Google Cut Off Meta's Gemini Access Because It Ran Out of Compute

Welcome back to SavvyMonk, your one-stop for AI and tech news that actually matters.

Google has spent the year telling the world it owns one of the largest pools of AI compute on the planet, and at the end of June it admitted it does not have enough of it. The company quietly capped how much Meta could use its Gemini models, then turned around and agreed to rent more than a hundred thousand chips from SpaceX to keep its own customers served.

Let's get into it.

Most AI tools ask you to trade your data for intelligence. Norton Neo doesn't. It's the first safe AI-native browser built by Norton, and it gives you powerful built-in AI without handing your privacy over to get it. Search, summarize, and write with AI built directly into your browser. Your data stays yours. Your context stays private.

Built-in VPN, anti-fingerprinting, and ad blocking come standard. No add-ons. No setup. No compromises.

Fast. Safe. Intelligent. That's Neo.

Download Norton Neo

Around March 2026, Meta went to Google wanting to buy a larger slice of Gemini, the model family it had come to depend on, and Google said it could not provide the full computing capacity the social media company was asking for. The decision capped Meta's access and slowed several of its internal AI projects.

Other Google Cloud customers ran into the same ceiling, yet Meta was hit harder than most because of the sheer volume it wanted, and the shortfall pushed Meta to tell staff to spend their AI tokens more carefully, trimming the unit of compute that every model query burns through.

The limits stayed private until the end of June 2026, when the arrangement became public and put an uncomfortable fact on the table. One of the richest companies in technology had been turned away by another over nothing more exotic than a shortage of chips, and the buyer being rationed was no small account but a rival large enough to build its own frontier models.

Why Meta Leaned on a Rival

The detail that makes this awkward is that Meta was paying Google at all. Meta has spent years promoting its open Llama models as a credible alternative to the closed systems built by its competitors, so renting compute from one of those competitors sits oddly with the public story. The reason is practical rather than sentimental, because Gemini proved better than Llama at the unglamorous safety work that keeps a platform running, the automated removal of harmful content and the hunt for scams across services used by billions of people. When the model that does that job well belongs to a rival, dependence follows, and dependence is exactly what the cap exposed.

That exposure has a cost beyond the delayed projects. Meta now has to run critical moderation systems on a supply it does not control, set by a supplier that also competes with it for the same scarce hardware, which is the worst position a company of its size can be in.

The Chips Google Had to Rent

The clearest sign of how tight things have become is what Google did on its own side of the ledger. In a regulatory filing disclosed on 5 June 2026, Google committed to pay SpaceX nine hundred twenty million dollars a month for access to roughly 110,000 Nvidia GPUs along with the processors, memory and components that surround them.

The full rate runs from October 2026 through June 2029 after a cheaper ramp-up period, and Google described the arrangement in plain terms as short-term bridge capacity to meet surging demand for its Gemini Enterprise platform.

Read that next to the Meta decision and the shape of the squeeze comes into focus. A company that could not spare enough Gemini capacity for a paying customer was at the same moment leasing a vast block of GPUs from a rocket builder, with the chips housed in SpaceX data centres and Google retaining ownership of its models and data. The filing even spells out an escape hatch, since Google can walk away if SpaceX fails to deliver the promised hardware by the end of September 2026, a clause that tells you how uncertain the supply of these chips really is.

The Money Behind the Crunch

None of this is happening because Google is short of cash to spend. Its cloud arm pulled in more than twenty billion dollars in the quarter that ended in March, up by roughly two thirds from a year earlier, and chief executive Sundar Pichai has said plainly that capacity limits stopped the unit from growing faster while its backlog of unmet demand swelled toward four hundred sixty billion dollars. The company has guided to well over one hundred eighty billion dollars of infrastructure spending across 2026, and it still cannot serve everyone who wants to buy.

That is the part worth sitting with. Money is not the bottleneck anymore, since the constraint has moved to physical things that money cannot conjure on demand, the chips themselves, the electricity to run them, the cooling, and the long lead times on every data centre that has to be built before any of it switches on.

What Meta Is Building Instead

Faced with a supply it cannot trust, Meta is doing the obvious thing and building its way out. The company has been shifting its safety and moderation workloads onto Muse Spark, a new internal model developed under its Superintelligence Labs division, so that the jobs it once handed to Gemini can run on hardware it owns.

The pivot fits a brutal stretch at the company, which cut 8,000 roles in May 2026 while reassigning around 7,000 workers into AI-focused teams and guiding to between one hundred fifteen and one hundred thirty five billion dollars of its own infrastructure spending for the year.

The Gemini cap did not start this shift so much as accelerate it. Meta was already trying to lower its reliance on outside frontier models, and being told to use fewer tokens by a rival that doubles as its supplier is the kind of nudge that turns a long-term plan into an urgent one.

A Pattern Bigger Than One Deal

The most useful way to read the Meta story is as one data point in a wider shortage. Anthropic signed its own compute agreement with SpaceX in late May, leasing the full capacity of a data centre packed with more than two hundred thousand GPUs at a figure reported above a billion dollars a month. So the same rocket company is now a major landlord to two of the largest names in AI at once, which would have read as a joke a year ago and now reads as a market.

Put the threads together and the picture is consistent across the industry. Demand for AI compute is climbing faster than even the most aggressive spending can supply it, the firms with the deepest pockets are renting hardware from whoever has spare racks, and a company is being told to economise on tokens by the very cloud provider it pays. The shortage is no longer a forecast, it is the thing shaping who gets to build what.

The Bottom Line

The headline is a rival being cut off, but the real story is that compute, not capital or talent, has become the hard limit on AI. When Google both rations Meta and rents 110,000 GPUs from SpaceX in the same stretch, the message is that nobody has enough, including the company sitting on one of the biggest supplies in the world. Watch where the chips flow over the next year, because in this phase of the race, access to hardware will decide more than any model benchmark does.

Category: Strategic Analysis

"Act as a procurement strategist for a company that depends on a third-party AI provider for a critical workload like [workload]. Walk me through the risks of that dependency by asking me about my current provider, my monthly token spend, my switching costs, and how central this workload is to the business. Then give me a ranked plan to reduce single-supplier risk over the next [number] months, with the highest-leverage move first and a rough cost and timeline for each step."

The instinct when a company gets cut off is to read it as a falling-out, but the more honest reading is a supply problem wearing the costume of a rivalry. Google did not refuse Meta out of spite, it refused because the chips were not there, and that distinction matters because it tells you the squeeze will not ease when two firms make peace. It eases only when the hardware catches up, and that is years of building away. Keep asking, when you read about an AI deal, whether you are looking at strategy or at scarcity.

Hit reply, I read every response.

See you in the next one.

— Vivek

P.S. Know a founder or engineer who is betting their roadmap on someone else's AI models? Forward this to the person most likely to get rationed next. They can subscribe at https://savvymonk.beehiiv.com/