Gigawatt delays push OpenAI and Anthropic toward 20 MW data centers

I.The biggest buyers in AI are shopping small

On 18 September, CNBC reported that Anthropic and OpenAI are hunting for smaller data center deals, casting the net wide enough to include sites of 20 to 30 megawatts. According to a summary of the sourcing, four people said Anthropic had explored deals in the UK and the Nordics, two said OpenAI had looked at Nordic sites, and one described US talks involving both.

The size is the story. In August, Anthropic agreed a six-year, $45 billion lease for 460 MW at Nscale's Monarch campus in West Virginia. OpenAI's Stargate programme has passed its original 10 GW target, with 3 GW added in Georgia and 8 GW in Ohio. A 25 MW site is a rounding error on either balance sheet.

Nobody is buying a rounding error for its size. They are buying it for its date. The Monarch systems are not due until late 2027. A powered hall that exists today can serve customers this quarter.

The capacity that matters is no longer the capacity you have contracted. It is the capacity that is energised.

II.Time-to-power replaced price as the scarce input

The market data explains the pivot. CBRE's H1 2026 report, published 27 August, put primary-market vacancy at 1.4%, with more than 80% of the 7,481 MW under construction already preleased. Fewer than 1,500 MW remain available to prelease across North America's primary markets, which CBRE equates to roughly six months of supply.

The detail that matters most for this signal: asking rents rose in every deployment size, and the steepest jump, 8.3%, was for users seeking 3 to 10 MW. Small, ready blocks are where the price pressure is sharpest. That is what a market looks like when large buyers start competing for space that used to go to enterprises.

Behind the shortage sits the grid. CBRE says power constraints are stretching completion timelines. In Britain, Ofgem's 29 July consultation proposed a commitment fee of £237,500 to £712,500 per megawatt to flush speculative data center projects out of a connection queue in which data centers account for at least 80 GW of a 125 GW demand surge. The consultation closed on 16 September with no decision date set.

Local politics compounds it. Data Center Watch counted at least 45 US projects worth $68 billion blocked or delayed by community opposition between April and June, with legislative or executive action introduced in 30 statehouses. TechSpot's account of the report counts 843 opposition groups across 49 states.

III.The honest read: a hedge, not a rupture

Three caveats keep this signal in proportion.

First, it is a report of talks, not deals. CNBC's sources describe exploration. No counterparty, price, delivery date or signed contract has been disclosed. OpenAI said only that it weighs sites on requirements, performance, reliability, timing and cost.

Second, the opposition numbers are moving the other way. Data Center Watch's Q1 2026 tally was 75 contested projects worth about $130 billion, so Q2 was lower on both counts. Coverage that framed the Q2 figure as projects "blocked" also overstates it: the report pools blocked and delayed projects and does not separate them.

Third, the delay narrative has a credible skeptic. In June, SemiAnalysis argued that claims that half of 2026 US capacity is cancelled confuse press-release megawatts with buildable ones. Its own year-end 2026 forecast for hyperscaler self-builds moved about 1% in six months. Energy analytics firm Currence, by contrast, told Network World that 30% to 50% of large capacity slated for 2026 will likely slip, with only about 5 GW of 16 GW visibly under construction. The two do not reconcile, and we do not average them. Both can be partly right: projects with permits, power and equipment are mostly on time, while the speculative layer is not.

The engineering point cuts in favour of the signal but limits it. A 20 to 30 MW hall cannot host a frontier training run, which needs tightly coupled clusters at campus scale. It can host inference, which parallelises across sites. The EPRI, Nvidia, Prologis and InfraPartners pilot, announced in February, targets exactly this: 5 to 20 MW prefabricated sites at substations with spare capacity, at least five in the US by the end of 2026. Small sites are a serving layer, not a substitute for the campuses.

IV.Who captures the capex now

Market structure. When the buyer's binding variable is the date power is live, bargaining power moves to whoever holds energised space. That favours existing colocation operators, neoclouds with powered shells and former crypto miners with grid connections, over developers holding land and a queue position. Anthropic's 20-year TeraWulf lease in Kentucky and its $11.6 billion, seven-year Akamai agreement of 24 September both buy from incumbents with infrastructure in place. Akamai's release specifies CPU workloads on a distributed network, not GPU halls, so it shows the pattern rather than proving it.

Prices. CBRE's 8.3% rent rise for 3 to 10 MW blocks is the first measurable price effect. Our estimate of the stakes: Anthropic's Nscale lease works out to about $16 million per MW per year ($45 billion over six years across 460 MW). If a 25 MW site leased at a comparable rate, it would be worth roughly $400 million a year. The assumption is generous, since that contract bundles next-generation chips, but it shows why a few dozen megawatts now attract the largest buyers in the market.

Regions. The UK and Nordics appear in the reporting for different reasons. The Nordics offer cheap, low-carbon power and cool air. The UK offers proximity to European customers but faces a reform-in-progress queue. Neither is guaranteed: OpenAI did not conclude a deal for Stargate Norway in April and paused Stargate UK, citing energy costs and regulation, while Microsoft took the Narvik capacity instead.

Capital. Fragmentation raises unit costs. Twenty sites of 25 MW need twenty power contracts, twenty security perimeters and twenty network builds where one 500 MW campus needs one. Labs are paying a premium for speed. That premium is rational only while demand outruns supply.

V.Second-order: the map of AI compute fragments

The first consequence is geographic. Inference capacity that follows spare substations rather than master-planned campuses spreads compute into secondary markets and smaller countries. That creates openings for regions with stranded grid capacity and quick permitting, including parts of southern Europe that do not feature in this round of reporting.

The second is political. Opposition has been aimed at visible, gigawatt-scale projects. Smaller sites at existing substations draw less attention, which is part of their appeal. If that becomes the dominant way to add capacity, local consent fights shift from planning hearings to utility tariff cases, where the question is who pays for grid upgrades.

The third is operational. A lab serving models from dozens of mid-sized sites runs a distributed systems problem, not a real estate one. Routing, model placement and failover become sources of margin, much as content delivery networks turned scattered points of presence into a business.

VI.Precedent: gigawatt plans have slipped before

The pivot has an internal precedent this year. In March, Oracle and OpenAI scrapped a planned 600 MW expansion at Abilene after financing talks failed and the local grid would not be ready for about a year, by which point the Blackwell hardware intended for it would be a generation old. The 1.2 GW campus and the wider 4.5 GW agreement stayed on track.

Epoch AI's April site-by-site review found Abilene running about 0.3 GW of its planned 1.2 GW, with other Stargate sites still at foundations, framing or roofing. The lesson is consistent: gigawatt programmes do proceed, but on construction time, while demand arrives on product time. Small deals bridge the gap between the two clocks.

VII.How it plays out: three scenarios

Probabilities are our analytical judgment, not forecasts from a model, and they sum to 100%.