← Back to News
ANALYSIS

Google Told Meta No: The Compute-Rationing Regime Goes Public

Around March, Google told Meta it could not supply as much Gemini capacity as Meta wanted, and Meta began rationing tokens internally. On May 17 the limits were formalized. Even a trillion-dollar buyer cannot pay its way past a supplier ceiling, and AI compute is now allocated, not simply sold.

By Michael Eakins•• min read
GoogleMetaGeminiAI ComputeGPU Supply

A detail from the Google–Meta relationship has quietly become one of the most important facts about the state of AI, and it is the kind of fact that only looks small until you sit with it. One of the richest companies on earth asked to buy more of a product, was told it could not have as much as it wanted, and responded by using less — because no amount of money would produce more of the product on the timescale it needed.

What happened

Around March 2026, Google informed Meta that it could not fully meet Meta's requested compute quota for the Gemini models Meta had been buying to power internal workloads. The restriction disrupted the timelines of multiple internal Meta AI projects, and Meta instructed employees to use AI tokens more sparingly and to improve efficiency per call. On May 17, 2026, Google formalized the squeeze, imposing compute-based usage limits on Gemini access broadly. The governing principle of the new regime is that access scales with available capacity, not with willingness to spend.

The context makes the cap more striking, not less. Google is spending on the order of 180 billion dollars on AI infrastructure this year and still could not satisfy a single strategic customer's demand — while simultaneously securing additional GPU capacity from SpaceX to feed its own needs. A company of that scale acting as both a rationer of compute to customers and a scavenger of compute for itself is the clearest signal available that capacity, not capital, is the binding constraint.

Google AI infrastructure spend this year

~$180B

Still not enough to meet all of Meta's requested Gemini capacity — while Google separately sourced GPU capacity from SpaceX for its own workloads

↑ 1%trillion-dollar buyer, Meta, rationed anyway

Why price stopped clearing the market

In an ordinary shortage, price rises until demand matches supply and the market clears. What is happening in AI compute is different: a rationing regime, where the supplier holds price roughly fixed and allocates a fixed quantity by priority, contract, and relationship rather than by the highest bid. The reason is that supply cannot respond to price on any timescale that matters. Hyperscaler forward orders for Nvidia's Blackwell generation consumed most of the available allocation through 2027; TSMC's CoWoS advanced-packaging capacity is booked through at least mid-2027; high-bandwidth memory is concentrated among a few makers running flat out; and data-center power is contracted a decade ahead. When every link in the chain has a multi-year lead time, the short-run supply curve is nearly vertical — you can pay double and get nothing more this year — so rational suppliers ration by hand instead of by price.

The mechanism that changed

Allocation, not price

With advanced packaging booked through mid-2027 and accelerator allocation spoken for years ahead, suppliers ration scarce capacity by priority rather than clearing it by raising the price

↑ 2027%year through which CoWoS packaging is already allocated

Allocation becomes a weapon

Once a supplier rations by discretion, the act of allocation is a strategic instrument — and the supplier is frequently a competitor of the customer it is rationing. Google competes with Meta across advertising, consumer AI, and the race for frontier models. The benign reading of the cap, that Google simply ran short and served itself first, is almost certainly correct, but that is exactly the point: a supplier that is also a rival will rationally serve itself before it serves you. The moment your compute supplier is also your competitor, your growth runs on their sufferance.

Meta's response shows how sharp the incentive to escape has become. Rather than renegotiate for more Gemini capacity, Meta accelerated a shift it was already pursuing — moving critical workloads such as content moderation onto its own internal Muse Spark model to reduce external dependence. The same reflex is visible across the industry: Anthropic is reportedly in talks with Samsung to build a custom AI chip, and the hyperscalers have spent years building their own accelerators precisely so their most important workloads never sit in someone else's allocation queue. The rational answer to being rationed by a rival is to stop depending on that rival — which is why the allocation turn is driving an aggressive wave of vertical integration.

Why it matters

The buildout is not failing — TSMC just posted record revenue on exactly this demand, which confirms the atoms are being made and sold as fast as they can be. But it is not keeping pace with demand, and until it does, AI compute will keep clearing on allocation rather than price. That reorders the industry: giants with secured supply or their own silicon on one side, and a mid-market crowded out of the on-demand queue on the other. For anyone building on AI, the practical lesson is to stop assuming elastic supply — secure reserved capacity, stay portable across providers and models, and treat compute as a scarce input you plan around rather than a commodity you summon.

For the full analysis of what the shift from price-clearing to allocation-clearing means for the structure of the AI industry, see the companion feature, The Allocation Turn: When Compute Is Rationed by Capacity, Not Price.

Sources