Meta declined to provide compute to Anthropic, WSJ reports, after Amodei asked Wang directly; about $10 billion in talks may be over

Meta declined to provide compute to Anthropic, the Wall Street Journal (WSJ) reported. According to the report, Anthropic CEO Dario Amodei asked Meta's Alexandr Wang directly this year to share compute with Anthropic. Meta reportedly said no because it needs the capacity for its own Muse. The two companies had reportedly been in talks worth about $10 billion, and the report says those talks may now be over. We have not been able to read the WSJ article itself. This account relies on summaries from an AI Weekly alert and Techmeme (October 10). We have also not been able to confirm whether Meta or Anthropic has acknowledged the report. Everything below is based on press reports.

What was reported: the CEO asked himself, and Meta said it needed the capacity for Muse

The report makes three main points. First, the request came from Amodei himself, and he made it to Meta's Wang. This was a direct approach between top executives, not a discussion between procurement teams. Second, Meta said it was turning Anthropic down because Muse needs the compute capacity. Third, the talks were reportedly worth about $10 billion.

That figure should be treated with caution. Reported values of AI deals have sometimes been far from the amounts later disclosed: one deal reported at $250 billion turned out to be $105 billion in an 8-K filing. The figure of about $10 billion should be treated as the WSJ's reporting, not as an established number. The summaries we have also don't say how far the talks had gone, such as whether the two sides were negotiating terms or had only had an initial approach.

Meta isn't a lender of compute: Muse is using the capacity

Meta keeps putting huge sums into its own data centers. From the outside, that made it look like a company that could lease spare compute to other labs. This report suggests Meta does not intend to give up that capacity.

Muse, the reason Meta reportedly gave, is the agent product Meta put at the center of its Connect keynote on September 23. Meta added a voice mode, computer use on the Mac, a dedicated email address and integration with Meta's glasses. Muse connects to services including Walmart, Best Buy, Sephora, Shop Pay, PayPal, GitHub and Notion. Meta says it plans to keep Muse free to use and to charge a small fee on transactions. Muse Realtime Avatar, announced at the same event, can handle only 12 concurrent sessions on a single GB200. Because Muse is free and earns money on each transaction, it needs more inference capacity the more people use it. It would not be surprising if Meta concluded it has no room to lease capacity to others.

Meta is also competing with Anthropic in the same agent market. Muse is marketed on computer use and app integrations, so it competes with Claude for enterprise deployments and customers. Leasing compute to a competitor would mean running that competitor's products on Meta's own infrastructure. If the report is accurate, Meta has decided its power and GPUs are a competitive advantage to keep, not a product to sell.

Anthropic ahead of its IPO: more promises that use a lot of compute

Anthropic is preparing to go public. Its own announcements from the past few weeks show a growing list of promises that will use a lot of compute.

Claude Haiku 5.5, released October 7, costs $0.10 per million input tokens and $0.50 per million output tokens for prompts of up to 100,000 tokens, and Anthropic says it is about 75% cheaper than Haiku 4.5. Artificial Analysis measured it producing about 162,000 output tokens per task. It is a cheap model that uses a lot of tokens. In Anthropic Cyber Mission on October 8, the company promised to scan open-source projects for free on request, using its most capable models, including Claude Mythos. On the same day, it announced $150 million in Claude and API credits over three years for the federal government's Genesis Mission. Claude Frontier Academy, announced October 2, plans to train 10,000 FDEs by the end of 2027, and the more enterprise deployments grow, the more inference demand rises.

All of these assume Anthropic can get the compute later. When the IPO prospectus comes out, one of the first things investors will check is where Anthropic rents the capacity to meet this demand, and on what terms. A report that a possible major supplier turned it down makes that harder to explain.

Fewer suppliers: every company with GPUs becomes a competitor

This case can be read as one example of a larger trend. Many companies with large amounts of compute now build their own frontier models or agents. Meta has Muse, SpaceX calls itself SpaceXSI and is marketing Grok Bot, and Google has released Gemini 4 Argon. Companies that used to sell compute are, one after another, competing with the companies that would buy it.

There are exceptions. Musk wrote on X that SpaceX's Grok Bot will also send some tasks to Claude Opus 5.5, depending on the task (a secondhand report). Reflection AI has reportedly arranged to get compute from Nebius and SpaceX. Using a competitor's model is a different decision from leasing infrastructure to a competitor, but not every relationship rules out deals entirely. Even so, with Meta reportedly staying out, Anthropic appears to have one less supplier it can negotiate terms with.

What we still don't know

We don't yet have enough information to answer the headline's question of whether the talks are over. There is only one WSJ report, and we have not been able to find comments from Meta or Anthropic. The timing of Amodei's request has been reported only as "this year." We don't know how many years the roughly $10 billion would cover, or what kind of deal it refers to, such as GPU time or space in a data center. The reports also don't say whether Meta might reconsider once demand for Muse levels off.

What is clear is that the idea that compute owners are choosing to use it for their own agents has now been reported as a specific negotiation. The next question is how Anthropic will explain its other sources of compute before the IPO. The prospectus and any future large deals will be the next things to watch.