ReadTube
Back to Dwarkesh Patel

Dylan Patel — The Single Biggest Bottleneck to Scaling AI Compute

Dwarkesh PatelMarch 13, 20262h 30m
Topics29
Big Tech CapEx and Compute Timeline0:00TSMC Allocation Preferences for Stable vs. High-Growth Markets27:06AGI-Pilled Commitment and Supply Chain Dynamics29:36Evolving AI Compute Bottlenecks34:32EUV Tools as Ultimate Bottleneck37:01Limits to ASML Expansion46:01AI Compute Power Demands and Supply Chain Constraints54:09Why 7 nm Fallback is Naive: Design Targets and Numerics Shift58:01Multi-Chip Scaling and Inter-Chip Communication Losses1:00:09China vs West: Process Tech vs Scale Crossover1:05:32Long-Term Projections and China Dominance1:09:34Memory Crunch and HBM vs Commodity DRAM1:16:01Nvidia CapEx and Memory Allocation1:23:23Memory Crunch Expectations1:23:30Memory Production Constraints1:29:05Elon Musk's TeraFab Plan1:32:40Power Scalability1:42:30Behind-the-Meter Power Solutions1:50:08Skepticism on Space GPUs1:54:30GPU Reliability and Deployment Challenges1:58:03Future Optimization2:03:31Scale-Up Domains Explained2:06:00Parameter Scaling and Scale-Ups2:09:00SemiAnalysis Spreadsheets and Clients2:14:01Investment Bets on Recent Trends and Memory Crunch2:16:32TSMC N2 Capacity and Apple vs. AI Customers2:18:31Huawei's Potential if Unbanned2:22:02Robots and Compute Implications2:24:33De-Risking Taiwan via Airlifting Engineers2:28:34
In a Nutshell

Big Tech's $600B CapEx this year deploys ~20-50 GW of AI compute, but the single biggest long-term bottleneck to scaling beyond 200 GW/year by 2030 is ASML's EUV tools (limited to ~100/year), constraining TSMC's advanced node production amid exploding HBM/DRAM demand that triples prices and halves smartphone volumes. AI labs like OpenAI/Anthropic race to lock 5+ GW via aggressive multi-provider deals (Nvidia, CoreWeave, Oracle) at $2-2.40/GPU-hour, prioritizing unipolar fleets for massive models while power/labor/data centers prove solvable via modularization, turbines, and behind-the-meter gas. US leads China in compute/revenue scaling (10 GW/lab next year), but Huawei could dominate sans bans; space GPUs remain impractical due to deployment delays and chip limits.

AI-Generated Notes

These notes were generated by AI and may contain inaccuracies.

Combined forecasted CapEx of big four—Amazon, Meta, Google, Microsoft—is $600 billion this year. Yearly prices of renting that compute equate to close to 50 gigawatts. Not all coming online this year; paying for compute coming online over coming years.

OpenAI raised $110 billion, Anthropic raised $30 billion. Compute coming online this year for labs on order of four gigawatts total. Cost to rent compute that OpenAI and Anthropic will have this year is $10 to $13 billion per gigawatt. Raises cover yearly compute spend, excluding revenue.

$600 billion CapEx across supply chain totals order of $1 trillion. Portion for compute online this year (chips, parts). Much is setup CapEx. 20 gigawatts incremental capacity added this year in America; portion not spent this year, some from prior year.

Google's $180 billion: chunk on turbine deposits for '28/'29, data center construction for '27, power purchase agreements, down payments for future scaling. Applies to all hyperscalers.

Anthropic and OpenAI at 2 to 2.5 gigawatts now, scaling larger. Hyperscalers' biggest customers are Anthropic and OpenAI.

Anthropic: $4-6 billion revenue added last few months. Straight line: add $6 billion/month, implying $60 billion revenue next 10 months. At reported gross margins, $40 billion compute spend for inference. At $10 billion/gigawatt, need 4 gigawatts inference capacity (assuming flat R&D training fleet). Need >5 gigawatts by year-end; tough but possible. OpenAI similar or higher.

Sign in to read the full notes

Get access to AI-generated notes, topic timestamps, and more.