This week, Dylan Patel put a number on it: by 2028, OpenAI and Anthropic will control most of the world's available AI compute. This isn't speculation — it's a trend extrapolation by SemiAnalysis (a semiconductor and AI infrastructure research firm) built on capex, chip capacity, and buyer cash flow. What we should care about: the byproduct may be a sovereign debt crisis.

What this is

Dylan Patel is the founder of SemiAnalysis and a long-time tracker of the chip and compute markets. He laid out several key judgments on his brother Dwarkesh's podcast:

  • By 2028, the available compute (FLOPs, think of it as "total AI man-hours") controlled by OpenAI and Anthropic will exceed that of all other players combined.
  • The reason: these two companies can monetize AI (subscriptions + APIs), so they keep outbidding rivals in the compute auction.
  • The monopoly is underwritten by massive capex: global AI-related investment will exceed $10 trillion in the coming years — roughly two years of the US federal budget.
  • Laboratories are shifting more compute from "inference" (answering user questions) to "R&D training," and the competitive focus has moved to next-generation models.

Industry view

The mainstream narrative treats this as a justified "winner-take-all" outcome: Nvidia's stock is soaring, hyperscalers (i.e., AWS, Azure, Google Cloud — the giant cloud companies) keep posting record capex, and the market is already pricing in this end state.

But dissent is real:

  • Sovereign debt risk: Hyperscalers are funding chip purchases with debt; if AI revenue falls short, rates could rise, squeezing countries that hold no AI assets.
  • Open source and outside players: DeepSeek (China) and Mistral (France) prove that efficient models don't need massive compute — centralization isn't irreversible.
  • Geopolitical constraints: China receives less than 10% of new global compute, yet produces competitive models with fewer resources, undercutting the "compute is everything" logic.
  • Antitrust possibility: Compute is increasingly viewed as national infrastructure, and regulation or export controls could rewrite the outcome.

Impact on regular people

For enterprise IT: In the coming years, enterprises sourcing AI services will be down to two core vendors, leaving less bargaining room than today.

For careers: AI training and infrastructure roles will continue concentrating in a few companies, but application-layer (building products and services with AI) job opportunities will actually grow.

For consumer markets: High-stakes compute bidding will push up AI product costs, ultimately passed on to consumers as subscription hikes or service fees.