After two years of financial euphoria, artificial intelligence enters a more demanding phase in 2026. The question is no longer whether AI is strategic, but whether the observed growth is based on independent final demand or whether it relies on loops internal to its ecosystem, in other words circular deals that fuel its start-up economywhich should not be confused with a market, at the risk of forming a speculative bubble.
Thus, the infrastructure needs of AI (computing, GPU, data centers, energy) require massive fundraising, very far from any profitability. A significant part of this capital is committed to players who are simultaneously suppliers, strategic partners and sometimes investors. If value circulates, it circulates above all in the same hands, the time to build the necessary foundations for innovative products still in search of markets.
The OpenAI / Microsoft case illustrates this foundation phase. Microsoft has committed more than $13 billion to OpenAI since 2019, while OpenAI has become one of Azure’s largest customers. The Redmond firm increased its stake in the company to 27% in return for a commitment to purchase services of more than $250 billion from OpenAI. The loop is simple: invested capital turns into cloud invoices, these invoices fuel Azure’s growth, and this growth justifies the continuation of the strategic partnership. The model holds as long as external revenues (Copilot, Azure OpenAI Services, enterprise contracts) cover the cost of computing. If they slow down, the circular part comes to light.
OpenAI also sealed a five-year, $300 billion deal with Oracle to build data center infrastructure and secure the energy supply essential to its models.
This logic is found, on another scale, around NVIDIA. In its 2025 fiscal year, the group generated more than $130.5 billion in revenue, largely driven by data centers and AI. A substantial part of this demand comes from hyperscalers and still loss-making, venture capital-funded startups. In many cases, 30 to 50% of the funds raised by these companies are dedicated to computing and GPUs, mainly Nvidia. Venture capital thus indirectly finances the revenues of the key supplier, the valuation of which in turn strengthens trust in the entire ecosystem.
Thus, some startups show reduced expenses for several quarters, even though the real cost of computing has not disappeared. It has simply been deferred, sometimes against warrants, exclusivity clauses or future commercial commitments, while their valuation is immediate. In specialized fund portfolios, circularity also becomes organizational, with infrastructure, model and application startups selling services to each other, generating a few million euros of cross-turnover, sufficient to support high multiples on still narrow bases.
The last indicator to consider is industrial. Major platforms have committed hundreds of billions of dollars to AI data centers over the period 2024-2026. On certain recent capacities, initial utilization rates remain below optimal, encouraging volumes to be secured via contracts concluded within the same financial ecosystem. Circularity then acts as a short-term stabilization tool, but increases vulnerability if final demand is slow to materialize.
Taken in isolation, none of these mechanisms is abnormal. Together, however, they outline a system where capital temporarily plays the role of client, while waiting for the market to take over. The tipping point is not technological, but financial. In 2026, the hierarchy of AI will probably be based on more prosaic criteria: income actually collected, concentration of customers, dependence on a single supplier-investor, ability to last twelve to eighteen months without refinancing. This is when the industry will move from a closed circuit to a market economy.
