Gigawatts, not GPUs: the compute deals of summer 2026 in one table
The summer of 2026 was not about GPUs. Every headline deal was denominated in gigawatts and decades, and the largest of them, reported by The Information on September 6, puts Anthropic at a minimum of 14.8 GW of contracted compute and up to $517 billion over ten years. That is roughly three times the ~$180 billion through 2029 the company was reportedly budgeting in December 2025. 14.8 GW is the output of ten-plus APR1400-class reactors. Here is the summer in one table, then what it does to your token price.
The table
| Deal | Reported | Size | Timing |
|---|---|---|---|
| OpenAI Project Camellia, Georgia | Jul 23, Axios | $30B, 3.2 GW | power phased 2028-2032 |
| Anthropic x Nscale, West Virginia | Aug 27, Bloomberg | $45B, 460 MW, Vera Rubin | from late 2027 |
| Anthropic, all contracts since Oct 2025 | Sep 6, The Information | min 14.8 GW, up to $517B over a decade | Amazon + Google cover 11 GW / $300B+ |
| Anthropic TPU with Google and Broadcom | resurfaced Sep 8 | ~$80B, 3.5 GW | capacity from 2027 |
| Figure x Nscale, Barstow TX | Sep 5 | $3.5B, may exceed $6B, up to 100k GPUs | H2 2027 |
| Crusoe | Sep 3 | $3B Series F at ~$30B valuation; reported $13B / 5-year GPU contract with Jane Street | multi-year |
| Qualcomm x AWS | Sep 8 | inference silicon, up to $60B, 1.6 Tbit/s optics | multi-year |
| Google, Finland | Sep 9-10 | nuclear power locked for AI datacenters | not disclosed |
Nvidia's quarter on August 27 is the demand-side confirmation: $96.2 billion revenue, $89 billion of it data center, up 117% year over year, with guidance of +70% for next year.
Power is the binding constraint
Look at the timing column. Camellia's power arrives in phases between 2028 and 2032. Nscale's West Virginia capacity starts late 2027. Figure's GPUs in Barstow land in the second half of 2027. Nobody is short of a purchase order; everybody is short of a substation. Axios put it plainly when Camellia was announced: power, not silicon, is the binding constraint on AI infrastructure.
The chip count tells you how fast demand is moving against that constraint. Epoch AI's estimate, reported by the New York Times on July 29, is about 20 million AI chips in service today, doubling roughly every nine months, on track for about 200 million by the end of 2028. More than three doublings of demand against a power supply that grows on a 2028-2032 timeline is the shape of a price squeeze, a capacity squeeze, or both.
When the scarce input is a gigawatt with a 2028 delivery date, the price of a token in 2027 is set by who signed a power contract in 2025.
What it means for inference price and availability
- Frontier tiers will not get cheaper on their own. Frontier list prices have converged at $10 per million input and $50 per million output across vendors, and the capital plans above are why. Cache pricing and volume discounts are where the movement is; /prices tracks what is live.
- Inference-specific silicon is the hedge the hyperscalers are buying. Qualcomm x AWS is inference only, not training, and the pitch is performance per watt. If your cost model assumes Nvidia-only supply, revisit it, and read Hopper vs Blackwell for how the generations differ per watt.
- Route aggressively. If power caps frontier supply, the teams that route routine work to cheap or local models keep their frontier quota for the work that needs it. The pattern is in the 99 percent cost architecture.
The honest limitation
Almost every figure in the table is a reported commitment, not a spend. 'Up to $517 billion over a decade' is a ceiling in contracts The Information saw, not a line in a filing, and Anthropic's ~$80 billion TPU deal was announced earlier in 2026 and merely resurfaced in the September 8 roundups. Figure's $3.5 billion 'may exceed $6 billion'. Treat the table as a map of intent with a two-year lag, and verify the numbers against primary filings before they go in a board deck.