Sep 22, 2025 · 1h 38m · a16z
Dylan Patel on the AI Chip Race - NVIDIA, Intel & the US Government vs. China
Dylan Patel · 1h 13m spoken Sarah Wang · 8m spoken Guido Appenzeller · 5m spoken Erik Torenberg · 28s spoken
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of the a16z Podcast, SemiAnalysis founder Dylan Patel joins Sarah Wang, Erik Torenberg, and Guido Appenzeller to analyze the fast-evolving AI semiconductor landscape. Together, they explore NVIDIA's market dominance and Intel collaboration, US-China chip export controls, hyperscaler capital expenditure, and emerging hardware architectures.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. The host holds 0.5% of the talking time here. How this is scored →
The host as informed peer 2.8 Guest teaching 5.8 Guest disagreement 1.3 The host pushing back 1.4
speaking balance: gold is the host, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 1:35:33 Cocaine Procurement Analogy
Dylan bluntly compares buying enterprise AI GPUs to calling around for illegal drugs, using informal and provocative framing.
Hardest push from the host ▶ 1:00:25 Data Center Infrastructure ChallengeGuido directly challenges Dylan's call on AWS re-acceleration by questioning whether legacy AWS facilities have adequate density, water access, and liquid cooling capabilities.
Biggest teaching moment ▶ 1:32:00 Disaggregated Pre-fill vs Decode MechanicsDylan walks through the mathematical and architectural distinctions between pre-fill and decode compute requirements to clarify why CPX strips expensive HBM.
The host holds their own ▶ 44:00 Intel's Mask Revision DisastersGuido leverages his background as former Intel Data Center CTO to demonstrate insider knowledge on how Intel reached 15 mask steppings (E2), causing multi-quarter delays.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | The host as informed peer | Guest teaching | Guest disagreement | The host pushing back | Why |
|---|---|---|---|---|---|---|
| NVIDIA's Investment in Intel and Future Product Collaboration | 3 | 4 | 1 | 1 | Erik opens with NVIDIA's $5B investment in Intel. Dylan offers analysis on PC chiplet collaboration and Intel's financial position, while co-host Guido contributes context as former Intel Data Center CTO. | |
| Huawei's AI Roadmap, US Sanctions, and Domestic Mandates | 2 | 7 | 1 | 1 | Dylan provides a comprehensive history of Huawei's Ascend roadmaps, US sanction workarounds, and domestic stockpiles. Guido asks a clarifying question regarding China's export negotiation strategy. | |
| HBM Supply Chain Bottlenecks and Manufacturing Yields | 2 | 6 | 1 | 1 | Dylan breaks down high-bandwidth memory (HBM) supply bottlenecks, tracking Chinese equipment imports like etch tools and through-silicon vias. The host listens as Dylan explains manufacturing yield learning curves. | |
| Jensen Huang's Huawei Strategy and Technological Galapagosing | 2 | 5 | 1 | 1 | Dylan outlines Jensen Huang's strategy against Huawei, introducing Noah Smith's 'technological Galapagosing' concept. The hosts prompt the topic and let Dylan detail the thesis. | |
| Hyperscaler CapEx Growth Projections and AI Value Creation | 4 | 5 | 2 | 2 | Dylan presents his aggressive $455B-$500B CapEx numbers versus Wall Street consensus. Guido engages in peer-level discussion regarding white-collar productivity tax and AI economic value creation. | |
| Jensen Huang's Leadership Style and NVIDIA's Historical Bets | 2 | 6 | 1 | 1 | Dylan explains NVIDIA's historical bet-the-farm culture, aggressive supply ordering before contract confirmations, and Jensen's CFO/CEO dynamic. | |
| NVIDIA's First-Pass Silicon Culture and Engineering Leadership | 4 | 6 | 1 | 1 | Dylan highlights NVIDIA's engineering discipline of first-pass silicon success (A0 steppings). Guido adds host expertise by validating how Intel's repeated mask revisions (E2 steppings) created severe market delays. | |
| Capital Allocation Strategy and Ecosystem Neutrality | 2 | 5 | 1 | 1 | Dylan analyzes NVIDIA's massive balance sheet cash reserves and discusses why Jensen avoids buying entire startup funding rounds to remain ecosystem-neutral. | |
| Power Infrastructure as the Ultimate AI Bottleneck | 3 | 5 | 2 | 2 | Dylan argues NVIDIA should invest directly into power and data centers rather than cloud software layers. Guido adds historical context on Intel's customer concentration issues. | |
| AWS Growth Re-Acceleration and Infrastructure Dominance | 4 | 6 | 2 | 3 | Dylan outlines SemiAnalysis's thesis on AWS revenue re-acceleration. Guido pushes back intelligently, questioning whether legacy AWS data center capacity meets modern high-density liquid cooling requirements. | |
| The Engineering Trade-offs of Custom AI Silicon | 2 | 6 | 1 | 1 | Dylan details the engineering hurdles of developing custom AI chips like Trainium, pointing out how frontier labs write custom low-level assembly kernels for specific models. | |
| Oracle's Cloud Expansion and Financial Engineering | 3 | 7 | 1 | 1 | Dylan explains how SemiAnalysis predicted Oracle's surge by tracking satellite imagery, electrical permits, and transformer supply chains to map out gigawatt data center deals. | |
| xAI Colossus II Buildout and Regulatory Maneuvering | 2 | 6 | 1 | 1 | Dylan discusses xAI's Memphis cluster buildout and Elon Musk's tactic of expanding across the state border into Mississippi to bypass local power and municipal permitting bottlenecks. | |
| GB200 TCO, Blast Radius Reliability, and SLA Architecture | 3 | 7 | 2 | 2 | Dylan details the total cost of ownership (TCO) and severe reliability 'blast radius' issues of GB200 NVL72 architectures, explaining how failure rates require splitting high and low-priority workloads. | |
| Disaggregated Pre-fill, Decode, and CPX Architecture | 4 | 7 | 1 | 2 | Guido asks if NVIDIA's newly announced Rubin CPX pre-fill chip cannibalizes existing GPUs. Dylan explains the architectural split between compute-heavy pre-fill and memory-bandwidth-heavy decode. | |
| GPU Procurement Dynamics and the Neocloud Market | 2 | 5 | 2 | 1 | Dylan candidly compares buying GPUs to procurement dynamics in illegal drug markets. He wraps up by analyzing current Hopper versus Blackwell market tightness. |
Statements from this episode (42)
Patel: Buying AI GPUs resembles buying cocaine via informal networks
“How you buy GPUs is like buying cocaine. You call up a couple people, you text a couple people, you ask, yo, how much you got? What's the price?”
Dylan Patel: Intel is crawling to NVIDIA in a full-circle reversal
“It's kind of poetic that everything's gone full circle and Intel's sort of crawling to NVIDIA.”
Patel: Integrated x86 and Nvidia graphics laptop would be best on market
“An x-ay six laptop with NVIDIA graphics fully integrated would be probably the best product in the market.”
Appenzeller: Intel lacks competitive AI chips and its Gaudi effort is done
“They can't, they don't have anything competitive, right? There was the Gaudi effort that's more or less done, right? There was the internal graphics chips, which never competed really at the high end, right?”
Appenzeller: Intel-Nvidia partnership means AMD is 'fucked'
“I think AMD is fucked, right? I mean, they're, you're just, If your two arch nemesis suddenly team up, that's the worst possible news you can have, right? They were already struggling, right? Their cards are good, their software stack is not, right? They were …”
Appenzeller: Intel-Nvidia alliance weakens Arm's core value proposition
“I think ARM is a little bit screwed as well, right? Because they are, their biggest selling point was sort of like, look, we can partner with everybody that doesn't want to partner with Intel.”
Patel: Huawei Sourced 2.9 Million TSMC Chips via Shell Entities
“They were able to acquire three million chips, 2.9 million chips from TSMC through these other entities, right? Roughly five hundred million dollars worth of orders”
Patel: NVIDIA Lost $20B+ China Revenue from H20 Ban
“Our revenue estimate for Nvidia in China for just H-twenty was north of twenty billion because that's what they were booking in capacity slash had to write off.”
Patel: China Will Mass Produce 7nm AI Chips Despite Sanctions
“Even though the government says they're for 14 nanometer, the actual equipment that's banned is only for below seven nanometer, and so they'll be able to make a lot of seven nanometer AI chips and maybe even get to five with, you know, using existing equipment…”
Patel: China Will Pause NVIDIA Bans During AI Chip Capacity Gap
“I think they'll be able to ramp. I think it'll take a little bit longer and there will be, like, a sort of a gap in between where China probably backtracks and says it's fine.”
Patel: HBM production capacity remains a bottleneck for Huawei
“I think production capacity wise, it is still absolutely a bottleneck. They certain types of equipment required for making HBM need to be imported. They're working on domestic solutions, but as far as we know, they have not imported enough equipment for this.”
Patel: China stockpiled lithography at 30-40% of equipment imports
“China, because they wanted, they sort of, like, wanted to stockpile lithography, and they were worried about the coming ban, they were importing lithography at a much higher rate than that, right? Like, 30, 40% of their equipment imports were lithography, and …”
Patel: China has only sampled HBM2 and not started HBM3 production
“Yield, they haven't even started production of high speed, of HBM-III, right? They've only done some sampling of HBM-II.”
Patel: AI end market will far exceed total semiconductor market
“The end market of AI is going to be way larger than the end market of semiconductors and equipment.”
Patel: Huawei surpassed Apple in TSMC orders and market share before bans
“Yeah, well, like, I mean, like, every other, like, Huawei's beat Apple, right? They passed Apple up in TSMC orders, they passed Apple up in phone market share not in the US, but, like, in many parts of the world. Before the bans came down, and then even now th…”
Patel: Wall Street projects $360B hyperscaler CapEx next year
“The consensus for the banks is three hundred and sixty billion dollars of spend next year across all of them.”
Patel: Hyperscaler CapEx will reach $450B-$500B next year
“And my number is closer to, like, it's like four 5500 and that's based on, like, you know, all the research we do on, like, data centers and, like, tracking each individual data center in the supply chains, right?”
Patel: OpenAI will not achieve profitability until 2029
“They're not turning a cash flow, they're not gonna be profitable until 2029.”
Patel: NVIDIA is the only $10B+ semiconductor company founded post-1990
“It's the only semiconductor company that's worth, you know, I think even north of ten billion dollars that was founded as late as it was. Like MediaTek was in the early nineties and then Nvidia and everyone else is like from the seventies mostly.”
Patel: NVIDIA almost always ships first-pass A0 silicon
“So, like, NVIDIA always ships A zero. Almost always. They sometimes ship A one.”
Patel: NVIDIA added Tensor cores to Volta months before fab
“There's a story about how Volta, Which was the first NVIDIA chip with Tensor cores. You know, they saw all the AI stuff on the prior generation P. 100 Pascal. And they decided we should go all in on AI. And they added the Tensor cores to Volta, like only a han…”
Patel: NVIDIA will accumulate hundreds of billions in cash balance
“He's gonna have hundreds of billions of dollars of cash on his balance sheet.”
Patel: Silicon Valley AI startups spend 75% of venture rounds on GPUs
“Most companies in the valley spend, what, 75% of their round on GPUs, right?”
Patel: Energy and data centers are NVIDIA's primary growth bottleneck
“Do you invest in data centers and energy? Yeah. Do you invest in, because that's the bottleneck for your growth really is, is A, well, how much people want to spend and can spend, and B, the ability to actually put them in data centers.”
Patel: Apple has lacked visionary innovation under Tim Cook for a decade
“The reason why Apple hasn't done anything interesting in like, you know, nearly a decade is, is, you know, they've got a not visionary at the head. Tim Cook's greatest supply chain. And they're just plowing the money into buybacks.”
Appenzeller: Intel's customer concentration allowed hyperscalers to push down prices
“One of the biggest problems we had was that our customer base sucked, right? I mean, we were selling to, most of the chips went to the large hyperscalers, you know, which they're way too concentrated, and they build their own chips, and so you can push down yo…”
Patel: AWS year-over-year revenue growth will re-accelerate above 20%
“This is the lowest AWS revenue growth will be on a year-to-year basis. For at least the next year, right? And it's re-accelerating to north of 20% again because of all these massive data centers they have online with Tranium and GPUs, right?”
Patel: Amazon will lose its position as largest data center operator within two years
“The company with the most data center capacity in the world, that, and still today, although they may get passed up in the next two years is Amazon. Actually, they will get passed up based on what we see is Amazon, but incrementally, Amazon still has the most …”
Patel: AWS Trainium remains difficult for developers to program
“Oh, it's still bad. You know, it's tough to use.”
Patel: Oracle does not physically build its own data centers
“Oracle doesn't build their own data centers either, right? By the way, they get them from other companies. They co-engineer but they don't, Physically build them themselves.”
Patel: NVIDIA GB200 system costs $50,000 all-in per GPU
“If I'm talking about a GV 200, right? Each individual GPU is 1200 watts. But when you talk about the CPU, the whole system, it's roughly 2000 watts. At the same time, you know, all in, everything, simplicity's sake, 50,000 dollars per GPU, right? The GPU doesn…”
Patel: OpenAI signed up to pay Oracle $80+ billion annually
“With OpenAI, it's not, and so there's gotta be some, like, error bars as you go further out in terms of, like, Will OpenAI exist in 28, 29 30, and will they be able to pay the 80 plus billion dollars a year that they've signed up to Oracle with, right?”
Patel: Oracle purchases GPUs only 1-2 quarters before leasing
“The GPUs they purchase one to two quarters before they start renting them. So they, they're not, you know, the downside risk is pretty low for them in terms of If they don't get the deal, well, they don't get the revenue, but they're not, it's not like they ha…”
Patel: Approximately ten 100,000-GPU AI clusters exist globally
“There are like 10 hundred K GPU clusters in the world.”
Patel: Elon Musk deployed 100,000 GPUs in Memphis within six months
“A hundred K GPUs in six months. He bought a factory in like February of 24 and had models training within six months, right?”
Patel: Musk bought a Mississippi power plant to bypass energy constraints
“His facility's like a mile away from Mississippi, and he bought a power plant in Mississippi and he's putting turbines there. Yeah. The regulation is completely different, right?”
Patel: NVIDIA GB200 yields 3x-4x performance per dollar on DeepSeek
“If you're running DeepSeq inference, the performance difference per GPU is like north of like six, seven X, and it continues to optimize you know, for DeepSeq inference. And so the, you know, then it's like, well, I'm only paying 60% more for six X, and it's l…”
Patel: Next-gen GPU failure rates are flat or getting worse
“GPU failure rates at best are the same and likely worse, right? Gen on gen, because everything's getting hotter, faster, et cetera.”
Patel: Multiple major neoclouds have sold out of NVIDIA Hopper capacity
“It turned out that multiple major neoclouds had sold out of hopper capacity.”