Mithril

Congrats to our partners at Sail Research for officially coming out of stealth 🚀. It's been great working with Neil Movva and team. Excited to see you build the best inference platform for long-horizon agents.

The data processing revolution has been chiefly concerned with making structured data as query-able as possible. In parallel, tabular AI models have given us accurate predictions over this data (for example, Prior Labs, Kumo, Probabl, Neuralk, The Forecasting Company). Now we have generative AI, which can operationalize superhuman intelligence given sufficient context. Yet, AI agents operating over these model outputs and structured data cannot be reliable if they do not have the ability to generate and verify the reasoning embedded in the AI models. The scale of structured data has restricted enterprises to analytics agents that access information through simple aggregates.  With reasoning, we can move beyond this intelligence layer to build agents that provide reliable recommendations and execute actions. I founded OuterProduct, along with Adityanarayanan Radhakrishnan and Mikhail Belkin, to solve this problem. At OuterProduct, we have built the Unified Reasoning Engine as a culmination of our decades of combined research. In our Science publication (https://lnkd.in/g7Wb36FP), we discovered a key abstraction for how general AI models learn from data beyond classical methods. This led us (along with David Holzmüller) to develop a first-principles tabular model (https://lnkd.in/ghykqVNp, ICLR) that can outperform even tabular foundation models across benchmarks (https://lnkd.in/gzRab6Sy). In another follow-up Science publication (https://lnkd.in/gquRbeav), we showed this technology can be applied to govern and monitor LLMs. I am incredibly proud of the technology we have built and the team that has joined us. My co-founders have always pursued the most ambitious problems, including, for example, Misha Belkin’s discovery of the now well-known double descent phenomenon (https://lnkd.in/gCiKj7i5). See our website and join our waitlist to stay up to date on our product (www.outerproduct.com).

Cloud Next last month was a clear statement of where Google Cloud sees TPUs going: split between training and inference. TPU 8t scales to 9,600 chips for training; TPU 8i delivers ~80% better price-performance for inference vs. Ironwood. Both are expected to GA later this year. The chip Google still points to as the inference price-performance sweet spot today, Trillium (v6e), has been hard to actually try. Committing to a long Calendar reservation upfront doesn't fit a benchmarking workflow. We offer TPUs in preview, with the same flexibility (Spot and Reservations) as we offer on GPUs: - bid a max and pay market on spot - reserve a single chip (or 4, or 8) without a long commit - relist unused reservation capacity back for credit Reply or DM if you want to be allowlisted. #TPU #Trillium #GoogleCloud

Excited to see our partners at uRun officially come out of stealth 🚀. It’s been great working with Keegan McCallum and team as they build the future of real time video gen. Looking forward to what’s ahead. Big congrats to the whole team!

We’re proud to share that Mithril has been named to Fast Company’s Most Innovative Companies in Artificial Intelligence for 2026. This recognition is a meaningful milestone for our team and our mission to build the AI omnicloud: a single platform for transparent, flexible access to the compute infrastructure powering modern AI. This includes our recent announcement that all reservations on Mithril are now "flexible" allowing you to resell compute that would otherwise sit idle. You can find our product announcement in the comments. Thank you to our customers, partners, and team for helping make this possible. See the full list here: https://lnkd.in/e_cM4yBC #FastCompany #AI #CloudInfrastructure #Innovation

Proud moment for the team: Mithril was highlighted by Oracle among its AI innovators in connection with its latest quarterly earnings materials: https://lnkd.in/e7QaNCTN We’re excited to keep building infrastructure that helps teams deploy and scale AI faster. Appreciate the recognition from Oracle and the partnership with OCI! #AI #CloudInfrastructure #OCI #GenAI #MLInfrastructure

First in our series on how Mithril’s unique offerings allow AI practitioners to minimize price and maximize flexibility. Quick video from our sales engineer Ben O'Neil on how you can use Mithril’s spot market to run a 12 hour job on an A100 for less than a dollar. https://lnkd.in/gkPjSZ7U"

Anthropic just launched a "fast mode" for Claude Opus 4.6 — 2.5x faster, same model, same intelligence, but at 6x the price. The developer reaction has been… polarized. Some are calling it usurious. Others are calling it one of the biggest productivity unlocks of the year. I think the visceral negative reaction from some devs is actually pretty interesting. Nobody is worse off for having more options — the standard mode is still there, same price, same quality. So why are people upset? I think ironically it's because they're worried fast mode will succeed. If it works well enough that others adopt it, you get a tragedy of the commons -- like everyone standing on their tippy-toes at a concert. Now you're forced to pay the 6x premium just to stay competitive. Zero-sum thinking. But here's what I think people are missing: even 6x greater cost for 2.5x speed can make some applications genuinely cheaper. If you're running a pipeline, or using Claude as an orchestrator in a compound system alongside customer or open-source models running on your own GPU servers, fast mode can unblock the entire system. The orchestrator is the bottleneck. Speed it up and you reduce idle GPU time across the whole pipeline. The bottleneck matters more than the line item. We ran an experiment at Mithril that illustrates this same principle at the GPU layer. We took some underutilized previous-generation GPUs (A100s) and listed them on our platform with a $0.50/hr floor price in an auction. That was already super profitable. Then we dropped the floor 50x — to $0.01 — and let the market purely determine the price. The results were fascinating (see chart). Earnings went UP by 2.5x. And the price dynamics were compelling: [1] during peak hours, preemptible GPU usage surges to $2-4.00/gpu/hr. [2] At night and on weekends/holidays, it drops to $0.04 or even $0.01. A 400x spread. The feedback from users has been that this is better for both sides: For users with high-priority workloads, it's better to have guaranteed availability at a variable price than variable availability at a guaranteed price (first-come, first-serve). For users with flexible workloads — say you need 4 hours of runtime within the next 6-14 hours — it's better to run at night and save 10-100x. Both cases — Anthropic's fast mode and our auction experiment — point to the same notion: there's enormous room beyond flat pricing for compute and its derivatives (like tokens). Better market design pushes the whole price-performance frontier forward. Price-sensitive customers save money. Performance-sensitive customers get guaranteed availability. High-value inference jobs can run anytime. Long-running training jobs can run flexibly with much better economics. It's a rare case where better mechanism / market design genuinely makes everyone better off. I hope we see more of it!

Researchers at the Broad Institute of MIT and Harvard use Mithril’s GPU Omnicloud to better understand gene expression. “Through Mithril, computations that once took up to a week on traditional infrastructure can now be completed in hours.”   ~ Dr. Adit Radha, MIT Mathematics and Broad Institute Read how the Broad Institute is leveraging Mithril to accelerate biological discovery in our latest case study. https://lnkd.in/ggXgshEw