Inference demand makes metro interconnection the scarce asset

OpenAI and Anthropic are in talks to buy data center capacity 20 to 30 megawatts at a time, a size class that barely registers against the gigawatt campuses that have set the price of AI infrastructure for two years. CNBC first reported the talks, which Data Center Dynamics carried: both labs are negotiating 20-30MW capacity deals at data centers, likely for inference workloads rather than the training runs that built the megasites.

OpenAI has discussed such deals with companies in the US and the Nordics, while Anthropic has held similar conversations in the UK and the Nordics, and the Nordic overlap on both lists says something about how the labs are shopping: for places where the load is not itself the contested variable. Consent is the product in every data center trade, and it is cheaper to buy where the queue and the planning posture already point the same way.

OpenAI described the smaller deals as portfolio construction rather than a change of tack. "We're building a diversified compute portfolio to meet growing demand for AI around the world," the company said, adding that different workloads need different infrastructure and that it weighs partners on performance, reliability, timing and cost. It said it does not comment on specific commercial discussions. The buyers who have spent two years signing gigawatt-scale deals are now saying, on the record, that a meaningful slice of their load wants something else.

Read next to the training book, the smaller ask looks like a rounding error. The report puts Anthropic's signed compute capacity leases over the past 11 months at a level of at least $517 billion — the report's figure, not an audited tally — and spread evenly across those months that is roughly $47 billion of capacity committed a month, a fleet assembled site by site rather than concentrated in a handful of anchor campuses. A tenant committing at that pace is not short of floorspace in the abstract; it is short of floorspace in the places its users are.

The gigawatt deal was the training era's unit of account: cheap electrons, distance from neighbors, a greenfield grid slot bought with the tenant's signature. Grid permission rather than land or capital is the underwriting asset in digital infrastructure, and OpenAI's Georgia Power grid deal is the clearest example, a customer-funded utility build that turns interconnection into a tollbooth. Inference is a different siting problem because serving queries means being near the people making them, which points this demand at metro campuses, powered shells and interconnection already in hand rather than the rural land banks the megaprojects were assembled from.

The immediate beneficiaries of that shift are operators already holding energized capacity in the markets the labs named — colocation landlords above all. A 20-30MW requirement likely fits inside an existing footprint without a new county, a new consent fight or a new transmission line, which is exactly the constraint visible from the power side. The greenfield pipelines underwritten around a single gigawatt anchor have a narrower set of possible tenants than their pro formas assumed.

European capacity has its own arithmetic, and Europe's grid bottleneck as the binding constraint for AI buildout on the continent is the place where a well-capitalized entrant still waits on a wire, which is one reason the Nordic overlap in both shopping lists deserves more weight than the UK or US mentions.

There is a delivery catch, and it runs against the arithmetic of scale: a 20MW hall likely carries a fixed operating minimum — technicians, security, cooling maintenance — that does not shrink in proportion to the load, so a portfolio of small sites multiplies staffing demand instead of concentrating it. The delivery constraint has moved into supplier slots and raw materials, with staffing the final bottleneck before energization, and a fleet of 20-30MW leases spreads that bottleneck across more doors at precisely the moment the labs want them open.

The diversification OpenAI describes is geographic, not credit: every counterparty in these discussions, as reported, is one of two companies, and their capacity to pay sits on the same demand forecast the training leases already lean on. Only anchor-contracted digital assets earn infrastructure pricing. The 20-30MW lab lease tests that rule, and my read is that it should not clear at the same price as a hyperscaler-anchored campus. A 20MW hall is easier to re-let than a bespoke megacampus, so the exposure is tenant concentration and the length of the commitment rather than the building's specificity.

Anthropic's own buildout shows which way it is leaning. It has been building an in-house data center arm — in September it hired the architect of Equinix's hyperscale joint ventures — and the structures that surfaced with that hire put the anchor tenant on the cap table, which solves the lease and stacks the credit on a single demand forecast. A portfolio of 20-30MW sites is exactly the sort of procurement a tenant brings inside rather than outsources, and it likely leaves the tenant, not a developer, negotiating each interconnection.

Colocation contracts are shorter, priced monthly, and leave the operator holding the residual; a long-dated lease with a lab's name on the credit is a different asset with a different buyer, and the first one signed in the Nordics will set the template for the class.