Recorded 09 October 2026 · 53 min
Pat Gelsinger, who led Intel and VMware and is now a general partner at Playground Global, talks with two former colleagues about what AI changes in chip design and what it leaves untouched. His central point is that AI has made design the easy part. A strong chip can be designed in 3 months, but it takes 9 months to turn that into usable silicon, packaging and a rack. The bottleneck has moved to fabrication, memory, power delivery and the grid itself.
The tension is speed against physics. AI compresses logic and software work, while lithography, yields, thermals and grid capacity run on clocks measured in years. That shapes his calls: roughly 100 AI chip startups will consolidate, memory is due its first real innovation in 30 years, optics will replace copper for scale-up around 2028-29, and energy will cap how far AI spending runs. One interviewer, Guido, pushes back that agents make chip diversity cheap.
Logic yields, analog doesn't — ▶ 7:32
Chip design is at a turning point like the 486 was for EDA. Many logic functions can now be done with AI tools "in pretty incredible ways" because transistor budgets are big and a few experts can guide the tools. Analog is not AI-able yet: it needs lots of silicon data, so designers either get very conservative or do it the hard way. His rule: whenever technology makes something easy, the bottleneck moves somewhere else.
3 months to design, 9 to use — ▶ 9:21
Design now takes 3 months, but fab needs at least 3 months even with supercharged flows. Advanced 3D packaging adds roughly 3 more, then it has to become a rack, because "nothing's a chip anymore. It's a rack." At 18 months to scale with software, workload understanding is stale; Graphcore was not a bad design, but the world moved on. His target is a month or two, which needs new lithography and flows without $50 million of mask cost before prototyping.
HBM and the 40% power tax — ▶ 11:54
HBM is "a hideous memory" and simply the best available: poor bit density, bandwidth capped by shoreline, power and thermal problems, and DRAM dislikes heat. AI is a memory-compute workload, so this is a critical bottleneck. Power is the other. There are no good 3D tools for power dissipation and hotspots, so designers guard-band nearly the whole power envelope and lose about 40% of their power. Power modeling, dissipation and the delivery model all have to improve.
100 chips will converge — ▶ 14:38
He expects the field to narrow for three reasons. Workloads keep migrating (prefill, midfill, decode, CPU-like reasoning, 64-bit precision returning for chemistry), so extreme specialization will not fit workloads 2-4 years out. Scale requires capital and workloads, and 100 vendors cannot all get them. And big buyers will pick winners: OpenAI, Nvidia and Anthropic will favor hardware whose software co-evolves with their models. He thinks LLMs are also nearing their limits, which will shift algorithms again.
Agents versus capital — ▶ 19:12
Guido argues that software complexity once limited the industry to a few architectures, but an agent swarm can now write an optimized kernel overnight, so more heterogeneity is plausible. Gelsinger's reply is time and money: 18 months to build, then "could I borrow 50 billion" for racks, data centers and power at scale. Nvidia's Groq deal shows his expected path: in 2 years you won't tell what Groq is inside Nvidia. The heterogeneity gets hidden by abstraction, not removed.
Zero new memories in 30 years — ▶ 22:50
He has been associated with at least 5 new memory architectures, Optane among them, and knows of about 100 proposed over 30 years. Major new ones: zero. Beyond DRAM, SRAM and flash, nothing. Memory made money about one year in five, but all three big vendors are now in the top 20 most valuable companies, with roughly $2.5 trillion of added market cap in 4 years. He expects ferroelectrics and stackable, non-capacitive dense cells to follow, and says he has a stealth memory company. He is cool on flash and MRAM.
Mid-rise stacks, not skyscrapers — ▶ 27:40
Stack value rises with height, so per-layer yield must improve exponentially. Spare lines help, but a cracked die is a bad die. He is "not a crazy 16 32 stack kind of guy": 2-4 high memory stacks are the sweet spot. Even then, once power delivery in the Z dimension, redistribution layers and optics are added, a canonical package is already about 8 layers thick. He calls it an engineering feat and a manufacturing nightmare, and says engineers are becoming plumbers, with liquid cooling and diamond as options.
No to optical memory, no to PIM — ▶ 32:13
Moving memory off-package over optics means more optical-electrical conversions and more power burned. Compute costs few femtojoules per operation, while communication per bit is comparatively 1,000x worse. Processing-in-memory fails differently: it forces workloads to push compute into memory. PIM ideas have been around 25-30 years and he expects them still hanging around in 25-30 more. He allows he could be wrong, but wants high-density, high-bandwidth memory close to compute as the more generalizable answer.
Optics for scale-up, 2028-29 — ▶ 34:54
Copper is a shrinking waveguide: making it work at 5 m now costs more than making optical work at 100 m. The holdup is an optical supply chain never tested at scale, with laser capacity, thermals and packaging unresolved. He expects co-packaged and near-packaged optics around 2028-29 for 10 million GPU scale. He calls the NVL72 an engineering marvel and a manufacturing nightmare that cost the industry an extra 1.5 years. Large, predictable AI flows suit optical circuit switching, which dissolves the scale-up and scale-out boundary.
Energy caps the euphoria — ▶ 43:22
The US spent 10-15 years flat, retiring coal as fast as it added renewables, and now grows capacity perhaps 4% a year. Gas turbine lead times are 8 years, renewables depend on China, and the last US reactor came online 20 years ago. He expects more data center projects to default because the energy won't be there, an upper bound on AI spending. Inside the building he points to 800V DC, vertical GaN converting 800V to 5V in one step, and a superconducting company, Snowcap, claiming 1,000x better power performance.
VMware for agents — ▶ 48:44
Every innovation leads to the next abstraction, and the virtual machine deserves a permanent place in the compute hierarchy. Agent swarms raise questions VMware once answered for hardware: who manages the agents, sets their security profiles and watches their performance? He wants every element of virtualization and management rebuilt, with vMotion-style migration for agents. Humans need to set policies, write "constitutions" and watch dashboards. The design constraint changes from abstracting hardware for applications to abstracting hardware and operations for agents.
AI has made chip design the easy part, and that moves the bottleneck to everything downstream. Pat Gelsinger gives the arithmetic. A team that understands the workloads can turn AI tools loose and have a strong design in 3 months. Then the real clock starts. Fab takes at least 3 months even with supercharged flows. Advanced 3D packaging adds roughly 3 more. Then the part has to become a rack-scale system, because "nothing's a chip anymore. It's a rack." Design takes 3 months. Getting to something usable takes 9.
The delay is worse than it looks, because the target moves. If it takes 18 months to reach scale with software on top, the workload understanding that shaped the design is already stale. Gelsinger points to Graphcore: not a bad design, but the world moved on. His fix is to compress the 9 months to a month or two. That needs new lithography and design flows that do not demand $50 million of mask cost before prototyping. AI also helps unevenly. Logic is now tractable. Analog is not, because it needs large amounts of silicon data. Designers either stay conservative or do it the hard way.
Memory is the next constraint, and Gelsinger is blunt about it. HBM is "a hideous memory." Bit density is poor, bandwidth is limited by shoreline, and power and thermals are ugly. It is simply the best one available. AI is a memory-compute workload, so this matters. He counts about 100 proposals for new memory over 30 years and zero major new memories: DRAM, SRAM, flash. The reason was economics. Memory made money one year in five. That has flipped. All three big vendors are now among the 20 most valuable companies on earth, and the industry has added roughly $2.5 trillion of market cap in 4 years. Capital now justifies attacking the physics. He expects the first new memory innovation in 30 years and says he has a new memory company in stealth. He points to ferroelectrics and dense, stackable cells, and is cool on flash and MRAM.
On stacking he refuses the hype. The value of a stack rises with height, so per-layer yield has to improve exponentially, and spare lines cannot rescue a cracked die. He expects 2-4 high memory stacks, not 16 or 32. Even then the package is about 8 layers once power delivery, redistribution layers and optics are added. He is equally unmoved by two popular alternatives. Moving memory off-package over optics burns power in conversions: compute is cheap in femtojoules, but communication per bit is comparatively 1,000x worse. Processing in memory has been around for 25-30 years, and he expects it to still be around in 25-30 more, because it constrains workloads. He admits he could be wrong.
He wants optics everywhere else. He declared the death of copper about 25 years ago and says he will be right eventually. The physics is now on his side. Making copper work at 5 m costs more than making optical work at 100 m. The holdup is a supply chain that has never been tested at scale, with laser capacity, thermals and packaging all unsettled. He expects co-packaged and near-packaged optics around 2028-29, working at 10 million GPU scale. He calls the NVL72 an engineering marvel and a manufacturing nightmare that cost the industry an extra year and a half. Because AI traffic is large and predictable, optical circuit switching fits, and the line between scale-up and scale-out dissolves.
Power is the hardest limit because it is slowest to fix. "Energy capacity equals economic capacity," and the US spent 10-15 years flat, taking coal offline as fast as it added renewables. Growth is now perhaps 4% a year. Gas turbine lead times are 8 years, renewable supply chains lean on China, and the last US reactor came online 20 years ago. Gelsinger expects more data center projects to default because the energy will not be there. He calls that an upper bound on AI spending, not a stop. Inside the building he wants 800V DC and vertical GaN that steps 800V down to 5V in one conversion. Power per teraflop has been flat for 5 generations of Nvidia chips, so he backs a superconducting company, Snowcap, claiming 1,000x better power performance.
All of this feeds his view of the accelerator market. Around 100 startups will not all win. Workloads keep migrating: prefill, midfill, decode, reasoning that wants CPU-like compute, chemistry that brings back 64-bit precision. Extreme specialization will not match what runs 2-4 years from now. Scale needs capital and workloads, and 100 vendors cannot all get them. And the big buyers will pick winners, as OpenAI, Nvidia and Anthropic favor hardware whose software co-evolves with their models.
The same logic of abstraction runs through his last topic. In 2 years, he says, nobody will be able to tell what Groq is inside Nvidia. The diversity survives, but it is hidden. He applies this to agents. Every innovation leads to the next abstraction, and the virtual machine stays foundational. Agent swarms need what VMware once built for hardware: management, security profiles, performance and migration. Humans set the policies and write the constitutions.
The strongest objection comes from his interviewer, and it is only half answered. Software complexity once limited the industry to a few architectures. Now an agent swarm can write an optimized kernel overnight, so supporting unusual chips is cheap, and heterogeneity should rise. Gelsinger concedes some room, for a small diffusion workload versus a 15 trillion parameter model. Then he answers with capital and time: 18 months to build, $50 billion to deploy, and a million-GPU fleet cannot tolerate 1,000 of one card and 2,000 of another. Notice what that does. The argument has moved from software to money. His list of shifting workloads cuts both ways too. If inference fragments into prefill, midfill and decode, someone will build for each. His answer is that Nvidia buys them, which is consolidation by acquisition, not proof that the market cannot sustain many designs. Whether lasting diversity lives in the market or inside a few vendors is left open.
The open question. Gelsinger's case for consolidation leans on workloads shifting too fast for specialized chips. Yet his own examples of that shift (early prefill, midfill, decode, CPU-like reasoning, 64-bit chemistry) are arguments for more specialization, which he resolves by saying Nvidia will absorb it, as with Groq. That is consolidation by acquisition, not proof that 100 designs cannot each find a niche. Watch whether OpenAI and Anthropic actually back a handful of accelerators, and whether the heterogeneity stays visible or gets hidden inside a few vendors.
Watch the full video at https://www.youtube.com/watch?v=1Q_7yU7FZ1k.
Comments ()