Wall Street Journal sources say Dario Amodei personally reached out to Meta's Alexandr Wang earlier this year to ask for more computing power, and that Meta said no. That is the claim in a report summarized on October 9, 2026; the article itself sits behind the WSJ paywall, so this post stays with what can be checked and labels the rest. If you use Claude and wonder why limits keep tightening, the story is a useful window into how scarce capacity really is.
TL;DR: claimed vs verified
| Claim | Status |
|---|---|
| Amodei spoke with Wang earlier this year seeking compute | Reported by WSJ, citing sources. Not confirmed by either company in material we could access |
| Meta declined | Same: sourced report, no public confirmation |
| Rivals are forming alliances and executives are intervening personally | WSJ's framing of the wider scramble |
| Anthropic was in preliminary talks to lease compute from Meta in July | Reported by CNBC (person familiar) and, per CNBC, the New York Times, which put a potential deal at about $10 billion. Meta declined to comment |
| Anthropic uses SpaceX's Colossus 1 data center | Announced by Anthropic; see our earlier coverage |
What the WSJ report says
A summary of the Wall Street Journal's reporting says that, according to sources, Amodei spoke with Wang earlier this year hoping to source more compute and Meta declined the request. The report's broader theme, in the outlet's words, is that "bitter rivals are forming alliances and executives are having to personally intervene over the scramble for resources to fuel the AI boom."
What we do not have: the date of the conversation, the amount requested, the form the ask took (lease, purchase, or capacity share), or Meta's reason. None of that appears in the summary and we are not going to guess. Wang, as Meta's chief AI officer, is a natural counterpart because he runs the group that decides how Meta's own capacity is spent; we covered his views on alignment and openness in Wang, Musk and the open-source fight.
What CNBC reported in July
On July 17, 2026, CNBC reported that Anthropic was in "very preliminary talks" to lease computing power from Meta, according to a person familiar with the matter. CNBC added that the New York Times reported a potential deal being discussed worth about $10 billion, and that Meta declined to comment. The talks came weeks after Anthropic announced its SpaceX arrangement for Colossus 1 to add capacity for paid subscribers.
CNBC also gave context for why Meta might be a seller. Mark Zuckerberg said in May that Meta was considering entering the cloud computing business to show investors that AI spending can earn money beyond its current products, and Meta could spend as much as $145 billion on capital expenditures, including AI infrastructure, in 2026. Dave Brown, a longtime senior executive at AWS, was set to join Meta, per CNBC. Zuckerberg had also said last October that companies regularly ask whether Meta has compute to sell at a premium.
Are the two reports the same story?
Possibly, but not provably. A July lease negotiation and an earlier, personal ask from Amodei to Wang could be one process seen from two angles: an initial executive-level request, then staff-level talks, then a refusal or stall. They could also be separate. The two reports do not mention each other in the parts we could read, and the WSJ's "earlier this year" is vague. Be wary of any headline that merges them into a single confirmed timeline.
Timeline of Anthropic's compute moves, as reported
- May 2026: Anthropic and SpaceX announce the Colossus 1 partnership, and Zuckerberg says Meta is considering a cloud computing business.
- Earlier in 2026 (date not public): per WSJ sources, Amodei asks Wang for compute and Meta declines.
- July 17, 2026: CNBC and the New York Times report preliminary Anthropic-Meta lease talks, with a potential deal of about $10 billion discussed.
- September 2026: reports of up to $517 billion in Anthropic compute commitments and a contracted-value figure tied to Nvidia circulate; both are covered in our linked posts.
- October 9, 2026: the WSJ account of the Amodei-Wang exchange is summarized by Techmeme.
Seen in order, the July talks and the WSJ refusal could be sequential or separate. The timeline shows how long the search for capacity has run, not that any single request failed.
Why do AI labs ask rivals for compute at all?
Frontier training and inference run on a small supply of Nvidia-class accelerators, memory and power. Building a data center takes years, while demand for tokens grows monthly, so a lab that hits a ceiling looks for any large block it can rent. Hyperscalers such as Amazon, Google and Microsoft sell capacity as a business. A company like Meta, which built its fleet for its own use, has surplus only if its internal plans change. That is why a lease from Meta would be unusual: it would turn a competitor into a supplier.
For Anthropic the logic is straightforward. CNBC notes that access to enough chips remains a challenge and that the company places usage limits on its most advanced models. Every additional gigawatt reduces the pressure to ration. For Meta the logic cuts the other way. Its leadership has said it wants AI to pay for itself, and renting capacity to a model maker would bring revenue but also feed a rival whose models compete with its Muse line and consumer assistants.
What would a Meta cloud business change?
If Meta did start leasing compute, three things would shift. First, price discovery: another seller beyond the big three clouds could soften rates for GPU hours. Second, alliances: labs would depend on suppliers who are also competitors, which is already happening with SpaceX and Anthropic. Third, regulatory attention: vertical entanglement between model makers and infrastructure owners tends to draw antitrust interest. None of this has happened, and a refusal like the one the WSJ describes suggests Meta is not yet ready to be that supplier. Our post on Nvidia's plan to make GPUs an asset class explains why financing structures are making compute easier to trade.
Why would Meta say no?
We have no sourced reason, so here are the plausible explanations, clearly labeled as inference:
- Meta needs its own capacity. Its Muse models and consumer agents compete with Claude. Meta even runs parts of its Hatch agent on Claude today, as we noted in Meta's Hatch agent and its Anthropic spending, so it already depends on Anthropic in other ways.
- Strategic reluctance. Handing a direct rival a large block of chips could constrain Meta's own roadmap.
- Price or terms. A cloud business needs contracts, service levels and pricing Meta has not yet built.
None of this is confirmed. Each is a hypothesis for readers to weigh.
Anthropic's compute picture
The Meta ask fits a pattern. Anthropic has been stacking capacity from many places. Our coverage includes the reported $517 billion in compute commitments, the Nvidia contracted-value story, and the SpaceX Colossus 1 partnership. Rivals are doing the same: see OpenAI's $750 billion compute spend plans through 2030.
Interconnected nodes representing alliances between AI labs in the compute scramble
The WSJ's "alliances" framing is consistent with this. Competitors now sell capacity to each other, as in SpaceX supplying Anthropic, or ask each other for it. Whether Meta joins that market as a seller is the open question, and a refusal to Anthropic would tell us something about how it sees itself.
What this means for Claude users and builders
Capacity is the constraint behind many things users feel: usage limits, model availability and price. We tracked that in the Claude usage limits timeline and the 17 percent Claude Code limit cut. A failed bid for extra chips does not change any plan today, but it implies Anthropic is still hunting for supply, so plan on limits staying tight rather than loosening soon.
Practical steps:
- Build for variability. Add retries, backoff and a fallback model so a limit change does not break your workflow.
- Match model to task. Use the largest model only where it matters; smaller tiers stretch your quota.
- Watch official channels. Plan and limit changes appear in Anthropic's docs and changelog before they appear in rumors.
- Do not trade on the report. This is one paywalled account from unnamed sources.
How to verify this story yourself
Open the WSJ piece if you have access and check the byline, the date of the conversation and any on-record quotes. Search for statements from Meta or Anthropic spokespeople. Compare it with the CNBC July report. Until a company confirms, keep the language conditional: "sources say," not "Meta refused." For more on how we handle contested reports, see our claimed vs verified coverage of the AI labs' catastrophe wargame.
What to watch next
- A confirmation or denial from Meta or Anthropic.
- Whether Meta formally launches a cloud or compute-leasing business.
- New Anthropic capacity deals, and any change to Claude plan limits.
- Whether other labs report similar refusals, which would point to a broader capacity freeze.
Chiplet tiles on one substrate representing the GPU supply behind the Anthropic Meta compute request
Related reading
- Anthropic's reported $517B compute commitments, explained
- Anthropic and SpaceX Colossus 1 partnership
- Nvidia and Anthropic $180B contracted value
- Meta's Hatch agent runs on Claude
- Wang, Musk and the open-source fight
- OpenAI's $750B compute plans
- Claude usage limits timeline
- Sources: CNBC, July 17, 2026, Techmeme summary of the WSJ report
Based on reports available on October 10, 2026; details may change as the companies comment.