Dylan Patel predicts OpenAI and Anthropic will dominate global compute by 2028
Dwarkesh Patel sat down with SemiAnalysis founder Dylan Patel to discuss the economics of AI labs. Core thesis: as recursive self-improvement (RSI) nears, training costs will exceed inference costs and compute will shift from inference to training; OpenAI and Anthropic, with stronger compute monetization, will win compute bidding and could control most of the world's FLOPs—and thus AI labor—in the coming years.
Confirmed
- Patel predicts compute will shift from inference to training, with training costs exceeding inference costs
- OpenAI and Anthropic monetize compute better, giving them an edge in bidding for it
- Figures: the two labs took 30% (about one-third) of incremental compute early this year, expected to reach 40–50% next year; global compute doubles annually while frontier-lab compute triples with more efficient chips—if trends hold, they could control most global compute by 2028
- @AccBalanced adds economic corollaries: revenue per MW may determine compute ownership; inference profits can fund larger training clusters; a lab's best customer may be itself; compute ownership becomes labor ownership (the "Dario index thesis")
- @bookwormengr argues AI is accelerating toward compute centralization given current scaling and algorithmic progress
- @toptickcrypto notes the market may underestimate frontier labs' conviction: "hyperscalers spending over 100% of cash flow on compute" sounded crazy years ago but is now reality
Why it matters
- If the prediction holds, two labs would monopolize both global compute and AI labor, reshaping the AI industry
- Frameworks like "revenue per MW decides ownership" and "inference subsidizes training" illuminate AI infrastructure investment and lab business models
- The discussion was amplified by industry figures including Hugging Face CEO Clement Delangue
2026-08-25 ~ 2026-08-27 · 8 related posts
Primary sources
- [source] Dylan Patel: Two Labs Will Soon Control Most of the World's Compute — Dwarkesh Patel · 2026-08-25
- Anthropic and OpenAI on Track to Control Most Global Compute — Borthwick · 2026-08-26
- Prediction: OpenAI and Anthropic to Control Most Global Compute by 2028 — SucceededMind · 2026-08-26
- [source] Dario's 'Exponential': Compute Ownership to Become Workforce Ownership — AccBalanced · 2026-08-26
- Dylan Patel: Top AI labs will control most global compute by 2028 — Traditional-Chip8339 · 2026-08-26
- [source] OpenAI and Anthropic take a third of incremental world compute, Dylan Patel says — bookwormengr · 2026-08-27
- Frontier Labs' Compute Spending: From Crazy to Reality — toptickcrypto · 2026-08-27
1 near-duplicate retellings: scaling01