AI Data Center Cost in 2026: AI Servers, GPUs, Cooling and Operating Costs

More efficient accelerators fit fewer units per megawatt, so GPU cost per MW does not simply track the sticker price. Rack-scale systems such as VR200 NVL72 and AMD's Helios MI455X rack push densities to 190 to 230 kW per rack, which only direct-to-chip liquid or immersion cooling can handle. AMD Instinct parts cost less per accelerator than their NVIDIA peers, so at the same rack power they lower GPU spend per MW; the configurator models five of them from MI300X to MI455X. MI300X and MI325X prices are estimates because neither has a public list price.

Even before GPUs, AI buildings cost more per MW. The configurator's hyperscale AI factory costs C$36.7M per MW for the facility (BC baseline), against C$17.6M for a powered shell, because of heavier structure, denser power distribution and direct-to-chip liquid cooling at C$3.7M per MW.

For a 100 MW AI factory, the configurator gives:

California is the target market for many of our readers, so it is worth knowing it sits at the expensive end on every axis: US$0.205/kWh modelled power, a 1.22× construction cost factor, a 48-month grid queue and no sales-tax exemption on equipment. Price your own AI build, with your choice of accelerator, in the free live demo.

Ongoing Operating Costs After the Build

The build is only the start. In the configurator's hyperscale AI factory profile, the yearly operating cost has five lines:

Operating Costs of a 100 MW AI Factory, per Year

Maintenance, not power, is the largest yearly cost because it is charged on a capital base that is more than half GPUs. California's US$181M power bill, at US$0.205/kWh, is almost three times Texas's. For the full 10-year picture with the build cost included, use the prefab AI factory TCO calculator.

The AI Build Configurator shows every line for your site, with EBITDA and payback.

An 8-GPU NVIDIA B200 AI server costs about US$380,000 (C$536,000) fully loaded, and a 72-GPU GB300 NVL72 rack about US$4.95M. The configurator prices every accelerator with its share of CPUs, network cards, storage, chassis and rack-scale interconnect, so GPUs × unit cost gives the price of the whole box.

These sit inside published market ranges: about US$250,000 to US$320,000 for an 8-GPU H100 HGX system, US$400,000 to US$500,000 for an 8-GPU B300 system, and US$3.7M to US$6.5M for a GB300 NVL72 rack on analyst estimates. A 100 MW AI factory holds about 62,800 B200-class GPUs, or roughly 7,850 eight-GPU servers, which is why servers make up just over half of its C$7.88B build cost. Swap the GPU class and see the whole budget move in the free live demo.