The Scarcest Thing in Tech Is Now a Computer

Silicon Valley has spent three years insisting that the only ceiling on artificial intelligence is human ambition. The real ceiling, it turns out, is a warehouse full of chips. Google has quietly told Meta that it cannot rent it all the computing power it wanted to run Google's Gemini models, capping one of its largest customers because there are simply not enough machines to meet the demand, according to the Financial Times.

Around March, Google informed Meta that it could not supply the full Gemini capacity the social-media group had hoped to buy. Several of Google's cloud customers were rationed, but Meta, whose appetite for the models was exceptional, was hit hardest, and some of its internal AI projects slipped as a result. Meta has since urged staff to be more sparing with tokens, the units in which AI consumption is measured and billed.

The supplier that ran dry

The episode is striking because Google is at once Meta's rival and its supplier. Meta builds its own open-weight Llama models, yet like most of the industry it also leans on competitors' systems for specific tasks, and even a company spending tens of billions of dollars a year cannot always get served. The bottleneck is no longer talent or data or even money. It is the physical capacity to run the models: the chips, the power, the land and the cooling that turn electricity into answers.

That scarcity has quietly become the organising fact of the industry. The firms that own data centres now hold leverage over those that merely rent them, and the line between customer and competitor blurs every time capacity is allocated.

A $190 billion answer

The strain is visible in Google's own accounts. Google Cloud passed $20bn in quarterly revenue for the first time at the start of 2026, up roughly 63 per cent on a year earlier, while its order backlog, committed business it cannot yet fulfil, roughly doubled to around $460bn. Chief executive Sundar Pichai was blunt about the cause on the company's earnings call.

"Obviously, we are compute constrained in the near-term. And as an example, our cloud revenue would have been higher if we were able to meet that demand."

Google's response is to spend its way out. It more than doubled quarterly capital expenditure to $35.7bn and guided full-year 2026 investment to between $180bn and $190bn, against $91.4bn in 2025. It has also begun renting capacity from rivals: in June it agreed to pay roughly $920mn a month, about $30bn over the life of the deal, for some 110,000 processors at data centres tied to Elon Musk's SpaceX and xAI, running from October 2026 to mid-2029. A month earlier, Anthropic had struck its own arrangement for about $1.25bn a month of xAI compute.

• Google Cloud quarterly revenue: more than $20bn, up about 63 per cent year on year.
• Order backlog: roughly doubled to around $460bn.
• 2026 capital spending: guided to $180bn to $190bn, up from $91.4bn in 2025.
• Google to SpaceX and xAI lease: about $920mn a month, roughly 110,000 processors, around $30bn in total.

Europe's thin slice

For European readers the lesson is uncomfortable. The same crunch that lets an American giant ration an American rival also defines Europe's dependence: the continent owns only a sliver of the world's AI compute, most of it rented from the very US hyperscalers now turning customers away. Brussels has noticed. In January the EU cleared the way for AI gigafactories, sites of roughly 100,000 advanced processors each, backed by an InvestAI fund of €20bn meant to seed up to five of them.

Luxembourg already has a foothold. MeluXina, the EuroHPC supercomputer at Bissen, was among the first AI Factories designated in late 2024, built to give startups, researchers and public bodies sovereign computing they do not have to beg an American cloud for. It is a fraction of the scale Google commands, but the Gemini cap is a reminder of why such capacity, owned rather than rented, is now treated as strategic infrastructure rather than a technical nicety.

The age of cheap, limitless cloud is over. What replaces it is a harder economy in which compute is allocated, queued and fought over, and in which whoever controls the machines controls the pace of the AI race.