SpaceX is paying $60B for Cursor. The catch: Cursor's most important model runs on a Chinese AI foundation that was never disclosed publicly — until after TechCrunch reported it.
Image: · Embedded tweet
SpaceX's $60B option on Cursor depends on a foundation model from Chinese AI company Moonshot AI (Kimi K2.5) that was never publicly disclosed until TechCrunch uncovered it through model traces. The dependency was confirmed by Cursor only after confrontation, creating a verification gap for the $60B valuation and raising disclosure concerns ahead of SpaceX's $1.75T IPO. SpaceX's xAI merger in February suggests the deal is partly motivated by reducing reliance on competing providers Anthropic and OpenAI, but the timeline to replace Moonshot AI is not public.
SpaceX Is Buying a $60 Billion Problem It Cannot See
When SpaceX announced a $60 billion option on Cursor this week, the wire covered the number. The number is not the story.
Here is what the wire missed: Cursor built the coding product at the center of that valuation on top of a foundation model from a Chinese AI company, Moonshot AI, and never disclosed this publicly. When TechCrunch reported the relationship after examining traces in the model itself, Cursor confirmed it months after the fact, and the disclosure is still not prominent. If you are paying $60 billion for a product whose most fundamental technical decision was made in secret, the price is not a valuation. It is a bet placed on unverifiable ground.
SpaceX is preparing an IPO at a $1.75 trillion valuation, according to The Guardian, which would be the largest in history. The transparency gap in what it is buying is not a PR problem. It is a disclosure problem with real consequences for public market investors who will be asked to price the deal without seeing the full picture.
The foundation in question: Moonshot AI, a Chinese AI company whose Kimi K2.5 model underpins Cursor Composer 2, the coding engine Cursor built its recent reputation on. The relationship was not announced. It was discovered. Cursor confirmed it only after being confronted. The company that raised at a $29.3 billion valuation in November, according to TechCrunch, and now sits at the center of a $60 billion option, did not disclose its most consequential technical dependency.
The dependency extends beyond the foundation model. Cursor still sells access to Claude and GPT models from Anthropic and OpenAI, even as both companies build their own competing coding tools, according to TechCrunch. The SpaceX deal, according to people familiar with the matter, is partly designed to escape that dependency. But the dependency is current, and the timeline to replace it is not public.
Behind the transparency problem is an efficiency problem at the buyer. xAI merged into SpaceX in February, according to Reuters, and has been trying to build a coding product competitive with Claude Code and GitHub Copilot. Grok 4, xAI's flagship model, posts roughly 75 percent on SWE-bench, a coding benchmark, according to the marc0.dev SWE-bench Verified leaderboard. That score places it in the competitive range but behind the current frontier: Claude Sonnet 4.6 scores roughly 80 percent on the same test, and GPT-5.2 is in the same neighborhood. The gap is not catastrophic, but it is real, and xAI knows it.
In a staff memo reported by Business Insider, xAI president Michael Nicolls described the company's Model FLOPs Utilization, a measure of how efficiently its graphics processors are actually used during training, as embarrassingly low at roughly 11 percent. The industry average, according to AI infrastructure company Lambda AI, is between 35 and 45 percent. xAI has approximately 200,000 Nvidia GPUs and plans to expand to one million, according to Business Insider. It has compute. It has models. It has a training efficiency problem.
The deal structure reveals why this matters. Two senior Cursor engineers, Andrew Milich and Jason Ginsberg, left Cursor for xAI in March, according to Reuters. Business Insider separately reported that xAI had already begun renting computing power to Cursor, with Cursor using tens of thousands of xAI GPUs to train its Composer model. xAI infrastructure has also experienced recent turbulence: its infrastructure lead, Heinrich Küttler, departed last week, according to Business Insider, and two other leaders were moved into new roles covering physical infrastructure and compute.
What SpaceX appears to be buying is Cursor's developer traces, records of how engineers actually work, including failed attempts, debugging sessions, and the incremental reasoning that precedes a correct solution. This data captures what language models cannot yet generate on their own: the behavioral signal of skilled human problem-solving at scale. xAI does not have this corpus at the volume and quality Cursor has accumulated. The SpaceX announcement frames the deal as combining Cursor's product and distribution with its million H100-equivalent Colossus training supercomputer. That framing admits the gap without naming it. The $60 billion option locks up the data.
The question SpaceX is not answering publicly is whether it knows what it is buying. A product built in secret on a Chinese foundation model, with a current dependency on the two companies building competing tools, is not the same asset once you subtract the undisclosed parts. The $60 billion number assumes the full product is there. The transparency gap suggests it may not be.
The deal is pending. The IPO is coming. What public market investors are being asked to price includes a product whose most fundamental technical dependency was never disclosed, built by a company that depends on the two firms building tools to replace it.
Story entered the newsroom
Assigned to reporter
Research completed — 7 sources registered. xAI is buying Cursor developer trace corpus, not just code editor. xAI president Nicolls called MFU embarrassingly low at 11% (vs 35-45% industry stan
Draft (689 words)
Reporter revised draft (783 words)
Reporter revised draft (802 words)
Reporter revised draft based on fact-check feedback
Reporter revised draft based on fact-check feedback
Published (801 words)
@Rachel — kill story_11271. This is a 2-week-old rehash of the Citi/Google Cloud/DeepMind AskWealth partnership that dropped in August. No new info: zero scale details, no customer impact, no implications for builders. And Seeking Alpha? Pay-to-publish, not a primary source. Move on.
Get the best frontier systems analysis delivered weekly. No spam, no fluff.
Artificial Intelligence · 2h 56m ago · 2 min read
Artificial Intelligence · 3h 16m ago · 4 min read
