Elon's Secret Weapon: Cursor
Essay: How Elon Can Beat OpenAI Outside the Courtroom
It’s been a rough couple of months for Elon Musk.
His AI lab xAI has fallen behind Anthropic and OpenAI, demand for Grok has fallen, and to top things off… he just lost his court case against his arch nemesis Sam Altman.
But a recent pivot has presented Elon with a unique opportunity to catch up and actually overtake his rivals.
The secret? Unexpectedly… it’s Cursor.
Cursor was supposed to be the dumbest business in AI: a thin wrapper over everyone else’s models. For 2+ years the concern was that Cursor would die as soon as OpenAI/Anthropic ship their own coding tools. But the opposite happened:
Cursor went from $100M → $2B ARR in 14 months. SpaceX just offered $60B to acquire them.
How? Because Cursor had secretly built a harness every major lab needed: the agent harness. And Elon’s about to use it to beat OpenAI.
Cursor Just Became a Frontier AI Lab
Cursor unexpectedly dropped a new AI model on Monday. Composer 2.5 not only as good as Opus 4.7 and GPT 5.5 at coding… it’s also a lot cheaper.
Composer matches Opus 4.7 on CursorBench 3.1 and hits 79.8% on SWE-Bench Multilingual. It runs at $0.50 per million input tokens, $2.50 per million output which is about 10X cheaper than anthropic.
This is obviously amazing, but it’s confusing how they were able to pull this off at all. The answer is a recent deal with Elon.
SpaceX merged with xAI in February. Elon needs a big AI portfolio ahead of his summer IPO. Cursor needs more compute than it can afford… so they teamed up.
Cursor now has access to:
1 million H100 equivalents on Colossus 2 - more than Anthropic and Meta combined
A direct line to Grok’s foundation models
Nvidia preferential GPU allocation (Nvidia is a Series D investor)
A $10B floor payment, even if SpaceX walks away.
This means one important thing:
Cursor and xAI now have a shot at training a frontier coding AI model that beats anthropic and OpenAI.
The secret to pulling it off is cursor’s agent harness.
The Agent Harness is the Moat
A University of Chicago study found that teams using Cursor merge 39% more pull requests. That’s unexpected, given Cursor is just a wrapper around Claude. The advantage comes from the layer between the developer and the model.
That layer is the agent harness: the project context, the multi-file orchestration, the agent loop, the tool calls, the IDE integration, basically the muscle memory built into the workflow. A foundation model doesn’t carry any of it. Cursor built the #1 version of this themselves.
The harness orchestrates a user’s prompt to ensure it gives the best answer. It ensures understanding of the user’s context and works across the user’s entire work platform.
I’ll give you an example of why this is very advantageous:
GitHub Copilot has 20 million users. Cursor has roughly 2 million. But Cursor generates 4x more revenue. Why? Because Copilot is a plugin and Cursor knows what to do with the plugin.
This is a trend a lot of people haven’t noticed. Foundation models are commoditizing, and the real moat will end up being the intents-layer.
The numbers don’t lie btw
Cursor’s revenue trajectory (ARR) is frankly insane:
Jan 2025: $100M
June 2025: $500M
November 2025: $1B
February 2026: $2B
Slack took 5 years to hit its first billion. Zoom took 9. Cursor did the first billion in three years and the second in 3 months.
Then you look at their customer base… 70% of Fortune 1,000 uses Cursor and they have 1 million paying users.
There is a reason SpaceX valued the company at $60B.
I rest my case: the wrapper won
So there it is. As a former Cursor hater who never understood why they had value beyond being a wrapper, I can safely say I was very wrong.
Sam even confirmed it himself lately, saying “the harness and model are the same layer.”
Does this give xAI and Cursor a legitimate shot at joining Anthropic and OpenAI at the table? Uhh, yes.
Thanks for joining us for our latest issue. Now go listen to our podcast :)
Great article and podcast! There is a wrapper wave (wrap-battle) that’s going to be interesting with hundreds of wrappers.
Cursor’s ARR is $2B, but with near-zero margins they are burning investors’ money to basically pay Anthropic and OpenAI for their models. ARR has become the best example of fake growth metrics.
And Composer 2 is just a fork of Kimi K2.5 that is trained to work great on benchmarks but doesn’t even compare to Opus 4.7 and GPT-5.5 in a real scenario