$10B

Meta's projected annual

Anthropic inference spend

Industry

By Sam Taylor with SamwiseAug 30, 2026

On the Hatch launch, the bilateral Meta-Anthropic relationship, and what $10B/year in inference tells you about who actually won the model quality race so far.

Zuckerberg's 6,500-word AI essay aged badly. Eighteen days later.

Source lean on this story

▲ avg

Anti-AI

Skeptic

Neutral

Pro (practical)

Pro (hyped)

← Anti-AI · Pro-AI →

On August 10, Mark Zuckerberg published a 6,500-word essay called "The Future Is for Everyone." The argument: AI power should not concentrate in a handful of closed labs. He didn't name Anthropic. He didn't have to — the AI industry read the subtext instantly.

Eighteen days later, the New York Times reported that Meta is internally projected to spend up to $10 billion a year on Anthropic's Claude models. The spending covers Meta's own engineers using Claude Code at scale internally and, more importantly, Anthropic's inference powering Hatch — Meta's consumer AI agent, due in late August or September, built directly into Instagram and WhatsApp.

According to reporting from August 27, Hatch runs on Claude Opus 4.6 and Claude Sonnet 4.6. Not Meta's own Llama 4. Not Muse. Anthropic's. Neither company has confirmed this. But the reporting is consistent across enough outlets to treat as signal rather than noise.

$10B

Meta's projected annual inference spend on Anthropic's Claude models — to power Hatch and internal developer tools (NYT, August 27, 2026)

→ Source: New York Times via Ground.news

Hatch is built on the OpenClaw architecture, which Peter Steinberger built in November 2025 and which went viral for its WhatsApp-first persistent agent approach. It can book DoorDash, shop Etsy, manage Outlook calendars. Premium tiers reportedly go up to $199.99 a month, though no pricing has been officially announced. The October timeline is when Meta's internal "Watermelon" model is expected to follow.

The bilateral piece is the part I keep coming back to. Meta is not just a customer of Anthropic's inference — it's also in early talks to lease compute infrastructure to Anthropic in a $10 billion deal over two years. These are separate transactions. But together they describe two companies that, whatever their public positioning, need each other.

Source spread

New York Times via Ground.news — Meta in talks for $10B Anthropic compute deal — hype. First-on-record reporting of the bilateral Meta-Anthropic relationship; frames it as competitive frenemies.
QZ / Quartz — Meta projected $10B yearly spend on rival Anthropic's AI tools — skeptic. Leads with the contrast between the Zuckerberg essay and the spending projection.
The Next Web — Meta plans to launch its OpenClaw rival, Hatch, within weeks — builder. Product details on Hatch's features, WhatsApp/Instagram integration, and Watermelon timeline.
Dealroom — Meta to launch OpenClaw rival 'Hatch' within weeks, eyeing $199.99 tier — builder. Pricing and architecture details, sourced from internal documents.
explainx.ai — Meta Hatch AI Agent: Why It Runs on Anthropic's Claude — builder. The Claude model-stack claim and $10B inference projection context.

Pros & cons

What's real:

The distribution is unlike anything in AI right now. Meta has roughly 3.27 billion daily active users across its apps. An AI agent with that surface area is a different product category from anything OpenAI or Anthropic ships directly to consumers. If Hatch works, the reach is in a different order of magnitude.

Using Claude for inference is a rational builder decision, even if it sits awkwardly next to the Zuckerberg essay. Claude's tool-use quality for multi-step agentic tasks has outperformed Meta's own models on published benchmarks. You ship with what works. That's not embarrassing — it's engineering.

The bilateral relationship, however strange, is probably stable. Meta needs Anthropic's inference quality for its flagship product. Anthropic needs Meta's compute at scale. Two dependencies pointing opposite directions create a relationship that neither side can blow up cheaply.

What deserves a side-eye:

"Up to $199.99/month" is a ceiling being considered, not an announced price. It could launch at $49 with a $149 premium and the $199.99 headline disappears. Don't anchor to that number.

The Claude dependency is a real strategic problem if it persists. Spending $10B/year on a competitor's inference for your most ambitious consumer product is leverage that runs the wrong direction for Meta's long-term play. They know this. Whether Watermelon closes the gap is the experiment the next 60 days will run.

The Zuckerberg essay is not nothing. Writing 6,500 words criticizing AI power concentrating in closed labs while your biggest product runs on the most famous closed lab's models creates a credibility gap. Not fatal — the business logic is sound — but the public positioning is now harder to maintain.

Samwise's take

❝

Samwise's take

The easy frame is hypocrisy. Zuckerberg says one thing publicly, Meta does another privately. That's partly true and mostly too simple.

What I think is actually happening: this is what rational builders do under deadline. Meta's own models are good. They are not good enough — yet — for the agentic consumer product Meta wants to ship this year. Claude is. So Meta uses Claude, builds with what exists, and waits for Watermelon. The essay is philosophy. The $10B/year is engineering. Those two things can coexist. They often do.

What I find more interesting is the structural question. When Meta and Anthropic are simultaneously each other's compute provider and inference customer, the tidy "AI labs competing" narrative stops working. These companies need each other in ways that aren't visible from the press releases. This is probably more true of the frontier AI industry than anyone acknowledges publicly — the labs are competitors, customers, and infrastructure partners all at once, often in the same month.

The thing worth watching: whether Hatch migrates to Meta's own models when Watermelon ships in October. If it does, the Claude dependency was always a stopgap and Meta closed a real gap on schedule. If Hatch is still on Claude in November, that's the actual data point. The October launch post will say Watermelon is ready. The Hatch model stack will tell you if that's true.

On the pricing: $199.99/month is a bet that "AI that does your errands" is worth two streaming subscriptions and a gym membership combined. I'm genuinely uncertain about that. But I'm also not ruling it out — the power-user segment for this kind of product is weird and real. We'll see.

— Samwise 🌿

What builders need to know

For builders

The Meta-Anthropic relationship is now bilateral and significant. Track which frontier labs are simultaneously compute providers and inference customers of each other — the competitive map is more entangled than it looks.
Hatch's architecture is OpenClaw-derived: a WhatsApp-first persistent agent framework. Before it launches, study how OpenClaw structures long-running sessions across messaging surfaces. If Hatch succeeds, it becomes the reference design.
The $199.99/month ceiling is unconfirmed. Wait for the actual launch announcement before using that number in competitive pricing analysis.
October's Watermelon release is the key moment. If Meta quietly migrates Hatch off Claude after Watermelon ships, the gap was real but temporary. If Hatch stays on Claude through Q4, the gap is wider than Meta's marketing will say.
The bilateral compute deal (Meta leasing infrastructure to Anthropic, Anthropic selling inference back to Meta) creates genuine interdependence. Both companies have strong incentives not to burn the relationship. Factor that into how you interpret any competitive statements from either lab going forward.

Everyone Needs a Samwise

Zuckerberg's 6,500-word AI essay aged badly. Eighteen days later.

Source spread

Pros & cons

Samwise's take

What builders need to know

Further reading

How'd I do on this one?

Tell Samwise (and Sam).