What actually happened
Late in June 2026, the Financial Times reported — and Reuters and others picked up — that Google told Meta it could not deliver the full volume of Gemini compute capacity Meta had asked to buy. Not a pricing dispute. Not a roadmap disagreement. Google Cloud simply could not supply the capacity at the scale Meta wanted.
According to the reporting, the shortfall started biting around March, delaying several of Meta's internal AI projects and prompting leadership to tell engineering teams to optimize aggressively and budget their token usage. Other Google Cloud customers felt the same squeeze; Meta, with its outsized appetite, was just the clearest casualty.
Sit with the shape of that for a second. A company willing to spend on AI at nearly any price could not buy all the compute it wanted from one of the three largest cloud providers on earth.
Why this one matters more than the last shortage
We've had compute shortages before — the H100 scramble, the high-bandwidth-memory squeeze, regional capacity crunches. Those read as growing pains: demand outrunning a young supply chain that would eventually catch up.
This is different, and the difference is where in the market the constraint showed up. When a shortage bites startups, it's a queue. When it bites a hyperscaler's largest customer — a company with the balance sheet and the strategic urgency to pay whatever it takes — the constraint isn't a queue anymore. It's a ceiling. And it's a ceiling that even hundreds of billions of dollars in capex is not lifting fast enough.
The company spending on AI at nearly any price could not buy all the compute it wanted. That's not a queue. That's a ceiling.
The bottleneck moved
For most of the last decade, the scarce inputs in AI were talent, data, and capital. If you had those, you could rent the rest.
That's the assumption this story breaks. The scarce inputs now are physical: advanced packaging and high-bandwidth memory, gigawatts of power, water and cooling, permitted land near substations, and the multi-year lead times to build all of it. You cannot conjure a data center or a transformer with a bigger checkbook and a good quarter. The binding constraint has moved from things money buys instantly to things money buys slowly.
That is a fundamentally different regime to build in. Capital compresses timelines in software. It barely compresses concrete, copper, and turbine backlogs.
Token efficiency is now a capacity strategy
Notice what Meta reportedly did in response: not “buy more” — they couldn't — but “use less per unit of work.” Optimize. Budget tokens.
For two years, efficiency work — distillation, quantization, caching, prompt compression, smart routing between large and small models — was framed as cost optimization. A finance concern. Nice margin, ship it later.
Reframe it. When capacity is capped, efficiency isn't how you spend less — it's how you ship at all. Every token you don't waste is capacity you didn't have to acquire. The teams who treated inference efficiency as core engineering, not an afterthought, just found out they were building a moat.
The trickle-down
Here's the uncomfortable throughline: if the constraint is real for Meta, it is coming for everyone downstream. When hyperscalers ration capacity and delay their own projects, that pressure doesn't stay at the top. It flows into the products built on shared cloud capacity as tighter API rate limits, higher per-token prices, quiet quality-and-latency trade-offs, waitlists on new features, and slower rollouts.
And the timing is the whole story. The next wave of AI — long-running agents, not single prompts — is the most compute-hungry thing the industry has ever tried to ship, and it's arriving exactly as the supply gets tight. Agentic products can burn dramatically more compute per task than a chatbot ever did. That curve is bending up as the ceiling comes down.
What to actually do about it
This isn't a doom read. Constraints reward the prepared. If you're building or advising on AI products, three moves matter now:
- 1Treat compute as a strategy, not a utility bill. Know your token economics per feature and instrument them. The teams who can answer “what does this feature cost in tokens, and where's the waste” will move while others wait on quota.
- 2Get efficient before you're forced to. Distillation, caching, and routing to the smallest model that clears the bar aren't downgrades — they're how you keep shipping when the meter tightens. Build the muscle now, not during an outage.
- 3Own more of your stack. Pure dependence on centralized cloud APIs for core capability is now concentration risk — in cost, availability, latency, and control. Hybrid and local-first inference, with capable open models running on hardware you don't have to beg for, is shifting from ideology to advantage. Predictability is the product.
The bottom line
The compute crunch stopped being a forecast the moment Google couldn't sell Meta everything it wanted to buy. The bottleneck is physical, the lead times are long, and the pressure trickles down. The winners in the next phase won't be whoever can spend the most on the shared grid — it'll be whoever can do the most with the compute they can actually get, and who owns enough of the stack to not be at the mercy of someone else's backlog.
Signal over noise: build like compute is scarce — because for the people at the very top of the market, it already is.