Anthropic is reportedly in early talks to lease up to $10bn of computing power from Meta over two years, according to DataCenterDynamics. The talks confirm Meta's drift from social network to compute lessor — a shift UK buyers weighing self-hosting LLM vs. cloud GPU cost decisions cannot ignore.
View the data behind this chart
| 2025 | 2026 | |
|---|---|---|
| Meta AI capex | $bn72 | $bn145 |
Meta's compute-leasing pivot: what's actually happening
DataCenterDynamics reports that Anthropic proposed the arrangement to Meta in June, with either party able to exit over the two-year term — a structure that looks more like a flexible capacity option than a conventional long-term cloud contract.
This follows Anthropic's much larger $45bn, three-year deal signed with SpaceXAI in May for access to the Colossus 1 supercomputer, and Mark Zuckerberg's own May comments that entering the cloud computing business was 'definitely on the table' for Meta.
Meta has since hired Dave Brown, a senior vice president from Amazon Web Services with 18 years there building out AWS's cloud business — a hiring signal that reinforces the direction of travel even before any Anthropic contract is finalised.
From social network to AI landlord: the Meta Compute model
Beyond the Anthropic talks, Meta is reportedly developing a broader initiative — dubbed Meta Compute — that would sell excess AI capacity under two possible models: hosted access to Meta's own closed-weight models such as Muse Spark, similar to AWS Bedrock, or raw GPU capacity rental akin to neocloud providers like CoreWeave.
Analysis frames this as Meta 'metamorphizing into an AI cloud for sovereigns', potentially working with sovereign governments and wealth funds on compute, storage and networking mixing open- and closed-source models — a wholesale model that sits apart from retail consumer cloud.
For buyers used to comparing AWS, Azure and Google Cloud, Meta now needs a seat at that same evaluation table when assessing pricing and supply security across the market.
Captive vs third-party: the calculus UK buyers now face
Meta is simultaneously buyer and prospective seller. It plans to expand its Hyperion supercluster in Louisiana from 2GW to 5GW, lifting planned regional investment past $50bn plus over $1bn in local infrastructure improvements — clear evidence of captive, owned-and-operated capacity at hyperscaler scale.
At the same time, Meta has committed up to $27bn over five years to Nebius AI infrastructure, including up to $12bn for early large-scale Vera Rubin GPU deployments from 2027, with Nebius selling any spare capacity to third parties while Meta backstops what's unsold. It has also expanded its CoreWeave commitment to more than $35bn through 2032.
That dual posture — build some, lease a lot — is the same trade-off UK buyers face internally, and it's worth modelling both paths using a Cloud vs. On-Premise TCO Calculator before committing capital or lease terms.
The financial engineering behind hyperscaler capacity deals
Bloomberg figures cited by DataCenterDynamics show Microsoft and Meta together committing around $50bn to additional data centre leases in a single recent quarter, pushing total hyperscaler future lease commitments beyond $700bn — with Microsoft at $155bn and Meta at $104bn.
Commentary on CoreWeave's business model notes that neocloud GPU providers rely as much on lease-back structures and long-term off-take agreements as on datacentre design itself. Meta's own multi-year commitments to CoreWeave and Nebius mirror that financially engineered approach, locking in capacity and pricing over five-to-seven-year horizons.
UK buyers negotiating GPU-as-a-service contracts should recognise that headline pricing often reflects financing structure as much as raw hardware cost — a distinction worth stress-testing with an AI GPU Calculator before signing multi-year terms.
Capacity bottlenecks are real — even for Google and Meta
Meta's diversification isn't just opportunistic. Google reportedly limited Meta's use of Gemini in March 2026, telling Meta it could not supply all the GPU capacity it wanted despite a six-year, $10bn-plus cloud deal signed in August 2025.
That episode is a reminder that even top-tier hyperscalers hit capacity walls, and it's part of why Meta is spreading its bets across Nebius, CoreWeave and its own Meta Compute build-out rather than relying on a single rival cloud.
Meta's CFO Susan Li has also flagged that 2025 capital expenditure, including finance-lease principal payments, would reach $64bn–$72bn, partly attributed to tariffs raising hardware costs — a cost pressure that ripples into every lease and off-take price UK enterprises will eventually see.
What UK infrastructure buyers should do next
Meta's capex is set to roughly double to as much as $145bn this year from $72bn last year, a scale that now rivals or exceeds traditional UK-facing hyperscalers. Buyers should treat Meta as a live fourth option alongside AWS, Azure and Google Cloud when comparing supply security and pricing, not a curiosity.
Before assuming third-party leasing is automatically cheaper, run the numbers properly — our AI workload repatriation business case guidance shows when bringing workloads back on-premise beats leased capacity, and our research AI servers coverage tracks how vendor roadmaps affect that calculus.
Where a captive build still makes sense, compare configurations directly through server configuration or vendor-specific paths like the Dell server configurator, HPE server configurator or Lenovo server configurator to benchmark against any leased quote.
- 01DataCenterDynamics — Anthropic considers leasing compute from Meta in $10bn deal · 17 July 2026
- 02Tom's Hardware — Meta reportedly plans to rent out its AI compute · 1 January 2026
- 03DataCenterDynamics — Nebius secures up to $27bn AI infrastructure contract from Meta · 1 January 2026
- 04The Register — Meta, Trump tariffs and AI capex · 2 May 2025
- 05DataCenterDynamics — Google limits Meta's AI use due to capacity constraints · 1 March 2026
- 06DataCenterDynamics — Microsoft and Meta commit to $50bn in additional data center leases · 1 January 2026
- 07The Next Platform — CoreWeave takes as much financial engineering as it does datacenter design · 9 April 2026
- 08The Next Platform — Meta Platforms metamorphizing into an AI cloud for sovereigns · 23 January 2026
- 09Tom's Hardware — Meta expands Hyperion AI supercluster to 5GW · 1 January 2026