Dylan Patel – Anthropic & OpenAI will have most of the world’s compute by 2028

What It Covers

Dylan Patel of SemiAnalysis projects that Anthropic and OpenAI will control the majority of the world's usable compute by 2028, driven by revenue per megawatt surging from $13M to potentially $100M+, accelerating centralization of AI infrastructure, and cascading effects on global interest rates, sovereign debt, and capital markets.

Key Questions Answered

• •Compute Centralization Timeline: Anthropic and OpenAI currently hold roughly 5 gigawatts each, representing approximately 30% of incremental compute added in 2025. By end of 2026, that share reaches 40–50% of new compute. By late 2028, on current trajectories, the two labs combined could control the majority of the world's high-performance compute, measured in effective FLOPS rather than raw watts.
• •Revenue Per Megawatt as the Key Metric: Tracking revenue per megawatt reveals lab economics more clearly than headline revenue figures. Anthropic has already reached $50M per megawatt in some segments, versus a $10–15M breakeven cost. Investors and compute suppliers should use this ratio to assess whether a lab can outbid competitors for scarce capacity, since whoever generates the highest revenue per megawatt wins the compute auction.
• •Inference-to-Training Reallocation Signal: Contrary to consensus expectations that inference will dominate compute allocation, Patel argues labs are already quietly shifting a higher fraction of compute toward R&D and training. When marginal inference revenue exceeds $30–50M per megawatt but internal AI research generates even higher returns, the rational move is to redirect capacity inward — a signal that external token availability may plateau even as total lab compute grows.
• •Supply Chain Bottleneck Arbitrage: A structural 100x gap exists between fab-level CapEx and end AI revenue generated. Roughly $6B in wafer fab tooling produces 1 gigawatt annually, which generates ~$100B in AI revenue over its lifetime. This gap creates arbitrage opportunities throughout the supply chain — turbine resellers, ASML tool holders, and substrate suppliers are already repricing upward, and the bullwhip effect means full supply chain repricing lags demand signals by 2–3 years.
• •China Compute Gap and Export Control Effectiveness: Export controls have compressed China's share of global AI compute from roughly 30% in 2022 to under 10% of incremental watts today. By 2028, China may reach 30 gigawatts total, but those gigawatts run on domestically produced chips that are 3–5x less efficient per watt than NVIDIA or Google hardware. This means a leading Western lab in 2028 could individually hold more effective compute than all of China combined.
• •Sovereign Debt and Interest Rate Risk: The $11T in cumulative AI infrastructure CapEx projected from 2024–2029 requires approximately $5T in new debt issuance across hyperscalers and their supply chains. This volume of corporate borrowing competes directly with government and consumer debt, pushing market interest rates up by an estimated 200–300 basis points. Countries with high debt-to-revenue ratios and short-duration bonds — particularly lower-income nations — face elevated default risk as capital reallocates toward AI infrastructure returns.