Google Can’t Even Supply Itself
Why It Capped Meta’s AI Access
The company building the AI just told its biggest customer there isn’t enough of it to go around
Around March, Google told Meta it couldn’t deliver the Gemini capacity Meta wanted to buy. Meta is one of the richest companies on Earth.
Sometimes a story sounds small until you notice who’s involved. Google capped Meta’s access to Gemini, its own AI model family, because it couldn’t supply as much computing capacity as Meta wanted to buy.
That’s not a startup running into limits. That’s two of the five most valuable companies on the planet, and one of them just told the other there isn’t enough AI to go around — even when money isn’t the obstacle.
Here’s what actually happened, why Google itself had to rent compute from SpaceX to cope, and what this shortage really says about the state of the AI industry.
Google told Meta in March it couldn’t meet Gemini demand
The restriction has remained in place since, delaying some of Meta’s internal AI projects.
Google itself pays SpaceX $920M a month for GPU access
Google called it “bridge capacity” to meet demand for its own Gemini Enterprise product.
Google Cloud’s backlog nearly doubled to $460 billion
That’s signed customer commitments waiting on supply that doesn’t exist yet.
Meta is shifting workloads to its own model, Muse Spark
After cutting 8,000 jobs in May, Meta reassigned 7,000 workers to AI infrastructure.
Google tells Meta it can’t deliver full Gemini capacity
Meta had been relying heavily on Gemini for several core workloads, including content moderation, customer service automation, advertiser chatbots, and internal coding tools, because it outperformed Meta’s own open-source Llama models at several of those tasks.
Around March, Google informed Meta it could not supply the full amount of Gemini capacity Meta wanted to purchase. The restriction has remained in place through June, according to the Financial Times.
Meta had to slow down and conserve
The cap disrupted and delayed some of Meta’s internal AI projects. Engineers were told to use AI tokens more efficiently, a notable reversal from earlier in the year, when Meta had encouraged aggressive AI tool usage, sometimes called “tokenmaxxing,” as part of performance evaluations.
Meta was reportedly hit harder than other Google clients because of the unusually large scale of its compute demand.
Google itself is renting compute from SpaceX
Google agreed to pay SpaceX roughly $920 million a month for access to about 110,000 Nvidia GPUs housed in xAI’s data centers, explicitly calling it “bridge capacity” to meet demand for its Gemini Enterprise product.
The company doing the rationing is, by its own admission, also short on supply. Google CEO Sundar Pichai acknowledged that Cloud revenue would have been higher if the company could meet existing demand.
We are compute-constrained in the near term.
Our Cloud revenue would have been higher
if we were able to meet the demand.
This looks like strength, not weakness, in demand terms
A bubble typically involves oversupply chasing too little demand. What’s happening here is the reverse: capacity is spoken for before it’s even built, and the two companies hitting the wall are among the best-capitalized firms on the planet.
That doesn’t mean AI spending is risk-free, but it’s a different risk profile than the “empty fiber networks” comparisons sometimes drawn to the dot-com era.
It accelerates the push toward in-house models
Meta has been shifting some Gemini workloads to Muse Spark, its own internal model under the Superintelligence Labs division, as a direct response to the unreliability of relying on a competitor’s infrastructure.
The broader pattern across the industry: companies that can afford to build their own compute and models are doing so, while smaller players remain more exposed to capacity rationing like this.
- This is a capacity problem, not a quality problem — Gemini wasn’t underperforming, it was oversubscribed
- Google rationing Meta doesn’t mean Google is struggling — its own backlog and revenue are both growing fast
- Watch for more in-house model investment — Meta’s Muse Spark push is the direct consequence
- Compute scarcity is an industry-wide pattern — Anthropic and Microsoft have also shifted to pay-as-you-go pricing amid the same pressure
⚠️ What This Doesn’t Mean
This isn’t evidence that the AI industry is overbuilt or in a bubble.
The defining feature of a bubble is oversupply with nowhere to go; what’s happening here is the opposite, with demand from the largest, best-funded companies outpacing what even Google can build.
It does mean that compute, not model quality, is increasingly the binding constraint on how fast AI products can scale.