Executive Summary
Dylan Patel delivers the most comprehensive technical analysis of AI compute bottlenecks available, revealing a mathematical ceiling that will constrain the industry by 2030. His core insight: while power and data centers dominated 2024-2025 constraints, the fundamental limitation shifts to semiconductor manufacturing—specifically ASML's EUV lithography tools. The math is stark: each gigawatt of AI compute requires 3.5 EUV tools, but ASML can only produce 100 tools annually by decade's end, capping global AI deployment at 200 gigawatts. This represents a 25% market share for Sam Altman's stated 52-gigawatt annual target—reasonable but constrained. Meanwhile, memory becomes the immediate crisis as HBM production requires 4x more wafer area than standard DRAM, forcing smartphone volumes to halve and iPhone costs to rise $250. Patel's supply chain tracking shows Nvidia secured early TSMC allocation while Google missed the boat, creating asymmetric advantages. The analysis reveals why Anthropic struggles with compute access despite $30B in funding—they were conservative on long-term contracts while OpenAI signed aggressive five-year deals. Most critically, the semiconductor supply chain hasn't embraced AGI-level demand scaling, with each tier applying a 'minus one' or 'divide by two' mentality to capacity planning. This creates arbitrage opportunities for those who understand the true scale of coming demand.
Key Insights
what Dylan Patel said“Three and a half EUV tools satisfies a gigawatt... 700 EUV tools by the end of the decade gets you to 200 gigawatts worth of AI chips”
This is a preview. Log in to see the full analysis including investment opportunities, risks, catalysts, and detailed insights.