Meta May Lease AI Compute to Anthropic in $10B Deal

Meta is planning between $125 billion and $145 billion in capital expenditure this year, roughly double last year's spend, and is now in early talks to rent some of that capacity to Anthropic in a deal the New York Times pegs at up to $10 billion over two years. That reported figure represents somewhere between 3.5% and 4% of Meta's two-year capex envelope at the midpoint. In other words, the headline number is large in absolute terms but small relative to the buildout it would sit on top of.

The more interesting number is 8,000: the headcount Meta said in April it would cut, about 10% of its workforce, partially to offset the cost of the same AI infrastructure it is now considering subletting. That is the framing that matters for anyone modelling Meta's AI economics.

The Numbers

Start with the capex line. Meta's most recent earnings report guided to $125 billion to $145 billion in 2026 capital expenditure, largely for AI infrastructure. Even at the low end, that is a doubling versus the prior year. For context, a hyperscaler at that spend rate is buying data center capacity at a pace that only three or four companies on the planet can match.

Against that, the reported Anthropic deal, as CNN reported, would be worth as much as $10 billion over two years according to three sources cited by the New York Times. CNN's own source characterised any specific numbers as speculative and the discussions as still early. Both Meta and Anthropic declined to comment. So the $10 billion figure should be read as an upper bound floated by leakers, not a signed contract value.

Do the arithmetic. $10 billion over 24 months is roughly $417 million a month in run-rate revenue at the high end. On $135 billion of annual capex at the midpoint, that recovers something like 3.7% of one year's infrastructure spend across two years. It is not, on its own, a meaningful offset to the buildout. It is, however, a meaningful proof of concept that the buildout has resale value at premium pricing, which is what Mark Zuckerberg flagged at Meta's May shareholder meeting when he said outside companies ask "almost every week" whether they can buy compute "at some premium to what we've bought it at."

Meanwhile Meta stock (META) is down more than 8% from a year ago, while capex is doubling. That is the tension the market is pricing: investors want the AI spend to show up in revenue, and a compute-leasing line item is one of the cleanest ways to do that, because unlike ad revenue attributed to AI features, cloud revenue is directly measurable.

What we do not know, and this matters for any model of the deal's actual economics, is the gross margin. The source does not disclose the cost basis of the compute Meta would lease, the utilization rate it currently sees on that hardware, or the term structure of the pricing. Without those, the difference between a 60% gross margin lease (competitive with AWS) and a 20% gross margin lease (barely covering depreciation) is invisible from the outside. The bound is knowable: if Meta's reported cost of capital and Nvidia GPU depreciation schedules hold, a $10 billion two-year deal at hyperscaler-comparable margins should throw off $3 billion to $5 billion of gross profit.

What's Actually New

Two things are genuinely new here and both deserve unpacking separately.

First, Meta is signalling a shift from purely internal compute consumption to a merchant model. Zuckerberg's May comment is the tell: "we haven't done that yet because we think that we have a use for the compute. But obviously if we get to a point where we feel that we have overbuilt, then that is an option that we have." The Anthropic talks suggest that point has arrived, or is close enough that pricing conversations make sense. Meta joining Amazon, Microsoft and Google as a compute merchant would make it the fourth hyperscaler by inference capacity, and depending on how aggressive it gets, potentially a genuine competitor rather than a niche supplier.

Second, and less remarked on, is the counterparty. Anthropic already has multibillion dollar compute licensing deals with Google, SpaceX, Microsoft and Amazon. Adding Meta would make it a five-cloud customer. For an AI lab, that level of supplier diversification is not accidental. It is a hedge against exactly the kind of platform-risk that a Google or Microsoft could exert if Anthropic's models started competing too directly with their own. Meta, notably, is building competing models: last month it released an upgraded Muse Spark that Meta said could rival the coding capabilities of models from OpenAI, Anthropic and others, and for the first time offered a paid tier. So Anthropic would be renting compute from a direct competitor. That is not unprecedented (Anthropic already rents from Google, which ships Gemini), but it does raise the same governance questions that engineering teams building on Claude should think about: where is the model actually running, and what contractual guarantees exist around data isolation between the tenant model and the landlord's own AI teams?

The paid Muse Spark tier is the second signal in the same story. Meta is monetizing AI two ways at once: selling model access directly, and selling the infrastructure that runs other people's models. That is the AWS-plus-Bedrock playbook, and it took Amazon a decade to build. Meta is attempting it in eighteen months.

What's Priced In for AI Development

Some of this the market and engineering community already expected. The compute shortage is real, has been real for two years, and every AI lab of consequence has been forced into multi-vendor sourcing. That Anthropic is talking to a fifth supplier is directionally unsurprising. That capex-heavy players would eventually monetize surplus capacity through leasing was also expected, though the timeline is faster than most had modelled.

What is less priced in, I'd argue, is the speed at which Meta specifically has moved from "we might overbuild" (May) to "we're in talks with Anthropic" (July). Two months. That compression suggests either the buildout is running ahead of internal demand faster than Meta expected, or the internal roadmap for AI features has slipped, or both. Either interpretation is material.

Also underpriced: the strategic optionality this gives Meta on Nvidia negotiations. A hyperscaler that leases compute externally has a fundamentally different relationship with its silicon supplier than a captive consumer does, because external revenue helps justify next-generation GPU orders on an economic basis, not just a strategic one. That subtly strengthens Meta's hand.

What is genuinely surprising is that Anthropic, which has publicly leaned on its Amazon and Google relationships as strategic, is now willing to add a competitor-cum-supplier. That is a signal about how tight capacity remains in mid-2026, even after two years of aggressive buildout across the industry.

Contrarian View

The consensus read on this story is: Meta is becoming a hyperscaler, this is bullish for META, watch the cloud-revenue line in future earnings. The contrarian read is that this is a signal of internal weakness, not strength.

Consider the sequence. Meta doubles capex. Meta lays off 8,000 people partly to fund that capex. Meta stock trades down 8% year over year. Then Meta starts shopping surplus capacity to a direct AI competitor within two months of Zuckerberg publicly floating the idea. That is not the timeline of a company confidently executing an infrastructure strategy. It is the timeline of a company that overbuilt against an internal AI roadmap that is not consuming compute as fast as the buildout produced it, and is now looking for someone, anyone, to fill the gap so the capex does not look like a mistake.

If that read is right, the $10 billion is not a business line, it is a hedge against Muse Spark and Meta's internal AI efforts underperforming. And the margins will show it: distressed capacity does not price like AWS on-demand.

If this plays out, we should see Meta disclose a cloud or infrastructure-services revenue line in its next two earnings reports, and the gross margin on that line will tell us which read is correct. Above 50% and it is a real business. Below 30% and it is inventory clearance.

Key Takeaways

• The reported $10 billion deal is roughly 3.5% to 4% of Meta's two-year capex, meaningful as a signal but small as an offset to the $125 to $145 billion 2026 spend.
• Anthropic adding Meta would make it a five-cloud tenant (Google, Amazon, Microsoft, SpaceX, Meta), which is a strong signal that mid-2026 compute remains supply-constrained even after eighteen months of hyperscaler buildout.
• The unknown that matters most is gross margin on the leased compute. The source discloses neither cost basis nor pricing, so the difference between a real cloud business and depreciation-recovery is invisible from outside. Bound: if margins clear 50%, this is a strategic line; below 30%, it is inventory clearance.
• Meta running Anthropic workloads while shipping a paid Muse Spark tier that Meta itself says rivals Anthropic's coding models is a contractual and governance question engineering teams building on Claude should be asking their account managers now.
• Testable prediction: within two earnings cycles, Meta will either disclose a new infrastructure-services revenue line or the Anthropic talks will quietly disappear. If it is the former, watch the gross margin; if it is the latter, the overbuild thesis gets stronger.