Colossus for coding models
- SpaceXAI and Cursor announced collaboration to train coding models on a Colossus supercomputer with massive GPU equivalence. (x.com) - Social coverage said the setup equals roughly one million Nvidia H100 GPUs and mentioned a potential $60 billion acquisition. (x.com) - Posts argue such scale could accelerate AI-written code capabilities and shift developer tooling economics if the project continues. (x.com)
SpaceX said on April 21 that it is teaming with Cursor to train coding models on xAI’s Colossus supercomputer. (techcrunch.com) The companies said the system combines Cursor’s software used by developers with Colossus, which SpaceX described as having compute equivalent to 1 million Nvidia H100 chips. SpaceX also said it can either buy Cursor for $60 billion later in 2026 or pay $10 billion to keep working together. (techcrunch.com) Coding models are artificial intelligence systems trained to read, write and edit software, and bigger training runs usually need far more chips, electricity and networking. xAI says Colossus was built in 122 days, expanded to 200,000 graphics processors, and has a roadmap to 1 million processors. (x.ai) Cursor had already been building its own models, including Composer, a software agent trained for coding tasks inside large codebases. Cursor said Composer was trained with reinforcement learning on real software engineering problems, and the company released Composer 2 in March 2026 with a technical report describing continued pretraining and large-scale reinforcement learning. (cursor.com 1) (cursor.com 2) The new deal lands in a market where AI coding tools are no longer side features inside chatbots. CNBC reported on February 24 that Cursor was updating agents that can test their own code changes and record work with videos, logs and screenshots as it competed with Anthropic, OpenAI and Microsoft. (cnbc.com) Cursor’s scale as a business helps explain the price tag. Bloomberg reported on March 2 that Cursor’s annualized revenue topped $2 billion in February, after doubling in three months, and TechCrunch reported that the company had been discussing a new round at a $50 billion valuation before the SpaceX agreement. (bloomberg.com) (techcrunch.com) The partnership also changes who controls the computing. TechCrunch reported that Cursor had been using models from Anthropic and OpenAI while trying to improve its own systems, and Decoder reported that Cursor said limited compute had held back training for its Composer models. (techcrunch.com) (the-decoder.com) That creates competing readings of the deal. SpaceX said the pairing of Cursor’s distribution with Colossus can produce “the world’s most useful models,” while InfoWorld reported that enterprise customers may ask harder questions about model neutrality and data contracts if a tool vendor and compute supplier are moving this close together. (pcmag.com) (infoworld.com) The immediate test is whether more chips turn Cursor’s in-house coding models into stronger products before SpaceX has to choose between a $60 billion acquisition and a $10 billion partnership payment later this year. (abcnews.com)
