Skip to content

AI Decision Support

Meta may lease a competitor $10 billion in computing power. Here is what business owners should know about AI capacity, not just AI price.

The New York Times reported on July 17, 2026, that Meta Platforms is in talks to lease Anthropic up to $10 billion in computing power over two years, an arrangement that would have Meta renting out data center capacity to a company that builds competing AI models. Reuters and CNBC independently confirmed the talks are real, though early-stage and not guaranteed to close. Anthropic proposed the deal in June. Both companies declined to comment on the terms. The direct answer for a business owner: the interesting fact here is not the dollar figure or which company ends up renting to which. It is that Anthropic, an IPO-bound AI lab already committing $50 billion to build its own data centers and already leasing capacity from SpaceX and TeraWulf, still needs more compute, and is willing to get it from a competitor to secure it. When the largest, best-capitalized AI buyers in the world cannot fully solve their own capacity problem, that tells you something about the market underneath every AI tool a smaller business depends on.

By Fabio Rabelo · Founder, ATLACIS ·

What happened

On July 17, 2026, The New York Times reported that Meta is in talks to lease computing power to Anthropic in a deal that could be worth up to $10 billion over two years, citing three people familiar with the discussions. Anthropic proposed the arrangement in June, and Meta is still considering it. Under the structure described, Anthropic would pay Meta in monthly installments over the two-year term, with specific terms still in flux, and either company could exit the agreement early. Reuters and CNBC each independently confirmed the talks through their own sources the same day. Reuters reported the negotiations have been complicated by the fact that Meta does not currently run a business selling its computing power, and that both Meta and Anthropic declined to comment when asked directly. Meta shares fell more than 2% that day amid a broader tech selloff, though they pared some losses after the report circulated. The arrangement would echo a deal Anthropic struck with Elon Musk's SpaceX in May 2026 for capacity at SpaceX's Colossus 1 data center in Memphis. The Verge additionally confirmed that Anthropic separately plans to invest $50 billion building its own data centers, on top of the capacity it already leases from both SpaceX and data center operator TeraWulf.

Why it matters for business owners

Most small and medium businesses will never sign a compute lease of any size, let alone one worth billions. The reason this story matters anyway is what it reveals about the market every hosted AI tool sits on top of. Anthropic is not a struggling company here. It is heading toward an IPO, already spending tens of billions on its own infrastructure, and already has multiple large capacity deals in place. And it still needs more compute badly enough to negotiate with Meta, a company whose own AI models compete directly with Anthropic's. That is a capacity story, not a cost story. A business evaluating an AI vendor typically asks about price per token, subscription tiers, or contract length. This story is a reminder that the harder, less visible constraint is whether the vendor actually has enough computing capacity to serve every customer at full speed, all the time, and that even a company with billions of dollars to spend cannot fully guarantee that for itself right now.

What owners should not misunderstand

This is not a done deal. Every outlet that reported it, including The New York Times originally, described the talks as early-stage and said they may not result in an agreement at all. It is also not a sign that Anthropic is in financial distress; the opposite reading fits the facts better; demand for Anthropic's models is high enough that its own data center buildout and two existing capacity deals are not enough, which is a demand problem, not a funding problem. It is also not evidence that Meta's AI models are falling behind or that Meta needs Anthropic's help with model quality. The two companies would remain competitors on models; this is purely an infrastructure and revenue arrangement, the kind of deal a company can strike with a rival's factory while still competing with that rival's products. Finally, this is a different story from Meta's own reported cloud ambitions covered here on July 2. That earlier report was about Meta's internal plans to build a compute-reselling business, with no named customer or price attached yet. This is a concrete, priced negotiation with an actual counterparty, which is a materially different and later stage of the same story.

The operational lesson

AI compute capacity, not just AI price, is a real constraint on the market right now, and it is not obviously loosening. If Anthropic, with an IPO on the horizon and tens of billions already committed to its own infrastructure, still has to shop for more capacity from a competitor, then smaller AI vendors further down the market are working with the same underlying scarcity, whether or not they say so. A vendor's marketing page rarely mentions capacity limits. A vendor's actual behavior during a high-demand period, rate limits, slower response times, usage caps on certain tiers, is where that scarcity shows up.

What a serious business should do next

Ask any AI vendor a business depends on directly what happens to service during periods of high demand: are there usage caps, does response quality or speed change, and is there any service level commitment on availability, not just uptime. Do not assume a larger or better-funded vendor has solved capacity for good; this story shows that even the biggest, most capitalized AI labs have not. Build workflows so that a slower response, a temporary usage limit, or a short outage from an AI tool does not stop the underlying business process, through retry logic, a fallback tool, or simply a manual path that still works. Revisit this story if the Meta-Anthropic deal actually closes, since a completed deal (with real pricing and capacity terms disclosed) would be a stronger signal than the current talks.

The Atlacis view

Atlacis evaluates AI vendors on more than price and model quality. Capacity and reliability under real demand are part of that picture, especially for a workflow a business cannot afford to have stall during a busy period. When we help an owner choose between a hosted AI tool, a private deployment, and an on-premise option, part of that conversation is what happens when the vendor's own infrastructure is under pressure, and whether the business has a fallback that does not depend on one vendor having solved a problem that, as this story shows, even the largest AI labs in the world have not solved for themselves yet.

The short version

  • Meta is in early talks to lease Anthropic up to $10 billion in computing power over two years, reported by The New York Times on July 17, 2026 and independently confirmed by Reuters and CNBC. The talks are early-stage and may not result in a deal.
  • Anthropic, an IPO-bound AI lab already committing $50 billion to its own data centers and already leasing capacity from SpaceX and TeraWulf, is the party seeking more compute here, from a company whose AI models compete with its own.
  • This is a capacity story, not a price or funding story. Demand for compute is outrunning even the best-funded AI labs' own infrastructure buildouts.
  • This is a later, more concrete stage of the story Atlacis covered on July 2 about Meta's compute-reselling ambitions. That earlier report had no named customer or price; this one does.
  • The operational lesson: ask any AI vendor what happens to service during high-demand periods, and do not assume a larger vendor has solved capacity for good. Build workflows that tolerate a slower or capped AI response rather than depending on unlimited availability.
Tags:AI infrastructureAI computevendor dependencyMetaAnthropicAI vendorsbusiness AIAI decision-makingAI buying decisionsprivate AI
FAQ

Common questions

Is the Meta-Anthropic compute deal confirmed?
No. The New York Times, Reuters, and CNBC all reported the talks as early-stage as of July 17, 2026, based on sources familiar with the discussions. Both Meta and Anthropic declined to comment, and every outlet noted the talks may not result in a signed agreement.
Does this mean Anthropic is running out of money or falling behind?
The reporting does not support that reading. Anthropic is heading toward an IPO and is already committing $50 billion to build its own data centers, on top of existing capacity deals with SpaceX and TeraWulf. Needing still more compute reads as a demand and capacity problem, not a financial one.
What should a business actually do with this information?
Ask any AI vendor directly what happens to their service during high-demand periods: usage caps, slower responses, or any service level commitment on availability. Do not assume a large, well-funded vendor has fully solved capacity, since this story shows even the best-capitalized AI labs have not. Build a fallback so a temporary slowdown or limit from one AI tool does not stop the business process behind it.
Keep reading

More from the blog

Meta is planning to spend up to $145 billion on AI infrastructure this year. Now it may rent you the leftovers. Here is what business owners should know.

Bloomberg reported that Meta is building a cloud business to resell its excess AI computing capacity, the same move SpaceX made with xAI's data centers earlier this year. Meta's stock jumped on the news while neocloud rivals CoreWeave and Nebius dropped. The story is not really about Meta's stock. It is a signal that the company behind your AI tool may increasingly be renting hardware from, or to, a competitor, which changes what businesses should check before committing to a cloud AI or private AI vendor.

The AI chip market is opening up. What business owners thinking about private AI should know.

On June 18, 2026, Bloomberg reported that Amazon is in active talks to sell its Trainium AI chips directly to companies for use in their own data centers, confirmed by Amazon AI chief Peter DeSantis. Google is making a parallel move with its own custom silicon. For businesses considering private or on-premise AI, the hardware vendor landscape is beginning to change in ways it has not before. Here is what that shift means and what it does not mean yet.

Google just missed its own launch date for its next flagship AI model. Here is what business owners should know before planning around any vendor's roadmap.

Google told the world at I/O in May 2026 that its flagship Gemini 3.5 Pro model would ship in June. It is now mid-July and the model still has not shipped. Bloomberg reports the holdup is disappointing coding performance, and Alphabet shares slipped on the news. The lesson is not about Google's engineering. It is that a vendor's announced launch date is a forward-looking statement, not a delivery date, and a business plan should not be built as if it were one.

Make better AI decisions, starting with one call.

Book a free AI Fit Call. We will tell you what to use, what to avoid, and where to start. No jargon, no pressure.