Close-up of a person’s hand holding an iPhone and using Google AI Mode, an experimental mode utilizing artificial intelligence and large language models to process Google search queries, Lafayette, California, March 24, 2025. (Photo by Smith Collection/Gado/Getty Images)
Gado via Getty Images
Google cannot get enough of its own artificial intelligence. Around March, it told Meta it could no longer supply as much Gemini capacity as Meta wanted, and Meta, one of the richest companies on the planet, had to tell its engineers to start conserving tokens. The company doing the rationing builds the technology itself.
That detail, reported by the Financial Times on June 28, breaks the story the financial press keeps repeating. Every week brings another warning about an AI bubble. A bubble is a specific thing. It is a glut: too much supply chasing too little demand, warehouses of product nobody shows up to buy. What is happening in AI infrastructure is the opposite. Capacity is spoken for before it is built, and the largest buyers on earth are being turned away. You cannot inflate a bubble in something the market is rationing.
When Google Rations Its Own AI
Meta had leaned on Gemini because it outperformed Meta’s own Llama models at the unglamorous, essential work of content moderation: catching scams, pulling down harmful posts, keeping the platform safe at scale. When Google capped the supply, Meta told staff to use AI tokens more efficiently and began accelerating its move back onto its own models.
The size of the two companies is the whole point. Google is spending more than $180 billion on capacity this year. Meta is one of the five most valuable companies in existence. Both have effectively unlimited cash. What neither can buy at any price right now is enough compute. When the constraint binds on the two richest participants in a market, it is a physical constraint, and it sits where we have said the bottleneck would sit for two years: in data centers, semiconductors, and the power to run them.
The bubble label gets pinned on AI because the numbers are enormous and the charts keep climbing. But the defining feature of a bubble is a surplus of supply, and a surplus is the one thing this buildout does not have. The dot-com telecoms laid fiber that sat dark for a decade. The homebuilders of 2006 finished houses no one came to buy. In each case production ran past demand and the excess had nowhere to go.
The AI buildout runs the other direction at every layer. Alphabet’s Google Cloud backlog, the future revenue already signed by paying customers, nearly doubled in a single quarter to more than $460 billion, with the company expecting about half of it to convert to revenue inside 24 months. Those are signed customer commitments: demand waiting on supply that does not yet exist. Meta, for its part, raised its 2026 capital budget to between $125 billion and $145 billion, close to double what it spent the year before. Companies do not commit sums like that to build inventory they expect to leave idle.
The Billion-Dollar Bridge
The sharpest evidence is what Google did about its own shortage. In June it agreed to pay SpaceX $920 million a month for roughly 110,000 Nvidia GPUs housed in xAI’s data centers, capacity it openly called a “bridge” to meet demand for its Gemini Enterprise product that was running higher than it could serve. A company spending more than $180 billion of its own this year is still renting nearly a billion dollars a month of someone else’s compute to cover the gap.
Google is not alone in writing that check. Anthropic signed its own SpaceX arrangement in May, at $1.25 billion a month. Nobody rents emergency capacity at those prices when supply is plentiful. Scarcity is the only thing that justifies the check. And when buyers bid up temporary capacity like this, the pricing power sits with whoever owns the capacity. The companies selling data center space, chips, and power are the ones setting the terms.
None of this makes the celebrated bears foolish. It means they are measuring the wrong quantity. A skeptic who sees record capex and calls it excess is pattern-matching to 2000, when the spending genuinely did outrun the demand. This time the demand is the part that keeps showing up: in the backlog, in the rationing, in the rent checks. The loudest of those skeptics still get treated as oracles, even as the downside bets behind them quietly decay toward nothing.
For a long-term investor, that reframing flips the emotion. Weakness in the shares of companies that own the scarce layer is a discount on capacity the entire economy is bidding for, not the first crack in a bubble. The names that benefit most are the ones that own the constraint rather than rent it. Alphabet sells the compute it cannot make fast enough and runs its own TPUs to widen that supply from the inside. Amazon owns hyperscale data centers and is bringing its own Trainium AI chips on stream to attack the same bottleneck. The logic runs down the rest of the stack too, to the companies that supply the chips, the networking, and the electricity, the layers that get paid first when capacity is the thing in short supply.
The bubble framing asks what happens when the spending stops. The more useful question is what happens while demand keeps arriving faster than the industry can pour concrete and wire substations to meet it. A market that rations its best product, signs its revenue years in advance, and rents emergency supply at a billion dollars a month is not running out of buyers. It is running out of room. What the headlines keep calling a bubble is the price of a shortage.