The Compute Crunch Is Real — Just Not Evenly Distributed Yet

Is the AI compute crunch real, or is it a forecast that hasn't arrived yet?

Every lab will tell you it feels urgent. Epoch AI published the numbers this week and they are stark: inference demand is growing roughly 10x per year while global compute supply is expanding at 3-4x annually. The gap isn't theoretical — it's arithmetic. The real question is who absorbs the squeeze first.

Here is the part that nobody in the open-weight ecosystem wants to talk about publicly: the economics are getting worse on two sides simultaneously.

On one side, the capability gap is widening. Since DeepSeek R1 in January 2025 — when open models briefly closed the gap to the closed frontier — the distance has grown every month. By the Epoch Capabilities Index, open-weight models now trail closed frontier models by an average of four months. On private benchmarks (the ones labs don't publish), the gap is 8-10 months. The average ECI gap of 8 points is roughly equivalent to the distance between GPT-5 and GPT-5.5 — a full generation. The best open-weight models score in the 50-54 range on the AA Intelligence Index; the closed frontier ceiling sits at 57, where Claude Opus 4.7, Gemini 3.1 Pro, and GPT-5.5 all land.

On the other side, inference costs that were supposed to decline are about to stop. Kimi K2.6 runs at roughly $3.50 per million output tokens — a fraction of the $25-30 per million that closed frontier providers charge. But that price advantage depends on supply growing faster than demand. When supply tightens, prices at the wholesale level rise, and open-weight model providers — who have no hyperscaler-level negotiating leverage — pay the increase first.

Meta, Mistral, and DeepSeek have each built their deployment businesses around the premise that inference gets cheaper over time. That premise is now contingent on chip supply expanding fast enough to outrun demand growth. Epoch's numbers suggest it won't.

Who captures the upside

The hyperscalers do. Microsoft, Google, and Amazon collectively spent $156.1 billion on capex in Q1 2026, beating Epoch's own projection of $155.1 billion. That number was $140.6 billion in Q4 2025. It will be $770 billion in 2026 and over $1 trillion in 2027 — roughly quadruple what they were spending when GPT-4 launched. They are building ahead of demand not because they are reckless but because the compute asymmetry between those who own the chips and those who rent them is now a durable competitive moat.

When inference supply tightens, the hyperscalers allocate capacity to their own model deployments first. API customers — especially those without enterprise priority contracts — get queued. The price of a token on the open market rises. The companies that locked in capacity agreements get priority. The cost advantage that drove open-weight adoption shrinks.

The SWE-Bench exception

One wrinkle: on coding tasks specifically, open models are nearly at parity with closed ones. SWE-Bench scores for the best open-weight models now match frontier closed models on the tasks that matter for software engineering workflows. If you are buying AI for code generation and code review, the open-weight case remains strong — the economics of running your own fine-tune on a Llama or Mistral checkpoint still beat the API costs of Claude or GPT-5.5 at scale.

But coding is one domain. For reasoning, agentic workflows, and multimodal tasks — the areas driving the bulk of enterprise AI spending — the closed frontier's lead is real and growing.

What this means for builders and buyers

If you are building on open-weight models because they are cheaper, the compute crunch is your problem. When wholesale inference costs rise, your per-token cost rises with it. If you chose open-weight for cost certainty (no vendor lock-in, fixed API pricing), that certainty has an expiration date tied to chip production curves you cannot control.

If you are a hyperscaler customer using closed frontier APIs, the crunch may mean priority degradation — your queries get slower or more error-prone — before prices rise. The labs have every incentive to protect their highest-revenue customers first.

If you are a lab that released an open-weight model in the past twelve months, you face the deepest squeeze: your model is falling behind the closed frontier on capability while your operating costs are about to rise. DeepSeek and Mistral have both positioned open-weight as a long-term strategic bet. That bet requires inference economics to keep improving. Epoch's numbers suggest they won't — not at the rate they have been.

The honest answer

The supply-demand imbalance Epoch describes is a structural trend playing out over years, not a visible outage you can point to this week. The GB200 and GB300 fleets are still ramping. Token prices at the API layer have not moved dramatically in most providers' public pricing tiers.

But the trend lines are pointing in one direction, and the capex commitments are real. The hyperscalers are spending $1 trillion next year on chips they expect to need. The inference demand curves are not theoretical. The open-weight developers who bet that compute would keep getting cheaper — because it always has — are now making that bet against a $156 billion-per-quarter counterparty that is actively building supply ahead of demand.

You will feel it. The only question is when.