Discover
The Open Compute Project Podcast
The Open Compute Project Podcast
Author: The Open Compute Project
Subscribed: 7Played: 87Subscribe
Share
© The Open Compute Project
Description
The latest news, developments, collaborations and in-depth look at the innovation that derives from the collaboration in the global OCP Community.
Join us as we discuss data centers, IT hardware, and infrastructure through expert interviews and insightful analysis to empower listeners with the knowledge to make informed decisions and deepen their understanding of the OCP. Your feedback shapes our content, so tune in, engage with our community, and be a part of reshaping hardware technology.
Visit https://www.opencompute.org/ to learn more.
Join our Facebook Group: https://www.facebook.com
Join us as we discuss data centers, IT hardware, and infrastructure through expert interviews and insightful analysis to empower listeners with the knowledge to make informed decisions and deepen their understanding of the OCP. Your feedback shapes our content, so tune in, engage with our community, and be a part of reshaping hardware technology.
Visit https://www.opencompute.org/ to learn more.
Join our Facebook Group: https://www.facebook.com
26 Episodes
Reverse
Recorded at the 2026 OCP Global Summit, host Rob Coyle sits down with Jacob Cummings, VP of Sales for data center markets at Americase. They talk about a part of the AI buildout most people never think about: how components worth $100K to $300K actually move through a data center. Jacob explains why today’s process for shipping and swapping spares is full of risky handoffs. His fix is to treat the case like a rack. Components clip into rugged, returnable cases that go from the OEM to the data hall and back again. The conversation also covers the real cost of cardboard on a hyperscale site and why lithium battery backup units (BBUs) need special handling. Jacob explains what a fully loaded case has to survive in drop testing, and why the rules for shipping hazardous materials start at the United Nations. Visit https://www.americaseinternational.com/Key takeaways●Packaging problems become handling problems. In Jacob’s years in material handling, most requests to lift or move equipment came back to packaging that wasn’t as good as it could be.●Today’s spares process is full of handoffs. Pallets of boxed spares get unpacked andshelved. Each part is carted out to the hall one at a time. Failed parts are repacked in whatever OEM packaging is still around and shipped back on a FedEx truck.●Treat the case like a rack. Americase’s cases are rigid, stackable, and rated for about1,000 trips. Components clip in the same way they would in an Open Rack. A technician moves a whole case with a pallet jack and swaps good parts out for bad ones in the hall.The full case then goes straight back on the truck.●Cardboard is expensive. One hyperscale customer spends about $600,000 a year on dumpster roll-offs just for BBU cardboard. Another Americase study put BBU receiving and handling labor at about $2 million a year.●Lithium changes the math. Because of thermal runaway risk, regulations cap how much battery capacity can sit in a given space and require dedicated dangerous-goodsrooms. Americase’s BBU cases keep one failing unit from setting off the others, and they contain the fire even if every unit fails. That earned a special permit allowing roughly five times the storage in the same footprint.●The drop test is the toughest. A fully loaded case has to land on its corner from 1.5meters without cracking, opening, or leaking.●Hazmat rules start at the UN. U.S. DOT rules and other national regulations flow down from the UN. HSC’s team is made up of former DOT officials and UN delegates. They aim to give the industry insight into rules 18 to 36 months before they arrive, and a voice in shaping them.Chapters 0:00 Welcome & meet Jacob 0:22 Americase's journey 1:35 From material handling to packaging 3:36 How spares get swapped today 6:42 Cases that work like a rack 9:25 The cardboard problem 11:43 Lithium BBUs & 5x the storage 14:44 Surviving the drop test 17:00 Shaping UN hazmat rules
Fresh off the floor at OCP Summit 2026, Matty Bakkeren joins us to unpack seven key observations that never made it onto a keynote slide. From the physics of wider racks and the reality of 800 VDC architectures, to smart silicon with Fujitsu Monaka, we take a close look at the growing gap between rapid hardware innovation and operational maturity. We also dive into standardizing management stacks vs. reinventing the wheel, and why proper operational editing and standardization matter far more than starting from scratch.Chapter Markers0:00 — Intro: Comparing Notes from OCP Global Summit 20260:58 — First Stop: The Google TPU Rack That Drew a Crowd3:33 — Chip to Grid: The Unit of Compute Keeps Moving7:00 — Who Is OCP For Now? When Specifications Become Standards11:49 — Open Rack Wide Is Everywhere14:22 — Wider Racks for General-Purpose Compute (and Why Air Still Matters)20:23 — Why Summit Matters: Learning in the Open22:55 — Market-Driven by Design24:37 — 800V DC: Is the Industry Ready?34:04 — Energy: Batteries, Grid Stability, and Data Centers as Grid Assets41:55 — Silicon Got Smarter: Fujitsu MONAKA and the 3D-Stacked Wafer49:16 — Standardize or Differentiate? Open's Strength and Achilles' Heel58:02 — Hardware Management and the Redfish Gap1:02:24 — Open Opens Doors, Geopolitics Closes Them1:12:57 — What We Didn't See: Data Centers in Space1:18:22 — What We Didn't See: Embodied Carbon and Sustainability1:22:33 — From Components to Systems, and What's NextGet Involved with OCPEverything OCP publishes is available under an open license. You can consume it, contribute to it, or lead it. Start with the Introduction to OCP courses on OCP Academy, catch up on the Global Summit 2026 recordings, and join Rob at the OCP EMEA Summit in Hamburg, April 6–8, 2027.If you enjoyed this episode, subscribe to the Open Compute Project podcast and share it with someone who missed the summit.
Chiplets aren't just a trend anymore, they're becoming the default architecture for the AI era. But as the industry stacks more dies into a single package to keep up with AI and HPC demands, it's inheriting a reliability problem nobody has fully solved: when four, eight, or more dies from different processes (and sometimes different vendors) share one package, a single weak link can disqualify the entire stack — and the cost of that failure is enormous.In this episode, OCP's Rob Coyle sits down with Evelyn Landman, co-founder and CTO of proteanTecs, to unpack why "known-good-die" testing is no longer enough and what it actually takes to reach "known-good-stack" confidence. They get into the hidden risks of multi-die packaging — thermal coupling, voltage droop, mechanical stress from through-silicon vias, and interface faults that traditional pass/fail testing simply can't catch — and how embedded telemetry is giving engineers real visibility inside the package, from bring-up and debug all the way through in-field operation.In this conversation:Why AI, HBM, and co-packaged optics are forcing the move to multi-die packagesHow "shift-left" testing catches faults before you commit to an expensive packageMix-and-match and binning: building a more consistent, higher-performing stackWhat embedded telemetry reveals that lab-condition pass/fail testing missesReal-time actuation and fleet-level predictive maintenance in the fieldHow power and thermal insight extends useful life across data centers, automotive, and physical AIThe role of OCP's Open Chiplet Economy in moving from siloed integration to open, interoperable chiplet operationsInterested in the chiplet ecosystem? Reach out to Evelyn and the proteanTecs team, and get involved with the Open Chiplet Economy work stream at the Open Compute Project.Chapters00:00 Intro: chiplets as the default AI-era architecture01:11 What's pushing the industry to modular dies03:22 The hidden cost: the problem shifts to the package04:56 Shift-left: catching faults before you package05:28 Single-die pass/fail vs. multi-die reality07:25 Where reliability breaks down in the field today10:06 What "known-good-stack" confidence looks like12:20 Mix-and-match and binning before assembly15:37 Embedded telemetry vs. traditional pass/fail17:30 Thermals, hotspots, and heat-sensitive optics & memory19:54 Real-time actuation to prevent imminent failure20:59 Board- and fleet-level predictive maintenance24:43 Scaling to millions of devices, tighter margins27:16 OCP's Open Chiplet Economy & the road aheadAbout the Open Compute ProjectThe Open Compute Project (OCP) is a collaborative community committed to redesigning hardware technology to efficiently support the growing demands on compute infrastructure. Its members span hyperscalers, ODMs, component manufacturers, and ecosystem partners — sharing open designs and working together on shared technical challenges that no single company can solve alone.
What if the biggest constraint on the AI era isn't compute power, software, or even money, it's how we cool the hardware that runs it all? In this episode, Rob Coyle sits down with Jean-Marc Denis, Chief Product Officer at Submer, to explore why immersion cooling has gone from an exotic idea in a Barcelona garage eleven years ago to what many believe will be the dominant cooling paradigm for AI infrastructure. Jean-Marc brings nearly 40 years of high-performance computing experience — from designing supercomputers to founding the European Processor Initiative — to a sweeping conversation about intelligence per watt, the power and water constraints reshaping data center strategy, and why the real money in AI is not in training mega-models but in inference at the edge.This episode covers both the technical and business dimensions of the AI infrastructure buildout, including why the six-month land-to-first-token expectation is forcing a co-design revolution, and how immersion cooling's simplicity — not its complexity — may be its most underrated advantage.What You'll LearnWhy the AI era has shifted the core efficiency metric from raw compute performance to tokens per watt and intelligence per wattHow immersion cooling's power overhead advantage — from a PUE of 1.3 down to as low as 1.02 — translates to dramatically more compute capacity within the same power envelopeWhy power availability, not cost, is now the binding constraint — and how that changes the calculus for data center operators at every scaleThe community pushback on large data centers and why immersion cooling's zero water usage is a strategic asset for permitting and local acceptanceThe difference between greenfield and brownfield deployments and why investors are now demanding a path from land to first token in six months or lessWhy the edge is where the real AI enterprise money is — and why inference at the edge demands a fundamentally different infrastructure philosophy than training at hyperscaleHow immersion cooling compares to direct-to-chip liquid cooling in terms of maintenance complexity — and why the comparison has flippedSubmer's co-design approach: turnkey nano, micro, and mini data center solutions from a few kilowatts to 100+ megawatts, designed around usage model firstWhy the Cambrian explosion of inference AI engines makes cooling flexibility — not optimization for one chip — the most important infrastructure property going forwardChapters0:00 — Introduction — Meet Jean-Marc Denis, CPO of Submer and HPC Pioneer0:53 — Submer's Story — From Barcelona Garage to Immersion Cooling Mainstream3:59 — Intelligence Per Watt — How AI Redefined the Efficiency Metric9:54 — Power Constraints at Scale — What PUE 1.02 Means at 1 Gigawatt13:26 — Water, Permits, and Community Acceptance — The Hidden Advantage of Waterless Cooling20:02 — The Maintenance Myth — Why Immersion Is Now Simpler Than Direct-to-Chip24:24 — Co-Design and the 6-Month Data Center — Land to First Token28:03 — The Inference Edge Opportunity — Where the Real AI Enterprise Money IsAbout the GuestJean-Marc DenisJean-Marc Denis is the Chief Product Officer at Submer, the Barcelona-based immersion cooling company he joined in November 2024. He brings nearly 40 years of high-performance computing experience, including designing large-scale supercomputers, building an HPC organization from zero to a top-four global ranking, founding and chairing the European Processor Initiative, and launching a microprocessor design company. About the Open Compute ProjectThe Open Compute Project (OCP) is a collaborative community committed to redesigning hardware technology to efficiently support the growing demands on compute infrastructure.
Every AI server rack deployed in a data center must pass electromagnetic compliance (EMC) testing before it can go live — it is a legal, regulatory requirement enforced by the FCC. But as AI hardware scales to megawatt-class racks with disaggregated compute, power, and liquid cooling systems, the test labs responsible for certifying that hardware are falling dangerously behind. There are only one or two labs in the world currently capable of handling the size, weight, and power demands of today's AI rack configurations — and the gap is widening with every new product generation.In this episode, Jaswanth Vutukury — Lead EMC AI Hardware Infrastructure Engineer at Meta — joins host Rob Coyle to break down exactly why this bottleneck exists, what it means for hyperscalers and ODMs trying to bring next-gen AI infrastructure online, and why he is using the OCP podcast as a call to action: for test labs to join the OCP ecosystem, for hyperscalers to begin collaborating on the problem, and for the industry to explore alternatives to traditional testing before this becomes a hard stop on AI deployment.What You'll Learn- What electromagnetic compliance (EMC) testing is, why it is legally required, and what failure looks like in a real data center environment- Why AI rack infrastructure — with disaggregated compute, power, and liquid cooling across multiple racks — presents an entirely different testing challenge than traditional IT equipment- The three specific bottlenecks breaking the current EMC testing model: weight and turntable capacity, power delivery (1MW+), and physical chamber size- Why there are only 1-2 labs globally capable of testing today's AI rack configurations — and the scheduling crisis this creates for the entire industry- Why the problem is structurally self-perpetuating: AI rack cycles run 12-18 months, but building or upgrading an EMC lab takes about a year — with no sign of the pace slowing- Jaswanth's direct call to action for EMC labs, hyperscalers, and ODMs — and why OCP is the right forum to bring all parties together- What alternatives to physical testing — including simulation — may need to be explored as the industry works toward a longer-term solutionChapters0:00 — Introduction — Meet Jaswanth Vutukury, Lead EMC AI Hardware Infrastructure Engineer at Meta0:33 — What Is EMC Testing and Why Does It Matter? The 30-Second Explainer2:00 — Why AI Infrastructure Is Different — Scale, Speed, and the Widening Lab Gap5:01 — Inside the Bottleneck — Weight, Power, and the Chamber Size Problem11:33 — Real-World Impact — Only 1-2 Labs Globally, and Everyone Is Competing for Them21:20 — The Path Forward — Bringing Labs into OCP, Industry Collaboration, and What Comes NextAbout the GuestJaswanth VutukuryJaswanth Vutukury is the Lead EMC AI Hardware Infrastructure Engineer at Meta, where he oversees electromagnetic compliance testing and certification for AI server rack infrastructure. His background spans multiple hardware-intensive industries — including autonomous vehicles and consumer electronics — giving him a uniquely broad perspective on why the AI infrastructure EMC challenge is unlike anything the industry has faced before. He came to the OCP podcast with a clear message: the bottleneck is real, the timeline is short, and the solution requires the industry to move together.About the Open Compute ProjectThe Open Compute Project (OCP) is a collaborative community committed to redesigning hardware technology to efficiently support the growing demands on compute infrastructure. Its members span hyperscalers, ODMs, component manufacturers, and ecosystem partners — sharing open designs and working together on shared technical challenges that no single company can solve alone.




