From Machine Brief, June 18:
Turing Award winner Yann LeCun labelled Elon Musk's xAI a 'failure' and warned AI labs risk a 'big bubble explosion' if costs aren't controlled. This guide covers xAI's $6.4 billion losses, whether an AI bubble exists, and what rational valuations look like in a post-Fable 5 market.
Introduction
Yann LeCun — one of the "Godfathers of AI," recipient of the 2018
Turing Award with Yoshua Bengio and Geoffrey Hinton, and founder of AMI
Labs after leaving his role as Meta's chief AI scientist — does not hold
back. On June 18, 2026, he labelled Elon Musk's xAI a "failure," said
he expects it won't be able to compete with OpenAI and Anthropic, and warned that AI labs are risking a "big bubble explosion" if they don't cut costs and raise prices.
This is not a new spat. LeCun and Musk have been trading public
criticism since at least 2024, when LeCun openly questioned Musk's
claims about xAI's technical progress and Musk responded by calling
LeCun "out of touch." But the June 2026 comments are different. They
come as xAI reportedly lost $6.4 billion in 2025 according to SpaceX's
IPO filing — the first public look at Musk's AI financials. They come as
AI company valuations have reached levels that make even true believers
uncomfortable. And they come from someone whose technical credibility
in AI is essentially unassailable.
This guide covers what LeCun said, why he said it, the xAI financial
picture, whether there's actually an AI bubble, and what rational
valuation looks like for AI companies in 2026.
What LeCun Actually Said
LeCun's comments warrant direct quotation because the force of his language matters:
On xAI specifically: he called it a "failure" and said it "won't be
able to compete" with OpenAI and Anthropic. The assertion is specific:
xAI's model performance hasn't matched the frontier, and LeCun doesn't
see a path to closing the gap.
On the broader AI industry: he warned that labs are risking a "big
bubble explosion" — his words — unless they cut costs and raise prices.
The implication: AI companies are burning cash on compute at rates that
cannot be sustained by current revenue, and when the capital markets
re-price the risk, the correction will be severe.
This is not a random critic. LeCun's contributions to AI — convolutional neural networks, backpropagation applications, self-supervised learning,
the energy-based model framework — are foundational. When he says an AI
company's technical approach won't work, it carries different weight
than when a financial analyst says the same thing.
The xAI Financial Picture — $6.4 Billion in Losses
The SpaceX IPO filing, which became public in May 2026, revealed
numbers about xAI that had previously been private: the company lost
$6.4 billion in 2025. This is the first hard data on xAI's financials,
and the figure is striking even in an industry known for large losses.
For context: OpenAI reportedly lost approximately $5 billion in 2024
and projects losses of $14 billion through 2026 before reaching
profitability. Anthropic's losses are not publicly known but are
estimated in a similar range. xAI's $6.4 billion loss for 2025 places it
in the same financial league — but without having achieved the same
model performance or enterprise traction.
Where did the $6.4 billion go?
Compute infrastructure. xAI built Colossus,
reportedly one of the largest GPU clusters in the world, in Memphis,
Tennessee. The cluster reportedly contains 100,000+ H100/H200 GPUs. At
roughly $25,000-40,000 per GPU plus networking, power, and cooling
infrastructure, the capital expenditure alone is in the billions.
Talent. xAI has reportedly hired AI researchers at
compensation packages competitive with OpenAI and Anthropic — meaning
$2-10 million annually for senior researchers. A research team of
several hundred people at these compensation levels costs $500 million
to $1 billion per year.
Training runs. Training Grok-3
and Grok-4 requires tens of thousands of GPUs running continuously for
weeks or months. Each frontier training run costs an estimated $100-500
million in compute alone, and multiple runs are needed as experiments
fail and parameters are adjusted.
Inference costs. Every query to Grok costs money in
compute. If Grok has significant usage — and it does, integrated
directly into X — those inference costs scale with usage.
The fundamental question LeCun is raising: is this spending producing
commensurate results? Grok-3 and Grok-4 have not matched GPT-5.4 or Claude Fable 5 on major benchmarks. No independent evaluation
has placed a Grok model at the frontier. If you're spending comparable
amounts to OpenAI and getting worse results, the business case gets
harder to make.
Is There Actually an AI Bubble?....
....MUCH MORE
On the other hand SpaceX, Xai's parent company, just signed another deal to lease GPUs, this one for $150 million per month following on the deal with Google for $920 million per month and
See NextBigFuture, June 22: SpaceX Has Another $150 Million Per Month Deal
There is also the $1.25 billion-per-month Anthropic deal but that is for a mix of Nvidia chips including older A100's, H100's and H200's.
