Meta's Money Play Isn't AI: It's Data Centers

Anthropic is reportedly negotiating to lease roughly $10 billion worth of Meta's data center capacity over two years, according to the New York Times. The talks are early and could collapse, but they land on top of a pattern that's too consistent to ignore. Meta has spent the past several weeks assembling something close to an actual compute-leasing business, and Anthropic just became its highest-profile potential tenant.

Key Points:

• Anthropic is reportedly negotiating a two-year, roughly $10 billion deal to lease Meta data center capacity, with an early-termination option similar to the one in its SpaceX contract.
• Meta has stood up an internal division called Meta Compute, led by infrastructure chief Santosh Janardhan, Meta Superintelligence Labs lead Daniel Gross, and president Dina Powell McCormick, built specifically to sell excess capacity and hosted models to outside companies.
• The playbook isn't new. SpaceX already leases Colossus capacity to Anthropic for about $1.25 billion a month and to Google for roughly $920 million a month, arrangements Bloomberg Intelligence estimates could generate more than $50 billion by 2028.
• Meta's own 2026 capital expenditure guidance jumped to $125 billion to $145 billion, up from $72.2 billion in 2025, which is the kind of spending that makes any idle GPU look expensive.
• Meta is simultaneously renting capacity it doesn't own, including a deal with Nebius worth up to $27 billion, while also preparing to sell API access to its own Muse Spark 1.1 model at 75% less than Claude's pricing.

The Deal Nobody Expected

Reportedly floated by Anthropic in June, the arrangement under discussion would have Anthropic pay to run workloads on Meta's infrastructure, likely on Nvidia hardware rather than Meta's custom MTIA 400 chips, since adapting Anthropic's workloads to different silicon isn't worth the engineering cost for either side. The reported number, $10 billion over two years, is a third of what Anthropic pays SpaceX per month on an annualized basis, which tells you this would be a meaningful but secondary leg of Anthropic's compute strategy, not a replacement for it.

This Isn't A One-Off, It's A Named Business Unit

The detail that turns this from a rumor into a strategy is the org chart. Meta Compute has actual executives attached to it, per Bloomberg's reporting, and Zuckerberg has said publicly that selling excess capacity remains an option if the company ends up building more than it needs. That's not hedge language from a spokesperson, that's a CEO describing a plan. Add in the Nebius deal, where Meta is the one leasing capacity from someone else's data centers, and you get a company positioned on both sides of the compute market at once: buying when it's short, selling when it's long.

Why This Might Be The Monetization Play

Here's the case for yes. Meta's capex nearly doubled year over year into 2026, chips depreciate fast, and demand for frontier-model training runs in bursts rather than a straight line. Idle infrastructure between training cycles is exactly the kind of asset that turns into revenue if you can find a renter, and Anthropic, OpenAI, and a growing list of AI labs need capacity faster than they can build it themselves. SpaceX proved the model works well enough that Bloomberg Intelligence is willing to put a $50 billion number on it. Meta building an entire division around the idea, with named leadership, is a strong signal it sees the same math.

Here's the case for caution. Meta's own statements still frame this as a hedge against overbuilding, not a stated primary revenue strategy, and the company's core ad business still funds the vast majority of its earnings. Renting spare capacity is a good way to reduce the sting of a $145 billion bet, but it isn't yet a business Meta has said it's counting on.

The Complication: Meta Competes With Its Own Customer

The stranger part of this story is that Meta and Anthropic compete directly in the same market. Meta just launched Muse Spark 1.1, a coding-focused model it plans to sell through an API priced 75% below Claude. Renting infrastructure to a rival while undercutting that rival's pricing on your own model is either a sign Meta doesn't think infrastructure and model competition are the same fight, or a sign the infrastructure money is simply too good to leave on the table regardless of who's paying it.

What This Means For Marketing And Growth Leaders

None of this changes what tool you use tomorrow, but it changes how you should read the next twelve months of AI pricing and availability news. Compute is becoming the actual product being sold in this market, and the companies that own data centers are increasingly willing to rent to their own competitors to pay for them. That has real downstream effects on model pricing, rate limits, and which vendors stay financially stable long enough to build a roadmap around. Building that kind of infrastructure literacy into how your team evaluates AI vendors is part of the growth strategy work we do with clients right now.

If you want a second set of eyes on which AI investments are worth betting your roadmap on, and which are just noise, that's exactly what our AI marketing services team is built for.