Yann LeCun has little patience left for Elon Musk’s xAI. In a sharp CNBC interview, the veteran AI researcher called the startup a failure, argued that it cannot realistically challenge the sector’s leaders, and warned that the broader AI business is drifting toward an AI bubble that will eventually have to deflate.
That is a pretty awkward verdict from someone often nicknamed the ”godfather of AI.” LeCun, now the founder of AMI Labs and formerly Meta’s chief AI scientist, said xAI has already been damaged by the departure of its founders and may struggle to recruit top talent after Musk’s handling of the previous team.
xAI’s expensive Memphis infrastructure
LeCun’s criticism is not just about reputation; it is about the bill. He pointed to a $2.5 billion loss in xAI’s segment at SpaceX in the first quarter and said the company now has a huge infrastructure footprint it must rent out to make the economics work. That footprint includes the Colossus 1 and Colossus 2 data centers in Memphis, whose capacity is reportedly being leased by Anthropic and Google.
That kind of pivot is telling. In AI, the loudest builders keep talking about scale, but the market is slowly forcing a more boring question: who pays for all the chips, power, and cooling once the hype wears thin? xAI is hardly alone here, but it is one of the more visible examples of how quickly infrastructure bragging rights can turn into a rental business.
The AI bubble problem
LeCun’s broader warning lands harder than the xAI shot. He argued that AI services are getting more expensive for customers even as the cost of running them falls, but not fast enough, so investors are still subsidizing usage. That arrangement cannot last forever, he said, and the obvious outcome is pressure on the big names to raise prices and cut costs.
OpenAI and Anthropic are the clearest examples of companies caught in that squeeze. They may have the strongest brands and the most attention, but the economics of serving AI at scale still look fragile, especially once the easy money stops pretending it is a business model.
Why LeCun wants model worlds, not just chatbots
On the technical side, LeCun said the industry needs to move beyond large language models and toward so-called world models. His argument is simple: LLMs are good at predicting the next word and useful for coding and math, but they are not built to reason about how the physical or simulated world actually works.
World models, by contrast, aim to encode objects, cause and effect, and actions. That is a more ambitious target and also a tacit rebuke to the current race for AI agents, which promise to autonomously handle complex tasks while still sitting on top of systems whose running costs are too high for many users to justify.
The uncomfortable question for Musk’s xAI and its rivals is whether the next phase of AI will be defined by better models or better economics. If the industry keeps charging like a luxury product while delivering something consumers treat like a utility, the correction may not be subtle.