By end of 2027, compute rationing will be an openly disclosed norm: three more named-customer caps or a top-cloud published capacity-tier allocation policy

The prediction

In 2026, compute rationing stopped being a back-office rumor and became a reported fact. Google told Meta it could not supply as much Gemini capacity as Meta wanted, formalized compute-based usage limits on May 17, and Meta responded by conserving tokens and accelerating its own Muse Spark model. My claim is that this was not an isolated dispute but the opening of a durable regime in which access to AI compute is allocated by capacity and priority rather than cleared by price — and that the rationing will increasingly be disclosed in public rather than hidden inside enterprise contracts.

By December 31, 2027, at least one of the following will be true: (a) three or more additional publicly reported instances — beyond the Google–Meta Gemini cap — will emerge in which a major cloud or frontier-model provider caps, throttles, or rations a named large customer's access to AI compute or a frontier model primarily because of capacity constraints rather than price, policy, or payment dispute; or (b) at least one top-five cloud provider (AWS, Microsoft Azure, Google Cloud, Oracle Cloud, or CoreWeave) will publish explicit capacity-tier allocation terms in which on-demand access is formally subordinated to reserved-capacity commitments — i.e., published terms stating that on-demand availability is not guaranteed and is served only after reserved tiers.

The Google–Meta cap is the baseline and does not count toward the three instances in (a). A qualifying instance in (a) must involve a material AI-sector provider, a named large customer, and reporting that attributes the cap to capacity rather than to a billing or policy dispute. A qualifying policy in (b) must be a published, generally-applicable term of service or capacity program, not a one-off private contract.

Why 78 percent confidence

The structural driver is strong and already in motion. Every link in the compute supply chain — accelerators, advanced packaging, high-bandwidth memory, and power — is gated by multi-year lead times, so the short-run supply curve is nearly vertical and cannot respond to price. In that regime, rationing is the rational clearing mechanism, and the incentives to disclose it are rising: providers want to steer customers toward reserved commitments, and enterprises increasingly demand contractual capacity guarantees they can plan around. Reserved-instance and committed-capacity programs already exist; formalizing them into explicit priority tiers where on-demand is openly subordinated is a small step, and several providers are visibly moving that way. On the reporting side, the Google and Meta story proved that these disputes now leak and get covered, so three more surfacing over eighteen months is a low bar as long as the shortage persists.

Confidence is held at 78, not higher, for two reasons. First, disclosure risk: much rationing happens inside private enterprise contracts and never surfaces in public reporting, so the countable instances in (a) could lag the underlying behavior. Second, a capacity easing — a faster-than-expected ramp of packaging and memory capacity, or a demand slowdown from an AI-capex pullback — could relieve the pressure enough that providers quietly stop rationing rather than formalizing it. The direction is clear; the uncertainty is whether the rationing becomes public and formal, or stays buried in contracts, before the window closes.

What would prove this right

Either branch resolves it. On the reporting side: three or more clearly-attributed instances, on or before December 31, 2027, of a major provider capping or throttling a named customer's compute or frontier-model access for capacity reasons — for example, another hyperscaler limiting a large enterprise's model quota, or a frontier lab rationing a major reseller's API capacity. On the policy side: a top-five cloud provider publishing generally-applicable terms that formally subordinate on-demand access to reserved-capacity tiers and state that on-demand availability is not guaranteed.

What would prove this wrong

The window closes with fewer than three qualifying reported instances in (a) and no qualifying published policy in (b). This would most likely happen if capacity eases enough that rationing recedes, if AI capex slows sharply, or if rationing continues but stays entirely inside private contracts and generally-applicable allocation terms — leaving the Google–Meta cap as an early, visible episode rather than the start of an openly disclosed norm.