① Yann LeCun stated that Elon Musk’s xAI has become a “failure case,” with the departure of its founding team making it difficult to attract top AI talent and challenging to compete with OpenAI and Anthropic.
② He warned that AI service costs remain high, and user payments are insufficient to cover losses. If prices cannot be raised or costs reduced, the industry may face a bubble burst.
Currently, both Google and Anthropic are renting computing resources from these data centers.
“I am not optimistic about xAI’s prospects,” said Yann LeCun. He stated that he does not believe xAI will be able to compete with industry giants such as OpenAI and Anthropic in the future. “A Massive Bubble Burst” In recent months, corporate spending on AI has faced growing scrutiny, as the cost of the technology is far higher than many had previously expected. According to media reports, OpenAI CEO Sam Altman said earlier this month during a company livestream that enterprises have begun serious discussions about the scale of their AI investments, and described AI costs as a “huge problem.” “AI service prices are rising, and while operating costs are declining, the pace of decline is far from fast enough. As a result, almost all of these companies are losing money, and the usage costs for the vast majority of users are effectively subsidized by investors. This situation cannot sustain itself over the long term,” said LeCun. He further pointed out that AI laboratories such as OpenAI and Anthropic must either raise prices or cut costs, otherwise the industry will eventually face a massive bubble burst. Questioning Large Models, Betting on World Models LeCun has long criticized the limitations of large language models (LLMs). Most current mainstream AI products are built on LLMs, whereas he favors the “World Models” approach. LLMs predict the next word by learning language patterns, making them particularly suitable for reasoning and programming tasks. World models, by contrast, attempt to understand the operational rules of the real world or simulated environments, including objects, causal relationships, and the connections between actions. At present, from Anthropic to OpenAI, multiple AI companies are focusing on developing AI agents—intelligent systems capable of autonomously executing more complex tasks. “Personally, I believe that we cannot have truly general and reliable agentic systems until they are based on world models,” said LeCun. LeCun acknowledged that LLMs are indeed highly valuable in fields such as programming and mathematics, but he emphasized: “The cost of running these systems at current performance levels is extremely high, and the amount users are willing to pay is far from sufficient to cover those costs.”
