Dylan Patel, CEO of SemiAnalysis, discusses the trillion-dollar semiconductor supply chain powering AI infrastructure. The conversation covers how big tech's $600 billion combined CapEx from Amazon, Meta, Google, and Microsoft translates to future compute capacity, the bottlenecks constraining AI scaling, and the geopolitical implications of semiconductor dependencies.

The discussion explores multiple constraint layers: from immediate power and data center limitations to longer-term bottlenecks in EUV lithography tools and memory production. Patel explains how these constraints shift over time, with clean rooms and fabs becoming critical in 2025-2026, before the ultimate bottleneck emerges in ASML's EUV tool production capacity by 2028-2030.

The conversation also examines the competitive dynamics between AI labs, particularly how OpenAI's aggressive compute procurement strategy contrasts with Anthropic's more conservative approach, and the implications for China's semiconductor ambitions versus Western infrastructure advantages in different AI timeline scenarios.

The $600 Billion CapEx Timeline and AI Lab Compute Needs

Big tech's $600 billion combined CapEx includes setup costs for future years: 'A big chunk of that is spent on turbine deposits for 28 and 29. A chunk of that is spent on data center construction for 27' - Dylan

Anthropic and OpenAI currently operate at 2-2.5 gigawatts but need massive scaling to serve revenue growth, with Anthropic adding $4-6 billion monthly revenue requiring proportional compute expansion

Anthropic's conservative compute strategy backfired: 'Dario was very conservative... but in reality, he's definitely missed the pooch in terms of going like OpenAI, which was let's just sign these crazy deals' - Dylan

Both labs will reach 5-6 gigawatts by year-end through direct capacity plus cloud partnerships via Bedrock, Vertex, and Foundry revenue sharing arrangements

GPU Pricing Dynamics and Last-Minute Compute Acquisition

H100 rental prices have surged from $1.40 to $2.40+ per hour as AI labs outbid other customers for scarce capacity over 2-3 year contracts

GPU depreciation cycles extend beyond traditional 3-5 years because model improvements increase value per chip: 'An H-100 is worth more today than it was three years ago' - Dylan

Last-minute compute acquisition forces labs to accept higher costs through neo-clouds, shorter-term contracts, or revenue-sharing arrangements with hyperscalers

The Alkin-Allen effect applies to AI: fixed cost increases in compute make users willing to pay higher margins for slightly better models rather than cheaper alternatives

The Memory Crunch Destroying Consumer Electronics

Memory prices have tripled, adding $150+ to iPhone costs and forcing smartphone volumes down from 1.4 billion to projected 500-600 million annually

HBM provides 2.5 terabytes/second bandwidth per stack versus DDR's 64-128 gigabytes/second in the same chip edge area - an order of magnitude difference

30% of big tech's 2026 CapEx goes to memory, with AI chips requiring 170,000 DRAM wafers per gigawatt of data center capacity

Low-end and mid-range smartphones bear the brunt of cuts: 'Xiaomi and Oppo are cutting low-end and mid-range smartphone volumes by half' due to memory cost increases - Dylan

The Ultimate Bottleneck: EUV Tools and ASML's Production Limits

Each gigawatt of AI capacity requires 3.5 EUV tools, with ASML producing only 70 this year, scaling to 80 next year and 100 by 2030

By 2030, 700 total EUV tools could theoretically support 200 gigawatts of AI chips, making Sam Altman's 52 gigawatts annually feasible at 25% market share

EUV tools are humanity's most complex machines, requiring 18 precision mirrors with molybdenum-ruthenium layers and components moving at 9 Gs with sub-nanometer accuracy

ASML's supply chain includes over 10,000 suppliers, with critical components from Zeiss (optics), Cymer (light source), and specialized stages manufactured across multiple countries

Power Solutions and Behind-the-Meter Alternatives

Power constraints are solvable through diverse sources: combine cycle turbines, aeroderivatives, reciprocating engines, ship engines, and fuel cells from 16+ manufacturers

Half of new capacity by decade-end will be behind-the-meter due to grid interconnection challenges, despite higher costs than grid-connected power

The US grid has 20% excess capacity for peak load management that could be unlocked for data centers through utility-scale batteries and peaker plants

Energy costs remain small fraction of total GPU TCO: doubling power costs increases H100 hourly cost from $1.40 to just $1.50, easily absorbed by model value improvements

China's Semiconductor Ambitions and Geopolitical Implications

China could achieve 100+ DUV tools annually by 2030 for indigenous production, but still lacks EUV capabilities for leading-edge processes

Fast AI timelines favor the West due to massive infrastructure investments, while longer timelines beyond 2035 could enable China's vertical supply chain advantages

Huawei with TSMC access 'would arguably be better than NVIDIA' given their superior AI research, networking expertise, and integrated capabilities - Dylan

Taiwan risk remains critical: losing TSMC would drop global AI chip capacity from hundreds of gigawatts annually to maybe 10-20 gigawatts through Intel and Samsung

Space Data Centers and Future Scaling Challenges

Space data centers face deployment delays, reliability challenges, and expensive optical interconnects that outweigh free power advantages in chip-constrained world

GPU failure rates of 15% requiring RMA, plus months-long space deployment, waste precious capacity during peak constraint periods

Space makes sense only after 2035 when chip production scales sufficiently and Earth resources become more contentious than manufacturing capacity

Modularization and factory integration will reduce data center labor requirements from current 5,000 workers per 1.2 gigawatts to more manageable levels