The economics of a data-centre megawatt

Most investors meet data centres as real estate: a building, a tenant and a lease. AI campuses behave more like infrastructure. The capital cost per megawatt is high, power is both the main input and the main risk, and the timing of energisation drives value more than any other variable.

This note sets out the unit economics as they stand in 2026, using published benchmarks, and then works through an illustrative 20 MW project. Figures labelled illustrative are assumptions for the example, not market quotes.

What a megawatt costs to build

JLL's 2026 benchmark puts construction at about $11.3m per MW of IT capacity, excluding IT equipment. The cost splits broadly into land and enabling works, the shell, electrical plant, mechanical and cooling systems, and fit-out. Electrical and mechanical systems typically account for the majority, and both rise with density: liquid cooling, higher-capacity distribution and more resilient backup all add cost per MW.

Where the grid cannot deliver on time, on-site generation is a second capital line. Rabobank estimates reciprocating gas engine plants at $1,700 to $2,000 per kW installed. BNEF puts new combined-cycle plant at about $2,157 per kW, up 66% since 2023. Fuel cells cost more, around $3,000 to $4,000 per kW, but can be deployed in about 90 days.

What the hardware demands

AI hardware sets the power envelope. A single GB300 NVL72 rack (72 GPUs) draws around 135 to 140 kW. Including cooling and distribution, roughly 1,000 GPUs of this class need about 2.6 MW of facility power. The next generation is expected to need around 5 MW per 1,000 GPUs, and rack-scale designs of 600 kW are on vendor roadmaps for 2028.

For investors, density matters in two ways. It raises capital cost per MW, and it shortens the useful life of facilities designed for yesterday's loads. A hall built for 10 to 20 kW racks cannot host current GPU systems without major electrical and cooling upgrades.

How revenue is earned

Data-centre revenue takes four main forms:

• Wholesale lease: a single occupier leases capacity, priced per kW per month, on terms of 10 to 15 years or more.
• Retail colocation: many customers lease racks or cages, at higher prices per kW and shorter terms.
• Powered shell or build-to-suit: the developer delivers power and building; the occupier fits out.
• GPU cloud: the operator owns the GPUs and sells compute by the GPU-hour or under multi-year capacity contracts.

Each carries a different risk profile. A hyperscale wholesale lease to an investment-grade tenant behaves like infrastructure debt. GPU cloud revenue carries technology, utilisation and pricing risk, but much higher revenue per MW.

Worked example: an illustrative 20 MW AI campus

Assumptions (illustrative, for the arithmetic only): 20 MW of IT capacity, wholesale lease at $150 per kW per month, PUE of 1.25, power at $95 per MWh from the grid or $130 per MWh on site.

1. 1.Build cost: 20 MW × $11.3m = $226m, excluding IT equipment.
2. 2.On-site gas engine plant for the full load: 20 MW × 1.25 PUE = 25 MW; with N+1 redundancy, about 28 MW installed × $1.85m = about $52m.
3. 3.Annual lease revenue: 20,000 kW × $150 × 12 = $36m a year, or $3m a month.
4. 4.Annual energy: 25 MW × 8,760 h = 219,000 MWh.
5. 5.Annual power cost premium for on-site generation: 219,000 MWh × ($130 − $95) = about $7.7m a year. In most leases this is passed through to the occupier, but it still affects competitiveness.
6. 6.Cost of delay: one month without energisation forgoes about $3m of revenue, while carrying financing on more than $226m of capital.

The conclusion is clear. If a grid connection is three years away and on-site generation can energise the site in 18 to 24 months, the 12 to 18 months gained are worth $36m to $54m in revenue alone, well above the annual cost premium and comparable to the generation capex itself.

The variables that move the model

• Energisation date: the single largest driver of value. Equipment slip and grid delay hit revenue directly.
• Power price: a $10 per MWh change on 219,000 MWh is about $2.2m a year at this scale.
• PUE: each 0.1 of PUE on a 20 MW IT load is about 17,500 MWh a year.
• Occupier credit and term: these set the cost of capital. A signed lease with a strong tenant transforms financing terms.
• Hardware generation: designing for 2026 densities only risks early obsolescence.

What this means for investors

• Ask for the energisation date, how firm it is, and what happens if it slips.
• Check that the capex budget includes generation, cooling and distribution sized for current GPU densities.
• Model the power cost on both grid and on-site cases, and confirm how it passes through the lease.
• Test the downside: six months of delay, a higher power price, and a weaker tenant.
• Value the equipment position: secured slots for generation and long-lead electrical plant reduce the largest risk in the model.
• Nodus screens opportunities for power, site and capital readiness before they reach investors.