The most expensive coding project in history

THE $10 BILLION that SpaceX will owe Cursor even if the $60 billion acquisition never closes is, nominally, a breakup fee. Read more closely, it is a research budget. The number converts the AI industry's loudest unanswered question — whether developer traces or compute is the scarce input in coding models — into a priced experiment, and Elon Musk is the one underwriting it.

The arrangement, reported by Bloomberg on Tuesday, gives SpaceX the right to acquire Cursor for $60 billion later this year, contingent on the integration of Cursor's Composer training recipe with xAI's compute stack. Cursor is unusually valuable to any buyer: roughly $2 billion in annual recurring revenue, Composer 2 scoring 73.7 on SWE-bench Multilingual against frontier competition, and a corpus of developer traces that Cursor claims is the richest in the industry. The company was in talks for a $2 billion funding round at a $50 billion valuation as recently as last week; the round has been quietly canceled because xAI is now picking up the compute bill it was meant to fund.

The deferred structure conveniently sidesteps SpaceX's imminent IPO, targeting a $2 trillion valuation and liable to be complicated by a $60 billion acquisition booked mid-filing. xAI has a visible problem the deal is designed to fix. Musk admitted last month that Grok had fallen behind on coding, and in March he ordered layoffs at the company and vowed to rebuild it from the ground up. xAI has been poaching Cursor engineers individually for months. Buying the company outright is the maximum version of that hiring effort.

Cursor's own calculus is less discussed but equally coherent.

The company's growth depends on frontier labs renting it their best models to wrap into a developer experience — an arrangement that works only so long as those labs prefer the margin from the rent over the margin from selling directly. Anthropic's release of Claude Code, OpenAI's of Codex, and Google's of Jules all point the same direction. Cursor's $50 billion valuation exists in a window that is closing; exiting to a buyer outside the frontier-model race, at the peak of its leverage, is the rational move. SpaceX is the available buyer with both the compute and the motive.

Test in production

Yet the $10 billion floor is the interesting number, not the $60 billion ceiling. Typical buyer-paid breakup fees in recent major tech M&A have run 2–5% of deal value; at roughly 17%, SpaceX's obligation is triple the upper bound and, in absolute terms, larger than the enterprise value of all but a handful of public software companies. The $10 billion is paying for what SpaceX learns while the deal is pending. In the arrangement's intended form, Cursor's Composer team trains its model on xAI's compute, and xAI then runs the same recipe on Grok. That gives SpaceX a direct test of whether Cursor's data and methods are portable — whether the recipe travels. Ten billion dollars is what SpaceX is willing to pay, at minimum, to find out.

The question is not rhetorical. Anthropic has reached the frontier on coding without owning an IDE or acquiring a developer-tooling company at Cursor's scale; Claude Code, built on Sonnet, is regarded as the reference product by much of xAI's own departed research bench. OpenAI's Codex has done the same through its partnership with GitHub rather than a Cursor-scale acquisition. Cursor's own Composer 2 technical report describes reinforcement learning layered on a Kimi K2.5 base inside sandboxed coding environments, an approach that depends as much on the training harness and RL infrastructure as on the traces themselves. If the corpus of developer interactions were the decisive ingredient, the competitive picture in coding would look very different. Multiple frontier coding models have emerged under materially different data regimes, which suggests compute, training methodology, and talent each have a plausible claim to the title of bottleneck.

Compounding assets

The deeper reason $10 billion is a defensible number has to do with what frontier labs now believe code is for. Increasingly, code is treated as the substrate for the next generation of models: systems that can reliably write and debug software can, in principle, improve themselves, and the reinforcement-learning environments used to train them are built out of the same material. Coding is the narrow slice of AI where a near-frontier team is raising the ceiling rather than catching up to it, and that makes Cursor one of the few acquisition targets in the field whose value plausibly compounds rather than depreciates. The $60 billion option is not only a bet that developer data is the bottleneck; it is a bet that developer tooling is the category.

Which is the bet the $10 billion is underwriting. If Composer's recipe ports cleanly to Grok, SpaceX has confirmed that Cursor's data pipeline is the asset and will pay the remaining $50 billion to own it. If the port fails — if Grok improves only at the margin, or if the Composer team cannot replicate its gains on unfamiliar infrastructure — SpaceX has still bought the answer, and Cursor has a $10 billion consolation prize. The wager is priced such that the loser, from SpaceX's perspective, still walks away with the information.

What cannot be bought

The variable no part of the deal structure can price is retention. Musk's management style has worked for the rocket and automotive engineers who have few comparable alternatives at comparable scale. AI researchers have options, which is why Meta, OpenAI, and a lengthening roster of well-capitalized neolabs keep poaching them. If Cursor's best researchers leave in the months between signing and closing, the $60 billion option will be exercised — if it is exercised at all — on something materially smaller than the asset SpaceX contracted to buy. That risk is the loudest silence in the deal structure. There is no escrow for talent.

The deferral is the elegant part. By waiting until after its IPO, Musk avoids booking a $60 billion acquisition against a company still being priced by the public market, and gives himself a full operating quarter of real data before deciding whether to exercise. The rest of the industry can keep arguing the data-versus-compute question at conferences for free. SpaceX has decided the answer is worth a ten-figure check to obtain first — and if the answer comes back in the affirmative, the rest of the market will be bidding for a category Musk has already cornered.