Google told Meta no — even Meta can't buy enough AI compute

Information Technology 5 min 60 sources

Google capped how much of its Gemini AI Meta could rent, because Meta wanted more computing power than Google could supply. The richest companies in tech are now rationing each other, and the chipmakers selling the shovels are having the year of their lives.

Key takeaways

Google capped how much of its Gemini AI Meta could rent because demand for data-center compute now outstrips what even the giants can supply — so the biggest tech firms are rationing each other.
The chipmakers selling the hardware are having a historic year: SK Hynix up 310%, Sandisk up 780%, and Samsung and SK Hynix reportedly planning $1.3 trillion in new capacity.
Even as share prices soar, the Bank for International Settlements warns AI "exuberance" could end in a long investment bust — the whole boom rests on the bet that compute demand keeps climbing.

The most telling tech story of the weekend is not a launch. It’s a refusal.

Google has limited how much of its Gemini AI models Meta can use, after Meta tried to buy more computing capacity than Google could supply, the Financial Times reported on Sunday [8][59]. Google told Meta around March that it couldn’t meet the full demand, and the shortfall delayed some of Meta’s internal AI projects [8]. Meta has since told staff to be more sparing with “tokens” — the units that meter how much AI work you do [8].

Sit with the shape of that for a second. Meta is one of the richest companies on earth, building its own giant data centers, and it was still renting AI compute from Google — its direct rival in advertising and AI. And Google said no.

Why even the giants are renting — and getting cut off

Running a large AI model takes enormous amounts of a specific resource: data-center compute, the racks of chips that do the actual calculating. Demand for it is outpacing supply across the whole industry, not just at Meta. Google’s own cloud chief, Sundar Pichai, said earlier this year that compute constraints were holding back growth and had nearly doubled the cloud unit’s backlog of unfilled orders in a single quarter [8]. Google Cloud revenue hit $20 billion for the quarter ending March, and the bottleneck was capacity, not customers [8].

So the companies racing hardest on AI all need more compute than they can build fast enough — and several of them quietly rent from each other to cover the gap. When the supplier is also a competitor, that arrangement holds right up until the supplier needs the capacity for itself. Several other Google clients were affected too, the FT said, just less than Meta, because Meta’s appetite was unusually large [8].

What changes for anyone building on these platforms: capacity is now a real constraint you have to plan around, not an infinite tap. “Be efficient with your tokens” went from a cost-saving tip to an operating order inside one of the biggest tech firms in the world [8].

The shovel-sellers are having a historic year

If compute is the scarce thing, the companies that make the chips are the ones cashing in — and their 2026 has been extraordinary.

Shares in semiconductor and memory-chip makers have surged this year as investors pile into the hardware under the AI boom, according to Guardian analysis of London Stock Exchange data [1]. South Korea’s Kospi index is up 125% — its strongest first half since at least 1990 — driven by Samsung (up 183%) and SK Hynix (up 310%) [1]. In the US, the storage maker Sandisk is up 780% this year and roughly 4,510% over twelve months; Micron is up 296% [1]. Memory chips, long the boring commodity corner of the industry, are suddenly the hottest seat in tech because AI models need huge amounts of fast memory to run [1][4].

The cause is simple supply and demand. “Demand exceeding constrained supply led to a surge in memory chip prices,” said Dan Coatsworth of investment platform AJ Bell, calling it gains in six months you’d normally expect over decades [1]. Samsung and SK Hynix are reportedly planning a combined $1.3 trillion in spending to build more capacity [3] — a number so large that investors actually sold the shares on the news, nervous about how much cash it will eat [3].

The strain even reaches your shopping cart: Apple raised MacBook and iPad prices by about 20%, blaming the rising cost of memory chips [1][44].

The doubt creeping in

For all the surging share prices, the people who watch financial plumbing for a living are getting nervous. The Bank for International Settlements — the bank that central banks use, a kind of referee for the global financial system — warned this week that AI “exuberance” risks ending in a lengthy investment bust, with the boom adding to global fragilities [45][58]. Some of the hyperscalers — the handful of firms running the biggest cloud data centers — have actually seen shares fall as investors rotate out of software and into hardware; Microsoft is down 24% this year [1]. The huge spending plans mean more borrowing and thinner cashflow, turning software firms into capital-heavy ones [1].

So two things are true at once: the chipmakers are booming, and serious institutions are warning the whole structure is overstretched. The thread running through all of it is the same — everyone is betting that demand for AI compute keeps climbing fast enough to justify trillions in spending [3][45].

The geopolitics of who’s allowed to compute

The compute crunch has a political twin. The US has been restricting foreign access to the most advanced AI models, and the fallout is now spilling into Europe: Austria is lobbying the EU to host Anthropic — the AI lab behind the Mythos models — inside the bloc, to get around US curbs that block foreigners from the top-tier systems [6][15]. Austria’s digitalization secretary wrote to the European Commission urging member states to explore “the strategic establishment” of Anthropic within the EU [15]. It’s an early sign that access to frontier AI is becoming something countries negotiate over, the way they once negotiated over oil pipelines.

The under-covered fight: small banks vs. stablecoins

Away from the chip headlines, a quieter battle is heating up over digital money. About 4,000 small US community banks have launched a six-figure ad campaign against the Clarity Act, a landmark bill setting the rules for crypto [57]. Their worry: the law would let crypto companies pay rewards to people who use “stablecoins” — cryptocurrencies pegged to the dollar — which the banks fear could pull deposits out of rural lenders that fund roughly $850 billion in loans to farmers and small firms [57]. “When crypto gets a free pass, communities pay the price,” the ad warns [57]. It’s a reminder that the biggest fights over new financial technology are often won or lost in who gets a seat at the table when the rules are written — and small-town lenders are betting they were left out.

02 · Lesson · why it matters

When you rent from your rival, your ceiling is their permission

The thing you didn't build is the thing that can be taken away — and the AI race is now mostly a race to not depend on anyone.

A no that says everything

The headline is small. Google told Meta it couldn’t have as much AI computing power as it wanted. Two of the largest companies in the world, and one of them got turned away at the counter.

The surprise isn’t that Google said no. It’s that Meta had to ask at all. Meta builds enormous data centers. It spends more on infrastructure in a year than most countries’ budgets. And it was still buying compute from Google — the company it competes with for ads, for attention, for the future of AI itself.

That’s the thread worth pulling. Not “compute is scarce.” The quieter fact underneath it: when you need something you can’t make fast enough, you end up depending on whoever can — even your enemy. And dependence has a ceiling you don’t set.

Owning a thing and being able to use it are different

We tend to think of a powerful company as self-sufficient. It isn’t. Every big AI firm is a stack of things it owns and things it rents. Meta owns the app, the data, the models. It rents — at least at the margin — the compute to run them.

The part you rent is the part someone else can ration. Google didn’t have to break a contract or do anything hostile. It just needed that capacity for itself, and Meta’s projects slowed down. No villain. Just the ordinary arithmetic of a shared resource: when the supplier and the customer both want the same scarce thing, the supplier serves itself first.

So a firm can look unstoppable and still have a hand on its throat that belongs to a competitor. The strength is real. So is the dependency. Both at once.

The race nobody names is the race to not depend

This reframes what the AI giants are actually doing with their trillions. The spending — Samsung and SK Hynix reportedly planning $1.3 trillion in new chip capacity, every cloud firm racing to build — isn’t only about going faster. A lot of it is about not having to ask anyone for permission.

Vertical integration is the unglamorous name for it: own every step, so no step can be cut off. Apple designs its own chips so it doesn’t wait on a supplier. Google builds its own AI hardware so it isn’t at the mercy of one chipmaker. The whole industry is quietly trying to internalize the things it currently rents — because the most dangerous line on the balance sheet isn’t a cost. It’s a dependency on someone who’d be glad to see you stall.

But here’s the catch the spending can’t escape: you can’t vertically integrate the laws of physics. There are only so many chip factories, only so much electricity, only so much memory coming off the lines this year. Even the firms that own the most still hit the same wall. The bottom of the stack — the silicon, the power, the physical plant — is shared whether they like it or not.

You are in this stack too

It’s tempting to read this as a story about giants squabbling over toys. It isn’t only that. The same shared scarcity reaches all the way down to you.

When Apple raised laptop and tablet prices by about a fifth, it didn’t blame greed. It blamed the price of memory chips — the same chips Meta and Google are fighting over. The AI race bid up a component, and the cost landed on someone buying a computer to do their homework. When an app you use feels slower, or a service quietly limits how much you can do, you may be feeling the far end of a capacity crunch decided in a data center you’ll never see.

You don’t own the stack. Almost no one does. We rent our piece of it — our cloud storage, our streaming, the model behind the chatbot — from a few companies who are themselves renting from each other, all sitting on a physical layer of chips and power that nobody can conjure on demand. The trillion-dollar firms are not above this web. They’re just nearer the front of the same queue.

That’s the humbling part. We picture the powerful as the ones who never have to ask. The truth is messier: everyone is renting something they can’t make, from someone who might one day need it back — and the bigger you get, the more the thing you can’t build becomes the thing that decides your limits.

03 · Lab · your turn

Build or Rent the Scarce Thing

Rehearse the trade-off between renting compute cheaply from a rival and building your own, and feel who sets your ceiling when the crunch hits.

04 · Hope · carry this

The same shortage that has giants rationing each other is also the reason a trillion dollars is now going into building more of it — and scarcity, when it's this visible, has a way of becoming the next decade's abundance.