# Dylan Patel — The Single Biggest Bottleneck to Scaling AI Compute Source: https://www.youtube.com/watch?v=mDG_Hx3BSUE Recap page: https://rapidrecap.app/video/mDG_Hx3BSUE Generated: 2026-03-13T17:03:35.193+00:00 --- ## Quick Overview The single biggest bottleneck to scaling AI compute moving forward, especially by 2028 and beyond, shifts from power and data centers back to the semiconductor supply chain, specifically the limited production capacity of ASML's Extreme Ultraviolet (EUV) lithography tools, which are essential for advanced logic and memory fabrication. **Key Points:** - The combined forecasted CapEx for Amazon, Meta, Google, and Microsoft is $600 billion, which translates to close to 50 gigawatts of potential compute, but much of this spending covers long lead time setup for 2027 and beyond, with roughly 20 gigawatts of incremental capacity deployed in America this year. - Anthropic currently operates at roughly two to two-and-a-half gigawatts and needs to scale to well above five gigawatts by the end of the year just to sustain projected revenue growth, implying they might reach five or six gigawatts by year-end through direct capacity and cloud services like Bedrock. - OpenAI has secured significantly more compute access than Anthropic by being more aggressive in signing deals with many providers, including CoreWeave and SoftBank Energy, while Anthropic's conservatism led to being constrained when compute demand exploded. - The economic value of an H100 GPU increases over time because advancements like GPT-5.4 allow the chip to serve more tokens of a higher-quality model, meaning its utility value, rather than its comparative performance to future chips, prices it in a constrained environment. - The depreciation cycle for GPUs may be longer than previously assumed (e.g., less than two years), as the utility value derived from rapidly improving models keeps the effective value of existing hardware high. - The ultimate long-term bottleneck is the manufacturing of logic wafers, requiring approximately three and a half EUV tools to fabricate one gigawatt of Nvidia's latest chips, equating to about $1.2 billion in tooling for $50 billion worth of data center capacity. - ASML's production capacity, limited by its complex supply chain components like Zeiss optics and Cymer sources, is projected to reach only a little over 100 EUV tools by the end of the decade, constraining total potential AI compute to around 200 gigawatts, even if fully allocated to AI. **Context:** Dwarkesh Patel interviews Dylan, the CEO of SemiAnalysis, focusing on the massive capital expenditure ($600 billion forecasted by the Big Four hyperscalers) directed towards AI compute infrastructure and the scaling challenges faced by leading AI labs like OpenAI and Anthropic. The discussion contrasts the immediate needs for inference capacity supporting current revenue against the long-term capital commitments made for future infrastructure buildouts, particularly concerning the semiconductor supply chain. ## Detailed Analysis The primary constraint on scaling AI compute is identified as the semiconductor supply chain, specifically the manufacturing capacity of ASML's EUV tools, which dictates the rate at which leading-edge logic and memory wafers can be produced. While hyperscaler CapEx is enormous, much of it covers long-term setup deposits for 2027 and beyond, with current deployment around 20 gigawatts in America this year. AI labs like Anthropic are struggling to meet their own explosive revenue-driven compute needs, requiring them to scale from 2-2.5 GW to over 5 GW by year-end, a feat made harder by their conservative early procurement strategy compared to OpenAI's aggressive deal-making across numerous suppliers. The conversation highlights an economic dynamic where the utility value derived from increasingly powerful models (like GPT-5.4) keeps the market price and effective lifespan of existing H100 GPUs high, contradicting bearish predictions of rapid depreciation. Critically, achieving one gigawatt of modern AI chip capacity requires about two million EUV passes, necessitating roughly three and a half EUV tools costing $1.2 billion, yet ASML can only increase output from 70 to just over 100 tools by 2030 due to the complexity and lack of expansion efforts in their multi-layered supply chain, including optics from Carl Zeiss and sources from Cymer, which are not AGI-pilled enough to rapidly scale their highly specialized components. ### Hyperscaler CapEx and Timeline - $600 billion combined forecasted CapEx by Big Tech - Much of this pays for setup deposits for 2028/2029, like turbine deposits for Google - Roughly 20 gigawatts of incremental capacity deployed in America this year. ### AI Lab Compute Requirements - Anthropic needs to reach well above five gigawatts by year-end for inference alone, assuming current revenue growth trajectory - OpenAI secured more compute by aggressively signing deals with many players including CoreWeave and SoftBank Energy - Anthropic faces difficulty due to prior conservatism, potentially forcing them to use lower-quality providers or revenue-share agreements. ### GPU Valuation and Depreciation - The value of an H100 is set by the utility derived from serving newer, better models like GPT-5.4, not by comparison to future chips - This implies that the depreciation cycle for GPUs may extend beyond five years, contrary to some bearish financial models. ### The Ultimate Bottleneck - The constraint shifts from power and data centers to the semiconductor supply chain, specifically logic and memory wafers - One gigawatt of Nvidia's latest chips requires 55,000 3nm wafers and approximately three and a half EUV tools. ### ASML and EUV Constraints - ASML can only increase EUV tool production from 70 now to just over 100 by the end of the decade, constraining total output to about 200 gigawatts annually - This slow expansion is due to the complex, specialized, multi-component supply chain (Zeiss optics, Cymer sources) that has not been aggressively scaled up. ### Market Allocation Dynamics - TSMC prioritizes allocation to more stable growth markets like CPUs (e.g., Apple, Amazon Graviton) over volatile high-growth AI chips, though Nvidia secures the majority of high-performance capacity due to early commitment and willingness to pay high margins.