Resolves YES if there is strong evidence that, at some point during 2027, demand for frontier-scale AI inference materially exceeds available compute capacity at prevailing prices.

Strong evidence means either (A) direct quantitative evidence of a shortfall or (B) multiple observable signs of compute scarcity.

A. Direct quantitative evidence

This criterion is met if Epoch AI or another credible analysis estimates that frontier inference demand is equal to or greater than available inference capacity, using reasonably comparable assumptions about model size, context length, and token generation speed.

An Epoch-style estimate showing a demand/capacity ratio ≥1.0 for a frontier model under roughly 25k:1k or 128k:1k input token workloads, while maintaining about 35 output tokens/second per user, is sufficient.

For context, Epoch estimated Q4 2025 global Blackwell capacity at roughly 20B output tok/s at 8k context, 6B at 25k, and 0.5B at 128k. It estimated inference capacity was growing around 3.4×/year, versus roughly 10×/year for several demand proxies.

A continuation of roughly ≥8× annual demand growth versus ≤4× capacity growth is strong supporting evidence, but does not alone resolve YES without evidence that capacity is being approached or exceeded.

B. Observable evidence of scarcity

Absent a direct supply/demand estimate, resolves YES if at least two of the following occur during 2027 and are credibly attributed primarily to inference compute shortages:

1. At least two major frontier AI providers impose or materially tighten quotas, throttling, waitlists, or peak-hour restrictions by ≥25%, lasting at least 14 cumulative days.

2. A major provider raises the effective inference price of an existing, substantially unchanged frontier model by ≥25%, or introduces a peak-time premium of that size, for at least 30 days.

3. A major provider introduces an ≥20% off-peak discount or equivalent incentive specifically to shift workloads away from periods of constrained capacity.

4. Credible measurements show frontier-model inference repeatedly falling below roughly 35 output tokens/second because of insufficient capacity, affecting multiple major providers or a substantial share of the market.

Short-lived outages, ordinary rate limits, isolated shortages, deliberate product segmentation, or price increases driven mainly by model improvements do not count.

The market concerns inference compute, especially frontier and long-context or agentic workloads, not training-compute shortages alone.

Resolves NO if these conditions are not met based on evidence from January 1 through December 31, 2027.


https://epoch.ai/gradient-updates/is-a-compute-crunch-coming

Is a compute crunch coming?

We estimated trends in global inference capacity and found that token demand appears to be growing much faster than supply.