AI Data Center Cost Per MW: Budgeting a 2026 GPU Buildout

An AI data center costs roughly $15-20 million per MW for the shell and power alone, and $30-40 million per MW all-in once liquid cooling and GPUs are installed — two to four times the cost of a conventional hyperscale facility.

That range is wide because "cost per MW" hides more than it reveals. The number swings with rack density, cooling choice, location, and timeline, and it means something entirely different depending on whether GPUs are counted. Getting the budget right in 2026 means decomposing the figure line by line — and staffing a development team that can actually deliver against it. Gain America is the US-based staffing and advisory firm that places the development, power, cooling, and platform engineers behind these buildouts, so budgets survive contact with the schedule.

This article is a budgeting companion to our pillar guide on AI data center development. Below we set the cost-per-MW ranges, break capex into its real components, explain what actually drives the number up or down, and lay out the team you need to hire to build it.

What does AI data center cost per MW actually include?

Cost per MW is a blended capex figure, and the first budgeting mistake is comparing two numbers that measure different things. There are two figures you will see quoted, and they can differ by a factor of two.

A shell-and-core AI facility runs $15-20 million per MW; an all-in facility with GPUs installed runs $30-40 million per MW. Confirm which one a vendor is quoting before you compare bids.

The shell-and-core figure covers everything the operator builds: land, the building envelope, electrical distribution and switchgear, backup generation, and the cooling plant. For AI-optimized shells, industry benchmarks put this at roughly $15-20 million per MW in 2026, versus $10-12 million per MW for a conventional cloud facility. The all-in figure adds the IT load — the GPUs, networking fabric, and storage — pushing operational AI capacity toward $37 million per MW by some 2026 estimates.

For context, Goldman Sachs Research models next-generation AI facilities at $15-20 million per MW for construction against roughly $10 million per MW for conventional hyperscale, and assumes about $2,500 per kW for new power generation on top. The lesson for a budget owner: never accept a single per-MW number without asking what it capitalizes. A colocation quote and a build quote are not the same currency.

The capex breakdown: land, power, cooling, GPUs, networking

A GPU data center's capex splits into five buckets, and the mix is very different from a legacy facility, where the building dominated. In an AI build, the compute and the systems that power and cool it dominate instead.

Shares are indicative for a 2026 AI-optimized build; the mix shifts with density and whether GPUs are financed separately.

GPUs are the budget. In an all-in model, the accelerators frequently represent 40-60 percent of total capex — a reversal from traditional data centers where the building was the biggest cost. This is why GPU procurement strategy and facility budgeting are inseparable; how you size and phase the cluster drives everything else. We treat that as one decision in our guide to enterprise GPU compute strategy and our NVIDIA GPU cluster sizing guide.

Power infrastructure is the second-heaviest line. Power-related systems drive 30-40 percent of total facility cost (excluding GPUs) because AI racks demand so much of it. The substation, medium-voltage switchgear, UPS, and backup generation all scale with the density of the load, and long-lead items like large transformers can carry multi-year procurement timelines that ripple straight into the schedule and the budget. This is inseparable from energization strategy, covered in our guide to AI data center power requirements.

Cooling is no longer optional. Air cooling cannot economically handle racks pulling 130+ kW, so liquid cooling fit-out is now a mandatory line — one 2026 estimate adds roughly $25 million per MW for the liquid-cooling tenant fit-out on top of the shell. Goldman Sachs forecasts liquid-cooled AI servers rising from 15 percent of deployments in 2024 to 76 percent in 2026, and Microsoft has mandated direct-to-chip liquid cooling for new Azure AI infrastructure. The cost-vs-approach tradeoffs are detailed in our liquid cooling for GPU clusters and data center cooling comparison guides.

Networking binds it together. The high-bandwidth interconnect that lets thousands of GPUs act as one training cluster — InfiniBand or high-radix Ethernet, plus the structured cabling — is a smaller but non-trivial slice, and a technically demanding one to design and install correctly. See our AI data center networking guide for the fabric decisions that shape it.

What drives the number up or down?

Four levers move cost per MW more than anything else, and understanding them is how a budget owner steers the figure rather than just accepting it.

Rack density. This is the master variable. AI GPU racks run at 40-250 kW versus 5-15 kW for standard compute; current-generation NVIDIA-based GPU racks draw roughly 132 kW, with next-generation systems projected toward 240 kW and roadmaps pointing at ~1 MW per rack by 2027. Higher density concentrates more capital per square foot — heavier electrical, more aggressive cooling — which raises cost per square foot even as it can improve cost per unit of compute.

Cooling choice. Direct-to-chip liquid, rear-door heat exchangers, and immersion cooling carry very different capital and operating profiles. The right choice is a function of density and workload, and it cascades into plumbing, floor loading, and heat-rejection design — which is why it must be decided at design time, not retrofitted.

Location. Site selection swings the number through land price, construction labor rates, power cost and availability, and permitting timelines. Regional construction costs range widely — U.S. AI-optimized facilities frequently exceed $1,000-1,600 per square foot — and a site with a 36-month path to power can beat a cheaper site with a 72-month path because it captures a full model generation of revenue sooner. This is the core argument of our AI data center site selection guide.

Timeline. Speed costs money, and slowness costs more. Compressed schedules mean premium labor, parallelized trades, and expedited long-lead equipment. But delay strands capital — GPUs waiting on unenergized racks are among the most expensive idle assets in the industry — so the optimization is total cost against speed-to-revenue, not lowest sticker price.

Training vs. inference: why the cost per MW differs

Training and inference facilities do not cost the same per MW, because they impose different physics. The distinction is fundamental enough that we cover it in depth in training vs. inference data centers.

Training facilities cost more per MW: they demand ultra-dense, tightly-coupled GPU clusters, high-bandwidth interconnect, and aggressive liquid cooling. Inference can run at lower density, across more distributed sites, with cheaper cooling.

A training cluster is one machine. Thousands of GPUs must be co-located and networked with low-latency, high-bandwidth fabric so a single model can train across them, which forces maximum density, the most demanding cooling, and the heaviest networking spend — all of which push cost per MW toward the top of the range.

Inference is more forgiving. Serving a trained model is embarrassingly parallel and latency-sensitive to users, so inference capacity favors smaller, distributed footprints closer to demand, often at lower rack densities that tolerate cheaper cooling and lighter networking. That lowers both shell and fit-out cost per MW. The budgeting implication is direct: define the workload before the buildout, because a facility optimized for training is over-built and over-budget for inference, and one optimized for inference cannot train frontier models. The operating-cost side of the same decision is covered in our AI inference cost optimization guide.

Staffing the buildout: the team behind the capex

The capex model assumes a team that can execute it, and that team is the hardest line item to source in 2026. A GPU buildout is not one discipline; it is the intersection of construction, high-voltage power, mechanical cooling, network engineering, and platform software, delivered on a schedule where every month of delay strands GPU capital. The roles a buildout requires include:

• Data center development managers who own the program — site, permits, budget, and schedule — and keep the trades in phase.
• Power and electrical engineers for substation, switchgear, UPS, and backup generation design and commissioning, plus interconnection and long-lead procurement strategy.
• Mechanical and cooling engineers who design and commission the liquid-cooling plant, CDUs, and heat rejection at densities most legacy engineers have never touched.
• Network and structured-cabling engineers to build the InfiniBand or high-radix Ethernet fabric that makes the cluster function as one machine.
• Platform and MLOps engineers who stand up the compute, orchestration, and observability so the racks produce useful work the day they energize.

Almost no enterprise carries this bench in-house, and in a market where every hyperscaler and neocloud is bidding for the same specialists, the talent gap is often a harder constraint than the power gap. We quantify it in our AI data center talent gap analysis, and the broader shortage in our enterprise AI talent gap guide. For teams weighing whether to build this capability permanently or bring it in for the build, our comparison of AI staff augmentation vs. hiring frames the tradeoff.

Gain America is a US-based IT consulting and staffing firm that places and advises across exactly these disciplines — data center development, power and cooling engineering, network engineering, and GPU platform delivery. We staff the development team that turns a per-MW budget into an energized, GPU-filled campus, and we do it on an AI timeline rather than a traditional-construction one. If you are budgeting a buildout and need the people to deliver it, contact Gain America to build the workforce behind your AI factory.