OpenAI and Anthropic Are Looking at Smaller Data Centers

Hero photo: A SmartCube modular data center — HGV1, Wikimedia Commons, CC BY-SA 4.0.

01 What changed this week

On September 18, CNBC reported — citing four people familiar with the conversations — that OpenAI and Anthropic are exploring data center deals in the 20-to-30-megawatt range, a fraction of the massive campus agreements both companies signed over the past year. The same reporting says the labs are looking at sites in the United States, the United Kingdom, and the Nordics.

Neither lab is abandoning its big campuses. Anthropic announced a $50 billion U.S. infrastructure investment in November 2025, anchored by Texas and New York facilities. The smaller deals ride on top of that — a second, faster lane for capacity. OpenAI told CNBC that different workloads need different infrastructure, and that it evaluates opportunities on performance, reliability, timing, and cost.

This is an analytical scenario based on current reporting and records, not a prediction. Figures discussed are potential future contenders per public reporting; no deployment is assumed.

02 Why the frontier labs suddenly want small

The logic is speed. Gigawatt-scale campuses face lengthy development timelines: power procurement, interconnection, construction, and energization can stretch across years. A 20-to-30-megawatt lease in an existing shell — often from data center operators or neocloud providers, which is how both labs typically rent compute — can reach live capacity in a fraction of that time.

That matters because AI demand is not waiting. Both labs are racing to serve inference traffic — the work of answering live user and agent requests — which grows continuously and cannot be deferred the way a training run sometimes can. When the constraint is how fast you can add serving capacity, a small deal you can use this quarter beats a huge deal you can use in 2028.

03 Inference is reshaping the data center itself

There is a structural reason this shift is happening now, not just a scheduling one. Training frontier models wants concentrated compute — tens of thousands of GPUs packed into one low-latency cluster. Inference wants distribution: capacity placed near users, spread across regions, sized to traffic patterns. As the industry's center of gravity moves from training runs to always-on serving, the ideal footprint changes shape.

The labs are not alone in noticing. Nvidia said in February it would work with data center stakeholders to study smaller facilities designed for distributed AI inference — an acknowledgment from the chipmaker whose GPUs defined the gigawatt era that the next buildout may look different.

04 What small sites give up

The 20-to-30-megawatt deal is a trade, not a free lunch. Small sites sacrifice the economics of scale: power-cost efficiency, staffing per megawatt, and the bargaining leverage that comes with being an anchor tenant on a gigawatt campus. They also fragment operations — more sites means more failure points, more network hops, and more complicated capacity management.

There is also a dependency story. Renting small from neocloud and wholesale operators means relying on counterparties' balance sheets and build quality. The same labs that spent 2024 and 2025 securing long-term dedicated capacity are now adding shorter, smaller commitments — flexibility that cuts both ways if demand shifts.

05 The grid is the hidden protagonist

Behind every one of these deals stands the same obstacle: power. Interconnection queues in major U.S. markets have stretched for years, and grid operators from PJM to ERCOT are struggling to keep up with AI-driven load growth. Small data centers sidestep part of that problem by moving into buildings that already have power — but they compete for the same electrons as everyone else.

That competition has a policy dimension. Regulators and utilities are already deciding how AI load gets treated in capacity auctions and rate cases. If frontier labs increasingly deploy as many mid-sized loads rather than a few giant ones, the questions change: how do you forecast a swarm? Who pays for grid upgrades when the biggest single customer is 30 megawatts instead of 1,000? The smaller-deal era makes AI infrastructure look less like a few mega-projects and more like ordinary — if unusually dense — commercial load.

06 What to watch

Three signals will tell whether this is a durable shift or a gap-filler. First, whether the labs sign repeat small deals — a pipeline, not a handful of stopgaps. Second, whether the big campus agreements stay on schedule; if gigawatt projects start slipping, small deals become the primary lane, not the fast one. Third, whether hardware follows: Nvidia's distributed-inference study and the rise of modular data center products both suggest suppliers see the same opening.

The verified facts: CNBC reported on September 18, citing four sources, that OpenAI and Anthropic are exploring 20-to-30-megawatt deals across the U.S., UK, and Nordics; both companies already hold large campus-scale agreements; Nvidia in February announced work on smaller facilities for distributed inference; OpenAI says different workloads need different infrastructure.