The fastest iteration won. And still had to sell.
In short: Cursor's investors tell its story as a victory of iteration speed over both intelligence and resources. That holds, but not entirely: the company that iterated fastest of all ended up selling to the one that had the compute. What transfers is not advice about speed but three mechanisms — proprietary data from distribution, treating model and product as one thing, and the shift of the bottleneck from writing code to vouching for it. The last of those reaches companies that will never train a model.
a16z published a piece — signed by Matt Bornstein and Martin Casado — on how Cursor ended up inside SpaceX. The argument is clean and fits in one sentence: the team that iterates fastest wins. Not the smartest, because everyone is smart, and not the one with the most resources, because resources can be caught up with.
Most of that argument is right. It just ends in a way that complicates it — and the complication is the interesting part.
What happened
Briefly, and with the caveat that all of it comes from one text: Cursor started as a product, not as a lab. Four people around Michael Truell first tried an email client and a CAD tool, then built the editor they wanted to use themselves. By 2024 it was the fastest software company in history to $100 million in annual recurring revenue.
Then the labs shipped agents — Claude Code, Codex — and with the end of 2025 came a generation of models that made the whole editor format look dated. Cursor answered with models of its own. Composer in October 2025, Composer 1.5 in February 2026, Composer 2 five weeks later, Composer 2.5 eight weeks after that. The last two built on an open base model, post-trained on proprietary data.
And then, when it became clear that the next step was not possible without compute, SpaceX bought Cursor. For $60 billion in stock, four days after going public.
Two readings of the same ending
Both readings hold up, and the first one is told by an investor with money in the company. That is not an accusation — it is context a reader should have before taking the conclusion away. The piece is written by people who describe the founders as "some of the most special founders we've ever met", and they are right and interested at the same time.
Our reading is a third and duller one: iteration speed gets you to the door of the frontier; compute is what gets you through it. Cursor did everything that can be done with cleverness and pace, and still had to solve one axis by acquiring a partner for it. If "iterate faster" were the whole rule, it would not have had to.
What actually transfers
"Iterate faster" is useless advice because everyone gives it and nobody says what on. Three specific mechanisms from the story do transfer, though — including to a company that will never train a model.
Data from your own distribution is an asset nobody can buy. Cursor had no compute, but it had something the labs did not: it could see how real code gets written, by developers all over the world. On that pair — an open base plus proprietary data — it built a model that came within sight of the frontier. The same reasoning holds two orders of magnitude further down: the cheapest route to a model that works better for you than a general one almost always runs through data only you have, not through a bigger model.
Model and product are worth building as one thing. Cursor's COO describes it as "not viewing the model and the product as separate things". In practice that means the shape of the product decides what the model learns, and the model's capabilities decide what the product looks like. Commission a model separately and bolt it onto a finished process, and both come out worse.
And above all: the bottleneck moved, it did not disappear. This is the part that reaches everyone.
That is exactly what our breakdown of agent cost is about: the fifth and sixth lines of the bill — the people who process the output and the harness that makes it trustworthy — are usually larger than tokens and appear on no pricing page. Cursor's answer was a set of "outer loop" products: review, integration, tests, shipping. It is the same thing from the other side. When production gets cheap, verification gets expensive.
What does not transfer
The starting conditions. Cursor had a product people loved, enough money for four model generations in eight months, and founders who could tell the company in January 2026 that meetings were cancelled and the hours would be brutal. Most companies have none of that, and copying the pace without the conditions only means burning out faster.
The quoted productivity numbers. Roughly 5 to 10 per cent from autocomplete, around 35 per cent from agents, about 65 per cent of production code from AI at enterprise customers. A nine-month migration in nine weeks. A $1.2 million, six-month project finished in a three-day hackathon. All of it may be true — and all of it comes from the vendor of the product being measured, published by its investor. These are not independent measurements and we will not present them as such.
The number that says something about your company is a baseline you record before deployment. Without it a saving cannot be shown, only claimed — and claiming it is what every supplier you let through the door will do.
An honest note on sources
Everything factual in this piece — dates, amounts, names, model versions and quotations — comes from a single source: the a16z essay on Cursor and SpaceXAI. We did not check it against press releases, IPO filings or financial statements, because that was not possible while writing. The source also has a financial interest in the company it describes.
There is deliberately no link to that piece here: it reached us as text without an address, and a guessed URL is worse than none.
We are saying so because the rule this site runs on is to trace every figure to the original publication, not to a piece summarising it. This article does not meet that rule, and writing that down is more honest than leaving it out. The argument we build on top does not rest on the exact amounts: whether the acquisition was sixty billion or forty, the shape of the story is the same.
Frequently asked questions
Why did SpaceX buy Cursor?
According to the a16z essay, it was a trade in which each side held what the other lacked: Cursor could squeeze a competitive coding model out of very little compute, and SpaceXAI had the data centres but could not close the gap on coding. The transaction is reported as a $60 billion all-stock acquisition. Those figures come from a single source and are not verified against a primary document in this piece.
Does this mean iteration speed decides everything?
It decides whether you reach the door, not whether you get through it. Cursor iterated faster than anyone and still hit a limit that speed could not route around — it needed compute it could not buy. And 'iterate faster' is advice everyone gives, so it carries no information; the useful question is what exactly you iterate on and where the feedback comes from.
What can an ordinary company take from the Cursor story?
Three things. That data from your own distribution is an asset nobody can buy. That model and product are worth building as one thing rather than two. And above all that when writing code gets cheap, the bottleneck moves to vouching for it — at one company in the cited text, volume went from roughly 150,000 to 800,000 lines a week and broke the review process.
Are the productivity numbers real?
The figures quoted — roughly 5 to 10 per cent from autocomplete, around 35 per cent from agents, about 65 per cent of production code from AI at enterprise customers — come from the vendor of the product being measured, published by its investor. They are not independent measurements. A baseline you record in your own company before deployment will tell you more than any of them.
About the author
Filip has spent his career between media and technology — filming for ŠKODA, Pilsner Urquell and Range Rover, then co-founding a startup in San Francisco, working with a YC-backed company and serving as CIO at Renato. He built ZEUS Legal AI and now runs the systems Salamon Capital operates on.
New analysis, when there is one
One message when a new analysis or article goes out. Nothing else — no sequence, no digest, no third party. The address is stored so there is somewhere to send it, and one click deletes it for good.