SemiAnalysis breaks down the compute shortage: OpenAI and Anthropic are snapping up all of the world's newly added compute capacity.

SemiAnalysis CEO Dylan Patel said at Runtime 2026 that OpenAI and Anthropic will absorb 25%–40% of all new compute capacity in 2027, with the two labs' combined footprint rising from about 4 GW at end-2024 to roughly 10 GW this year and more than 20 GW next year, against global capacity that has nearly doubled from 60 GW to over 100 GW. The squeeze has lifted H100 rental rates from a low of about $1.45–1.70/hour last October to $2.40–3.20/hour now, and Patel expects B300 rentals could exceed $7/hour, arguing that open-model token inference is profitable enough to keep bidding up GPU prices even at 30% utilization. Financing is the binding constraint: with the 10-year Treasury yield up from 3.96% to 5.31% in early October and roughly $7 trillion of the $11 trillion in AI capex through 2029 expected to be debt-funded, emerging cloud providers struggle to borrow for 1 GW builds and must pre-sell capacity to hyperscalers, leaving 65%–70% of next year's contracted compute with hyperscalers and top labs. Nvidia has backstopped about 6.5 GW of clusters, with a stated ceiling of 45 GW, via structures such as SoftBank Energy's 9 GW Ohio site reserved for OpenAI and Firmus rent guarantees. Watch H100 and B300 spot and contract pricing, prepayment terms and availability at CoreWeave and Nebius, and whether debt markets reopen for neoclouds.