xAI Holds Talks With Mistral and Cursor to Chase Anthropic

xAI has held recent talks with French startup Mistral and AI coding firm Cursor about a three-way partnership, Business Insider reported Wednesday, on top of the alliance SpaceX disclosed when it secured a $60 billion option to buy Cursor on Tuesday. Mistral cofounder Devendra Chaplot joined xAI last month to lead pretraining, and Cursor has already started training models on xAI's Colossus supercomputer in Memphis. Elon Musk has repeatedly told engineers he is worried about Anthropic's lead in coding and agents, the people said, and xAI president Michael Nicolls said this month that the company is "clearly behind" rivals.

The visible number is $60 billion. SpaceX, which absorbed xAI in February in a $1.25 trillion merger, said Tuesday it can buy Cursor outright by year-end or pay the coding startup $10 billion for joint work if it walks. The structure is a call option, not a closed sale. That keeps SpaceX's confidential IPO filings undisturbed and lets the rocket company finance any later purchase with public stock.

Mistral's entry would change the math. The Paris lab has frontier-class researchers and a brand built on being Europe's independent answer to American labs. Folding it into a Musk-led trio would hand xAI model expertise it lacks, give Cursor a second model supplier so it does not depend only on Colossus, and give Mistral the GPUs it has spent two years scrounging for. Two years of scrounging. None of the three companies has confirmed the talks.

Cursor sits at the most exposed point in the AI coding stack. Its product runs on top of Claude and GPT, the same models Anthropic and OpenAI are now wrapping inside their own competing editors. Anthropic in January blocked xAI from getting Claude through Cursor, an early sign the relationship was breaking. You don't keep buying tokens from companies that want your distribution.

Theory Ventures' Tomasz Tunguz framed the option in plain terms. "For $10 billion, SpaceX buys a call option on the distribution it couldn't retain, and Cursor wins the independence it hasn't yet secured." Cursor says its in-house Composer 2 model has reached "frontier-level performance at a fraction of the cost of other models," and that Composer 2.5 is the version Colossus will now train at scale. Whether the benchmark holds in production is a different question. Aadit Sheth of The Narrative Company was sharpest: "Distribution without a defensible model underneath is a rental. We'll know in 6 to 12 months whether that $60B bought a moat or a rental."

Chaplot's move to xAI follows a pattern. Two senior Cursor engineering leads, Andrew Milich and Jason Ginsberg, joined SpaceX in March and now report directly to Musk and Nicolls. That is not how partners behave. That is how an acquirer behaves before the price is set. By the time SpaceX signed the option agreement, key engineers were already inside the parent. Cursor still locked in $10 billion of joint-development work as a floor, but the talent flow had been one-way.

Mistral's calculus is harder. Joining a Musk-led coding alliance against Anthropic would help with compute. It would complicate the independence story. Paris reads those headlines too. Brussels does. Mistral's whole pitch has been independence from American labs, and that pitch gets harder to deliver next to Musk's logo.

SpaceX is racing a deadline. The confidential IPO filing went in early April. Bankers expect a price tag near $1.75 trillion and a $75 billion raise. Both figures, if they land where the rumor mill says, would set records. By stitching coding software, Mistral talks and an existing AI lab onto the rocket and Starlink business, Musk is asking public investors to apply software multiples to a company that mostly sells launches and broadband.

Andreessen Horowitz stands to clear about $6 billion on its Cursor stake at the $60 billion price, Thrive Capital another $4.2 billion. They will not lobby for restraint. Not at those numbers. Anthropic still owns the coding revenue, the user data and the editor shelf space that any three-way alliance is built to chip away at. Twelve months from now we will know whether a French lab, a Memphis supercomputer and an editor for vibe coders can be glued into something that looks like a frontier model company. Or whether the option expires for $10 billion and a press release.