For a 1 GW AI hyperscale data center, servers (chips, board, chassis, in-rack fabric) absorb $5.0 B/yr, or 60% of an $8.5 B/yr total cost of ownership when CapEx is annualized over asset life [S2].
Facility shell, mechanical and electrical infrastructure add $1.4 B/yr (16%), while energy, the largest OpEx line, runs $594 M/yr, or 7% of TCO [S2]. Independent analyst commentary pegs chips plus networking at roughly 80% of lifetime cost, with everything else as rounding error [S4].
Up-front CapEx split: $38 B for 1 GW, two thirds to servers
Epoch AI's stylized 1 GW US hyperscaler model prices up-front CapEx at $38 B, of which servers are $21 B (55%), facility $12 B (32%), network infrastructure $4.6 B (12%), with utility works and land under $40 M combined [S2]. The model assumes 1 GW of nameplate IT load and converts CapEx to annual cost using each asset class's lifespan and a cost-of-capital discount rate.
The implication for sourcing: a single 1 GW site consumes roughly 4,000 to 6,000 MW of facility and electrical build for every 1 MW of IT load, an overbuild ratio that drives the medium-voltage transformer, switchgear and busway lead times now stretching into 2027-2028 across US hyperscalers. The $21 B server line alone, at an Nvidia DGX H100 list near $269,000 per 8-GPU system, corresponds to tens of thousands of GPU nodes per GW-class campus, matching Masley's order-of-magnitude estimate [S4].
Chips plus networking sit near 80% of lifetime cost
Semiconductors and the in-rack and spine-leaf fabric that ties them together account for about 80% of total AI data center cost including power, with non-chip costs likely falling as campuses scale and chip costs flat or rising [S4]. The breakdown puts the next tier as facility construction (skilled labor, building shell, energy and cooling plumbing), then electricity, then everything else.
On the same 1 GW TCO, network infrastructure is $1.17 B/yr (14%), facility is $1.39 B/yr (16%), and energy is $0.59 B/yr (7%); maintenance ($120 M), taxes ($143 M) and labor ($40 M) together add under 4% [S2]. This concentration matters for capital planners: a 20% swing in GPU pricing moves the site's lifetime cost by roughly 12%, while a 20% swing in electricity moves it under 2%.
What "building and power" actually costs
Facility CapEx (shell, mechanical, electrical) is the second-largest line at $12 B, or about one third of the $38 B up-front bill [S2]. Within facility, the dominant cost drivers are medium-voltage transformers, generator yards, UPS strings, chillers or liquid-cooling distribution, and the structural slab; civil and land are a small slice. Land runs $13 M, or 0.2% of site cost, and a 100 MW site fits on roughly 80 acres at 0.81 acres/MW [S4].
Power OpEx, often discussed as the next crisis, is $594 M/yr on a 1 GW site, or 7% of TCO and 65% of total OpEx [S2]. US data center load is forecast to climb from roughly 30 GW in 2025 to 90 GW or more by 2030 (22%/yr), and AI's share of total data center power rises from 14% in 2025 to 27% by 2027 [S3]. Even at that growth, the dollar share of electricity in TCO is bounded by the underlying split.
What is not in the 60% number
Epoch's 60% server share counts only the IT equipment rack, not networking; network is broken out separately at $1.17 B/yr [S2]. If network is folded in, IT stack reaches roughly 74% of annualized TCO, consistent with the 80% chip-plus-network figure cited independently [S4]. The remaining 20-25% covers everything that keeps the chips running, which is exactly the busway, transformer, power quality analyzer, digital multimeter and metering chain that the B2B catalog is built around.
Lead times on dry-type transformers are now the binding constraint at many US campuses, with hyperscaler orders crowding utility and renewables demand; see the Data Center Construction Overtakes Office and Retail Spend in 2026 read for the capex cycle framing.
Selection implications for non-IT equipment buyers
If a project is sized for a 1 GW campus at $38 B up-front, the addressable non-IT spend is roughly $13-17 B per site: facility $12 B, utility works $20 M, plus the network fabric (cabling, optics, switches) that often sits with electrical contractors [S2]. For an equipment vendor, that translates to a few hundred million to over a billion dollars of transformer, switchgear, UPS, busway, chiller and data logger content per GW build, depending on liquid versus air cooling mix and on-site generation share.
Three signals to track through 2026-2027: (1) DGX-class system pricing, which dominates the 60% server line and is the single biggest cost swing variable; (2) US hyperscaler power capex, projected to push campus power capacity from 30 GW in 2025 toward 90 GW by 2030 [S3]; and (3) facility-cost-per-MW trend, which Masley argues is declining as campuses scale, even as chip dollars per site climb. For most B2B buyers outside the GPU supply chain, the practical headline is that you are selling into a smaller share of a much bigger pie, and the binding constraints are lead time and grid interconnection rather than dollar budget.