Meta's $10B Anthropic tab outs the AI race's awkward economics

For months, Mark Zuckerberg has cast Anthropic — and OpenAI — as the closed, centralized dangers his open-source movement exists to fight. His own spreadsheets tell a flatter story. Sources told The New York Times that Meta's internal projections assumed it could spend as much as $10 billion a year on Anthropic's models, at the same time Zuckerberg was publicly trashing the lab as a rival. The report, published Aug. 27, frames the pair as the industry's most cosy frenemies: competitors in the sell-to-everyone race, and — on paper at least — deeply interdependent customers.

That tension is the real headline here, and it cuts deeper than a headline about two companies who don't like each other. It exposes a truth the AI industry keeps trying to talk its way around: at the frontier, model quality still gates the whole stack, and nobody's own models are reliably good enough to bet the future on — not even the one spending the most to build its own.

The projection, placed in context

The $10 billion figure — projected annual spend on Anthropic's tools — is an internal planning number, not a signed contract. That matters, and we'll get to why in a moment. But it is not abstract. The Times report lands less than two months after a separate set of talks between the same two companies, in which Anthropic was reported to be leasing up to $10 billion of computing power from Meta's data centers over two years. That earlier deal, first reported in July, had Meta playing landlord to a rival's silicon needs. The new figure points the other way — Meta as the customer buying a rival's model output at scale.

Read those two data points together and a pattern emerges: Meta's planners have priced in dependence on Anthropic whether it is selling compute to the lab or buying models from it. Neither direction is a casual hedge. Ten billion dollars a year would make Meta one of Anthropic's largest customers in waiting, sitting alongside the hyperscalers that burn cash on model access, and it recasts Meta's own multibillion-dollar buildout as a bridge — expensive insurance against the possibility that its own homegrown models never close the gap.

The economics underneath the bravado

The why is the interesting part. Meta has acknowledged, in its own words, that it may build more data centers than it needs relative to the demand it can generate on its own. It is on track to spend as much as $145 billion this year, much of it on AI infrastructure — more than double last year's figure. And its public messaging around its models has been quietly defensive: reports this year have described Meta's proprietary models trailing Anthropic and OpenAI on coding, reasoning, and writing. When the lab you dismiss in public keeps beating your own work where it counts, the engineering pride collides with the quarterly planning cycle — and the planning cycle usually wins.

Anthropic's side of the ledger is just as revealing. The lab is assembling compute from every landlord it can find because it cannot build fast enough to keep up with demand for Claude and Claude Code. We've tracked that scramble closely — Anthropic's $9.1 billion compute deal with Bitcoin miner Riot and the even bigger commitments to SpaceX and Nscale were all part of it. A company this desperate for capacity being courted as a model supplier to a giant like Meta shows how the frontier's economics now run both directions: whoever owns the best open weights, the cheapest compute, or the most-wanted model becomes everyone's reluctant partner.

This is also the deeper structural point that our Deep Dive on Nvidia's quiet takeover of the AI stack laid out this morning: whoever controls the scarce layer keeps getting paid by everyone, competitor or not. Meta wanting Anthropic's models is the flip side of Amazon and Microsoft happily hosting OpenAI's and Anthropic's workloads that compete with their own. Compute scarcity and model scarcity have made "amicable codependence" the industry's default operating state.

Who wins, who loses

If the projection ever hardens into a contract, Anthropic is the obvious winner. A customer of Meta's scale backstops its revenue story at a delicate moment — with an expected IPO looming, and investors pricing a valuation reportedly near $1 trillion. A $10-billion-a-year anchor from a deep-pocketed rival would read as a market telling analysts it doesn't actually need to beat OpenAI to monetize; it needs to survive the oligopoly intact. It also strengthens Anthropic's hand in the compute negotiations, because a guaranteed giant customer justifies more capacity commitments.

Meta's win is subtler, and it's largely about its own investors. Staring down a $145 billion capex bill that Wall Street keeps questioning, a large, credit-worthy buyer for either its idle compute or for frontier models it isn't building well internally turns an expense into a strategic hedge. The loser, awkwardly, is OpenAI. Its two biggest rivals cozying up — Meta as Anthropic's customer, Anthropic depending on Meta — is precisely the kind of alliance that could marginalize the third player in enterprise deals where Meta has distribution reach that Anthropic doesn't.

The skeptical view

Before we crown this a done deal, hold the counter-case. The number is, by the Times's own framing, an internal projection — what Meta's planners believed it could spend, not what it has agreed to spend. Projections of that kind are often downside hedges, contingency plans for the scenario where Llama and Muse keep underperforming and Meta has no choice but to buy frontier capability elsewhere. They are not contracts; a company that beats its own model problem rescinds them overnight. There is also the genuine corporate-strategy friction: if Meta ships a model business built partly on Anthropic's weights, it hands a direct competitor a cut of whatever it earns — funding the roadmap of the rival it wants to beat.

Analysts who've watched this relationship note that neither company is new to these arrangements, and that such deals carry exit clauses precisely because they're provisional. The safest reading: the $10 billion figure captures how seriously Meta's planners rate the possibility that they'll need Anthropic — not certainty that they will.

What to watch next

The number worth watching is whether this projection turns into an actual contract, and what happens to OpenAI as Meta and Anthropic's economics interlock further. Watch Anthropic's IPO documents for disclosed customer concentration; watch any Meta earnings-call question about whether its "internal" model spend already budgets in a rival's capacity. And keep an eye on whether this disclosure, first surfaced by the Times, prompts either company to publicly correct — or quietly confirm — the scale.

The gap between what frontier labs say about each other and what their planning documents budget is now public and measurable. Do you think a $10 billion projection is a real bet, or boilerplate cover for a model that might not arrive? Tell us in the comments.

Sources: The New York Times · Techmeme · CNN Business (July compute-lease context)