SemiAnalysis: OpenAI and Anthropic Are Claiming Nearly 30% of Global New Compute Capacity
nashnova research
OpenAI and Anthropic together are booking 25%–40% of the world's new compute capacity, adding roughly 12 GW by 2027 — squeezing everyone else out of the chip market.
Two companies taking 30% of new compute — can anyone else still buy?
OpenAI and Anthropic's combined compute grew from about 4 GW at end of last year to 10 GW this year, with 20+ GW projected next year.
Global new compute in 2027 is roughly 40 GW. The two companies account for about 12 GW of that — approximately 30%.
This means → everyone else — other tech giants, cloud providers, startups — is fighting over the remaining 70%. SemiAnalysis founder Dylan Patel put it bluntly: "It is now impossible for a lot of people to buy compute."
What is all that compute being used for — training or inference?
Primarily training new models, not serving users. Anthropic currently has about 4 GW of compute; the entire Claude Code product uses only about 1 GW for inference.
In plain terms = the compute consumed by users is dwarfed by what the labs spend on teaching models — "training" eats far more chips than "answering questions."
This reflects a structural dynamic: even if user growth plateaus, training budgets will keep locking up chips. The bottleneck is supply being absorbed by labs, not a demand-side explosion.
How much have rental prices risen, and how profitable is selling inference?
H100 one-year contract prices climbed from a low of about $1.70/hr last October to $2.40–$3.20 now, with a midpoint around $2.80. Patel expects H100 could top $3 and B300 could exceed $7.
The economics of inference drive that willingness to pay: building a B300 cluster costs roughly $9M/MW/year; renting from a cloud provider costs about $14M/MW/year. But running DeepSeek V4.1 Flash on that megawatt at 60% utilization generates around $70M/year in revenue.
In plain terms = $14M in rent yields $70M in inference revenue. Even at 30% utilization, the margins are extraordinary. That is why rental prices keep climbing — buyers pay up because they earn far more.
Patel disclosed that SemiAnalysis itself operates a compute brokerage and has access to actual transaction prices, which constitutes a potential conflict of interest.
Why not just build more data centers to ease the shortage?
Expansion requires capital, and capital costs are rising. The 10-year U.S. Treasury yield climbed from 3.96% to 5.31% by early October; AI-related debt issuance cooled visibly in September.
This means → new entrants face a near-impossible scaling path. Patel described a scenario: a 100 MW emerging cloud provider trying to reach 1 GW would have to bundle its capacity and sell it to a hyperscaler, which then borrows in the debt markets. "None of that compute reaches the open market."
Nvidia has backstopped roughly 6.5 GW, including a land-power-building guarantee for SoftBank Energy's 9 GW Ohio facility — set for ~2030 and fully reserved for OpenAI — a floor-price arrangement through Firmus, and residual-value guarantees. But even if Nvidia's backstop ceiling reaches 45 GW, "that's still a small fraction of new compute through 2030."
What does this mean for startups?
About 65%–70% of next year's contracted compute is flowing to hyperscalers and AI labs.
Startups leasing clusters must pay half up front and wait three to nine months for delivery, on top of the near-impossibility of securing a one-year contract.
Patel's closing line: "I hope they can get compute, otherwise we're all slaves to Anthropic and OpenAI."
Are cloud providers even reliable — if you get the compute, can you trust it?
SemiAnalysis's ClusterMAX benchmark tracks 323 cloud providers, evaluated 77, and gave positive ratings to just 19. CoreWeave and Nebius scored highest, but both are sold out for the next six months.
Security gaps were the most alarming finding: SemiAnalysis attacked these clouds with AI models and found vulnerabilities already public on the internet — models could locate and exploit them to access other customers' storage.
In one test, an open-source model gained full Slurm scheduler access to an entire cluster, revealing job names belonging to a nation's defense industry and national-security intelligence agencies. SemiAnalysis stopped the test immediately and published a security report.
This reflects a structural contradiction between compute concentration and supply quality — the tighter capacity gets, the less choice users have, while many providers' security and operations lag far behind their expansion pace.
市场有风险，内容仅供研究参考，不构成投资建议。