ThisCalculator

Everyday

Estimate AI Data Center Build Cost

Updated · Standard formula, shown below

Estimate AI Data Center Build Cost

This guide walks you through estimating the total capital cost of building an AI data center, from land and shell construction to the specialized power, cooling, and GPU clusters that dominate the budget. You’ll end with a realistic per-megawatt and total figure you can use for planning or budgeting.

What You’ll Need

  • Target IT load in megawatts (MW) — the power your GPUs will actually draw.
  • Number and type of GPUs (e.g., H100, A100) and their typical power draw per unit.
  • Local construction cost per square foot or per MW (varies by region).
  • Utility interconnection costs or quotes from your local power provider.
  • Cooling system type (air vs. liquid) and associated cost factor.
  • Land cost or lease rate for your chosen site.
  • Contingency percentage (usually 10–20%).

Step-by-Step

  1. Define the IT load. Decide the total power your GPU racks will consume. For example, 1000 H100 GPUs at 700W each = 700 kW of IT load. Add 10–15% for network switches and storage.
  2. Calculate the facility power overhead. Cooling, power distribution, and lighting typically add 30–50% on top of IT load. Multiply IT load by 1.3 to 1.5 to get total facility load. For 1 MW IT, that’s 1.3–1.5 MW total.
  3. Estimate shell and core construction cost. Use a per-square-foot or per-MW figure. In the US, shell and core for a data center runs roughly $800–$1,200 per square foot, or $8–$12 million per MW of IT load. Multiply by your IT MW.
  4. Add power infrastructure. Include transformers, switchgear, generators, UPS, and electrical distribution. This typically costs $3–$5 million per MW of IT load. If you need new substation or long utility lines, add $1–$3 million per mile.
  5. Add cooling system. Air-cooled systems cost about $2–$3 million per MW; liquid cooling (direct-to-chip or immersion) runs $3–$5 million per MW. Choose based on your GPU density.
  6. Include the GPU cluster cost. Multiply the number of GPUs by the unit price (e.g., $25,000–$40,000 for H100-class). Add 20–30% for servers, networking (InfiniBand or Ethernet), and storage.
  7. Add land and site work. Land cost varies wildly. Use your local price per acre (typically 5–20 acres needed). Add site preparation, fencing, and security.
  8. Sum all components. Add shell, power, cooling, GPU cluster, and land. This is your base estimate.
  9. Apply contingency. Multiply the base by 1.15 (15% contingency) to cover unforeseen costs.
  10. Compute per-MW cost. Divide the total by your IT MW to get a per-MW figure. This is useful for comparing to industry benchmarks (typically $10–$20 million per MW including GPUs).

Common Mistakes

  • Confusing IT load with total facility load — forgetting overhead leads to under-sizing power and cooling.
  • Using GPU list price without adding server, networking, and storage costs.
  • Ignoring utility interconnection fees, which can be substantial.
  • Underestimating cooling costs for high-density AI racks that require liquid cooling.
  • Not including contingency or escalation for construction inflation.

Tips

  • Use a range, not a single number — costs vary by region and supply chain.
  • Check current GPU prices; they fluctuate with demand.
  • Get quotes from at least two general contractors for shell and core.
  • Consider colocation vs. building — sometimes leasing is cheaper for a pilot.
  • Factor in operating costs (electricity, staff) separately from capital cost.

Troubleshooting

If your estimate seems too high or low, check these:

  • Recheck the IT load. Did you include all GPUs and their actual TDP? Did you add network gear?
  • Verify the overhead factor. For high-density liquid cooling, overhead can be lower (20–30%) than for air (40–50%).
  • Compare with industry benchmarks. If your per-MW cost is outside $10–$20 million, review your assumptions.
  • Get a professional cost estimator. For a real project, engage an engineering firm early.

Frequently Asked Questions

What is the biggest cost in an AI data center?

Typically the GPU cluster itself, often 50–70% of total capital cost. Power and cooling are next.

How much does a 1 MW AI data center cost to build?

Roughly $10–$20 million including GPUs, depending on location and cooling type. Excluding GPUs, it’s about $5–$8 million.

Does liquid cooling cost more than air cooling?

Upfront, yes — liquid cooling adds $1–$2 million per MW. But it allows higher density and lower operating costs, which can offset the initial premium.

How long does it take to build an AI data center?

Typically 12–24 months from design to operation, but utility interconnection can add months or years.

Can I use a rough cost per square foot?

Yes, but per-MW is more reliable because AI data centers are power-dense. Per square foot varies widely with density.

What is a reasonable contingency?

10–20% is standard. For AI data centers with new technology, use 20% to be safe.

Related

Method: the standard formula for this calculation, rounded as each result states. See how the calculators work or report an error.