The myth of infinite scalability is starting to crack. Behind the spectacular advances of ChatGPT, Sora, and other frontier models, OpenAI now faces a tougher, less glamorous challenge: building a viable business around artificial intelligence itself.

In a recent investigation, Le Monde highlighted growing doubts about the company’s economic model — and the uneasy relationship between state support, massive infrastructure costs, and the elusive quest for profitability.

The illusion of infinite growth

OpenAI’s user base is enormous, its cultural footprint undeniable. But none of that guarantees a sustainable profit engine. Like social networks in their early days, OpenAI’s products have become global habits long before they’ve become viable businesses.

The article notes that OpenAI’s leadership has asked Washington to extend chip-industry tax breaks to AI data centers — and even considered forms of state guarantees for future mega-projects. Sam Altman quickly denied any request for “bailouts,” but the signal was clear: compute is expensive, margins are thin, and the company is looking for policy relief.

AI may be the new electricity — but it’s also becoming a very expensive utility to run.

Margins at the mercy of computation

Every question asked to ChatGPT, every generated image or video, consumes an immense amount of energy, silicon, and capital. Unlike software, large-scale AI does not scale linearly: costs grow with usage.

That’s the paradox haunting OpenAI’s model. Its consumer success drives ever-higher compute bills. Even with billions in revenue and Microsoft’s backing, the economics are fragile. As Le Monde put it, the company’s expansion “conceals structural fragilities” — heavy reliance on subsidies, limited pricing power, and fierce competition.

It’s not the code that’s expensive; it’s the electricity behind it.

The politics of “privatised gains, socialised losses”

What truly alarms observers is the emerging pattern: private firms capturing the upside of AI innovation while lobbying the state to shoulder systemic costs.

The Le Monde piece evokes a familiar criticism — the fear that AI giants may reproduce the logic of Big Tech monopolies, this time under the guise of infrastructure. If governments start underwriting GPU parks and data centers, the frontier of intelligence risks being financed by taxpayers rather than markets.

And yet, the counter-argument has weight: if AI becomes a public-good infrastructure — powering schools, administrations, hospitals — shouldn’t public policy support it?

We are witnessing the birth of a new debate: not whether AI should be subsidised, but under what conditions — and with what public returns.

From model provider to platform economy

OpenAI’s path forward depends on one thing: value capture.

If base models become commoditised — as rivals like Anthropic, Mistral, and Google DeepMind flood the market — the differentiation will shift upward to the integration layer: agents, orchestration, domain-specific applications.

That’s where the next economic frontier lies. Those who build usable systems, not just powerful models, will control the margin.

OpenAI seems to know this. Its push toward “ChatGPT Teams,” GPT-based agents, and the coming wave of AI operating systems reveals a clear strategic pivot: moving from infrastructure to ecosystem. But the economics of that transition remain uncertain — and time is short.

The real question: who pays for intelligence?

The deeper issue isn’t OpenAI’s balance sheet; it’s the societal one.
When “intelligence” becomes a resource — consumed like bandwidth or electricity — who pays for it, and who controls it?

The Le Monde investigation serves as a reminder that the cost of thinking machines is not abstract. Every query, every token, is backed by real-world capital, energy, and political trade-offs.

If the future of cognition depends on subsidies, we must decide collectively what kind of intelligence we’re funding: one that serves private profit, or one that amplifies public value.