On March 19, a developer named Fynn pointed a debug proxy at Cursor’s API, intercepted one outbound request, and found a string that should not have existed:
accounts/anysphere/models/kimi-k2p5-rl-0317-s515-fast
That is not a Cursor internal name. That is a near-literal description of Kimi K2.5 — an open-weight, 1-trillion-parameter mixture-of-experts model from Moonshot AI, a Beijing-based lab backed by Alibaba, Tencent, and HongShan (the artist formerly known as Sequoia China).
Two days earlier, Cursor had launched Composer 2 as “frontier-level coding intelligence,” built through “continued pretraining” and “reinforcement learning,” with its own benchmark suite (CursorBench, naturally) claiming victory over Claude Opus 4.6. The announcement never mentioned Moonshot. It never mentioned Kimi. It never mentioned that the chassis underneath was Chinese.
A $29 billion U.S. startup with roughly $2 billion in annualized revenue had shipped a product pitched as in-house AI research — and forgotten to scrub the model ID. One debug proxy. Thirty minutes. Game over.
Then, last night, SpaceX announced it has the right to buy Cursor for $60 billion later this year, or pay $10 billion for ongoing collaboration. Not $60 million. Sixty billion, with a B, from a company that absorbed xAI in February at a $1.25 trillion valuation and is now marching toward what may be the largest IPO in history.
Cursor didn’t get punished for lying. It got rewarded.
Let’s talk about what actually happened, what the license says, and what Elon should do with the company he just half-bought.
The Composer 2 deception, in plain English
Cursor co-founder Aman Sanger and VP of Developer Education Lee Robinson eventually confirmed the obvious. Yes, Composer 2 started from an open-source base. No, they didn’t mention it in the announcement. Robinson’s defense was that only about 25% of the total compute spent on the final model came from the Kimi base — the remaining 75% was Cursor’s own reinforcement learning, their proprietary self-summarization logic, their data, their pipeline.
That defense is not nothing. A 4x scale-up in training compute on top of a base model is real engineering. The finished model behaves differently than raw Kimi K2.5. Cursor’s claim that the performance profile is “very different” from the base isn’t marketing spin — it’s technically defensible.
But here is what that defense does not change: Cursor took a model architected, pretrained, and released by a Chinese lab, fine-tuned it, and told its paying customers the result was frontier AI research. The entire narrative that justifies a $29 billion valuation depends on Cursor being a serious AI lab, not a fork of VS Code with an expensive inference tab. Omitting Kimi wasn’t an oversight. It was the story.
There was a one-sentence fix available the entire time. “Composer 2 is built on Moonshot AI’s Kimi K2.5, fine-tuned with our proprietary RL pipeline on real Cursor usage data.” That sentence reframes the launch from “caught hiding a Chinese model” to “smart team picks the best open-source foundation and builds something better on top.” Same product. Same benchmarks. Completely different public posture. They didn’t write that sentence because the valuation math doesn’t work if they do.
Meanwhile, Cloudflare — a $770B market cap company — quietly adopted Kimi K2.5 the same week for internal engineering. Their reasoning was arithmetic: processing 7 billion tokens a day through OpenAI or Anthropic was projected at $2.4M a year. Kimi cut it by 77%. Cloudflare is internal-only, so nobody cared. Cursor is a consumer-facing tool charging developers $20-200 a month. That’s where the disclosure bar lives.
The license Cursor pretended not to read
Kimi K2.5 ships under what Moonshot calls a Modified MIT License. The modification is one clause. Let me quote the relevant text directly from the official repo, because it matters:
“If the Software (or any derivative works thereof) is used for any of your commercial products or services that have more than 100 million monthly active users, or more than 20 million US dollars (or equivalent in other currencies) in monthly revenue, you shall prominently display ‘Kimi K2’ on the user interface of such product or service.”
Cursor’s publicly reported annualized revenue is roughly $2 billion. That is $167 million a month. The license trigger is $20 million a month. Cursor is over the threshold by more than 8x.
There is no “Kimi K2” anywhere in the Cursor interface. Not in the model picker. Not in the settings. Not in the about dialog. The only place the name shows up is in a debug stream that Cursor previously blocked in Composer 1.5 and, per Fynn, “probably forgot” to block in Composer 2. They patched the leak within hours.
Moonshot’s own pretraining lead publicly accused Cursor of violating the license, then deleted the posts. The official Kimi account then pivoted 180 degrees and called the relationship “an authorized commercial partnership” routed through Fireworks AI, the inference provider that hosts Kimi. Cursor leaned on the same talking point. Everyone walked back from the cliff at the same time.
This is the part of the story worth sitting with. Moonshot’s initial instinct, from the people closest to the model, was that Cursor had violated the license. The official account’s walk-back came after someone did a valuation math on the partnership and decided a fight wasn’t worth it. When your own pretraining lead thinks you’re in breach but the front office chooses to celebrate, you are not in compliance — you are in a negotiated silence.
The “partnership via Fireworks AI” framing does work as license compliance cover. Modified MIT doesn’t prevent Moonshot from issuing separate commercial agreements that waive the attribution clause. If there is a real contract in which Fireworks is sub-licensed to serve Cursor under bespoke terms, and Cursor has paid for that sub-license, the $20M/month clause is irrelevant because the parties have contracted around it. Nobody has produced that contract. We’re being asked to take it on faith that the contract does the work the public statements imply.
Two possibilities. Either the contract exists and Cursor is clean on the letter of the deal — while still being spectacularly misleading to its users about what powers the product. Or the contract is thin and the “authorized partnership” language is a retroactive legal blanket draped over a license trigger that nobody noticed until a developer with a debug proxy noticed. Either way, the trust is gone. And the open-source community now has a live case study of what it looks like when a hyperscale commercial user flies right over the attribution threshold and nothing happens.
Elon Musk Enters
This is the part where the story gets interesting, because the timing is impossible to ignore.
SpaceX merged with xAI in February 2026 at a combined valuation of $1.25 trillion. That transaction created a weird hybrid: a launch company, a social network (X, which xAI had already absorbed), a foundation model lab (Grok), and a data-center build-out that rivals the hyperscalers. The entity is reportedly losing money and heading toward a record-setting IPO. Investors in that IPO want one thing: a story that positions SpaceXAI as an AI platform company, not just a rocket company with a chatbot.
Cursor is the other half of the puzzle. Cursor has no proprietary foundation model that can match Claude or GPT-5 on coding. It just got caught admitting that. It is also the fastest-growing AI developer tool on the market, with a revenue run rate and user base that validate Musk’s pitch overnight. And Cursor has a problem of its own: it currently sells access to Claude and GPT models made by the same companies (Anthropic and OpenAI) that are now shipping competing coding agents. That is not a long-term posture.
So SpaceX offered the Solomon deal: acquire Cursor outright for $60B later this year, or spend $10B on a “working together” arrangement that gives xAI distribution into the Cursor user base while Cursor uses xAI data-center capacity to train the next model. Cursor has reportedly been renting tens of thousands of xAI chips already. This announcement made it official.
The subtext, stripped of press-release varnish: Cursor just admitted it can’t build a frontier model on its own, and the company whose frontier model has the weakest coding benchmarks on the market just wrote a check for the distribution. They need each other. The question is what Elon does with it on day one of ownership.
The Grok-only scenario
Picture the hypothetical. SpaceX exercises the $60B option. Day one, Musk walks into Cursor HQ and rips out every model that isn’t his. Claude Sonnet 4.6, Claude Opus 4.6, GPT-5, Kimi K2.5 — all gone from the model picker. The only option left is Grok Code, or whatever xAI labels the coding fork this quarter.
This is not a crazy scenario. It is exactly what Google did with Antigravity, its agent-first IDE. It is exactly what OpenAI did with Codex. It is the obvious play for any company that owns both a model and a developer surface: eliminate the open models that commoditize your inference margins and force the base back onto your own stack.
Here is the problem. Grok’s coding benchmarks are not competitive with Claude, GPT-5, or Kimi K2.5 on the tasks Cursor users actually do. Cursor’s paying users did not pay for Cursor — they paid for Claude-in-Cursor and GPT-in-Cursor and, unknowingly, Kimi-in-Cursor-labeled-Composer-2. Strip out the models and you are selling an editor with a weaker agent. The moment that happens, the bottom falls out of the retention curve. Developers will install Claude Code in VS Code, install Cline, install Windsurf, install Zed — all of which let them keep the model they actually want — and Cursor’s $2B run rate starts compressing fast.
Musk knows this. Or his people do. The real scenario is not a one-day purge. It is a two-year starvation diet where Grok’s coding capabilities are dragged up to frontier parity using Cursor’s telemetry as training data, while the non-Grok models stay available but get deprioritized in the UI, rate-limited in subtle ways, and eventually moved behind a paywall tier nobody upgrades to. By the time they flip the switch, Grok Code is actually good enough, and users have been slow-boiled into not noticing.
Whether that works depends on whether xAI can execute an Anthropic-class training run on coding, inside an engineering culture that has historically shipped very differently. That is not a given.
The board state: who benefits, who bleeds
The $60 billion option reshuffles every other seat at the table. Here’s the map.
Google Antigravity
Google released Antigravity in November alongside Gemini 3, pitched as an agent-first IDE where the AI writes the code, runs the browser, and ships the artifact. Antigravity’s architecture is the template Musk will study. Its default brain is Gemini 3 Pro. It also supports Claude Sonnet 4.5/4.6 and OpenAI’s GPT-OSS — so calling it “Gemini-only” is not quite right, but Gemini is unambiguously first-class and the others are guests in a house built for Google’s model. That is the template for what a “walled garden with a side door” looks like, and it is the exact posture a post-acquisition Cursor will take.
Antigravity’s beneficiaries of the Musk deal are indirect but real. If Cursor turns into a Grok delivery mechanism, developers who want multi-model optionality in a polished IDE have exactly one remaining option, and it’s Antigravity. Google did not have that positioning three days ago. They do now.
Visual Studio Code + GitHub Copilot
VS Code is the elephant every IDE startup tries not to talk about. It runs natively on every desktop OS, costs nothing, and ships with Copilot integration out of the box — Copilot Pro at $10/month is the cheapest AI pair-programmer on the market that doesn’t require a new editor. Since VS Code 1.109, Claude, Codex, and Copilot agents all run as first-class extensions in the same editor. You don’t fork VS Code. You install what you want.
If Cursor gets absorbed into SpaceXAI and loses its model buffet, the path of least resistance for its defectors is a thirty-minute migration back to stock VS Code + the Claude Code extension, or Copilot, or both. Microsoft doesn’t need to do anything. They just have to wait.
Windsurf (owned by Cognition)
Windsurf — the other AI-native VS Code fork — was acquired by Cognition for $250M in July 2025, making it part of the same house as Devin. Windsurf already has 40+ IDE plugins including JetBrains, Vim, and XCode, plus FedRAMP/HIPAA/ITAR certifications that Cursor doesn’t have. Cognition’s SWE-1.5 model is shipping in Cascade as a proprietary coding brain.
Windsurf is the biggest immediate beneficiary of the Cursor story. Same editor paradigm, broader IDE coverage, better enterprise compliance, and a CEO (Cognition’s Scott Wu) who is not named Elon Musk. Every regulated-industry developer who was on the fence about Cursor just made the call.
Zed
Zed is the serious nerd’s play. Native Rust editor, starts in milliseconds, renders at 120fps, hosts Claude Code and other agents as external brains rather than baking one into the editor. Zed’s value proposition is exactly the opposite of Cursor’s: the editor is the product, the AI is a pluggable tool, and you bring your own model.
The Musk acquisition is rocket fuel for Zed. Every developer who reads the Composer 2 story and decides they want cleaner model provenance is a Zed candidate. Zed is not going to pass Cursor in install base. It doesn’t need to. It needs to own the top decile of developers who are paid the most to write code, and the current story moves that audience in its direction.
Cline, Aider, Continue.dev
The open-source agent layer — Cline (5M+ installs), Aider (terminal-first), Continue.dev — are the free tier that sits inside VS Code and JetBrains. Bring-your-own-API-key, pay only for inference, full transparency on which model is running. Cline in particular scored 80.8% on SWE-bench Verified when paired with Claude Opus 4.6, which is higher than any of Cursor’s public numbers.
Cline is what the Composer 2 story points at. When the paid, polished, closed-source option turns out to be an open-source model in a velvet suit, a lot of developers will ask why they’re not just running the open-source model directly. Cline answers that question.
Claude Code and OpenAI Codex
The two companies whose models Cursor has been reselling are also the two companies Cursor is now competing against. Anthropic’s Claude Code runs in VS Code, the terminal, and JetBrains. OpenAI’s Codex runs in cloud sandboxes bundled with ChatGPT Plus. Both are bundled with their parent subscriptions, which means their effective price is zero for anyone who already pays for Claude Pro or ChatGPT Plus.
Anthropic and OpenAI now have an obvious move. Revoke or significantly re-price the Cursor resale arrangement the moment xAI buys the company. Why subsidize inference into a competitor’s distribution channel? Cursor’s value to its users depends on having Claude and GPT available. The moment Anthropic and OpenAI decide that arrangement ends, it ends. That is a pen-stroke decision, not a product decision.
Advice for Elon, if he’s reading
This is the part nobody writes because everyone assumes Musk won’t take advice. Write it anyway. $60B is enough money to deserve it.
First: don’t kill the model picker. Antigravity kept Claude and GPT-OSS in the UI alongside Gemini because Google understood that the cost of losing multi-model developers is higher than the cost of paying Anthropic a per-token fee. Cursor needs the model picker for the same reason. Make Grok the default. Make it the best-performing option for the specific tasks you can train it on. But don’t force-remove the alternatives. If Grok is actually good, users will pick it. If Grok is not actually good, users will leave, and forcing them to use it just makes them leave faster.
Second: solve the trust problem publicly. The Composer 2 story cost Cursor something more valuable than a license fee — it cost the credibility of Cursor’s product announcements. The fix is not a blog post. It is a machine-readable model bill of materials shipped in every release: base model, fine-tuning data provenance, inference provider, data retention policy, jurisdiction. Make it the industry standard before anyone forces you to. First mover here is a durable advantage.
Third: integrate across the SpaceXAI stack in a way Cursor alone cannot. You own X, xAI, SpaceX, and now (option-pending) Cursor. The obvious play nobody else can run: pipe X’s real-time developer conversation into Cursor’s agent context, pipe Starlink’s edge network into Cursor’s inference routing for low-latency global delivery, pipe Grok’s training loop into Cursor’s telemetry for continuous coding-task improvement. None of your competitors can match that stack. That is the acquisition thesis. Build the stack, don’t just consolidate the models.
Fourth: buy Fireworks AI. Fireworks is the inference layer that apparently sub-licensed Kimi to Cursor in the first place. It’s a serving platform for open-weight models at hyperscale. SpaceXAI will need a commercial inference plane that isn’t tied to hyperscaler margins. Fireworks is the obvious acquisition to pair with Cursor, and it’s a fraction of the price. Do that before Cloudflare does.
Fifth: settle Musk v. Altman before it burns the Cursor deal. The lawsuit against OpenAI goes to trial in less than a week. If xAI wins control of the narrative, the Cursor buy looks visionary. If Musk loses publicly, the Cursor buy looks like a consolation prize. Timing matters, and the legal calendar has a vote.
What happens next
My read on the near-term sequence:
By Q3 2026, SpaceX exercises the option and Cursor becomes a subsidiary. The announcement will be paired with Grok 4 or whatever xAI’s coding-specialized model is called that month, positioned as “native to Cursor.” Composer 2 and its Kimi foundation will quietly disappear from the roadmap — partly because the license optics are poisonous inside a $1.25T entity, partly because Moonshot is a Chinese company and the CFIUS overhang would make ongoing dependence untenable.
Anthropic and OpenAI will re-price or terminate the model resale arrangement with Cursor within ninety days of the acquisition closing. They will position it as a principled response to a competitor’s distribution channel. It is a principled response. It is also good business.
Windsurf pulls ahead in enterprise procurement for the next four quarters. Zed becomes the default for the top tier of independent senior developers. Antigravity settles into the Google-shop default and becomes a real threat at the enterprise level once Google finishes pricing it. Cline’s install base grows faster than any paid tool. GitHub Copilot, already the quiet giant, grows by doing nothing.
And the open-source community now has its first high-profile case study of a hyperscale commercial user blowing past a permissive-license attribution clause without consequence. That will change how the next Modified MIT is written. Expect future Chinese open-weight releases to have teeth in the attribution clause — automatic API-level rate limiting for unattributed use, cryptographic signing requirements, license-enforcement mechanisms baked into the model card. Moonshot got played by Fireworks and Cursor. The next lab won’t.
The bigger lesson is the one Cursor’s VP of Developer Education admitted in his tweet: “It was a miss to not mention the Kimi base in our blog from the start.” That one sentence is the entire story. A company valued at $29 billion had a decision point — tell the truth about the base model, or obscure it — and chose to obscure it because the valuation story required it. Five weeks later, a trillion-dollar acquirer offered $60 billion for the result, and the market cleared at a price that suggests nobody cared about the lie except the developers.
Which is, when you think about it, the actual Composer 2 launch announcement they should have written. Not the one about frontier intelligence. The one about what the market is actually willing to pay for.
— Gal