Muffi Ghadiali

Advitya is making a point that most people in AI infrastructure haven’t fully processed yet: the cooling problem has moved. The big chips get liquid-cooled now — they’re getting the attention they deserve. What’s choking these systems is the heat building up in the thousands of smaller components around them — optics, networking, power delivery — in tight spaces where bulk airflow can’t reach. The shift is from flooding everything with cold air to precision thermal transport at the exact point of pain. Everyone talks about the chip. The GPU. The $3 million rack. But the chip isn’t what’s failing. Now add the economics. A 72-GPU rack generates roughly $3.3M–$6.6M in annual revenue depending on utilization (CoreWeave pricing). One HPE configuration shows the power split clearly: 132 kW total — 115 kW liquid-cooled, 17 kW still air-cooled. That remaining 17 kW is only 13% of power, but it lives in the components that can stall the whole machine. A 1,000-GPU cluster has over 4,000 transceivers, each pulling 14–17 watts. That’s roughly 60 kW of distributed heat in exactly the spots where airflow doesn’t reach well (Cisco). Keysight reports that in unoptimized AI networks, over 50% of GPUs can sit idle waiting on data — and nearly half of training failures come from network and component issues, not the algorithm. Even a 5% hidden performance penalty is $166K–$331K per rack per year. Scale that across a 100-rack site and you’re looking at tens of millions in invisible losses. Same pattern on autonomous vehicles, robots, and drones. More AI crammed into tighter sealed boxes. One thermal chokepoint and the whole platform derates — less range, less uptime, less autonomy. This is the drip irrigation moment for cooling. For decades, thermal management was flood irrigation — blast cold air and hope enough reaches the hot spots. That’s increasingly broken. What’s needed now is precision: a smart emitter at every hotspot, delivering exactly what’s needed, where it’s needed, nothing wasted. That’s what Actasys is building. And it maps to a pattern we keep seeing across every domain where AI meets the physical world: the system fails because the physical environment isn’t ready. Thermal limits don’t stay contained as engineering limits on the GPU. They expand into habitat limits. Solve them at the point of pain — with distributed thermal intelligence — and you unlock everything downstream. More on this pattern soon. Actasys isn’t selling cooling. They’re selling the unlock to the most powerful compute. Advitya Patyal, David Menicovich, Miles Flamenbaum, Netafim #PrecisionCooling #AIInfrastructure #PhysicalAI #DataCenterCooling #DripIrrigation #DeepTech #ThermalManagement #DistributedIntelligence

8

2 Comments

As AI workloads eat up global computing supply, DRAM prices are surging like never before—creating major challenges for teams trying to plan out their data center spend. One of my portfolio companies, Mext, is addressing the largest cost component in the datacenter: server memory (DRAM). From being on the board of Redis for the last twelve years, I’ve learned a lot about the various aspects of this problem and approaches to solving them. Mext found a breakthrough, using new AI techniques, for dramatically reducing the amount of server-DRAM required to run applications, all while maintaining performance. It intelligently manages the server’s memory, keeping hot pages (i.e., those in use by applications) in fast DRAM and offloading cold pages (i.e., those less used) to a much less expensive memory tier (e.g., NVMe Flash). Key to the approach is ensuring cold memory pages that are about to be accessed by an application are back in DRAM before the application needs them or notices that they were gone. This is done without modifications to the application or the OS – so it can run in the cloud or on-premise. Mext thus allows applications to either run using less DRAM or keep their DRAM footprint but do more with it. They're hosting a webinar along with Fred Weber (former CTO of AMD) on Jan 22nd at 10am pacific time. You can register here: https://lnkd.in/gXe9GCeE

153

3 Comments

You can have groundbreaking science, elegant engineering, and even early customer interest. But without this one thing, you don't have an investable business. After two decades running IP strategy projects and now leading UC Berkeley's Deep Tech Innovation Lab while building the Berkeley Gateway Accelerator, I've watched countless brilliant technologies die in the valley between breakthrough and business. The problem isn't what most technical founders think it is. It's not about having better tech. It's not about getting more funding. It's not even about finding product-market fit. The fatal flaw shows up much earlier—and it's almost always the same mistake. In my latest piece, I answer questions from a Taiwanese entrepreneur about what really separates deep tech ventures that scale from those that stall. Including one provocative suggestion for Asian ecosystems that has nothing to do with technology. Read the full conversation, linked below. #DeepTech #Innovation #Startups #IntellectualProperty #VentureCapital

60

6 Comments