Musk: only three organizations have enough coding data
The chip suppliers and the biggest buyer of chips are suddenly telling the same story: capability is no longer the bottleneck — money is. Elon Musk spelled out what he thinks the scarcest training input actually is, and the chiefs of Intel and Broadcom spent a day in Singapore arguing that access to capital may decide who survives the next phase.
Elon Musk says only three organizations in the world have accumulated enough coding training data — and xAI was not one of them. Appearing by video link at X Takeover 2026 in Austin on October 10, Musk was asked why SpaceX paid roughly $60 billion for Cursor, and his answer was a blunt admission: to build competitive superintelligence you first have to crack coding, and by his knowledge only OpenAI, Anthropic and Cursor have stockpiled enough programming training data to do it — with Google, in his telling, "barely half" a member of the club. The list pointedly contains no xAI and no Grok; the acquisition that closed on August 14 was, on this telling, Musk buying the missing ingredient rather than waiting for synthetic data to catch up. If the world's most aggressive builder of AI compute thinks the scarce resource is a private corpus of real code nobody else can replicate, that reframes the Cursor deal from a product acquisition into a data land-grab — and it suggests the next frontier fight is over who owns training data nobody can legally or practically scrape twice.
Intel and Broadcom chiefs say capital — not chips — is now the AI industry's biggest constraint. Speaking at a financial conference in Singapore on October 12, Broadcom's Hock Tan said access to capital "might be the biggest constraint" facing the sector, predicted the data-center chip shortage would not ease until 2028 or 2029, and noted that much of the AI labs' enterprise revenue still comes from pilot projects — customers who, in his arcade analogy, spend their ten dollars of tokens in ten minutes and learn nothing. Intel's Lip-Bu Tan agreed that "it's capital and the cost of capital" plus the long-run ROI question, while adding that Chinese open-source models are proving "good enough" for budget-conscious users. Both men run companies that sell the picks and shovels, so a capital-constraint warning doubles as a demand warning — and it lands the same week Broadcom helped OpenAI finance its custom chip buildout, which is exactly the kind of spending Tan is implicitly questioning.
Anthropic users suspect a stronger, unannounced Claude model is already running in production — the community is calling it "Claude Next." Over the weekend, testers circulating on Chinese and English AI channels reported outsized quality from existing Claude labels: one said a four-line prompt ran two hours and burned 27 million tokens on a Claude Max plan to produce a rendered animation, others said Fable 5.1 requests began returning noticeably smarter answers that handle recent events without invoking web search — a probe some are using to flag a suspected new checkpoint. None of this is confirmed: "Claude Next" is a community nickname with no API identifier or model card behind it, Anthropic has not commented, and the viral "3D movie" demos are code-rendered animation rather than a video model. What keeps it on the radar is timing and precedent — Anthropic holds a pre-IPO investor day on October 14, and a gray release during peak attention would not be the first time a lab quietly improved a model behind an existing label.
What to watch: whether Anthropic names anything at tomorrow's investor day, and whether any of the "Claude Next" testers can produce a verifiable API string.
If coding data is the real moat, is every lab without a proprietary code corpus quietly out of the race? Tell us in the comments.
