The founder of rival AMI Labs says OpenAI, Anthropic, and similar frontier AI labs are selling access below cost while investors foot the difference. OpenAI CEO Sam Altman has separately called AI operating costs a 'huge issue' for businesses.
Frontier AI labs are selling access to their most powerful systems for less than it costs to run them, and outside investors are covering the difference. That structural gap is the substance underneath Yann LeCun's "big bubble explosion" warning to CNBC about OpenAI, Anthropic, and the rest of the field, and it is a problem that OpenAI CEO Sam Altman has separately acknowledged in less dramatic terms (The Decoder).
LeCun, who left Meta's AI research lab to start his own company, told CNBC that AI service prices keep climbing while operating costs are not dropping fast enough. "All of these companies are losing money," he said, and investors are effectively subsidizing usage. The market is, in his framing, one cost shock or one price correction away from a "bubble explosion" (The Decoder).
Two things make LeCun's warning worth reading carefully rather than dismissing as a rival's argument. The first is the stake he has in the answer. His new company, AMI Labs, has raised roughly $1 billion, according to The Decoder's reporting, and is built around a different technical bet: "world models," or systems that try to build an internal model of how the physical world works, in contrast to the large language models that dominate at OpenAI and Anthropic. If the LLM-led labs are right about the path to more capable AI, AMI's bet is wrong. LeCun has reason to want the field to reckon with that possibility, and a reader should hold that motive in view when weighing his numbers.
The second is that the cost critique is not his alone. Altman, who runs the largest of the labs LeCun named, has publicly described AI operating costs for businesses as a "huge issue." When a critic and his target agree on the diagnosis, even if they disagree on the prognosis, the diagnosis deserves a closer look than either party has a reason to grant on its own.
What would have to be true for the math to work? Three things, at least: inference costs per query have to keep falling faster than the labs raise prices, the labs have to find new revenue lines that compensate for money-losing flagship products, or the gap has to stay small enough that the people writing the checks remain willing to absorb it. The bubble framing is LeCun's, not the industry's, but the underlying question, which is whether selling access below cost can persist long enough for costs to come down, is a real one. The honest answer right now is that no one outside the labs' finance teams can verify the gap with precision, and that is itself part of the problem.
A sidebar from the same interview that is worth keeping in proportion: LeCun called Elon Musk's xAI "a kind of failure," arguing that the founding team has left and Musk can barely recruit top talent, and that he does not expect xAI to compete with OpenAI or Anthropic. LeCun and Musk have clashed publicly for years, largely over Musk's political views. The shot is a personal one, and the economics story does not need it to land. It is, however, a useful reminder that the unit-economics argument is being made by someone with his own reasons to want the LLM camp to falter.
What to watch from here: the prices labs actually quote enterprise customers relative to the inference cost of serving them, whether consumer or agent products can break even at scale, and whether the next round of funding tightens terms in a way that forces prices up rather than extending the runway. None of those signals are predictions. They are the things that would tell a reader whether LeCun's framework is being borne out, or whether the gap is closing faster than the critics expect.