Anthropic and OpenAI are on track to become the marginal buyers of the world's AI compute. Who pays the bill?

Anthropic and OpenAI are on track to become the marginal buyers of the world's AI compute within a few years. Dylan Patel, the chip analyst who runs the SemiAnalysis newsletter, laid out that argument on the Dwarkesh Podcast this week, and the rest of the story is mostly mechanism: how the concentration happens, who pays for it, and what might break it open.

Patel's claim is concrete: the two labs will set the price for the rest of the market. The reason is monetization. Compute is the input to AI products, and AI products are the fastest-growing software revenue stream in the market. When the buyer of the input can charge customers enough to recoup and outpace the cost of the buildout, that buyer can outbid every other customer for limited fabrication and memory capacity. Patel pegs Anthropic's third-quarter 2026 profit above $1 billion and frames a $6 trillion valuation in the same SemiAnalysis newsletter, which is how a research-only lab stays in the bidding against Microsoft, Google, and the sovereign AI clouds.

The dollar scale of that bid is what moves this from chips to capital markets. Patel puts aggregate AI capex above $10 trillion by the end of the decade. That is not revenue. It is buildout: fab capacity (the factories that print advanced processors), data centers, and the power plants behind them. The financing already leans on the balance sheets of the hyperscalers (Microsoft, Google, Amazon, Meta) rather than operating cash flow, and the second-order effect Patel names is not technological. It is monetary. If those cloud giants borrow at the scale the capex demands, their debt pushes up long rates, and the rate move lands on sovereign borrowers, mortgages, and any equity that is not part of the AI complex.

The drivers Patel names inside the labs are the same ones chip buyers have watched for two years: economies of scale that compound with parameter count, hard physical scarcity of leading-edge fab capacity, and the longer-term pressure from continual learning and recursive self-improvement, the idea that a model can use its own outputs to retrain a stronger successor. Each one tilts the playing field toward the operator with the most compute. The first two are already priced in by chip buyers. The third is the variable most readers will hear about for the first time when a lab announces a self-improvement result, and it is the part of the thesis that turns a forecast into a deadline.

Four counter-pressures follow from Patel's mechanism. Open-weight model efforts, including Meta's Llama line and the Chinese open-weight ecosystem, give any buyer a usable base model without paying frontier prices, which caps the labs' pricing power on commodity tasks. Distributed training research, where the work of training a model is split across many smaller data centers, attacks the assumption that frontier training requires a single 100,000-GPU cluster. Sovereign compute programs in the EU, UK, India, and Saudi Arabia are explicitly trying to build an outside bid, even if the dollars do not yet match hyperscaler commitments. Capital-market discipline, in the form of debt downgrades or bond-fund redemptions, is the pressure that arrives last and lands hardest, because it is the same channel that transmits the cost to the rest of the economy.

The watch item for the next two years is the gap between announced capex and delivered capacity. Patel has a track record of being early on the supply side of this cycle, including the $6 billion fab estimate for a single leading-edge process node that is now treated as the floor rather than the ceiling. If the buildout runs ahead of model demand, the capex curve bends down and the financing pressure eases. If the buildout runs behind, the marginal-buyer problem gets sharper. The reader does not have to predict which model wins the next benchmark to track the outcome. The story is being priced in chip fabs, in treasury auctions, and in the gap between the two.