Anthropic wants to lease Meta’s spare hardware, and the deal under discussion could reach $10 billion over two years, according to CNBC report July 17 citing a person close to the talks. Anthropic reportedly discussed the idea in June. Nothing is locked, the payment would be spread over monthly installments and both parties would retain the option of leaving.
The dollar amount is not the story of the founders. The seller is. Meta has been operating as a computational buyer for years, and a lease would turn it into a seller, reshaping the market that will ultimately set the price of your Funding for AI infrastructure and your monthly inference bill.
What’s Really on the Table
Start with the warnings. No deal exists, Anthropic declined to comment and Meta remained silent when asked by Reuters. The $10 billion should be interpreted as the outer limit of a proposal rather than a signed figure.
The size becomes clearer with a comparison. Anthropic committed to a much larger deal with SpaceX in May, and the Meta proposal would represent only a fraction of that.
| Counterparty | Declared value | Term | Status |
|---|---|---|---|
| EspaceX | 45 billion dollars | Three years | Signed in May |
| Meta | Up to $10 billion | Two years | In talks |
Meta’s reasoning is quite simple to follow. Advertising still funds almost everything the company does, and leasing unused capacity would open up a second revenue stream while putting it in competition with specialty providers like CoreWeave and Nebius.
A buyer who becomes a seller
Startup teams almost never sit across from a hyperscaler. The consequences are still affecting them. As new capacity arrives at the top of the market, availability improves and prices decline as the downward effect is felt.
Scarcity follows the same path in the opposite direction. Large, multi-year commitments require the latest hardware first, leaving smaller accounts to queue longer and pay more. Anyone whose inference costs have doubled in the last year should follow supply news with the same attention they give to model launches.
Defensive movement is optional. A team capable of shifting lighter workloads to local AI models negotiates from a stronger position, simply because leaving poses a credible threat.
Where is the AI spending curve?
Deal Size Context: Gartner expects global AI spending to exceed $2.5 trillion in 2026up 47% year-on-year. In this context, $10 billion is significant but not extraordinary.
Meta also publicly hinted at this decision. Speaking to shareholders in May, Mark Zuckerberg described cloud services as something the company was actively considering, and mentioned that other companies continued to approach Meta to purchase access to the model or excess capacity.
Seen in this light, the July report confirms a direction rather than revealing one. The open question is whether Meta can spare enough hardware to keep the economy going while its own models continue to consume more of it.
Questions to ask your cloud provider
Vendors rarely advertise their flexibility, so increase it yourself. Ask how your rate responds if usage is way above or way below the forecast you committed to, because this clause decides whether growth rewards or punishes you.
Portability also deserves a direct question. Find out if you can retrieve refined weights, integrations and logs yourself, without opening a ticket and waiting a fortnight. Hesitation on this point reveals exactly how much influence you will have at renewal.
Then establish your position in the queue. Contracts often push smaller customers toward enterprise accounts when supply gets tight, and knowing this in advance allows you to plan launches based on what’s realistically available.
Three markers to follow
Housekeeping first and foremost. Pull out your compute contracts, note the commitment length and exit conditions, then calculate what you’re spending per active user each month. Without this number, you can’t tell if a price cut helps the company or if it just subsidizes waste.
Few companies do it well. Budgets have increased in many companies without any evidence of performance, which explains the disappointing results. AI KING reported in large companies. A lower unit price will never save a feature that no one wanted.
Over the next quarter, three things will tell you where costs are going. Does Meta officially launch a cloud product, do specialty providers adjust their prices accordingly, and do smaller customers start reporting shorter wait times for capacity? These answers matter far more to your 2027 budget than any overall assessment.
