SpaceX's $60B Option on Cursor (Anysphere): A Paid Experiment, Not M&A
A chocolate coconut cream pie, a $60 billion option on Cursor, and a paid experiment dressed up as M&A. $10B is what SpaceX pays to find out. $60B is what the answer is worth if it comes back yes.
I'm Tyler Kirwan. I've spent 15 years in procurement in big tech. I'm building datacenterindex.ai as a public notebook while I learn about the data center buildout. This is not investment advice.
I was picking up a chocolate coconut cream pie to celebrate my son's six-month birthday when I pulled up yesterday's news on my phone. SpaceX has an option to acquire Cursor's parent company, Anysphere, for $60 billion later this year. If SpaceX doesn't exercise it, they pay Anysphere $10 billion for "our work together." Cursor, which has reportedly cleared $2 billion in annualized revenue, is used by more than a million developers every day, and now reaches roughly 70% of the Fortune 1000.
The headlines this morning are mostly about what the deal means for the AI model race. But I keep thinking about a different implication. SpaceX has ready access to one of the largest GPU clusters on earth. They already supply Cursor with compute. They are not short on chips or capital. So why is one data source worth $60 billion?
The pie on the passenger seat was analogous to what SpaceX is doing. My wife and I already know we love the coconut cream pie from our favorite pie shop. The chocolate coconut cream was the experiment. We knew the coconut cream worked. We did not yet know whether the chocolate layer would make it better or just make it busier. SpaceX has the same question, scaled up by about $70 billion.
The part of the deal that actually costs $60 billion
SpaceX and Cursor have been working together for a while. xAI has been supplying GPUs to Cursor for model training for months already. The "$10 billion for our work together" line in the press release does not describe a future collaboration. It is paying for a relationship that already exists.
A post this morning from @anandnk24 on X sharpened how I was reading the option structure. The cleanest description of the option, once you sit with it, is a paid experiment dressed up as M\&A.
Cursor is going to train Composer 3, its in-house frontier coding model, on Colossus. xAI is going to run the same training recipe on Grok. Both parties find out, at the same time, whether Cursor's developer traces are actually the scarce input in frontier coding models. If the training work ports over and Grok closes the gap on Claude Code and Codex, SpaceX exercises the option and owns the pipeline. If it does not, SpaceX pays $10 billion for the experiment and walks away with a better Grok anyway, plus an answer to a question xAI could not answer internally.
Read that way, the two numbers do something I had not appreciated before. $10 billion is the experiment's running cost, no matter what happens. $60 billion is the value if it comes back, yes. The gap between those two numbers reflects how highly SpaceX values the question itself, and it is very large.
In dessert terms: SpaceX already has the coconut cream. Colossus is coconut cream at industrial scale. $10 billion is the cost of tasting the chocolate with the coconut cream. $60 billion is the price of owning the recipe if the chocolate turns out to be what makes the pie even better.
What Cursor's data actually is
Cursor sees more developer behavior than any other single system in the world. Billions of code completions served every day. Accept-reject signals on every suggestion. Human corrections. Chat threads explaining intent. Multi-file refactors. Debugging sessions across monolithic codebases. Math concepts turned into working code. Scraped GitHub is output. Cursor's pipeline is a process. It captures the path a developer took to get to the output, which is the signal that teaches a model to be useful rather than just syntactically correct.
That is the reason the experiment is worth running at all. xAI has been leaning heavily on X (the social network) for training data, and X is good at tone and not much else. Grok has been losing the coding and reasoning benchmarks to Claude and GPT for most of the last year, and the gap was never going to close on more X posts. Whether Cursor's specific flavor of developer-trace data is the ingredient that closes it is exactly what the $10 billion buys the right to find out.
What this reframes for me
I came into this morning thinking about AI spend as a compute story. More models, more training runs, more GPUs, more data center capacity. That is still directionally true. But the Cursor deal is a market signal that the frontier labs do not yet agree on where the next bottleneck actually is. Compute is getting abundant. The open question is whether specialized, high-fidelity, process-rich data is what makes the next model better. SpaceX is now paying $10 billion to find out, with $60 billion more riding on the answer.
The $60 billion valuation for a pre-IPO company with around 50 employees and a two-year-old product is not a typical acquisition multiple. It's a conditional price, payable only if the experiment comes back with the result that developer traces are, in fact, the moat. Read literally, SpaceX is saying: if the answer is yes, $60 billion is a bargain. If the answer is no, $10 billion for the experiment and a better Grok is the preferred outcome. Both sides of that option are coherent, which is what makes the structure as clean as it is.
Another implication for the physical buildout. More data still has to be processed somewhere, and somewhere is always a cluster. If every major lab now has to replicate something like the Cursor data flywheel, the training workloads that result still land in data centers. The buildout I am trying to learn about keeps getting more runway, not less.
What I'm watching
The specific things I am tracking next, in order of how much I think they tell me:
Whether OpenAI or Anthropic announces a comparable data deal in the next ninety days. If they do, the "data is the new moat" thesis firms up fast. If they don't, it's possible SpaceX overpaid for a problem the rest of the market thinks it can solve more cheaply.
Whether SpaceX actually exercises the option later this year. If they do, that is a real data point that vertical AI stacks (launch, compute, model, distribution) actually work at the top of the market. If they take the $10 billion path instead, I will read the option as more defensive than the announcement implied.
Whether Cursor's data pipeline stays exclusive. Anysphere continues to run Cursor independently under the current deal. If a large share of the data starts flowing to SpaceX in a way that degrades the product for other users, it has real second-order effects on enterprise adoption and on the comparable deals that follow.
I am also keeping the buildout thesis alive in a separate notebook. I will write about that soon.
A $60 billion option on one data pipeline is a very large number. SpaceX could have spent the same money on more compute instead. They didn't. They priced an experiment instead, and set the spread wide enough to tell you they really want to know whether the chocolate layer makes the pie. My wife and I have not cut into ours yet. I will report back on the smaller experiment.
Follow along at datacenterindex.ai as I learn my way through the rest of this. If you have been close to the data-infrastructure side of AI (building training pipelines, licensing proprietary corpora, or investing in data-centric AI businesses), I would love to talk.