GPU 租金半年翻倍，Tomasz Tunguz 解析 2027 年 AI 实验室的久期匹配难题

The essential ingredient to AI has more than doubled in cost in the last 6 months.

GPUs, the chips that power AI, have increased in rental price per hour from $3 to $12 to $13 on short-term contracts, pushing Anthropic’s blended cost per GPU-equivalent hour from $3.10 to $5.65 (within a $3.45 to $8.29 confidence band), according to data from the updated Anthropic S-1/A¹².

Anthropic, like other AI providers, is trying to rent as many accelerators as possible on multi-year contracts that fix prices around $2 to $4 per hour. When that long-term capacity is exhausted, AI vendors must buy access to GPUs in the short-term burst market, where prices have surged¹.

This cost spike changes the unit economics of a frontier lab. A doubling of GPU cost must be met by a doubling of model efficiency, a doubling of price to the customer, or a combination of the two to preserve margins.

With 40% to 50% of data center projects delayed³ & billions of dollars of GPUs sitting idle in warehouses waiting for power⁴, plus transformer & turbine lead times measured in years⁵, this supply pressure shows no sign of easing in the near term.

Meanwhile, the price war across frontier & mid-tier models⁶ makes raising token prices perilous : any lab that hikes rates cedes share.

Labs are responding on three fronts : engineering step-function gains in model efficiency like Opus 5.5’s 40% reduction in cost per task over Opus 5¹, tightening subscription allowances like OpenAI’s move to pool ChatGPT & Codex into a single shared token limit⁷, & pushing enterprise buyers from serverless API consumption into multi-year committed contracts.

Financial engineering in 2027 means closing that duration mismatch : selling a Fortune 500 buyer a multi-year minimum commitment at a locked-in rate to match the duration of the GPU lease underneath it.

In 2026, the race was to find GPUs at any price. In 2027, the winners will be the labs that match the duration of their customers’ contracts to the duration of their compute.

1. Anthropic, PBC Amendment No. 1 to Form S-1 (Form S-1/A), SEC EDGAR CIK 0002130357, Registration No. 333-299633, accession 0001628280-26-068676, filed November 2, 2026 (pp. 138–140, 160, 174–176) : https://www.sec.gov/Archives/edgar/data/2130357/000162828026068676/anthropicpbc-sx1ax1.htm . In the third quarter of 2026, Anthropic spent $5.0b on short-term compute alone : 51% of its $9.86b quarterly compute bill : flipping Q2’s $974m GAAP operating profit into a $567m Q3 operating loss even as quarterly revenue grew $5.40b sequentially to $16.96b. ↩︎ ↩︎ ↩︎

2. Per-hour contract rates & confidence bounds combine Anthropic S-1/A contract commitments (pp. 139, 174–176) with public capacity disclosures : SpaceX Form 424B4 final prospectus (filed June 12, 2026, CIK 0001181412, accession 0001628280-26-042639, disclosing $1.25b/month initial fee & ~325,000 NVIDIA GPUs across Colossus 1 & Colossus II : https://www.sec.gov/Archives/edgar/data/1181412/000162828026042639/spaceexplorationtechnologi.htm ; $3.1b/month amended SpaceXAI run-rate across 325,000 GPUs = $13.07/GPU-hour), Anthropic’s May 6, 2026 Colossus 1 announcement (220,000+ NVIDIA GPUs : https://www.anthropic.com/news/higher-limits-spacex), & the settled October 30, 2026 Ornn B300 Compute Price Index at $12.29/GPU-hour (https://data.ornn.com/). ↩︎

3. Goldman Sachs Commodities Research puts on-time data center delivery over the next one to two years at 50% to 60% (40% to 50% delayed), & JPMorgan (cited by the Wall Street Journal, June 2026) finds that more than 60% of data center capacity planned for 2027 is not yet under construction & a further 7% is delayed ; see Tomasz Tunguz, The Gap Between a Price & a Revenue, October 21, 2026. ↩︎

4. Business Insider, October 20, 2026 : under AWS’s “Early Rack Receipt” program, Amazon took delivery of roughly $5b of Nvidia GB300 racks in August 2026 & ~$3b per month thereafter before facilities had the power to activate them (https://www.businessinsider.com/amazon-nvidia-ai-gpu-data-centers-power-shortages-2026-10). ↩︎

5. Tomasz Tunguz, The AI Bullwhip, August 23, 2026 : generator step-up transformers average three-to-five-year lead times & GE Vernova & Siemens Energy have sold out gas turbine production through 2029. ↩︎

6. Tomasz Tunguz, The Most Important Market in AI Is the Middle, September 23, 2026, & How GPU Prices Can Double While AI Gets Cheaper, September 28, 2026. ↩︎

7. OpenAI Help Center (updated November 2, 2026) & Paul Sawers, OpenAI’s new shared usage limit pits ChatGPT against Codex, The New Stack, November 2, 2026 : starting in December 2026, ChatGPT & Codex on Plus, Pro, & Business plans will draw from a single shared token allowance. ↩︎