Dylan Patel — The Single Biggest Bottleneck to Scaling AI Compute

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.