IMAGE: KEVIN DIETSCH/GETTY IMAGES
THE VALUE of Microsoft, one of the world’s biggest companies, fell by $440bn on January 29th—a loss larger than the GDP of most countries. The trigger was a disclosure that nearly half of its future revenue depends on one customer: a startup that loses $2bn a month. That fact has moved a niche debate about artificial intelligence into the mainstream. Can the standard-bearer of the AI revolution really go bust?
OpenAI is testing the limits of corporate physics. Its revenues are soaring but its costs are rocketing faster. It took in over $20bn last year, up from $2bn in 2023, a pace of growth without modern precedent. Google took eight years to reach comparable scale; OpenAI did it in three. Yet it lost $13.5bn in the first half of 2025 alone. Full-year losses are projected at $25bn, rising to a cumulative $115bn by 2029.
That extraordinary burn rate comes from what Sam Altman, its boss, calls the “scaling law”: the belief that better intelligence comes directly from more computing power. Hence a plan to spend $1.4trn on infrastructure over eight years. The trouble is that the evidence for this law is weakening. Recent vastly more expensive training runs have yielded only marginal improvements. What if the cost of creating super-smart machines never falls enough to make them profitable?
The firm’s income provides little reassurance. Most of its money comes from consumers with subscriptions to services like ChatGPT Plus. But its real hope was high-margin enterprise contracts. Here it is being beaten. Anthropic, its main rival, now has 40% of the corporate market for foundation models, against OpenAI’s 27%. Meanwhile its application programming interface (API) business, which lets developers plug its models into their own products, is caught in a price war. Competitors including open-source models are offering similar performance for a fraction of the cost.
Faced with that OpenAI has doubled down. Its response is “Project Stargate”, a $500bn plan to build the world’s most powerful AI cloud. The first data centre, in Texas, is already online. The snag is the money. Only $52bn in equity has been secured. The rest is meant to come from debt that has yet to be raised. Trade tariffs have pushed up construction costs. The energy required would suck up 2% of America’s entire electricity generation. And even if the cash is found, the firm’s financial structure looks circular and risky. Microsoft and Nvidia invest billions in OpenAI, which then spends that money on Microsoft’s cloud services and Nvidia’s chips. The loop works splendidly when capital flows. It could be deadly if it dries up.
The firm is also losing its lead. Its share of global AI web traffic fell from 87% to 65% in the year to January, as Google’s Gemini gained ground. Top engineers are eight times more likely to leave for Anthropic than to join from it, forcing OpenAI to dole out lavish stock awards to keep them. Its legal troubles are mounting, too. A lawsuit from Elon Musk, an early backer, could lead to damages of over $100bn. New rules in Europe and America bring more costs and complexity. Public trust is feeble. A single big error could bring a regulatory crackdown that would hit the industry leader hardest.
Is there a way out? OpenAI is betting on “agentic” AI: systems that can perform complex, multi-step jobs. The idea is to shift from selling intelligence to selling labour, with pricing based on results rather than usage. Early tests are promising. Moderna, a drugmaker, has given ChatGPT Enterprise to thousands of staff, who now use it constantly. If OpenAI can keep its technological edge in reasoning, it might justify the steep prices needed to cover its vast costs.
But even that hopeful path has pitfalls. When Figure AI, a robotics firm, decided to build its own AI rather than use OpenAI’s, its valuation soared. This hints that in the world of physical AI, the real money may go to those who control both hardware and software not just the brains.
Comparisons to the dot-com bubble are inevitable but flawed. Unlike the flops of the 1990s, today’s AI leaders have real revenues and are woven into the operations of serious companies, suggesting genuine utility and not just hype. Yet the financial parallels are unsettling. The dizzying rise and sickening fall of the Nasdaq two decades ago is a warning for a sector where stocks have leapt by 600% in five years.
So will OpenAI survive? It is not a normal startup but a systemically important entity. Its collapse would shake the global tech industry. Yet with losses of $25bn a year it is entirely reliant on a constant stream of private capital. If a planned $100bn funding round fails or if it cannot raise debt for Stargate, it could be insolvent by 2027. The most likely end is neither triumph nor collapse but absorption. Crushed by losses and dependent on Microsoft’s infrastructure, it may eventually be taken over by the tech giant or a consortium of big cloud firms. OpenAI may yet be worth a trillion dollars. It probably will not be an independent company when it is. ■
