Anthropic and OpenAI Shift Toward Small Data Centers, Seeking 20-30MW Compute Deployments
nashnova research
Anthropic and OpenAI are negotiating 20-30MW small data-center leases across Europe and the U.S., opening a distributed compute supplement alongside their gigawatt-scale megaprojects — driven by the rapid rise of inference workloads.
Why are two leading AI labs suddenly shopping for "small" facilities?
CNBC reports, citing multiple sources, that Anthropic has approached providers in the UK and Nordics for 20-30MW compute deals; OpenAI is exploring similar opportunities in the Nordics.
A separate source says both companies are also negotiating deployments of the same scale inside the United States.
This means → the procurement playbook is shifting from "bet everything on one mega-campus" to a mixed portfolio of large and small, spread across many sites.
What makes small sites attractive?
Structure Research head Jabez Tan told CNBC the core appeal is "faster access to available capacity" — securing a few megawatts at an already-energized site is far quicker than waiting for a massive single campus to come online.
In plain terms = a mega-project is like building a new city — long lead times, many unknowns. A small site is like leasing a ready-made office — sign the contract and move in.
This reflects an urgency so acute that AI companies cannot afford to wait out big-project construction timelines.
Why can inference be split into small clusters while training cannot?
Training a large model requires massive numbers of chips working in tight coordination, with data flowing between them at high speed — physically hard to distribute.
Inference — the process of serving users after a model is trained — is different: each request is relatively independent and can be spread across multiple smaller clusters.
This means → inference is a natural fit for "many small sites," while training still depends on concentrated, ultra-large clusters.
How big will inference's share grow?
A JLL report shows that in 2025, inference accounts for 9% of global data-center AI workloads; training accounts for 14%.
By 2030, inference is projected to surge to 37%, while training drops to 13%.
In plain terms = in five years, nearly four in ten units of AI compute will run inference, and training will be the minority — that is the structural driver behind the explosion in demand for small sites.
Are the mega-projects still going ahead in parallel?
Anthropic previously signed a roughly $45 billion cloud-computing deal with Nscale, leasing about 460MW of capacity in West Virginia.
OpenAI's Stargate infrastructure project has exceeded its original 10GW commitment, adding 3GW in Georgia and 8GW in Ohio.
This means → mega-projects anchor long-term training capacity; small sites fill short-term inference gaps — the two tracks run in parallel, not as substitutes.
How will the competitive landscape for compute suppliers change?
Data-center builder Crusoe is pivoting to smaller, faster-to-build facilities, which the Wall Street Journal reports are cheaper and face fewer delays.
Crusoe announced $3.9 billion in funding the same day, reaching a post-funding valuation of $30.9 billion.
This reflects a capital-market bet on the "small and fast" compute-supply model — for traditional large-scale data-center developers, nimble mid-sized players are carving out a new slice of the pie.
市场有风险，内容仅供研究参考，不构成投资建议。