Funding & Finance

The Biggest Gamble in Business History

Nvidia has committed up to $100 billion to OpenAI. OpenAI is spending much of that money on Nvidia chips.

By Conor Healy

June 30, 2026

Nvidia has committed up to $100 billion to OpenAI. OpenAI is spending much of that money on Nvidia chips. The cash leaves one company, travels a short loop, and arrives back where it started, a little more valuable each lap. Once you notice this pattern you start seeing it everywhere, and it is the strangest feature of the largest spending spree in the history of business: the AI capex boom.

AI capex is the money Big Tech pours into the physical guts of artificial intelligence, the chips, the data centres, the power lines and cooling systems that make a chatbot answer. The five companies doing most of the building, Amazon, Alphabet, Meta, Microsoft, and Oracle, are on track to spend roughly $725 billion on capital expenditure in 2026, around three-quarters of it aimed at AI. The same five spent about $162 billion in 2022. In four years the number has more than quadrupled, and it is still climbing.

To grasp how unusual that is, look at what it does to the companies doing it. Meta is now spending more than half of its revenue on capital projects. Oracle is above 57%. Microsoft and Alphabet sit in the mid-to-high forties. Capital intensity at that level, for firms this size, has no real precedent.

Amazon’s free cash flow is expected to turn negative this year, an extraordinary thing for the most relentless cash machine in retail. These are not reckless startups but the most profitable companies on earth, and they are wagering their balance sheets on a future that has not arrived yet.

The $600 Billion Hole

Here is the problem that keeps analysts awake. The spending is enormous and certain. The revenue meant to justify it is neither.

David Cahn, a partner at the venture firm Sequoia, ran the arithmetic and found a roughly $600 billion gap. That is how much annual AI revenue the industry needs to conjure, every year, to justify the infrastructure it is building, and the gap has been widening rather than closing.

JPMorgan reached a similar place from another direction, estimating that the buildout needs about $650 billion in annual revenue just to clear a modest 10% return.

The revenue that exists is real and growing fast, which is what makes this argument so slippery. Amazon’s cloud arm runs at around $150 billion a year and grew 28%. Google Cloud is near $80 billion and growing 63%. Microsoft’s Azure AI business is smaller but expanding at triple digits.

Among the model builders, OpenAI ended last year near a $20 billion annual run rate, and Anthropic reached roughly $9 billion, up from about $1 billion twelve months earlier. None of that is fake, and none of it is slowing.

It is also nowhere near $600 billion. The bulls and bears are not really arguing about whether AI makes money, because it plainly does. They are arguing about a subtraction problem, and the number left over after you subtract today’s revenue from tomorrow’s required revenue is the largest in the history of corporate finance.

The Money Moves in Circles

The detail that should give everyone pause is not the size of the spending but its shape.

A lot of the money is moving between the same small group of companies. Nvidia invests in OpenAI, which buys Nvidia chips. OpenAI takes a stake in AMD while Nvidia takes a stake in OpenAI. Oracle signs vast cloud commitments with OpenAI, then buys the chips to fulfil them.

Each deal is defensible on its own, and the people doing them describe it as a virtuous circle that locks in scarce supply. Viewed from a height, though, it looks like a closed loop in which demand is partly manufactured by the suppliers themselves.

The clearest pressure point is CoreWeave, a company that rents out GPU computing. It has raised around $28 billion in equity and debt in twelve months, derives about two-thirds of its revenue from Microsoft, and counts OpenAI and Meta among its largest customers.

In other words, it borrowed heavily to buy chips so it could rent them back to the very ecosystem whose spending underwrites its loans. As long as the giants keep spending, the loop holds. The question is what happens to the borrower if they pause.

The Ghost of the Fibre Boom

We have watched a version of this film before, and the ending was not kind.

In the late 1990s, telecom companies laid fibre-optic cable across the world on the promise of limitless internet demand. Equipment makers fuelled the building by lending their own customers the money to buy equipment, a practice called vendor financing that looks a great deal like today’s chip-for-equity deals.

The cable was real and the demand did eventually come, years later. But the timing was wrong, the debt was heavy, and when growth came in below the forecasts the structure folded. Between 2000 and 2002, twenty-three telecom companies filed for bankruptcy. The fibre they buried is still in the ground, much of it only lit up a decade after the firms that laid it had died.

The lesson was not that the technology was fake but that being right about the technology and right about the timing are two different bets, and the second is the one that bankrupts you.

Nobody Is Sure It Works Yet

The uncomfortable backdrop to all this spending is that the productivity it promises is hard to find in the data.

A study from MIT’s NANDA initiative found that 95% of enterprise generative-AI pilots produced no measurable impact on profit and loss, despite tens of billions in corporate spending. The chief economist at Goldman Sachs concluded that AI added “basically zero” to the US economy in 2025. A National Bureau of Economic Research study found that around 90% of firms reported no productivity effect from AI in their workplaces.

Adoption is wide, with most large companies using AI somewhere, but it is shallow. Fewer than 40% have pushed it past the pilot stage into anything that moves the numbers.

This is the part that makes the gap feel dangerous rather than merely large. The spending assumes a wave of productivity and revenue that the measurements have not yet caught. The capital is being committed now, on the strength of a payoff that remains, for the moment, mostly a forecast.

The Case That This Is Fine

A fair piece has to put the other side, and the other side is stronger than the doom headlines suggest.

Goldman Sachs points out that AI capex sits at roughly 0.8% of GDP, against 1.5% or more at the peak of past technology booms, which implies room to spend rather than a ceiling already breached.

The hyperscalers also insist they are supply-constrained, not demand-constrained. Their problem is that they cannot build fast enough to serve the customers they already have, which is the opposite of the empty-fibre problem. And unlike the dot-com darlings burning venture money on a vision, these are profitable giants funding the build largely from operating cash, with real customers expanding real contracts quarter after quarter.

There is also a simple historical point. Every transformative technology, from railways to electricity to the internet, was overbuilt in its infancy, and most of them were worth it in the end despite the wreckage along the way. Overbuilding is what societies do when they sense something important and cannot yet price it precisely. The question is never whether some money gets wasted but whether the thing being built matters enough to justify the waste.

What Amodei Is Betting On

The most interesting voice in this argument belongs to someone with every reason to be a cheerleader and a habit of sounding like a warning label.

Dario Amodei, chief executive of Anthropic, has said total industry compute spending could approach $1 trillion by late 2027, and that frontier companies may eventually need $800 billion to $1 trillion in annual revenue to justify their commitments.

He frames the dilemma as a “cone of uncertainty.” A data centre takes one to two years to build, so a company has to decide today how much capacity it will need in 2027. Buy too little and you lose customers to rivals who can serve them. Buy too much and you are staring at bankruptcy if the demand is late.

His sharpest line captures the whole gamble: even a one-year slowdown in tenfold growth, he has warned, could make trillion-dollar infrastructure bets unsustainable.

That is the bet, stripped to its bones. The companies spending $725 billion this year are not betting that AI is useful, because they can already see that it is. They are betting that it becomes essential, fast enough, to a world that is still mostly running pilot programmes.

If they are right, the buildout will look in hindsight like the railways, an overbuild that built the future. If the timing is wrong by even a year or two, it will look like the fibre, right about everything except when. Nobody, including the people spending the money, knows yet which film they are in.