects, which prohibit non-commercial use.

The crawler read "practically [its] entire catalogue" of 18,926 projects in a span of 34 hours over a total of 26,429 requests. Once it was done, the crawler downloaded 1,027 files from 712 projects before it "suspended operations." Files downloaded include 364 files, or 54.8GB worth, from 224 projects hosted on SmutBase, OpenLab's sister site focused on adult content. A further 345 projects of adult-marked SFMLab projects, or 62.6GB worth, were also taken.

According to Open3DLab, the crawler's "main limiting factor" was its use of free download servers that capped speed and concurrent downloads. The scraping was detected on the same day, and blocked by the following morning. Further download attempts were detected on October 19, "neatly within office hours, only briefly pausing for an hour around lunchtime," and despite being blocked, "did not reduce its download attempts."

A final attempt on October 20, with requests tapering, "likely in an effort to test... rate limiting," was made, with requests stopping completely later that day. Open3DLab made another request at this point for comment, although without reply.

The Roblox user agent mentions "3D foundation model research." As Open3DLab's blog post notes, Roblox has been building such a model since March 2025, when it introduced Roblox Cube, its "core generative AI system" for 3D modeling for character rigging, with characters and avatars listed as future goals. Further AI tech that could build in-game models using natural language prompts, called "4D generation," was beta launched in February 2026. The email address listed in the user agent matches one of the researchers listed on the Roblox Cube research paper, according to Open3DLab, although it redacted the name here.

Noting Roblox's "track record," Open3DLab's "most likely explanation" is that the crawler was collecting "training or evaluation data" for its 3D foundation model, although it notes that it can't prove it.

Roblox is a gaming platform and game creation system with over 33 million users and billions of "game experiences" for its player base, 40% of whom are reportedly aged under 12 years old. Open3DLab is clear to state that it "doesn't think Roblox set out to collect adult content" but that "as far as [its] logs show, nobody checked," which, it believes, is "not a good position" for Roblox to be in.

On October 21, during GTC Berlin, Nvidia announced several massive, multi-billion-dollar deals with supply chain partners across substrates, advanced packaging, semiconductor testing, and optical engines to expand production capacity. Aimed at overcoming severe bottlenecks restricting the current and future AI infrastructure buildouts, the deals feature commitments and production roadmaps stretching through 2028, securing capacity for Nvidia and ensuring resiliency and redundancy in its AI accelerator supply chain. “We are in the midst of the largest infrastructure buildout in human history. Demand is extraordinary. Capacity is the constraint,” said Nvidia CEO Jensen Huang. “Our supply chain partners are the world’s best manufacturers. I’m grateful they are expanding capacity around the globe to scale production and help us deliver the AI supercomputing systems powering this new industrial revolution.”

The deals cover two major AI accelerator supply chain categories: hardware interconnects and OSAT (Outsourced Semiconductor Assembly and Test). Under hardware interconnects, Nvidia is securing capacity for both substrates and co-packaged optics. Nvidia has signed deals with Ibiden, Unimicron, and Samsung Electro-Mechanics — three of the world's five largest substrate manufacturers — to support its production capacity expansion and reserve significant portions of that capacity for the chipmaker. The company says it plans to secure 2.5x dedicated substrate capacity over the next three years. Nvidia has also signed agreements with SPIL (Siliconware Precision Industries Co., Ltd) to expand manufacturing capacity for the optical engines used in co-packaged optics by 15x by mid-2027.

On the OSAT side, Nvidia and Amkor are targeting a 300% increase in interposer wafer capacity by the end of 2028 through expanded advanced packaging capacity in South Korea, to support higher production volumes for Nvidia's platform. The company also plans to double its semiconductor testing capacity by 2027, investing $1.1 billion to reserve capacity from KYEC (King Yuan Electronics Co., Ltd.), the world's largest semiconductor testing company, which currently handles nearly all of Nvidia's chip testing.

Nvidia currently leads the AI chip market by a considerable margin, with its extremely powerful accelerators sitting in a vast majority of AI data centers globally. These deals target major bottlenecks in the AI chip supply chain, suggesting Nvidia doesn't plan to relinquish its position soon. So, why these specific deals? What are the exact details? What's the broader play? And where do Nvidia's competitors and the rest of the AI industry stand?

Uncorking bottlenecks

The substrate base, a specialized intermediary layer in every accelerator package that connects the tiny silicon chips above it to the board below, is one of the most critical bottlenecks in the AI chip supply chain. As we detailed in our comprehensive substrates roadmap, this indispensable physical layer, manufactured by only a handful of companies, is expanding in both the X-Y and Z axes as chip makers package an increasing number of components onto one board. Nvidia's Blackwell generation places two reticle-sized GPU dies and eight HBM3E stacks on a single package. Its Rubin and Rubin Ultra architectures mount even more components. As a result, each substrate now consumes much more material and production capacity, even as demand for individual substrates soars.

Fubon Research estimates that Nvidia would supply around 12.4 million data-center AI GPUs in 2028, up from a projected 8.2 million in 2026. Each one would feature a substrate layer. Nvidia is staying ahead of the curve, with a planned 2.5x capacity expansion over the next three years, led by Ibiden. Earlier this year, the company signed a deal to provide a multibillion-dollar capacity reservation advance to support Ibiden's Ono Plant in Japan, targeting 3x the plant’s current individual capacity.

To reach its 2.5x target, Nvidia has also brought on Samsung Electro-Mechanics, Samsung's substrate-manufacturing subsidiary, as a supplier. Beyond the immediate capacity from Samsung Electro-Mechanics’ existing operations, the deal will support facility upgrades and new plants in Busan and Sejong. Nvidia is projected to account for a significant portion of the substrate manufacturer’s output by 2028. To build even more redundancy, Nvidia also signed a capacity reservation agreement with Unimicron, supporting the latter's substrate manufacturing expansion in Taiwan. The expansion will reach full capacity by mid-2028 but is expected to start providing capacity by Q2 2027.

Next on Nvidia's to-expand-capacity list announced on October 21 is co-packaged optics. As next-generation speeds surpass 224 Gbps per lane, traditional copper traces on the board consume considerable power to move data even a few inches from the processor to the edge of the board. To address this, the industry is moving the optical engine into the accelerator package itself via co-packaged optics. Nvidia had long detailed its plans for silicon photonics and co-packaged optics as the standard for future AI accelerator packages. The company has signed a deal with Taiwanese optical engine manufacturer SPIL to expand production capacity by 1,400% by mid-2027, with plans to double again by the end of 2028.

A 4x increase in packaging, double the testing, and $1.1 billion specifically for U.S. facilities

Interestingly, the world's biggest semiconductor company, Nvidia, doesn't own a single Fab. After TSMC manufactures the logic chip and binds it with memory chips on a silicon interposer, final-assembly companies such as Amkor package that integrated module with the substrate and protective casing into the final accelerator package. These final steps are increasingly incorporating co-packaged photonics chips. Nvidia’s deal with Amkor aims to expand the latter's interposer wafer capacity by more than 300% by the end of 2028, through advanced packaging capacity expansions in South Korea. Amkor confirmed the deal is worth $1.3 billion, bringing Nvidia's total investment in Amkor across its Arizona and South Korea facilities to $3 billion.

After final assembly, finished Nvidia AI accelerator packages are shipped to KYEC for extreme testing. KYEC currently handles nearly all of Nvidia's chip testing. It therefore comes as no surprise that Nvidia expects semiconductor testing capacity to double from calendar 2026 to 2027 through buildouts across KYEC’s facilities in the U.S., Taiwan, and Singapore. On October 13, Nvidia signed a deal to pay KYEC $1.1 billion to reserve testing capacity at KYEC's new $1.4 billion U.S. facility, where high-volume production is slated to begin in April 2027. The funds will build domestic testing capacity and bolster the domestic supply chain. At the same time, KYEC is aggressively expanding operations in Singapore, where it expects high-volume production to begin in February 2027.

The bigger picture

Nvidia has been aggressively securing production capacity and long-term supply commitments to meet anticipated demand for its next-generation AI accelerators. According to its August 2026 financial disclosures, the company had accumulated $279 billion in outstanding supply and manufacturing-capacity commitments by July 2026. The company has since signed several additional deals. These commitments cover future component purchases and manufacturing capacity reservations, particularly for memory and chip production. Combined with its more recent agreements, Nvidia is securing nearly every stage of its AI accelerator supply chain, from semiconductor fabrication and HBM memory to substrates, advanced packaging, optical interconnects, final assembly, and testing.

Of course, the rest of the industry isn't standing by. On May 21, 2026, AMD announced over $10 billion in investments across Taiwan's semiconductor ecosystem, specifically targeting advanced packaging and manufacturing partnerships. On October 6, CEO Lisa Su confirmed that AMD intends to increase those investments further. The company is working with TSMC, packaging partners, substrate manufacturers, and memory suppliers to increase production substantially in 2027. Google and Amazon are also aggressively expanding their custom TPU and Trainium accelerator programs.

Intel, meanwhile, is positioning its foundry and advanced packaging operations as alternatives to TSMC's increasingly constrained production capacity. The company is expanding its EMIB-T packaging technology, with reports indicating that Google has already reserved Intel capacity to package more than three million TPUs in 2028. Intel is also advancing its 18A and 14A manufacturing processes as it attempts to attract major external customers.

Samsung is expanding its semiconductor operations as well. The Korean company, which held approximately 5.9% of the global foundry market by revenue in Q2 2026, is ramping up its 2nm manufacturing technology and preparing its Taylor, Texas, facility for volume production in 2027. As detailed in our comprehensive Samsung foundry roadmap, the company is investing heavily in its manufacturing operations as it attempts to close the considerable gap with TSMC.

Perhaps the most ambitious player is new entrant Elon Musk's Terafab, a joint semiconductor manufacturing project involving Tesla, SpaceX, and xAI. The project aims to eventually manufacture the equivalent of one terawatt of AI computing capacity annually, an extraordinarily ambitious target given the industry's current scale. Intel joined the project in April, and Musk later confirmed plans to use the company's 14A fabrication technology. The project aims to deliver its first experimental chips in 2027 and begin high-volume production by the third or fourth quarter of 2028.

Nvidia and other semiconductor firms are continuously fighting against emerging bottlenecks, but compute doesn't come online overnight, and fabs aren't built in a day. So, while almost every inch of the semiconductor supply chain remains constrained, it remains likely that big-name players will continue to invest heavily to ensure that products continue to flow into neoclouds and data centers alike.

Microsoft has offered voluntary buyouts to around 60,000 employees across the United States, Canada, India, and the United Kingdom. The employees, who must have an experience level of 70 or below, make up around 27% of the company's total workforce of 223,000 people, a CNBC report notes. This move comes hot on the heels of a similar move in April when the company announced its first-ever voluntary retirement program. That program was open to around 7% of the company's U.S.-based workforce, but it pales in comparison to this latest move. That time around, more than 30% of those eligible for the program took advantage of it.

The "Voluntary Exit Offer" was announced via a post on the Microsoft blog, penned by Amy Coleman, Microsoft's executive vice president in charge of people. "The Voluntary Exit Offer reflects an important part of how we’re evolving our approach to change: creating greater flexibility and, where possible, more ways for employees to shape their next chapter," Coleman explained, going on to say that the move "builds on what we learned" following the retirement program that was launched earlier in the year. Operationally, Coleman added that those who are eligible for the new program will receive an email with all of the relevant details today.

While the blog post points out that the Voluntary Exit Offer is a time-limited, voluntary program, the fact that it comes so soon after April's retirement program is notable. Companies run programs like this when they are looking for ways to slash their wage bills, something that Microsoft knows all about. The company is in the midst of an Xbox "reset" that includes a plan to cut 3,200 jobs by the end of the year. The company has also seen some studios close, with others merging and yet more being cut adrift to pursue life as independents.

As for why Microsoft has chosen now to make cuts, all fingers point to its ongoing AI data center plans. Last year, it announced plans to open the "world's most powerful" AI data center. Fairwater, in Mount Pleasant, Wisconsin, is now operational, with Microsoft lauding its closed-loop cooling system amid ongoing concern over the impact of data centers on local communities. Since 2023, Microsoft, Amazon, Google, and Meta have spent more than $1 trillion on AI infrastructure. AI data center investment is expected to skyrocket to $32 trillion by 2050, with Microsoft among those spending big.

All of this money has to come from somewhere. For Microsoft, at least some of that looks set to come from global salaries. It's unlikely this will be the last cost-cutting measure put forth by Microsoft execs, with Coleman adding that this voluntary exit program is just one "of the ways we’re trying to meet different business needs while continuing to create opportunities for our own people to grow and contribute."

Claude maker Anthropic has filed for an initial public offering (IPO) that could value the AI company at around $2 trillion, setting the stage for one of the biggest potential stock market debuts in the technology industry, the company announced today.

As per Form S-1 registration statement with the U.S. Securities and Exchange Commission, the firm intends to list its Class A common stock on the Nasdaq Global Select Market under the ticker ‘ANT’, although the number of shares to be offered and the price range have yet to be determined. A report from Reuters suggests the company could seek to raise as much as $100 billion, but those figures are not confirmed offering terms.

On to numbers, Anthropic generated $16.3 billion in revenue during the first half of 2026, up sharply from approximately $1.2 billion in the same period last year. The company’s operating performance also improved, although its reported net loss reached $80.7 billion. Despite the losses, it disclosed enormous commitments to AI computing infrastructure and outlined risks that investors will have to weigh against its rapid growth.

That net loss, however, requires context. It should not be interpreted as $80.7 billion in cash spent running the business, since accounting adjustments related to financing instruments contributed substantially to the reported figure. Even so, Anthropic cautions that its revenue growth rate “may not be sustainable” and could decline.

The company also warns that it may need additional financing sooner than expected and does not anticipate paying dividends in the foreseeable future. Those disclosures will matter to investors considering a company whose expansion depends on access to increasingly expensive computing resources.

Beyond figures, Anthropic’s filing contains warnings about the technology underpinning its business. The company says advanced AI could pose “catastrophic or existential risks to humanity,” including scenarios involving dangerous capabilities, deceptive behavior, and systems that become difficult for humans to control.

“Our development of highly advanced models, platforms, and applications and expansion of use cases could further increase the risk that our models cause harm,” Anthropic’s S-1 prospectus stated.

However, the firm said it is developing safeguards to mitigate AI risks through pre-release testing, red-team assessments, and systems designed to detect harmful outputs. It also said it may delay training or restrict model releases when safety measures are insufficient, while acknowledging that its efforts may not succeed.

Bloomberg Intelligence technology research head Mandeep Singh discussed the disclosures in a Bloomberg segment following the filing. He characterized worst-case scenarios as typical of risk disclosures for emerging technology companies preparing to go public, noting that such language helps address potential liabilities. He also pointed to Anthropic’s improving compute efficiency and an improvement in gross margins as a positive signal for investors. “Gross margins are inching closer to 50%, whereas just a quarter back… it was like… sub 20%.” He stated.

Meanwhile, Anthropic expects to spend up to $120.5 billion on compute involving SpaceXAI and Nvidia, alongside up to $114.5 billion with Google through 2033. Notably, while those massive infrastructure commitments are essential to training and running advanced AI models, they could become a costly burden if demand falls short of expectations.

OpenAI, the maker of ChatGPT, has also been preparing for a potential IPO, doubling down on efforts to scale revenue and computing capacity ahead of a possible debut. However, OpenAI CEO Sam Altman said in September that the company would not go public in 2026, citing concerns about AI safety and saying the company had more work to do before pursuing a listing. Elsewhere, Cerebras, which develops wafer-scale processors for AI workloads, completed its Nasdaq IPO in May after filing for a second time.

Still, Anthropic acknowledges that its growth may slow and that its financing needs could arrive sooner than expected. That said, Anthropic is reportedly targeting a mid-November IPO, with November 19 or 20 floated as possible listing dates. However, the company has yet to confirm a timetable for the offering.

Have you been watching many heist movies lately? Here's a Hollywood-ready script for a new one: Organized crime teams have taken to stealing data center gear off or with their transport trucks, to the tune of tens of millions per shipment. They employ hacking and phishing, vehicular ramming, GPS signal spoofing, and have ties to international crime networks. An extensive report on NBC sheds light on this new trend, remarking that over $100 million dollars of computing and related gear were stolen via transit interception in various guises since June 2026 alone.

This type of heist has been growing over the past couple of years, but just this year the one incident in Reno made $38 million in gear disappear. In the same city, $15m of semiconductors and Apple products got nabbed in July 2025, according to Fortune. The month of June 2026 was particularly profitable for thieves, as they made off with $13m of gear around the Mojave Desert area, and $10m worth of server rack components headed to Texas.

The stolen gear includes servers, fiber-optic and copper cabling, RAM modules, drives, routers, and other miscellaneous gear. Although NBC doesn't mention GPUs by name, stolen data center servers headed for AI buildouts are likely to have some.

Even copper wire is worth a lot these days, as one snagged trailer had $300,000 worth of the now-precious metal. As Keith Lewis from Verisk Cargonet puts it, "it's like stealing pure gold bars [...] it's so much easier to steal [data center shipments] than to rob a bank."

The methods used employ technological smarts and brutality in equal amounts. Each heist is unique, but broad patterns are apparent. Thieves gain access to online accounts of freight brokers and carriers, and from there on they can look up cargo manifests and shipping schedules, as well as impersonate shipping companies. Then they track a shipment or issue a false one, and eventually collect the truck with the precious cargo.

In a few recent cases, the heists involved thieves literally ramming escort cars so that the truck executes a trouble-free getaway. In those cases, the truck driver was likely in on the play, and probably got in the driver's seat thanks to the prevalence of AI-generated IDs coupled with poor personnel vetting practices.

And as for GPS trailer tracking, the malfeasants thought of that too: in one case, they seemingly used GPS beacon impersonation and broadcast fake coordinates so that neither human nor automated system believed anything was amiss.

The first instinct is to assume that all of the gear would end up on the second-hand market. But in an additional twist to the story, the thieves already have overseas buyers ready before the heist actually takes place.

Law enforcement officials believe that the gear is slurped up by entities operating in countries under the U.S.' tight AI export controls, including China. Verified Carrier, a platform for vetting carriers, says that this year, it's seen suspicious activity linked with U.S. cargo theft in India, Uzbekistan, Pakistan, Colombia, and Moldova. Scott Cornell of TAPA Americas estimates that six to seven out of 10 stolen cargo shipments (data center gear or otherwise) end up overseas.

David Warrick, of supply management firm Overhaul, describes the scale of the problem: "These are international syndicates. They run like big corporations…They have penetrated the supply chains at all levels, in terms of drivers, in terms of warehouse operators. They know where all the high-value goods are; they know who transacts them, and they look for the weak points in the supply chain to be able to exploit those."

Elon Musk has confirmed that SpaceX closed $40 billion in debt financing for graphics processing units (GPUs), with the chips expected to arrive in the first half of 2027. He disclosed the deal on X, October 25 (embedded below) while responding to investor Chamath Palihapitiya, who had warned that mounting financing pressures and delays in data center construction could leave AI computing capacity sitting idle next year.

Musk agreed with that assessment, but argued that SpaceX is better positioned than many of its competitors because it can borrow against revenue from businesses outside AI. “Pretty much,” Musk replied to Palihapitiya on X explaining that debt markets had largely shut off funding for companies without substantial revenue independent of AI data centers.

He further said SpaceX had closed $40 billion in debt for GPUs because it has investment-grade status and revenue from its Starlink satellite internet and space businesses. “That means the 200+ neoclouds are going to see guppy-level death and consolidation. Some will make it through, but not many,” Musk wrote.

It is still, however, unclear what Musk meant by “closed” or whether investors had agreed to the final debt pricing. The disclosure comes as SpaceX prepares for another enormous spending period. CFO Bret Johnsen had stated during SpaceX’s Q3 2026 earnings call that the company expects capital expenditure of $35 billion to $40 billion in Q4 “because of the additional capacity coming online this quarter.”

Musk also clarified that the borrowing was primarily callable debt for GPU purchases expected to arrive in the first and second quarters of 2027. “This is primarily callable debt for GPU purchases that will arrive in Q1 and Q2. Not a major factor this quarter,” Musk wrote.

Earlier this month, it emerged SpaceX had proposed a $40 billion debt package for Nvidia hardware, estimating that the proposed funding could cover roughly 360,000 Rubin GPUs across 5,000 Nvidia NVL72 racks. The proposed package reportedly included $10 billion in bank loans and $30 billion in investment-grade debt. Musk has now said the financing has closed, but has not confirmed the specific GPU models or quantities covered by the deal.

The spending fits into Musk’s broader push to expand AI computing capacity. In August, xAI outlined plans to increase its data center capacity sevenfold by 2027, targeting 10 gigawatts. In September, the publication also reported that SpaceXAI planned to add another 660,000 Nvidia GPUs during 2026, potentially bringing its total to around 1.44 million if the planned deployments proceed.

And GPUs are only part of the bill. Data centers need power, cooling, networking equipment, and physical infrastructure before those accelerators can do useful work, which underscores SpaceX’s latest move.

SpaceX, meanwhile, has begun selling access to its own infrastructure. In May, Anthropic secured access to SpaceX’s Colossus 1 supercomputer, which includes more than 220,000 Nvidia GPUs and over 300 megawatts of computing capacity. Later in June, SpaceX struck a separate deal with Google, which agreed to pay $920 million per month for access to approximately 110,000 Nvidia GPUs and related hardware, with payments scheduled to run from October 2026 through June 2029.

Cadence, one of the largest electronic design automation (EDA) companies, has unveiled Maestro, its new portfolio of domain-specialized AI models tailored for its chip design agents, in a press release sent just ahead of earnings. In initial results compared with general-purpose AI models, Cadence claims up to a 30% improvement in the quality of results and that Maestro effectively uses fewer tokens for the same task, with up to 5x the token efficiency. The first models “are available to select early access customers” in Q4 2026, and customers may run them on Cadence’s cloud, public clouds, or customer-controlled on-premises environments and post-train them on their own “proprietary data, designs, and workflows.” To learn more about Maestro, we spoke with two Cadence executives: Rob Knoth, senior group director of strategy and new ventures, and Paul Cunningham, senior vice president and general manager.

This approach is distinct from Synopsys’ own specialized model, GPT-Synopsys, which will run on OpenAI-hosted infrastructure, without using customer data to train the model. The models of both EDA giants, announced less than a month apart, now split on where a customer’s design data goes. Where they might not differ is why: Paul Cunningham, senior vice president and general manager at Cadence, told Tom’s Hardware Premium that agent costs have “really started to explode,” adding, “And I think the world realized so.” That makes it a natural step, he said, to move more of the agents’ iterating “into the weights of the models themselves.”

Maestro takes this step through Cadence’s AI Super Agents, which “interpret design intent, decompose tasks” and “orchestrate the Maestro portfolio with other leading frontier and open-weight models,” evaluating trade-offs to guide decisions with domain expertise. It also uses reinforcement learning (RL), where Cadence’s “deterministic solvers evaluate outputs against engineering objectives, constraints and ground truth,” and, importantly, the results feed back into the models for improvement.

The models include Nvidia’s Nemotron, and customers can run the customer-data training step themselves with Cadence’s fine-tuning harness, potentially “on their own GPUs, on their own network,” Cunningham said. He added that Nemotron, specifically, was “giving really solid results and actually improving very, very fast.”

One constant that keeps popping up with agentic EDA is the need to nail down what the numbers actually measure. In chip design, quality of results usually refers to the design’s power, performance, and area, though Cadence didn’t specify what its 30% figure measures. Both it and the 5x token efficiency figure are measured against “a general-purpose foundation model,” which Cadence told us is the open-source model Maestro started from before its training. The company wouldn’t name the foundational AI model that Maestro is derived from beyond mentioning that it is “one of the leading open source models that’s very competitive with commercial frontier models.”

Cadence credits the token savings to balancing its EDA and system design automation (SDA) “tool calls with LLM invocations.” As an analogy, Cunningham said, “You don’t need to drive to Safeway in the jet engine,” and you want to use “the right model at the right time for the right task.”

Runtime does matter, too: as Rob Knoth, senior group director of strategy and new ventures at Cadence, told Tom’s Hardware Premium in September that a tool call could take anywhere from hours to days to return a value, while the AI industry’s training loops are “used to minutes.” Efficiency could also lower costs, since Maestro use counts toward Cadence’s agent pricing, a subscription with a consumption allowance and charges for use beyond it, Cunningham said. He didn’t say whether the token savings cut a customer’s bill.

Being flexible for the customer is a key element of any solution, and for Cadence’s competitor Synopsys, this lies in Autopilot, which Synopsys says allows customers to “choose among commercial, open-source, and fine-tuned language models” across “Synopsys-managed and customer-managed deployments,” with cloud or on-premises infrastructure. For agent security, such as sandboxing, Cadence says it’s working with Nvidia on an integration with Nvidia’s OpenShell Reference Architecture, which Nvidia launched last week, naming both Cadence and Synopsys. These options can help protect customer intellectual property.

Cadence has had a busy 12 months with customers and agents alike. During Cadence’s July earnings call, CEO Anirudh Devgan said ChipStack, its design and verification agent, “has more than 20 customer engagements,” and ViraStack, its custom and analog agent, “more than 25 customer engagements.” Socionext said just last week that it was applying InnoStack. Silicon Labs has also credited ChipStack with accelerating debugging, and both companies say they are watching Maestro next.

Cadence's agentic consolidation began with its acquisition of ChipStack back in November 2025. In June 2026, the company claimed that its agent had reached Level 5 autonomy, which we explored in detail. In September, Knoth teased Maestro’s ingredients to Tom's Hardware Premium a month early, two days before the GPT-Synopsys announcement: “reinforcement learning loops, domain-specific post-training of open-weight models, model routing.” Maestro is now sitting under all four super agents, targeting a similar year-end window to Synopsys’ Autopilot; the table below has the full timeline.

This rapid pace of progress means that a clear distinction has to be made between early access and general availability. Both Cadence and Synopsys can point to customer engagements, and, at least in the latter’s case for the AgentEngineer portfolio and Autopilot platform, general availability remains a goal by year’s end; Cadence hasn’t given a general availability date for Maestro. For GPT-Synopsys, though, only “early technology engagements are underway,” according to its partners. The platforms and agents remain works in progress.

For Maestro, Cunningham hopes for “more case studies, more quantifiable benefits” through Q4. The real story, as he aptly put it, is that “we’re going to bring agents and agentic AI natively into our vertical, not that our vertical is going to get embedded into general-purpose agents.” Only customer results on real workflows will test Cadence’s 30% number, but going by what both companies have shown so far, their own early figures are extremely promising and exciting for the EDA industry.

China is spending more time loitering near undersea cables in the Pacific Ocean than any other nation, according to a report by the Center for Strategic and International Studies, a U.S.-based think tank in Washington, DC. The analysis compared the behavior of 200 research vessels from 47 different nations, including the U.S. and Russia, and found that Chinese vessels spent a "statistically significant larger share of time" near critical infrastructure.

The analysis, reported by Indo-Pacific Defense Forum, notes that Chinese activity was particularly suspicious in this regard. It notes that the "persistent and disproportionate activity" means that China "may be leveraging its research fleet for underwater data collection that may aid... analysis of terrain for sabotage or submarine operations."

Civilian research vessels provide a unique cover for states looking for "plausible deniability" for their research gathering activities. The report mentions the activity of a ship from the Xiang Yang Hong research fleet which, between July and October 2024, spent time near three undersea cables that connected the U.S. with Australia, Indonesia, Singapore, Guam and others.

The ships move in a "lawnmower" pattern which, it notes, is associated with surveying to create accurate seafloor maps. While the ships themselves aren't part of the PRC's navy, the data they collect could be used to serve China in future conflicts.

These "gray zone research activities" continue in an era of Chinese conflict with its neighbors as it asserts territorial claims in the South China Sea against states like the Philippines and Vietnam, not to mention its claims over Taiwan, which Beijing considers a breakaway province.

It also follows a pattern of authoritarian states looking to dominate undersea infrastructure. Last week, a report in The Telegraph found 500 incidents of Russian monitoring near undersea cables over the last five years. Meanwhile, Iran has threatened to attack undersea cables in the Strait of Hormuz in the event of further assaults by the United States if conflict erupts again.

Intel and AMD have spent decades competing for the processors inside PCs and data centers, but the rivalry hasn't stopped the two companies from working together when their interests align. And now, Intel CEO Lip-Bu Tan has reportedly discussed potential areas of cooperation with AMD chief Lisa Su, adding another possible partnership to the mix as Intel tries to regain its footing in the semiconductor industry.

According to a recording (embedded below) of Tan’s October 24 fireside chat with Chinese American Semiconductor Professional Association (CASPA) President Jimmy Cheng, the Intel CEO outlined his plans to rebuild the company’s CPU business, develop chips tailored to specific customer workloads, and attract more customers to its foundry operations. However, Tan did not publicly detail the reported areas of potential cooperation with AMD CEO Lisa Su.

Intel has traditionally built its business around its own processor roadmaps, but Tan says the company needs to be more flexible as customers look for silicon designed for particular workloads. He pointed to Intel's Xeon processors as one area where that approach is changing, alongside continued development of GPUs and accelerators.

"We have to reclaim the leadership in CPU. That is something that I am very determined to do." Tan said.

He also described a change in how Intel approaches customization. "Intel is so proud they never want to modify for customer needs," he said. "I'm modifying it for customer needs, and so that, I think, to me, is a fast-growing area, and I'm double, triple down on it."

The strategy extends beyond conventional CPUs, with Tan identifying application-specific integrated circuits (ASICs), co-packaged optics (CPO), physical AI, and advanced packaging as potential growth businesses. He wants two or three of these areas to eventually contribute around 10% of Intel's revenue each.

Intel's foundry business is another major part of Tan's plans. He said the company's 18A manufacturing process had reached volume production, while development of its next-generation 14A process was progressing strongly. Intel has already released version 0.5 of the 14A process design kit (PDK), which provides chip designers with the technical information and tools needed to develop products for the process.

Tan said version 0.9 was coming next, followed by version 1.0 for volume production in early 2027. "We have multiple customers [who] really want us to really scale their business," he said, referring to interest in Intel's foundry services.

However, Intel did not identify those customers or disclose the size of any potential manufacturing orders during the discussion. Notably, in January, Intel had two prospective customers evaluating 14A test chips, with firm supplier decisions expected during the second half of 2026 and the first half of 2027. Tan also described foundry manufacturing as a business where trust matters as much as technology. Customers hand over wafers that will eventually become products on which their own revenue depends, making missed deadlines or manufacturing problems costly. "The moment we take on the foundry, take on the wafer is a lot of responsibility to make them successful," he said.

He recalled one customer who had waited six to eight months hoping he would become CEO before presenting 15 areas where Intel had fallen short. Tan said he addressed the problems individually, after which the customer agreed to deepen its relationship with Intel.

That said, the prospect of further cooperation would not be entirely new. In October 2024, Intel and AMD established the x86 Ecosystem Advisory Group alongside other technology companies to coordinate improvements to the x86 instruction set architecture and make it easier for hardware and software developers to support both companies' processors.

That cooperation has since moved into AI software. Last week, it emerged that AMD was integrating its ZenDNN optimizations into Intel's open-source oneDNN library, bringing the companies' CPU optimization work for PyTorch AI inference under a common software path. The companies have also worked together on Advanced Matrix Extensions and other x86 capabilities through the advisory group, showing that they cooperate on shared technologies while continuing to compete for customers. For now, however, neither company has announced a new agreement arising from the discussions.

The AI-powered inrush of software vulnerability reports over the last year has shined a light on how many critical pieces of open-source software are chronically underfunded. The latest victim is PHP, the scripting language powering about 71% of websites, making the news due to a critical remote code execution vulnerability. The PHP Foundation, which develops and oversees the language, reportedly counts only one full-time person and 14 "mostly part-time" contractors. That's bad enough, but the real problem goes far deeper thanks to the internet's lazy development practices and systems administration.

In an interview with Reuters, the PHP Foundation's Executive Director Elizabeth Barron spells it out clearly: she's the only full-time person on staff, and works with the contractors' help. Pertaining to the software itself, the team's job is to develop and maintain the PHP core, including taking care of security fixes. The core team accepts contributions from the world at large, in an arrangement not too dissimilar to Linux kernel development — except that project is far better-manned, with many full-time contributors and company-employed developers.

Naturally, the PHP foundation collaborates with other companies and many cybersecurity firms, but the dev team was stretched thin even before the increase in security vulnerabilities with the help of countless AI agents. Barron does note that there was no reported exploitation of PHP2Shell before the security fix came out.

But the lack of a higher funding level likely led to the "structural failure" to address the underlying vulnerability in a timely manner, as the out-of-memory condition (OOM) that eventually led to the 2026 exploits was originally discovered in 2024. Given how widespread PHP is, and how many third-party extensions there are, the work on fixing the OOM problem "stalled and restarted repeatedly over months." It didn't help that the Chainguard security firm broke the PHP2Shell embargo too early by accident.

Even still, the team is thankful for the money it has, as "without the Foundation’s funding it would have been impossible to have multiple people work on this issue full time, eventually through weekends and around the clock to coordinate and communicate with other involved parties." One would hope that this latest incident releases the purse strings of more companies that make their living off of PHP.

And yet, this failing is far larger than just bugs in the core language, which, notably, were only one point update away. In theory, any server running the current 8.x brand of the software should be upgraded in minutes with the released patches, but reality was far different, as websites went down for hours or days out of fear of exploitation.

The real problem reared its ugly head: nearly 73% of PHP installations were running versions well past their end-of-life (EOL) date, meaning that obtaining fixes is now far harder. Many professionals who worked with the language, including yours truly, had spent decades warning of this reckoning, even considering many Linux distributions were backporting fixes.

Many of these servers were still running PHP 7, a version that was completely EOL'd four years ago, with many years of advance warning, and that originally came out in late 2015, almost eleven years back. If that doesn't sound silly enough, let it sink in that around 7.7% of PHP sites were still running version 5.x, which first saw the light of day 22 years ago in 2005, and saw the final sunset in 2018. It's safe to say that everyone had plenty of time to prepare.

To further illustrate the ridiculousness of collective complacency: the jump from PHP 5 to 7 and then from 7 to 8 required changing sections of website code, but the vast majority of said changes are simple to find and apply, and even before AI, there were already automated code-patching tools around. A couple years ago I mapped them all out for a PHP 5-to-8 codebase conversion myself.

Most of the alterations were about dropping support for ancient functions and libraries, which themselves had had long-standing warnings of incoming deprecation. Not a lot of core language features underwent modification, and a lot of convenient syntax was added to help clarify fiddly, error-prone code, too. That the overall situation got to this point beggars belief and illustrates how complacency, laziness, and cost-cutting are a far bigger force than a cadre of AI bots digging for flaws.

A hacker has successfully breached a vulnerable server, claiming a prize pot worth $2,021.22 in Bitcoin after exploiting PHP2Shell vulnerabilities, a posting on the Steal My Bitcoin website has shown. The challenge's creator confirmed the win on October 25, ending an experiment designed to test claims that exploiting critical software vulnerabilities is becoming easier.

The challenge, hosted on the website a day earlier and announced through a forum post, initially offered approximately $502 worth of Bitcoin. Donations subsequently increased the pot to 0.02511679 BTC, worth $2,021.22 at the time of the win. The creator's website recorded the successful claim at 15:47 UTC on October 25, after the challenge had been running for approximately 19.28 hours.

The experiment follows the disclosure of CVE-2026-109618 and CVE-2026-109619, two critical PHP vulnerabilities that can allow remote code execution under certain conditions. The flaws received critical severity ratings, and fixes were released on October 21. Turning the vulnerabilities into a reliable attack, however, requires more than simply running a publicly available exploit against an arbitrary server.

The challenge's creator, posting under the name ‘cloudlandsdev’, had said the project was designed to see whether anyone could use the vulnerabilities to compromise the server and claim its Bitcoin. “Made this because I'm curious if anyone out there has figured out how to actually use this PHP RCE, we spent the last few days panicking. Have fun!” the creator wrote.

PHP2Shell refers to two vulnerabilities in PHP’s request-handling and memory-management processes. Under vulnerable configurations, an attacker could potentially exploit these flaws to execute code with the privileges of the web server account.

The experiment quickly attracted attention, but the resulting traffic also became a problem. In an update, the creator said visitors were exhausting the server's available computing resources shortly after each capacity increase. Limited storage also prevented the server from retaining all its logs, complicating efforts to document a successful attempt.

The winner, ‘KillaBoi’, published a proof of concept on GitHub describing the attack against the live challenge and providing a local test environment. The exploit took advantage of PHP's handling of multipart uploads and session data, where, by pushing a PHP worker beyond its memory limit, the attacker could leave internal state in an inconsistent condition. Subsequent requests then exposed memory contents through session cookies, providing information about the locations of PHP's memory structures and executable code.

Notably, the attacker used those addresses to construct a forged PHP data structure that triggered command execution during cleanup. The final request returned evidence that the command had run inside the PHP process. Although the server was successfully compromised, the recovered private key did not match the advertised prize address. However, on-chain data shows the funds were subsequently sent to an address “bc1qr4…lk7a9”, labeled as a Kraken cold wallet.

Bitcoin recovery efforts have previously involved reverse-engineering software rather than exploiting a live server. In 2024, hardware hacker Joe Grand, known as ‘Kingpin”, and researcher Bruno recovered access to 43.6 BTC belonging to an owner identified as Michael, who had forgotten a 20-character wallet password.

The researchers reverse-engineered an older version of RoboForm's password generator and identified a flaw that made the password predictable. They recovered the wallet in November 2023, when the Bitcoin was worth approximately $1.6 million. Its value had climbed to nearly $3 million by the time the recovery was reported in May 2024.

A more recent case involved a Bitcoin holder who recovered five BTC worth roughly $400,000 in May 2026 after losing access to the wallet for years. The pseudonymous owner, known as ‘cprkrn’, said Anthropic's Claude AI helped identify a backup file from December 2019, which could be decrypted using a password the owner had previously written down in a notebook. The backup contained the keys needed to access the Bitcoin.

The latest event comes as security teams face pressure to patch PHP installations following the public disclosure of PHP2Shell. A separate report on the vulnerability disclosure controversy detailed how premature publication disrupted the coordinated release process. That said, for now, evidence of the latest successful attack suggests that exploiting these PHP vulnerabilities may not be that hard after all.