Expand menu

On August 29, OpenAI published a note with a bureaucratic title: "Our decision on Cursor following its acquisition by SpaceX." The content is anything but bureaucratic. OpenAI is terminating the contract that gave Cursor access to its models. Cutoff date: November 12, 2026. Seventy-five days' notice.

The reason OpenAI gives has nothing to do with what Cursor does. It has to do with who bought Cursor. In its words, the company "cannot be confident that SpaceX will use our technology within our terms of service, based on our experience with Elon Musk's companies violating contracts." It cites two precedents: the contract breach that followed the Twitter acquisition, and Musk's admission under oath that xAI violated OpenAI's terms on data handling. It also notes that it needs strict compliance oversight ahead of its next model, Astra.

The context: SpaceX agreed to buy Anysphere, Cursor's parent, for sixty billion dollars. The deal was announced on June 16 and closed in August. It is the largest startup exit on record.

Cursor breached nothing. Cursor got bought.

And there is the fact worth looking at once you strip away the noise of the Musk-Altman feud: a company lost access to a critical input because of a corporate transaction it was not party to and had no vote on.

The number almost nobody is looking at

Michael Truell, Cursor's co-founder, answered publicly with a single figure: OpenAI's models account for roughly 5% of the platform's traffic. The Information reported it the same day.

That 5% is the important part of this story, and not because it is small.

Think about what had to happen for Truell to be able to say that number on the day of the announcement. He did not calculate it that Saturday afternoon. He already had it. At some earlier point, and with nobody forcing them to, Cursor built two things: the instrumentation to know exactly what share of the product runs through each model provider, and the architecture to move that traffic elsewhere. Which is why a cutoff that would have been an existential crisis at another company is, here, a migration with a calendar.

The naive read is "Cursor got lucky." Luck had nothing to do with it. What Cursor had was the discipline to treat the model as an interchangeable input rather than a foundation. That decision gets made years before the day you need it, and you pay for it in complexity every day in between.

Why this changes the risk calculation

Two years ago, the risk analysis for a model provider had three reasonable lines: they raise the price, they degrade the service, they go out of business. All three are covered by contract or by provision.

What happened this week adds a fourth line that no contract covers: your provider decides it does not like your owner. Or your investor. Or your partner. The cutoff was neither technical nor commercial. It was corporate governance, and it happened one floor up, between two companies that are not yours.

Add the macro moment. That same weekend, Kevin Warsh used his Jackson Hole keynote to say that PCE inflation is running at 3.7% over twelve months and that the Fed's focus should be prices; the market moved the odds of a rate hike from 35% to 57% in an afternoon. In that same speech, the Fed chair noted that annualized token sales at the leading labs exceed one hundred billion dollars, more than 500% above a year ago. Translated: the input we all depend on is growing fast enough to be a macroeconomic variable, and the money financing that race may get more expensive. Model providers are going to have less patience with customers who cost them margin, and more incentive to decide quickly who they work with.

Three questions you can answer this week

At Indrox we work with companies that are putting models inside processes that can no longer be switched off. These are the three questions I am left with after reading OpenAI's note, and none of them needs a committee:

One. What share of your product runs through each provider today? If the answer is not a number you can say out loud this week, the problem is not solved — it is hidden. Truell could say 5% because somebody built the dashboard before it was needed.

Two. How many days does it take you to switch? Not in theory. In a production client, with its evals, its prompts tuned to one specific model, its contractually committed latency and its trained people. If the number you get is greater than seventy-five days, the notice Cursor received would not have been enough for you.

Three. What did you sign about termination? Not the availability SLA. The termination-for-convenience clause, the notice period, and whether there is any transition obligation at all. Almost every API contract I have read leaves that door open on the provider's side.

My read

The industry spent two years selling the idea that the model is a utility, like electricity. It is not. Electricity does not pick its customers by who their majority shareholder is.

A frontier model today is a concentrated provider with a dominant position, in a market where three or four players also carry personal rivalries and cross-litigation. That is not a utility: it is a commercial relationship with political risk. Treating it like a wall socket is the most expensive architecture decision a team can make right now, and the bill does not arrive on the day you make it.

I am not saying you should avoid dependency. That is impossible, and it would be bad for the product. I am saying dependency gets managed: you measure it, you spread it, and you give it an exit plan with a date. Cursor did that, which is why this week it is uncomfortable instead of dead.

The question I leave with anyone running a technology company is not whether this can happen to them. It is how many days of notice they would need to survive it.