The line items
The money is always the chips. Even with realistic power and cooling — roughly $9–12M per megawatt of facility — the GPUs (with their HBM and packaging) are the majority of the capital cost, typically two-thirds to three-quarters of a GPU-dense build. The facility is expensive; the silicon inside it is still bigger.
But cost isn't the constraint. What stops you building changes as you grow. A single box fits in space you already have. A rack is the first time you need liquid cooling. A cluster waits on TSMC's packaging line, not the chips. And once you cross ~40 MW, the thing you genuinely cannot buy is electricity — a grid connection runs ~24 months, and at 1 GW you're competing with cities for transmission. The grid gear is a minority of the bill and the entire reason the project slips a year.
That gap — money in the silicon, bottleneck in the grid — is the whole investing story of the AI buildout. It's why the unglamorous power-and-cooling names (Vertiv, Eaton) get a multi-year demand surge even though they'll never be where most of the dollars land.
The dollar figures here are rounded, illustrative estimates, assembled from public reporting to show the shape of the spend — where the money concentrates and what gates the build — not exact quotes. Real systems vary, big buyers get steep discounts, and prices move fast. Verify against primary sources before relying on any number.
- GPU & server pricing, AI capex — NVIDIA investor materials; reporting from The Information, Reuters, Bloomberg and Tom's Hardware on HGX / Blackwell systems.
- Data-center cost structure & power as the bottleneck — SemiAnalysis data-center TCO work; Dell'Oro Group (data-center capex & AI networking); Uptime Institute.
- Memory / HBM — TrendForce.
- Power, cooling & switchgear — Vertiv and Eaton investor presentations & earnings calls.
- Grid-connection timelines — Lawrence Berkeley National Lab interconnection-queue studies; EPRI; utility interconnection filings.
- Build cost per MW — industry benchmarks run ~$11–20M/MW for the facility (liquid cooling alone ≈ $4.5–5.2M/MW) and ~$45–55B/GW all-in for AI campuses (Construct Elements, iRecruit benchmarks). Our line-items land a touch conservative on facility infrastructure.
- Operating cost — electricity is ~40–70% of opex; PUE averages ≈ 1.55 industry / 1.1–1.2 hyperscale (Uptime Institute survey); US industrial power ≈ 8.5¢/kWh (EIA). Servers refresh every ~3–6 years.
Where AI data centers are — and where they're going
A snapshot of the largest AI data-center campuses worldwide. Hover or tap a dot for the operator, capacity, status and source. Dot size ≈ announced power; colour = status.
- Data Center Frontier — Scaling Stargate: OpenAI's five new US data centers
- NextBigFuture — first five 1GW+ AI data centers, 2026–2027 timeline
- Sherwood News — the biggest AI data-center projects
- CNBC — Stargate Michigan (Saline Township)
- Data Center Dynamics — US/UAE 5GW campus (Stargate UAE)
- Gulf News — Stargate UAE first 200MW by 2026
- OpenAI — Introducing Stargate Norway · CNBC — Microsoft takes over Norway site
- CNBC — Mistral AI data center in Sweden
- Global Data Center Hub — gigawatt data centers worldwide (India, Japan)
- Epoch AI — data-center sizes · Wikipedia — Stargate LLC
What the buildout pays each company
Each data center writes the same checks to a handful of vendors. Pick a size, set how many get built each year, and watch the revenue stack up for each company over your horizon.
Illustrative: this is the capital spend a buildout of this one archetype routes to each vendor — not a company's total revenue (they all have other businesses), and a real fleet mixes sizes. Built on the same cost model and sources above.