Frenemies at Scale: Meta Internally Projected Up to $10 Billion a Year on Anthropic's Models

The most revealing number in the AI economy this week came not from an earnings call but from a single sentence in a New York Times report: at one point this year, Meta internally projected that it could spend as much as $10 billion annually on AI models from Anthropic, according to two people familiar with the matter. The August 27 report, pointedly headlined “frenemies,” confirms what the industry has suspected for months — the company spending the most aggressively to build frontier AI is also becoming one of its chief rival’s most important customers.

The scale deserves a beat of reflection. Anthropic’s annualized revenue run rate topped $65 billion at the end of July, according to Reuters and CNBC — a sevenfold increase from roughly $9 billion a year earlier. A single customer projecting $10 billion a year in model spend would represent a meaningful slice of that trajectory. Meta has become, by the Times’ accounting, one of Anthropic’s largest customers, even as Mark Zuckerberg’s Meta Superintelligence Labs races to close the capability gap with in-house models like Muse Spark and the recently released Glimmer.

Why Meta is paying the competition

The immediate driver is product, not research. Meta’s forthcoming consumer AI agent Hatch — reported to be weeks from launch, with tiered subscriptions reaching $199.99 per month — currently runs on Anthropic’s Claude models rather than Meta’s own. The Information first reported the pricing structure; the Times report adds the financial gravity behind it. When your flagship consumer agent needs frontier-grade reasoning, tool use, and reliability on day one, you buy the best available inference — even from the competitor your superintelligence lab is trying to beat.

The deeper driver is time. Meta spent most of 2025 and 2026 reorganizing its AI effort around Meta Superintelligence Labs after its earlier in-house attempts underperformed. Muse Spark, released April 8 as MSL’s first model, and its July 1.1 update closed much of the gap for standard assistant workloads. But the hardest agentic tasks — long-horizon multi-step execution, robust coding, reliable tool orchestration — remain exactly where Anthropic’s Claude family is strongest. Buying Claude inference buys Meta quarters of time it would otherwise spend catching up.

The Microsoft parallel

The Anthropic spending is not an isolated case. A Bloomberg report on August 20 revealed that Meta has quietly become one of Microsoft’s largest AI customers, paying hundreds of millions of dollars a year to access AI models through Microsoft’s Azure platform — which, in a further twist, largely means resold OpenAI models. Meta runs trillions of tokens a week through rented rival infrastructure, per the report.

Put together, the pattern is striking: the company that pledged hundreds of billions of dollars for its own AI datacenters is simultaneously renting compute and models from Microsoft, OpenAI, and Anthropic. Some of this is classic benchmarking — you cannot beat a rival you have never run at scale. Some is genuine production dependency. The uncomfortable question for Meta investors is which proportion is which, and how fast the balance is shifting toward self-sufficiency.

What it means for Anthropic

For Anthropic, a $10 billion-a-year customer projection is both a blessing and a strategic complication. The company’s revenue growth — from $9 billion to $65 billion annualized in twelve months — is the fastest of any frontier lab, and enterprise demand is the engine. But concentration risk cuts both ways: Meta has an explicit migration plan. Reporting indicates Hatch is slated to move from Claude to Meta’s in-house Muse Spark models as they mature, meaning today’s largest customer is openly planning to become tomorrow’s departed one.

That dynamic shapes Anthropic’s pricing power, its IPO narrative (the company is widely reported to be moving toward a public listing as valuation talk circles $350 billion and beyond), and its willingness to sign deep custom agreements with a direct competitor. It also explains the parallel talks, reported in July, around Meta leasing compute to Anthropic in a potential $10 billion arrangement — a deal that would make the frenemy relationship bidirectional: Anthropic’s models running on Meta’s infrastructure, Meta’s products running on Anthropic’s models.

The industry’s new normal

The Meta-Anthropic axis is the clearest example yet of a structural shift in how AI competition works. In the cloud era, hyperscalers rarely bought core technology from one another. In the AI era, the frontier is narrow enough — and product deadlines tight enough — that the top labs are simultaneously each other’s suppliers, customers, benchmark references, and talent competitors.

Three implications stand out:

1. Model quality is now a procurement variable, not just a marketing one. When a lab with Meta’s resources chooses Claude over its own models for a flagship product, that is a cleaner signal of relative capability than any leaderboard.
2. Revenue at frontier labs is more concentrated than run-rate figures suggest. A handful of mega-customers — Meta, and reportedly others among the big platforms — account for outsized shares of the growth story.
3. Vertical integration is a race against dependency. Every quarter Meta pays Anthropic is a quarter of revenue Anthropic must assume will eventually disappear — and a quarter Meta spends closing the gap with Muse Spark’s successors.

None of this is secret from the participants; the Times piece makes clear both sides negotiate with open eyes. But the $10 billion figure crystallizes where the AI industry stands in late 2026: competitors so entangled that the clean narratives — open versus closed, Meta versus Anthropic, winner versus loser — no longer describe the actual wiring. The frenemy economy is here, and it is measured in billions per year.