14 September 2026
Heard In AI

OpenAI's Cursor cutoff and two theories about what it is really for

OpenAI has proposed ending the agreement that supplies its models to Cursor, now owned by SpaceX, on 12 November. On the Moonshots panel, one guest read the move as OpenAI betting on its own enterprise stack; another argued the real prize is reasoning traces — the working a model shows while solving a problem. Both explanations lead to the same awkward conclusion: Elon Musk and Anthropic now need each other.

A briefing reports one development when it happens. We correct or clarify it later; a new development gets a new briefing. How our formats work

"This week, the drama between Sam Altman and Elon Musk bubbled up once again," Peter Diamandis told his panel on Moonshots, before reading out the line that had set off the week: OpenAI would stop supplying its models to Cursor.

Cursor is a coding environment — the software a developer works in — that calls on large AI models to write, explain and change code. For years it leaned heavily on OpenAI's models, among others. What changed is ownership. Cursor announced on 14 August that SpaceX had completed its acquisition, presenting the deal as a route to the computing capacity needed for stronger and cheaper models and extending a model-training partnership struck in April. Diamandis put the purchase price at $60 billion, a figure he gave on the podcast rather than one drawn from the companies' announcements.

Two weeks later, OpenAI said it intended to wind down the agreement, proposing 12 November as the date access ends. The company pointed to a limited cancellation window that opens when a customer changes hands, and to what it described as previous contract violations by Musk's companies. The published decision is narrower than "we're shutting it down": OpenAI said it would keep existing access running for the maximum notice period its contract requires, while withholding models it has not yet released, naming Astra among them.

The quotation Diamandis read from the announcement was blunter about motive. OpenAI, he said, wrote that it was making the choice because "we cannot be confident that SpaceX will use our technology within the terms of service," citing its experience with Musk's companies. Musk's public reply was that he couldn't care less, and that OpenAI's founders were untrustworthy people who had, in his words, stolen an open-source non-profit.

Within hours, Diamandis said, Anthropic publicly offered Cursor whatever support it needed from Claude, its model family. The framing that followed on X — repeated by the panel as it circulated — was that Altman was now fighting alone against a strategic alliance of Musk and Anthropic's Dario Amodei.

The first theory: OpenAI now has a stack to sell

Dave, asked first, did not think the timing was accidental. A strong new OpenAI model, which he called Sol, had shipped shortly before, and Codex, OpenAI's coding agent, had become genuinely good. A year earlier, he argued, cutting off a major distribution channel would have been dangerous; now OpenAI had a competitive product of its own to put in the gap.

His case was about corporate customers rather than individual developers. Codex, running on Sol, hosted through Amazon Bedrock — Amazon's service for renting access to AI models inside a company's own cloud account — was, in his description, "phenomenally good for enterprises" as of the previous month or so. He said he uses both that combination and the Cursor-plus-Anthropic route side by side on his laptop every day, consuming large volumes of tokens in each. Against a single vertically integrated stack, he called the Cursor-and-Anthropic arrangement a "Rube Goldberg machine": a contraption assembled from parts that were never designed to fit together.

Dave also thought Anthropic had a commercial weakness that OpenAI could exploit. In his account, using Anthropic's models through Amazon's cloud still sends a customer's prompts and outputs to Anthropic for a 30-day review period, which he said the company justifies on safety grounds and which corporate buyers dislike. That description of Anthropic's data handling was his claim on the podcast, not a policy either company set out in the episode's source documents. He extended it into a character argument: Amodei is "a little bit trapped in his ethics," having hired a workforce that worries about AI escaping containment, while Musk's post implied that OpenAI's leaders carry no such hang-ups.

Altman, Dave noted, was late to turn from consumer products toward corporate revenue. His read on the cutoff was that OpenAI now has the whole stack lined up and has decided to run with it — a move he compared to Microsoft in the Bill Gates era, pairing its operating system with its own Word and Excel.

The second theory: it is about the traces

Alex offered what he called an alternative theory of the case. "It's always about who gets the reasoning traces," he said.

A reasoning trace is the intermediate working a model produces on the way to an answer, rather than the answer alone. Those step-by-step records are valuable training material: they can be used to teach another model how a strong system approaches a problem.

Alex cited allegations that Chinese frontier labs have used proxies to siphon frontier reasoning traces from Anthropic, and a precedent from the year before: Anthropic cutting off Windsurf's access after Google DeepMind hired away that company's team — probably, he said, to protect the traces. In his telling, the same logic now runs in reverse. Cursor sits between users and several frontier models, so whoever owns Cursor sees how those models reason. He believes gaining that dataset was close to the entire technical justification for SpaceX buying the company, and that OpenAI is concerned its users' traces would flow through Cursor to SpaceX.

That remains his interpretation. OpenAI's published reason is contractual confidence, and Cursor's own account of the acquisition emphasizes computing capacity rather than trace collection.

Why Musk and Amodei may be stuck with each other

The panel then asked how long an alliance between two such people could last. The answer that emerged was about dependencies rather than affection.

Anthropic, on the panel's account, has been buying compute from Musk's Colossus data center in Tennessee — paying through the nose for it, and exposed to losing it at short notice. But with OpenAI's models leaving Cursor, the practical choices left inside the product narrow to Anthropic and Grok; Chinese models are unattractive to many corporate users, and the panel said Gemini does not work well there. That gives Anthropic something to hold against Musk's control of computing capacity. SpaceX, for its part, was described as needing the revenue that comes from becoming a hyperscaler — a large-scale cloud provider — with anchor tenants, ahead of an IPO. "I think the outcome is pretty overdetermined at this point," one of the guests said, "but I think this ends with essentially everyone getting their own Dyson swarm."

On personalities, the panel's view was that two large egos with opposing politics will not be friends in two years. On dependency, the view was that the arrangement might hold anyway.

Diamandis pushed on the obvious escape route: could Grok become a strong enough coding platform that Cursor drops Anthropic for it? Alex called Grok "such a mushy concept at this point," saying that from the outside "today's grok seems like yesterday's cursor," and that yesterday's Cursor looked to him like a Chinese open-weight model post-trained on Claude's reasoning traces. He thought a simpler path existed: Musk could strike a deal with Anthropic, which now needs him for data-center capacity, white-label its model and call the result Grok 10.

Dave doubted Musk would build his way out. Frontier model work, as he described it, comes from small, tight-knit groups of a handful of brilliant researchers — Anthropic's DNA — while Musk's is enormous infrastructure buildouts like Tesla, SpaceX and Colossus. A panelist warned against betting against Musk all the same, recalling conversations in which Musk said he was dependent on nobody, and pointing ahead in the episode to Musk's response to shortages of gas turbines and solar supply: build them himself.

That left one scenario the panel returned to at the end. If Musk grows large enough, a guest suggested, he could simply buy Anthropic — a trillion or two trillion dollars — and fold it into the empire. Board members and investors would like that; Amodei would not. Which is why, the panel concluded, the unresolved question of supervoting control at Anthropic, shares that would let its founder outvote outside owners, is the pivot point on whether any of it can happen.

Share this article

Go to the original

Sources & further reading

  1. 01
  2. 02

Connected ideas and articles

From the conversation

Podcast episodes

Article history

Updates to this article

Tags

Altman says economic inertia slowed AI's impact. His podcast panel disputes the cause

On Moonshots with Peter Diamandis, the panel watched Sam Altman explain that he expected GPT-4 to put software businesses up for grabs far sooner than it did, and that the economy's inertia has made the transition "smoother and slower." Salim Ismail blamed institutions that move at a different speed from the technology, Alex pointed instead at the abstraction layers of the economy and prescribed vertical integration, and Emad Mostaque objected that the models simply were not good enough until recently.

6 min read

Why Graylin says distillation cannot explain all of China’s AI gains

Asked about allegations that Chinese labs extracted capabilities from Claude, Alvin Graylin argued that access to another model’s answers cannot explain every engineering advance. The Moonshots exchange turned on three distinctions: legitimate distillation versus prohibited extraction, query bills versus development costs, and learning from outputs versus improving the machinery behind them.

6 min read

Why a Moonshots panel thinks China's AI tokens go to video and America's to code

Alibaba's Wan 3.0 and a relayed claim that 70% of Chinese AI token use goes to video sent the Moonshots panel into an argument about money: one guest said American labs chase revenue per token while Chinese labs give their weights away, another said video is the only market that will trust a Chinese model. They ended up disagreeing about whether world models or text models reach self-improving AI first.

6 min read

Ed Zitron's 2027 forecast: OpenAI runs out of cash, and the losses spread

On The Diary of a CEO, critic Ed Zitron laid out a sequence he expects to start with OpenAI failing to raise its next round and end in ordinary retirement accounts. He traces the chain from a delayed stock-market listing to SoftBank's paper holdings, cloud growth forecasts and the concentrated US indexes — while Amazon's own filings and Andy Jassy's shareholder letter offer a different account of why the spending is happening.

8 min read