TCO & Capital

TCO Watch

The Real Economics of AI Infrastructure

Hardware is the minority of cost. Power geography, depreciation and utilisation dominate the five-year total cost of ownership — hardware is typically only 30–50% of the all-in five-year figure, with the balance in power, cooling, networking and operations.

Model TCO on goodput and cost-per-million-tokens. Whether to rent or build depends on sustained utilisation, workload profile and power access — there is no universal break-even threshold. Idle GPUs are the largest avoidable waste.

30–50%

Hardware share of 5-yr TCO

CUDO calculation CALC-001

$50.8k–$300k+

Monthly power, 1,000 GPUs

By region · CUDO calculation CALC-002

$176B

Understated depreciation

From 5–6 yr hyperscaler schedules

-99.7%

Token price collapse

Since the GPT-3 era

5%

Enterprise GPU utilization

95% of silicon investment sits idle

~$1.15tn

Hyperscaler capex 2025–27

Goldman Sachs estimate (CLM-TCO-006)

Breakdown

Where the five-year money goes

Five-year TCO components

Hardware is only 30–50% of the total

Power cost geography

Monthly electricity for a 1,000-GPU cluster ($ thousands)

Rent vs. buy by utilization

Relative TCO index (cloud = 100)

Global AI infrastructure market

$B · $90B (2026) to $465B (2033)

Analysis

The economics that decide

Across a five-year ownership horizon, hardware acquisition is an estimated 30–50% of total cost of ownership (CUDO calculation CALC-001); power, people and operations dominate the rest. Power & cooling are 15–25% (cooling alone ~32% of infrastructure spend), personnel 20–30% (a GPU infra engineer is ~$275,000/year), and maintenance 10–15%. A single all-in five-year figure is held pending a reproducible scenario.

Turn these signals into a placement decision

Discuss power, location, cluster architecture and TCO with a CUDO AI infrastructure specialist.