In March, Elon Musk acknowledged on X that xAI had to rebuild from the ground up. No one expected that just four months later, the company would not only complete its reconstruction but also catch up with the industry’s top players in the critical AI coding arena.<\/p>\n
In July, xAI released Grok 4.5, propelling its model capabilities back into the top tier. According to data from AI evaluation firm Artificial Analysis, Grok 4.5 achieved a coding agent score of 76, matching OpenAI’s GPT-5.5 Codex but at only half the cost. Meanwhile, Grok Build was open-sourced, Workflows began orchestrating hundreds of agents, and Build Mode enabled users to generate and publish websites directly within the chat interface. Grok also progressively integrated into mainstream development and productivity platforms including GitHub Copilot, Microsoft Office, and Google Workspace.<\/p>\n
A company that was tearing everything down just four months ago had suddenly found its rhythm. And the starting point of this turnaround traces back to a $60 billion acquisition.<\/p>\n
A Critical Reinforcement After the Founding Team Evaporated<\/p>\n
In March, xAI was mired in severe internal and external challenges. In the most closely watched AI coding market, Anthropic’s Claude Code and OpenAI’s Codex loomed like two mountains, leaving xAI struggling to breathe. More fatally, xAI’s founding team had almost completely departed\u2014aside from Musk, all 11 original founding members had left. Among them were top industry talents including Igor Babuschkin, who had previously worked at DeepMind and OpenAI, and Christian Szegedy, a core author of Google’s Inception neural network architecture.<\/p>\n
A model company less than three years old had lost nearly its entire founding technical team. Musk commented on the situation: “xAI’s first build failed. We have to start over.”<\/p>\n
The turning point came on March 12. Just as xAI’s founding team was approaching zero, two senior product engineering leads from AI coding tool Cursor\u2014Andrew Milich and Jason Ginsberg\u2014announced they were joining xAI and SpaceX, reporting directly to Musk.<\/p>\n
On April 19, SpaceX signed a compute power and joint development agreement with Cursor. On June 16, SpaceX further exercised its acquisition rights, planning to use approximately $60 billion worth of SpaceX stock to buy Cursor’s parent company, Anysphere. Since xAI had already been integrated into SpaceX’s structure, the acquisition was nominally SpaceX buying Cursor, but the collaboration’s focus consistently pointed toward Grok and AI models.<\/p>\n
Process Data: An Asset More Valuable Than Code Itself<\/p>\n
SpaceX stated in its announcement that Cursor possesses a mature coding product and professional developer users, while SpaceX provides the Colossus supercomputer. Cursor CEO Michael Truell also indicated that the primary bottleneck for Composer’s continued training was compute power, and that after connecting to Colossus, model training scale could continue to expand.<\/p>\n
Cursor gained compute support from xAI, but what xAI received from Cursor far exceeded expectations.<\/p>\n
According to the cooperation agreement disclosed by SpaceX, Cursor contributed not only personnel but also datasets, product documentation, workflows, prompts, technical specifications, and software code. Models jointly developed by both parties are co-owned and can be used by either for commercial products.<\/p>\n
The core value of these assets lies in their uniqueness. Public code repositories typically only show final results, whereas Cursor preserves the entire development process\u2014what requirements users proposed, which files the agent read, what commands it executed, why the first modification failed, which code developers accepted, and which operations they reverted. Training models with this process data enables AI to learn how to investigate problems, modify code, and handle errors in real projects.<\/p>\n
Furthermore, Cursor has a base of power users willing to entrust real projects to AI, rather than casual visitors dropping in to experiment. This means xAI gained not just static data but a developer ecosystem capable of continuously generating high-quality work feedback. xAI’s strongest resources previously were compute power and the X platform\u2014the former could train models, the latter provided information and traffic, but neither could reliably produce stable, high-quality work feedback. Cursor filled precisely this gap: real software projects, developer operation records, and a set of agent products already in use.<\/p>\n
The first outcome of this collaboration was Grok 4.5. The model was jointly trained by both parties, using trillions of tokens of Cursor data, encompassing interactions between developers and codebases, agents, and software tools. From signing the cooperation agreement to the release of Grok 4.5, less than three months elapsed.<\/p>\n
What the $50 Billion Premium Buys<\/p>\n
The initial agreement left SpaceX with two paths: if it abandoned the acquisition, it would need to pay Cursor a $1.5 billion termination fee and $8.5 billion in deferred service fees, totaling $10 billion; if it exercised the acquisition right, it would use approximately $60 billion worth of SpaceX stock to buy Cursor’s parent company, Anysphere. SpaceX chose the latter.<\/p>\n
The additional $50 billion clearly cannot be attributed solely to Grok 4.5. The $10 billion corresponds to collaboration results already produced; the other $50 billion buys Cursor’s team, products, developer relationships, and the data and feedback that can continue to be used for the next model.<\/p>\n
The reason this deal appears staggeringly priced is largely because the market is still pricing Cursor using the previous generation’s division-of-labor logic in the AI industry chain. Under the old model, OpenAI and Anthropic were responsible for training models, while Cursor connected to APIs and packaged models into products developers were willing to pay for.<\/p>\n
But the logic of AI competition has changed. The focus of large model competition is extending from pre-training to post-training. Pre-training can acquire knowledge from the internet, books, and public code; post-training requires tasks, processes, and outcomes. Agents have an even greater demand for this type of data\u2014what they need to learn is not “how to answer this question” but “how to get this thing done.”<\/p>\n
Cursor happens to preserve this process data. This data can be used for post-training and can also directly evaluate whether agents are genuinely improving. Model companies are beginning to move downstream into applications, while application companies are starting to move upstream to train models. Claude Code and Codex compete for developer entry points, while Cursor uses its own data to train Composer. What each party is competing for is not just subscription users but also the training materials continuously generated as users work.<\/p>\n
By the old method, Cursor sells an AI coding software; by the new competitive logic, it also produces raw materials for training the next generation of models. The $60 billion SpaceX paid for Cursor cannot be valued solely based on a software company’s revenue.<\/p>\n
What Cursor brings to xAI also includes a methodology for building agent products. Grok Build executes tasks, Workflows handles decomposition and allocation, Automations enables tasks to start automatically, Build Mode generates the final output, and Copilot and office suites deliver these capabilities into the places where users work every day. Previously, Grok primarily existed within X and chat interfaces, responding to user queries one at a time; now, Grok reads files, invokes tools, handles errors, generates outputs, and completes tasks.<\/p>\n
However, Grok 4.5 is merely the first report card from this collaboration. What comes next is whether Grok 4.6 can launch on schedule and continue improving in agent execution, tool invocation, cost, and error recovery. Model shelf life is getting shorter and shorter\u2014today’s number one may be forgotten in a few months. What is truly valuable is a company’s speed in obtaining feedback, correcting errors, and training the next generation of models.<\/p>\n","protected":false},"excerpt":{"rendered":"In March, Elon Musk acknowledged on X that xAI had to rebuild from the ground up. No one…\n","protected":false},"author":2,"featured_media":131757,"comment_status":"","ping_status":"","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[10],"tags":[53,16464,2798,19885,6541,140,47109,65752,157,658,2899],"class_list":["post-131756","post","type-post","status-publish","format-standard","has-post-thumbnail","category-xai","tag-anthropic","tag-anysphere","tag-claude-code","tag-colossus-supercomputer","tag-cursor","tag-elon-musk","tag-grok-4-5","tag-michael-truell","tag-openai","tag-spacex","tag-xai"],"_links":{"self":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/131756","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/users\/2"}],"replies":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/comments?post=131756"}],"version-history":[{"count":0,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/posts\/131756\/revisions"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media\/131757"}],"wp:attachment":[{"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/media?parent=131756"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/categories?post=131756"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.europesays.com\/ai\/wp-json\/wp\/v2\/tags?post=131756"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}