Terrestrial Compute Total Cost of Ownership

What is the amortized TCO per kW_IT/year for terrestrial compute?

What is the amortized total cost of ownership per kW_IT per year ($/kW_IT/year) for terrestrial AI compute?

Answer

Terrestrial TCO ranges from 6,764 $/kW_IT/year (optimistic, 2026) to 17,493 $/kW_IT/year (conservative, 2030). The central estimate is approximately 10,762 $/kW_IT/year in 2026, remaining nearly flat through 2040 at 10,698 $/kW_IT/year. The remarkable stability of terrestrial TCO reflects the dominance of GPU hardware cost (which is time-invariant in the model) over the more volatile energy cost component.

Terrestrial TCO is composed of amortized GPU cost, amortized infrastructure cost, annual energy cost, and non-energy opex. GPU cost is the dominant component at roughly 69-77% of total TCO across the modeled scenarios and years.

Inputs

Analysis

TCO Formula

Terrestrial TCO = gpu_amortized + infra_amortized + power_capex_amortized + energy_annual + non_energy_opex

Where:

• gpu_amortized = gpu_cost_per_kw × CRF(terrestrial_wacc, gpu_useful_life)
• infra_amortized = terrestrial_infra_cost × CRF(terrestrial_wacc, 15 years)
• power_capex_amortized = terrestrial_power_asset_capex × CRF(terrestrial_wacc, 20 years)
• energy_annual = variable_energy_cost × 8,760 hours × PUE (fuel + O&M + grid procurement; BTM capex tracked separately)
• non_energy_opex = $500-$1,100/kW_IT/year (staff, maintenance, property tax, insurance; page-backed)

CRF (Capital Recovery Factor) converts one-time capex into an equivalent annual cost accounting for the cost of capital. At the central terrestrial WACC of 7%, CRF is larger than simple 1/n division — for example, CRF(0.07, 5) = 0.244 vs 1/5 = 0.200. This reflects the real financing cost of capital assets. See the terrestrial WACC page for the derivation.

Component Breakdown

GPU cost (amortized) is the largest single component:

GPU cost is time-invariant (same across all years) and dominates TCO in all scenarios. In the central case, GPU cost is 7,926 $/kW_IT/year, representing 74% of total TCO. This dominance is the fundamental reason why terrestrial TCO is so stable over time -- the largest component does not change.

Infrastructure cost (amortized) over a 15-year facility life is modest:

Infrastructure amortization is a small fraction of TCO (6-10%) because the 15-year facility life spreads the cost over a long period. This is a structural advantage of terrestrial deployment: facilities outlast the GPU hardware they house, allowing multiple generations of compute to amortize the same building.

Power-asset capex (amortized) captures the capital cost of behind-the-meter generation:

Power-asset capex is a small fraction of TCO (0.2-0.4%) but makes explicit the capital investment required to achieve the low energy rates assumed in BTM-heavy scenarios. The optimistic scenario has the highest power-asset capex because it assumes the most aggressive BTM deployment (70% by 2040).

Note: The optimistic scenario has the highest power-asset capex because it assumes the most aggressive BTM deployment, which lowers variable energy cost but raises generation capex. The net effect on total TCO is positive (lower total cost). See the power-asset capex page for the full explanation.

Energy cost (variable) represents fuel, O&M, and grid procurement — excluding BTM generation capex which is tracked separately:

Variable energy cost plus amortized power-asset capex together represent the total cost of energy. In the central case, total energy (variable + power capex amortized) is approximately $728/kW_IT/year in 2030 — about 8% of total TCO. This is consistent with the widely cited observation that energy is "only 5-15% of total data center cost." The low share of energy in TCO is the single most important finding for the orbital comparison: orbital compute's primary advantage (free solar power) eliminates only ~8% of terrestrial costs.

Non-energy opex covers staffing, maintenance, property tax, and insurance — now a page-backed input rather than a hardcoded assumption:

Non-energy opex varies by scenario to reflect differences in labor markets, tax jurisdictions, and automation levels. It represents 6-9% of total TCO across scenarios.

Total TCO Trajectories

The trajectories are remarkably flat. The central estimate varies by only ~$50/kW_IT/year across the entire 2026-2040 period (less than 1% variation), driven by small fluctuations in energy cost. The optimistic trajectory declines modestly as energy costs fall; the conservative trajectory shows a slight hump as energy costs peak around 2030 before declining.

Why Terrestrial TCO Is Hard to Reduce

The dominance of GPU hardware cost creates a structural floor for terrestrial TCO. Even with zero energy costs and zero infrastructure costs, terrestrial TCO could not fall below GPU amortization + non-energy opex:

• Optimistic: 4,925 + $500 (non-energy opex) = ~$5,425/kW_IT/year
• Central: 7,926 + $750 = ~$8,676/kW_IT/year
• Conservative: 12,619 + $1,100 = ~$13,719/kW_IT/year

Actual terrestrial TCO is only 15-25% above this theoretical floor, indicating that infrastructure and energy are secondary cost drivers. This makes terrestrial compute surprisingly cost-efficient relative to its theoretical minimum.

Sensitivity Analysis

The three most impactful parameters are:

1. GPU cost per kW_IT (highest sensitivity): A 20% change in GPU cost shifts central TCO by ~$1,300/kW_IT/year (15% of total). This dwarfs all other inputs.

2. GPU useful life (moderate sensitivity): Extending useful life from 5 to 6 years reduces central TCO by ~$1,083/kW_IT/year (12%). This is why hyperscalers care deeply about depreciation schedules.

3. Energy cost (low sensitivity): Doubling energy cost from $0.075 to $0.15/kWh increases central TCO by only ~$723/kW_IT/year (8%). Energy is a small lever.

Infrastructure cost and PUE have even lower sensitivity due to the long facility amortization period and already-low PUE values of modern liquid-cooled facilities.