/ Insights

The 1-10 MW gap: AI inference needs capacity hyperscalers won't build

Illustration of a large multi-building data center campus beside a substation, seen from the air

Hyperscalers build 100 MW campuses that are committed years in advance. AI inference needs 1 to 10 MW close to users, and we believe that is the least served part of the market.

The data center market is out of space. CBRE put vacancy across the eight primary U.S. markets at a record low of 1.4 percent at year-end 2025. Supply grew 36 percent year over year to 9,432 MW, and vacancy still fell, because net absorption hit a record 2,497.6 MW in 2025, up from 1,809.5 MW in 2024. New capacity is absorbed as fast as it is completed.

The pipeline does not fix it. Capacity under construction in those eight markets was 5,994.4 MW at the end of 2025, down from 6,350.1 MW a year earlier, the first decline since 2020, as permitting, zoning, and power procurement delayed projects. And the pipeline is spoken for before it is delivered. JLL reports that 92 percent of the capacity under construction across North America is already precommitted (leased or owner-occupied), with vacancy at 1 percent for a second consecutive year on JLL’s broader measure, and likely to stay in the low single digits through 2030.

Where the capacity goes matters as much as how much there is. Northern Virginia, the largest market, held 4,039.6 MW of inventory at year-end 2025, up 37 percent in a year, and led all primary markets for net absorption in 2025 with 1,102.0 MW. The eight primary markets CBRE tracks are Northern Virginia, Dallas-Fort Worth, Silicon Valley, Chicago, Phoenix, New York Tri-State, Atlanta, and Hillsboro.

The industry builds for the biggest tenants first, because that is the product it knows how to finance. A hyperscale facility is typically 40 to 100 MW or more, leased to a single tenant on a 10 to 15 year term. Wholesale colocation facilities, typically 5 to 40 MW, house multiple enterprise tenants. Retail colocation facilities, typically 1 to 10 MW, house many small tenants. Edge data centers are smaller still, typically 1 to 5 MW, located close to population centers to reduce latency. The hyperscale product is a campus. The edge product is a building. They are not the same business, and the capital goes to the campus.

AI inference does not fit the campus. Training a model is a centralized job. It happens once, in one place, at massive scale, and the location barely matters. Inference is the opposite. It runs continuously, everywhere the model is used, and it has to sit close to the user, the sensor, the store, or the hospital. A hospital system running ambient documentation, a retailer running computer vision in stores, or a manufacturer running quality inspection on a line does not need 100 MW in Northern Virginia. It needs 2 MW or 5 MW in a specific metro, on a timeline measured in months, not years.

That is the gap as we see it. The market data shows capacity is tight everywhere, and we believe small, power-ready inference capacity close to users is the least served part of it. The hyperscalers build campuses. The large colocation providers build halls for tenants taking 10 MW and up. The regional providers that do exist tend to serve general enterprise colocation, not AI inference specifically. A tenant that needs 5 MW for inference in a metro where its users actually are, is shopping in a market that was not built for them.

We build for that tenant. Edge AI Site Solutions is a global provider building 1 to 50 MW build-to-suit data centers for AI inference. The scale ladder is simple: a 1 to 3 MW inference hall, a 5 to 10 MW build-to-suit, or a campus of 10 to 50 MW. We have several hundred sites identified, screened for power, fiber, and access, with 7 GW of power identified across them.

The market numbers say capacity is scarce and getting scarcer. Record absorption, record-low vacancy, and a pipeline that is 92 percent spoken for before it is delivered. If you need 5 MW close to your users, the hyperscale pipeline was not built for you. Someone has to build the small end of the market. That is what we do, from 1 MW up to 50 MW.

Sources

All insights

Talk to a specialist.

Tell us the market, the megawatts and the date. We reply within one business day.