VC
Value Add VC
โšกHomePulseโšกHelpful Apps๐Ÿ“Blog
โ† Value Add PulseAI$10B compute lease

Anthropic in Talks to Lease $10B of Meta's Compute

Anthropic and Meta are in early talks for Anthropic to lease up to $10B of Meta's AI compute over two years, letting Meta monetize its buildout while Anthropic diversifies beyond Amazon and Google.

Up to $10B
Deal size
2 years
Term
June 2026
Proposed
SpaceX Colossus 1 lease
Precedent
TC
Trace Cohen
Early-stage VC & angel ยท Founder, New York Venture Partners
July 17, 2026
2 min read
ShareXLinkedInEmail
THE RUNDOWN
1

Meta and Anthropic are discussing a compute lease deal worth up to $10 billion over two years, with Anthropic paying in monthly increments and either side able to exit early, according to Reuters and CNBC reporting July 17

2

Anthropic proposed the arrangement in June and talks remain early-stage with no guarantee of a signed deal, but the scale -- comparable to Anthropic's own May agreement to tap SpaceX's Colossus 1 data center in Memphis -- shows Anthropic diversifying compute sourcing beyond its anchor backers Amazon and Google

3

For Meta, leasing spare AI infrastructure to a rival lab would open a new revenue line from its enormous capex buildout and put it in direct competition with neoclouds like CoreWeave and Nebius, a notable strategic pivot for a company that has spent 2025-2026 building compute primarily for its own models

4

The talks land the same week Microsoft's Nadella publicly criticized Anthropic's access policies and the White House began asserting more control over frontier-model rollouts, meaning Anthropic is simultaneously fighting for compute independence and navigating new political scrutiny of who gets access to its models

TC
The VC Read ยท Trace's TakeTrace Cohen

Compute is the new cap table -- Anthropic leasing from a company it also competes with on model capability tells you the infrastructure layer has gotten so scarce that 'frenemy' is now a normal vendor relationship in AI. Any founder building a model-dependent product should read this as confirmation that compute access, not just model quality, is going to keep being the binding constraint through 2027, and multi-sourcing it the way Anthropic just did with SpaceX and now Meta is quickly becoming the standard playbook, not the exception.

Meta and Anthropic are in early talks for Anthropic to lease up to $10 billion of Meta's AI computing capacity over a two-year term, with Anthropic paying in monthly increments and both sides able to exit the agreement early, according to Reuters, CNBC and CNN reporting published July 17. Anthropic proposed the arrangement in June; a person familiar with the talks cautioned they remain early and may not result in a signed deal.

The move echoes a strategy Anthropic already deployed with Elon Musk's SpaceX, striking a deal in May to tap the full computing power of SpaceX's Colossus 1 data center in Memphis, Tennessee. Together, the two arrangements show Anthropic deliberately diversifying its compute supply chain beyond its two anchor backers, Amazon and Google, both of which have poured billions into the company in exchange for cloud commitments.

โ€œFor Meta, a compute-leasing deal with a direct AI-lab rival marks a genuine strategic pivot.โ€

For Meta, a compute-leasing deal with a direct AI-lab rival marks a genuine strategic pivot. The company has spent 2025 and 2026 building out enormous AI infrastructure primarily to train and serve its own Llama-successor models; renting spare capacity to Anthropic would generate a new infrastructure revenue line and put Meta in direct competition with neocloud specialists CoreWeave and Nebius, both of which have built entire businesses renting GPU capacity to labs that don't want to build their own data centers.

The backdrop matters here. This is landing the same week Microsoft CEO Satya Nadella publicly criticized Anthropic's access-control policies as not making sense, and the same week CNBC reported the White House is now asserting direct control over which partners get access to frontier models -- including reportedly blocking and then reinstating access to Anthropic's own Claude Mythos 5 and Fable 5 last month over national-security concerns. Anthropic is simultaneously trying to secure compute independence from its two biggest backers while navigating new political scrutiny over who it's allowed to work with at all.

The bear case: a two-year lease with an early-exit option isn't the multi-year commitment that typically justifies multi-billion-dollar data-center buildouts, and Meta extracting revenue from a company it also competes with on model capability is a more complicated relationship than a pure infrastructure landlord. What to watch next: whether the deal actually gets signed, and whether Amazon or Google respond to Anthropic's diversification by sweetening their own existing compute commitments.

ShareXLinkedInEmail
More onAnthropic โ†’Google โ†’Meta โ†’

Originally reported by CNBC. Analysis and editorial commentary by Value Add Pulse.

โ† Back to Pulse

THE WIRE in your inboxโ€” Tech, startup & VC news with Trace's take. Free, no spam.

Read Next

AI

China's Open-Weight Wave Forces an Enterprise Rethink

Kimi K3's benchmark-topping debut is accelerating enterprise interest in open-weight models, forcing US buyers to weigh self-hosted Chinese models against closed, subscription-priced offerings from Anthropic and OpenAI.

AI$960,000

Jensen Huang's Leather Jacket Sells for $960K

A leather jacket worn by Nvidia CEO Jensen Huang sold for $960,000 at Sotheby's, nearly 20 times its pre-sale estimate, with proceeds benefiting a philanthropic initiative for young tech builders.

AI$31.5B valuation (reported)

Moonshot's Kimi K3 Erases America's AI Lead

Chinese startup Moonshot AI released Kimi K3, a 2.8-trillion-parameter open model that topped a leading benchmark ahead of Claude and GPT-5.6, prompting Axios to declare China has erased America's AI lead.

@Trace_Cohenยทt@nyvp.com