“Kind of a failure, frankly.” That was Yann LeCun on CNBC’s Squawk Box Europe on June 18, talking about Elon Musk’s xAI with the ease of a man who has spent years betting against the room. The Turing Award and a career spent building the field give a line like that some weight. So does the fact that he now runs a lab selling the opposite bet, which matters more than it looks, and which we will get to. Either way, the claim is specific enough to check against the books.
Every founding member of xAI not named Elon Musk has left over the past year. The last holdout, Ross Nordeen, was reportedly cut off from the company’s systems and dropped from the internal group chat back in March, a strange way to part with the person who stuck around longest. This is the strongest leg of LeCun’s case and the hardest to wave off as rivalry. His read is that Musk has torched the bridges he would need to rebuild the bench. “Elon is now in a position that is very, very difficult for him to hire top people in AI,” he said, pinning it on how Musk treated the founding team. A few researchers can set the direction of a whole lab, and losing all of them while struggling to backfill at the frontier is a momentum problem, not a staffing one.
SpaceX absorbed xAI in February at a combined valuation of $1.25 trillion, and the AI side of that business reported a $2.5 billion operating loss in the first quarter. A number like that is survivable for a growth-stage company. How xAI is covering it is what should make you wince.
Colossus 1 and Colossus 2, the two Memphis data centers, were the moat. The pitch was proprietary compute at a scale that would hand Grok a structural edge over OpenAI and Anthropic. Instead, xAI is renting that capacity out to Anthropic and Google, with Google alone signing on for roughly $920 million a month. The hardware built to win the race is now a colocation business for the labs it was supposed to bury. That is the contradiction LeCun points at when he says xAI “will not be able to compete with OpenAI and Anthropic on the frontier.” You cannot run a frontier lab and serve as your rivals’ landlord at the same time.
From there, he widened the lens to the entire frontier industry, which he argues runs on a subsidy that arithmetic eventually collects. “The prices of those AI services are going up, but the cost of running them is going down, but not nearly fast enough,” he said. The shortfall gets covered by investors, and in his telling, that “can’t go on for very long.” Either the prices climb, or the costs finally fall fast enough to meet them. If neither happens, in his words, “there’s going to be a big bubble explosion.”
He is not wrong that inference keeps getting cheaper. Hardware efficiency improves, and distillation trims compute per query. It just does not fall fast enough to catch the subsidized pricing the labs lean on to grab and hold users, and venture money covers the difference. What makes this hard to dismiss as LeCun being contrarian for sport is that Sam Altman called AI costs a “huge issue” this month. When the CEO of the most valuable lab alive and its loudest outside critic land on the same diagnosis, it is worth sitting with.
In March, LeCun’s own startup, AMI Labs, closed $1 billion at a $3.5 billion pre-money valuation, and its entire premise is that large language models, the architecture under everything OpenAI, Anthropic, and xAI ship, are a dead end. He has argued for years that LLMs cannot reach general intelligence and that “world models,” the thing his lab is building, are the road forward.
Which means every beat of the xAI takedown doubles as a pitch for AMI Labs. If the frontier gets squeezed into raising prices and the talent bleeds hard into a real capability gap, his deck writes itself, and a genuine bubble would confirm the whole thing. The facts underneath still hold up. The departures are documented, and the Q1 loss is a reported number. The computer rental to rivals is not spinning either; it is just how the business runs right now. But the person arranging those facts into the word “failure” is also raising money on the bet that the paradigm behind them is broken.
The feud underneath this is old, too. Musk has publicly dismissed LeCun’s relevance to modern AI more than once, which makes it easy to read the xAI takedown as a grudge wearing a spreadsheet. It is probably both at once. The exodus and the losses are not opinions, and they would land on Musk no matter who said them out loud. Choosing “failure” over “a company under real pressure” is a choice, and the choice carries history.
SpaceX and xAI did not respond to requests for comment.
A $1.25 trillion valuation prices in things the Q1 numbers do not support yet, and a $2.5 billion quarterly loss softened by leasing compute to its rivals, which ~$920 million a month from Google, especially, is a company spending its strategic crown jewels to steady the books. Strip the rhetoric, ic and that is what is left, a lab caught between its ambition and its arithmetic. Talent is harder to put a number on, but every founding researcher walking out and a hiring problem standing behind them point the same way.
That is not the same as defunct. Grok still ships, and Colossus, awkward as its current role is, is real and enormous. LeCun’s claim is narrower than “failure” makes it sound. It is about position at the frontier, and on that narrow question,n the evidence bites. You do not close the gap with OpenAI and Anthropic after your best people walk and your compute advantage ends up leased to the very labs you are chasing.
The bubble callout lasts longer than the xAI fight, which is why it is the part worth keeping. It goes beyond any one company and is co-signed by the industry’s own leadership. Subsidized pricing held up by investor-funded inference, with costs falling more slowly than revenue models require, is a structural problem, and no single lab can patch it alone. LeCun’s fix is sitting right there in his cap table: drop LLMs in favor of something else. He might even be right about the architecture. What snags me is dumber than the architecture debate. He is probably correct about the numbers, and he has also got a billion dollars riding on me believing him, which is the kind of thing that makes me want to recount the money twice even when the tally looks clean.