Dwarkesh Podcast: Dylan Patel — Deep Dive on the 3 Big Bottlenecks to Scaling AI Compute

Date: March 13, 2026
Host: Dwarkesh Patel
Guest: Dylan Patel (CEO of SemiAnalysis)

Episode Overview

This episode is a rigorous, deeply technical, and often provocative exploration of the key constraints impacting the rapid scaling of AI compute. Dwarkesh Patel interviews Dylan Patel—one of the industry’s best-connected and most forthright semiconductor analysts—about where the bottlenecks truly lie as the AI industry, hyperscalers, and labs like OpenAI and Anthropic pour unprecedented amounts of capital into data centers, chips, and supply chains. The conversation spans capex spending, the economics and physics of semiconductors, supply chain strategy, geopolitical competition, energy, labor, and speculative futures for both Earth- and space-based computing.

Key Themes & Discussion Points

1. The Scale and Timing of AI Compute Expansion

Timestamps: [00:14], [01:41], [04:01]

• The combined 2026 CapEx of Amazon, Meta, Google, and Microsoft is forecast at ~$600B; total AI-centered capex globally is nearing $1T.
• Expansion isn’t all immediate: big tech CapEx is staggered—today’s spending on turbines or power agreements is for capacity coming online years ahead.
• OpenAI and Anthropic have already raised record sums, but much of their forecasted compute—measured in gigawatts—not available immediately.
  • “Anthropic needs to get to well above 5 GW by the end of this year. And it’s going to be really tough for them to get there. But it’s possible.” — Dylan [03:36]
• Labs that move first and boldest in locking up compute (OpenAI) gain pricing and availability advantages over more conservative rivals (Anthropic).

2. How Labs Secure Compute, and Downstream Effects

Timestamps: [04:20], [06:17], [09:20]

• OpenAI has “YOLO”-ed their compute procurement, signing massive, sometimes risky, multi-year deals with hyperscalers and lesser-known neoclouds.
  • “OpenAI has kind of got way more access to compute than Anthropic by the end of the year.” — Dylan [04:20]
• When demand exceeds capacity, late buyers must go to less-preferred providers and pay higher prices, sometimes 50%+ above baseline.
• Compute supply contracts vary (5-year vs. spot purchases); as AI demand surges, short-term contracts cost more, and labs find themselves fighting over the scraps.
  • “I've seen deals...as high as $2.40 for two to three years for H100s, which if you think about the margin…those margins are way higher.” — Dylan [07:36]
• Financially, this implies gross margin advantages persist for those who locked in early; latecomers are squeezed.

3. Depreciation Cycles and Value of a GPU

Timestamps: [10:03], [11:21], [15:08]

• Debate over GPU depreciation cycles: Michael Burry et al. predicted 3 years or less; Dylan argues much longer cycles are possible—even well over 5 years—because AI demand outstrips chip production and hardware is far from commoditized.
  • “The other lens is...what is the utility you get out of the chip? ...Because you are so limited…the prices these chips is not...what’s the comparative thing I can buy today? It’s actually what is the value I can derive out of this chip today, right?” — Dylan [13:29]
• As models get more powerful and efficient, the utility and thus market value of “old” GPUs goes up, not down.

4. CapEx Economics, Lock-in, Market Power

Timestamps: [19:37], [21:01], [24:51]

• Alchian-Allen effect: as the fixed cost of compute rises (GPU price inflation), buyers put more value on higher quality outputs (i.e., better AI models).
• Who captures the value?
  • Labs with locked-in cheap compute enjoy windfalls. Similarly, chip, memory, and foundry suppliers (esp. Nvidia, TSMC, SK Hynix) accumulate substantial margin.
  • “Who is able to accrue all the margin dollars is...potentially the cloud, potentially the chip vendors and the memory vendors. Until TSMC or ASML break out and they're like, no, actually we're going to charge a lot more.” — Dylan [22:22]

5. Semiconductor and Supply Chain Bottlenecks

Timestamps: [24:51], [26:07], [34:51], [36:51], [37:03], [41:06], [46:07], [51:31]

• Chips/Logic Production: Nvidia outcompetes on “AGI-pilled” long-term orders with TSMC; Apple (once dominant) is rapidly losing its strategic position.
• Memory Crunch: HBM demand, compounded by AI workloads’ need for large and fast memory, has driven up prices and crowded out the consumer market.
  • “A third of their capex is going to memory.” — Dwarkesh [83:33]
• Wafer Fabs and EUV Tools: Fabs (factories) and tools (esp. ASML’s EUV lithography machines) are the ultimate global bottleneck.
  • “By 2028-29, the bottleneck falls to the lowest rung on the supply chain, which is ASML.” — Dylan [37:03]
  • Physical and organizational complexity means even large prepayments can only accelerate fab/tool production so much.
• Margins and Power: Memory market squeeze elevates AI, depresses smartphone/PC markets—“people are going to start hating AI even more...PCs, smartphones, getting incrementally worse.” [83:41]

6. China vs. the West: Geopolitics and the Future of Compute

Timestamps: [65:36], [66:21], [66:29], [69:29], [70:36]

• China’s massive state capacity could, by 2030+, rival or overtake the “West” in pure chip production, especially if “timelines” for AI progress prove slower than expected.
  • “Fast timelines, US wins; long timelines, China wins.” — Dylan [75:49]
• But for now, China’s reliance on imported ASML tools and its lag in capex/networking means US labs have pulled ahead.
• If progress is rapid, as in the last 12 months, US and allied companies will enjoy a period of massive returns on infrastructure investments.

7. Energy, Labor, and Space-based Compute

Timestamps: [93:26], [103:19], [110:28], [111:37], [114:38]

• Power supply: No sign that energy infrastructure will bottleneck AI scaling before chips do. The US grid can be flexibly leveraged, alternative turbines, batteries, engines, and modular/behind-the-meter solutions abound.

• Labor: Labor to physically build datacenters and power systems is a growing constraint, but could be alleviated by modularization and importation of skilled workers.

• Space Centers: Elon Musk’s idea of space-based GPU arrays only makes sense when terrestrial energy or permitting becomes a hard constraint; in the current paradigm, deploying the scarce chips instantly, on Earth, is much more valuable.

  • “All that matters in a chip constrained world is get these chips working on producing tokens ASAP in a world...space data centers will eventually be a 10x game, potentially as Earth resources get more and more contentious. But that’s not this decade.” — Dylan [122:13]

Notable Quotes & Memorable Moments

• “There’s this meme: having the best model is a depreciating asset, but the reason it’s important [to have it first] is because you can sign these deals and lock in compute.” — Dwarkesh [10:03]
• “H100 is worth more today than it was three years ago.” — Dylan [15:09]
• “Who is AGI-pilled enough to buy compute in long timelines at levels that seem ridiculous to people who aren’t AGI-pilled?” — Dylan [29:30]
• “If smartphone volumes are halved...the percentage of the BOM that goes to memory and storage is much larger and the margins are lower. So there’s less capacity to even eat the margins.” — Dylan [84:09]
• “Chips are the bottleneck. You want them deployed, working on AI the moment they’re done being manufactured.” — Dylan [122:13]
• “Fast timelines, the US wins. Long timelines, China wins.” — Dylan [75:49]
• “If in 2019 that issue that Huawei was not banned from using TSMC, Huawei would have already eclipsed Apple as the biggest TSMC customer... It’s very arguable that Huawei, if they had tsmc, would be better than Nvidia.” — Dylan [142:25]

Key Technical Insights

Bottleneck 1: Semiconductor Manufacturing Tools (EUV)

• ASML EUV machines: 70 built this year, 100/year by 2030; each supports only a few gigawatts of AI compute per year.
• Upstream supply chain (Zeiss, Cymer, etc.) is artisanal, slow to scale.
  • “You can't just train random people for this in the snap of a finger.” [51:32]

Bottleneck 2: HBM/Memory

• AI chips need HBM, which is 3-4x less bits per wafer than standard DRAM; caps how quickly memory supply can scale.
• As memory for AI grows, consumer device markets shrink.

Bottleneck 3: Power/Data Centers/Physical Buildout

• Energy bottleneck now seen as less severe; creative solutions in turbines, batteries, permitting, and grid management.
• Labor may be limiting, but is partially addressable with modular construction and global talent.

China vs. US: Who Wins the Compute Race?

• If AI progress is rapid, the consolidation of capex, supply chain lock-in, and scale in the US/West is a major, perhaps enduring, advantage.
• If timelines slow down (to 2035+), China can leverage central planning, scale, and eventual 100% domestic supply chain to catch up or surpass.

Technology Futures: Robotics, Cloud Centralization, and More

• Humanoid Robots: Centralized cloud compute will drive many, with only action/interpolation at the edge, to conserve semiconductors and power.
• Space Compute: Not viable this decade—deployment lag and reliability challenges with current chip supply.
• Apple’s Downfall: As TSMC’s capacity shifts, Apple will no longer be the dominant customer; AI accelerators will take precedence.
• Potential Wildcards: Technologies such as 3D DRAM could shift bottlenecks, but timelines are uncertain and require major retooling.
• If Taiwan is lost: Airlifting engineers is not enough; global compute capacity crashes, growth slows massively.

Structured Timeline of Major Topics

• [00:14-04:01]: Scale and timeline of CapEx & compute for labs/hyperscalers
• [04:01-09:32]: How OpenAI/Anthropic are securing compute, competition over supply
• [09:32-15:09]: The value lifecycle and depreciation of GPUs, implications for margins
• [19:37-24:51]: CapEx logic, lock-in, who captures margins, Alchian-Allen effect in AI
• [24:51-46:07]: How Nvidia locked up the chip/memory supply chain, logic, and memory—TSMC, SK Hynix, Samsung, etc.
• [46:07-54:42]: Why EUV tool supply can’t easily scale, component-level bottlenecks, ASML supply chain complexity
• [55:00-65:36]: Can “old” process nodes step in as a fallback, real limits of non-leading-edge chips
• [65:36-75:49]: China’s potential to catch up; “fast timelines, US wins; slow timelines, China wins”
• [75:49-91:42]: Memory bottleneck’s implications for AI scaling and the consumer market
• [93:26-111:37]: Power infrastructure scale-up, labor constraints, modularization, global logistics
• [114:38-124:39]: Space-based compute: why it’s not imminent; edge devices vs. centralized brains
• [126:08-134:06]: Differences in compute network topology (Nvidia vs. Google vs. Amazon), scale-up domains
• [134:06-144:36]: SemiAnalysis’s data, how and why it’s used for financial advantage, Apple’s waning dominance
• [144:36-end]: Robotics, centralized intelligence, implications for supply chain risk, the Taiwan scenario

Final Takeaways

• Chips (and the tools to make them) are the ultimate bottleneck for AI scaling—far more so than power, land, or even memory, though all play a role.
• Labs and companies that moved first and are “AGI-pilled” enjoy compounding economic and competitive advantages.
• Geopolitics and industrial policy (esp. in China and the US) will determine who “wins” as the field centralizes.
• Space, robotics, supply chain modularization, and even “slow-mode” AI all have roles to play in the battle to maximize useful compute.
• The landscape is shifting so rapidly that even inside players and their suppliers can misjudge the pace and magnitude of change. As Patel observes: “Our numbers are always too high—until suddenly, they aren’t.”

Listen to the episode here for more: The full discussion is packed with industry anecdotes, contrarian insights, and an on-the-ground look at what’s really driving the next wave of AI infrastructure.