INSIGHTS // PRIME_DOMAIN_CAPITAL

Data Center Power Costs: The Hidden Variable in AI Infrastructure

Power decides data center economics more than fiber or location. Electricity costs, grid constraints, and the nuclear pivot reshaping AI infrastructure budgets.

For forty years, commercial real estate developers ranked site factors the same way: transportation access, labor, tax incentives, fiber connectivity. In 2026, that ranking collapsed. Power availability is now the dominant variable in data center site selection — CBRE's 2026 outlook and investor survey named it the industry's top challenge for the third consecutive year, and it now outweighs fiber connectivity outright. The industry spent a decade optimizing for bandwidth. The binding constraint turned out to be electrons.

The numbers explain the inversion. Former Google CEO Eric Schmidt told Congress that data centers may require an additional 29 GW of electricity by 2027 and 67 GW by 2030. Grid-connection timelines in many US markets have stretched to 24–48 months or longer. Hyperscalers are now seeking individual sites capable of supporting 250+ megawatts — the equivalent of a small city's load, delivered to a single campus. In Ireland, data centers consumed 22% of national electricity in 2024, prompting a moratorium on new connections that was only partially unwound in late 2025. Power is no longer a utility bill. It is the gating item on the entire AI buildout.

The grid is the bottleneck, not the GPU

Walk through what this means for pricing. Northern Virginia — the largest data center market on earth — saw colocation vacancy fall to 0.3% in Q1 2026, with pricing at $170–$215 per kW-month per CBRE. That vacancy figure means the market is effectively closed to new large deployments: a 50 MW requirement cannot be placed in Ashburn at any price on a normal timeline, because the power to serve it does not exist yet. Atlanta's vacancy fell from 3.6% to 1% in a single year. Across North America's top four markets, available supply rose by just 1.4 MW year over year in Dallas–Fort Worth and declined in the other three — while net absorption, the actual occupied capacity, rose 34% to 2.23 GW.

This is why the construction statistics mislead. Under-construction capacity surged 25% in H1 2026, and Europe's total market is on track to grow 20% to 13 GW by year-end, with neocloud AI signings quadrupling to 420 MW in the first half. But most of that capacity is pre-leased, hyperscaler self-build, or waiting on grid interconnection. The vacancy rate that matters — power available for lease, right now — keeps falling. Equipment shortages and limited grid availability have constricted the buildout even as cranes multiply.

For infrastructure planners, the consequence is a geographic arbitrage that did not exist five years ago. Land near substations, transmission corridors, and natural-gas infrastructure now commands premiums that can exceed the value of traditional logistics-oriented industrial real estate. Developers are targeting secondary and tertiary markets where utility constraints are less severe and permitting is faster — the Dakotas and the Pacific Northwest on cheap hydro and wind, the Mountain West, parts of the Southeast. The site-selection question has flipped from "where is the fiber?" to "where is the power, and when can we have it?"

What power actually costs, metered and bundled

Strip away the real estate and look at the electron economics. US commercial-industrial power for data center use typically runs around $0.10–$0.12 per kWh in the benchmarks infrastructure analysts use for total-cost modeling — the Dakotas at the low end, California and the Northeast at multiples of that. At $0.12/kWh, a single 1 MW of IT load costs roughly $87,600 per month in raw electricity before cooling overhead, redundancy, and the operator's margin. Apply a typical PUE of 1.4 and the facility draws 1.4 MW to deliver 1 MW of compute — the cooling tax is real and immediate.

This is why the per-kW-month colocation rate is really a power price in disguise. Retail colocation power components run $150–$250/kW-month in 2026 market reporting. AI-ready space in Ashburn is quoted around $180/kW air-cooled and $225/kW liquid-cooled at 100 kW-plus densities — the liquid premium of roughly $45/kW-month, or $1,000–$2,000 per rack-month in retail structures, reflecting both the cooling infrastructure and the density it unlocks. A 135 kW rack at $200/kW-month costs $27,000 a month in power alone. At AI scale, the power line item dwarfs space, cross-connects, and remote hands combined. The rack is packaging. The power is the product.

Metering structure decides how much of that product you actually pay for. Flat-rate power plans charge the committed allocation whether you use it or not — predictable, but punitive for spiky AI training workloads. Metered plans charge actual consumption, which rewards efficient scheduling and punishes idle GPUs. With an 8-GPU H100 node drawing roughly 10 kW at full tilt, the difference between a well-utilized metered contract and a flat-rate allocation can be thousands of dollars per month per rack. Negotiate the billing model with the same intensity as the rate; the model often moves the total more.

Bring-your-own-power and the nuclear pivot

Faced with four-to-ten-year grid waits, the hyperscalers stopped waiting. As of May 2026, tech giants had signed 13 nuclear projects totaling over 9.8 GW of committed capacity. Microsoft secured a $16 billion, 20-year power purchase agreement to restart Three Mile Island Unit 1 — 835 MW, with full power targeted for 2027. Google committed to 500 MW from Kairos Power's fluoride-salt reactors by 2030. Amazon invested $700 million in X-energy. Meta leads the pack with up to 6.6 GW across TerraPower, Oklo, and others. The NRC issued its first advanced-reactor construction permit to TerraPower in March 2026. Behind-the-meter gas generation and fuel cells are proliferating as bridge solutions.

This is the most important structural development in data center economics in a decade: the separation of compute capacity from grid capacity. For the next several years, the winners in AI infrastructure will not be the companies with the best GPUs — GPUs are commoditizing — but the companies with secured, contracted, deliverable power. Enterprises planning 2027–2028 deployments should be asking their colocation providers one question above all others: show me the power, the interconnection agreement, and the date. Everything else is interior design.

Compute is rented by the hour. Power is paid forever — and in 2026, it is power that sets the price of everything downstream.

The authority on data center economics will be the name buyers type when the power bill arrives. Datacenterprices.com is available for acquisition by that company — inquiries to info@primedomaincapital.com.

// ACQUIRE

datacenterprices.com
$15,888

Offered by Prime Domain Capital. Secure transfer via Afternic. Inquiries: info@primedomaincapital.com