Cursor Drops Composer 3, a 1.5T-Parameter Model Trained on SpaceX's Supercomputer

At its inaugural Compile conference in San Francisco, Cursor came out swinging with three announcements that collectively reframe what the company is building. The headline-grabber was already known to insiders: SpaceX had exercised its option to acquire Cursor (officially Anysphere Inc.) in a $60 billion all-stock deal. But the product announcements made clear that Cursor isn't waiting around to be absorbed. It's building infrastructure.

The Three Announcements

The Compile keynote surfaced three distinct launches, each targeting a different layer of the development stack:

• Origin -- a Git hosting platform built from the ground up for AI agents, positioned as a direct GitHub competitor
• Composer 3 -- a new frontier model being trained from scratch on SpaceX's Colossus supercomputer, with no third-party base model underneath it
• Cursor Mobile -- an iOS beta that lets developers prompt agents, edit code, and remotely control desktop Cursor sessions from their phone

Origin: When GitHub Wasn't Built for This

At the conference, Cursor announced Origin, a code storage and Git hosting platform fundamentally redesigned to treat AI agents as first-class citizens. Built on re-architected technology from its December 2025 acquisition of Graphite, Origin is purpose-built for the high-frequency cloning, pushing, and parallel branch merge conflicts that characterize AI agent workflows.

Performance metrics include 296,000 clones per hour and sub-400-millisecond global synchronization latency. On the main screen at Compile, Cursor displayed roughly 81,000 pushes and 22.6 commits per second on a single repository, with automatic failover completed within 10 milliseconds. The key insight here is that when agents are generating commits at machine speed, the bottleneck shifts from the developer to the infrastructure. GitHub was built for humans typing. Origin is built for fleets.

Automated merge conflict resolution is Origin's most radical differentiating feature. In traditional Git workflows, when two developers edit adjacent sections of the same file and create a conflict, human intervention is required to decide whose changes to keep. With agents running in parallel across dozens of branches, that human-in-the-loop model breaks down entirely. Origin is expected to officially launch in fall 2026, with a waitlist now open. Pricing and detailed enterprise features have not yet been disclosed.

Composer 3: The First Model Cursor Trained Alone

The model announcement is where things get technically significant. At Compile, Cursor announced a new model called Composer 3, described as being in the same size class as Claude Opus and GPT-5.5, trained from scratch with no Kimi base model, and built with 10-20x more compute than Composer 1. That last part is a meaningful departure.

Composer 2, Cursor's previous in-house model, was built on top of Moonshot AI's Kimi K2.5 -- a Chinese open-weights model. Cursor ran continued pre-training on a code-heavy data mix, followed by large-scale reinforcement learning in realistic Cursor sessions. About 75% of total compute came from Cursor's own training, not the base model. Composer 3 skips that foundation entirely.

The model is 1.5T+ parameters, retrained on 100k+ GPUs, and aimed at intelligence beyond coding. That last phrase is worth flagging: "beyond coding" suggests Cursor is positioning Composer 3 as a general-purpose frontier model, not just a code-specialized tool. For the past few months, SpaceXAI has been jointly training this model with Cursor, and it will be released in Cursor and Grok Build.

The compute backing this is substantial. SpaceX gave Cursor access to the GPUs operated by its xAI unit, with capacity equivalent to the combined processing power of 1 million H100 chips. That's the Colossus supercluster in Memphis -- the same infrastructure that trains Grok.

The Deal Behind the Demos

The product launches didn't happen in a vacuum. SpaceX exercised an option to buy Anysphere in an all-stock transaction widely described as the largest acquisition of a venture-backed startup ever recorded. The financing mechanics are what make this story unusual.

SpaceX debuted on Nasdaq at $135 per share, raising $75 billion in the largest IPO in history. Four trading days later, the stock closed at $192.46 on the day the Cursor deal was announced, adding roughly $740 billion in market capitalization. That run-up is what made an all-stock deal painless.

At $60 billion, SpaceX is paying about 15 times revenue -- one of the largest multiples ever paid for an AI software company. Yet Cursor's market share per Ramp corporate spending data fell from roughly 41% in June 2025 to about 26% by May 2026. Growing revenue and losing share simultaneously is the paradox at the center of this deal. Competition from Claude Code and OpenAI's Codex has been real.

The partnership agreement gives SpaceX the option to buy Cursor for $60 billion by year's end. If SpaceX had opted against an acquisition, it would have paid $10 billion for the collaboration. They didn't opt out.

Why SpaceX Needed This

All 11 xAI co-founders had left by the end of March 2026, weakening xAI's independent AI capability. Acquiring Cursor backfills the engineering bench SpaceX could no longer assemble internally. The Cursor team brings not just a product with millions of users, but a model training operation that has been iterating faster than most dedicated AI labs.

SpaceX told investors that it sees a potential $2.4 trillion AI infrastructure business and a $22.7 trillion opportunity in enterprise applications. Cursor is the product that makes those numbers credible. For SpaceX, owning Cursor could mean faster internal software development across rockets, satellites, autonomy, and manufacturing.

According to Bloomberg, Cursor scrapped a planned $2 billion fundraising round on account of the deal with SpaceX. Cursor had raised $3.38B since its 2022 founding from investors including Thrive, a16z, the OpenAI Startup Fund, and Nvidia. Those investors are now holding SpaceX stock instead of Cursor equity.

What Closes, What's Coming

The acquisition is expected to close in Q3 2026. The jointly trained model will be released in Cursor and Grok Build soon, and SpaceX looks forward to working closely with the Cursor team to advance frontier AI capabilities. For users, Cursor's product roadmap continues: Cursor's 2026 releases added Background Agents, Cloud Agents, Composer 2.0, multi-workspace and parallel-branch agents, a vulnerability scanner, and Bugbot PR review. Always-on Automations run agents 24/7, triggered by Slack, Linear, GitHub, PagerDuty, or webhooks.

The bigger question isn't whether the deal closes -- it's what Cursor looks like inside SpaceX's orbit. A company that built its reputation on model-agnosticism (supporting Claude, GPT, Gemini, and its own Composer models) will now be owned by a company with a competing model line. Whether Cursor stays open or gradually tilts toward Grok is the thing to watch.