The article argues that the critical constraint for frontier AI labs in 2027 is no longer finding GPUs, but matching the duration of customer contracts to the duration of compute leases. Short-term GPU rental prices have surged from roughly $3 per hour to $12–$13 per hour over the last six months, according to the updated Anthropic S-1/A12. That pushed Anthropic‘s blended cost per GPU-equivalent hour from $3.10 to $5.65, with a confidence band of $3.45 to $8.29.

AI providers try to lock in most accelerator capacity through multi-year contracts at fixed prices around $2–$4 per hour. Once that long-term capacity is exhausted, they must buy access on the short-term burst market, where prices have surged. The piece frames this as a unit economics problem: a doubling of GPU cost must be offset by a doubling of model efficiency, a doubling of customer prices, or some combination of both. Raising customer prices is dangerous, however, because a price war across frontier and mid-tier models means any lab that hikes rates cedes market share.

Supply-side pressures are expected to persist. The article cites that 40% to 50% of data center projects are delayed, billions of dollars of GPUs sit idle in warehouses waiting for power, and transformer and turbine lead times are measured in years. These factors keep the near-term supply pressure from easing, even while demand continues to grow.

Labs are responding in three ways: engineering step-function efficiency gains, tightening subscription allowances, and pushing enterprise buyers into long-term committed contracts. As examples, the article points to Opus 5.5‘s 40% reduction in cost per task over Opus 5, and OpenAI‘s move to pool ChatGPT and Codex into a single shared token limit. The enterprise push is described as financial engineering: selling a Fortune 500 buyer a multi-year minimum commitment at a locked-in rate that matches the duration of the GPU lease underneath it.

The central conclusion is that 2026 was about finding GPUs at any price, while 2027 will be won by labs that close the duration mismatch between customer contracts and compute leases.