Behind the meter: how AI campuses are buying power in 2026

Until recently, on-site generation at data centres meant backup diesel. In 2026 it increasingly means prime power: gas engines, turbines and fuel cells running continuously to supply the campus while the grid connection arrives, or instead of it. The shift is driven by grid waits of five to eight years in the US and UK, and by occupiers who will not sign leases against uncertain energisation dates.

This note looks at the deals that define the market, the economics, and what investors should look for.

The deals that define the market

The past 12 months have produced a run of gigawatt-scale orders, reported by Data Center Dynamics and company announcements:

Two points stand out. First, many headlines describe "turbines" when the equipment is in fact reciprocating engines. Second, the buyers are not utilities: they are developers, neoclouds and hyperscalers acting as their own power companies.

Why engines are winning the near term

Heavy-duty gas turbines offer the best efficiency at scale, but the order books are full. GE Vernova reported about 83 GW of gas-power backlog at the end of 2025, and large frames are effectively unavailable before 2030. Aeroderivative turbines are quoted at 12 to 40 months, depending on the source and model.

Gas engines, typically 2 to 20 MW per unit, can be delivered in 12 to 24 months, scale in modular blocks, hold their output better in hot climates and handle the rapid load swings of AI training clusters, often paired with battery storage. That combination of speed and modularity is why they dominate 2026 orders.

The economics

The published deal values give a rough sense of price per kW. Meta's El Paso project, at $473m for 366 MW, implies about $1,290 per kW. HD Hyundai's $676m for 1 GW for Corban implies about $680 per kW, likely for engine supply rather than a complete plant. Rabobank's benchmark for a fully installed reciprocating engine plant is $1,700 to $2,000 per kW, while BNEF puts new combined-cycle plant at about $2,157 per kW.

On the operating side, on-site generation typically produces electricity at $100 to $165 per MWh, against $90 to $95 for grid supply. For a 25 MW facility load, a $40 per MWh premium is about $8.8m a year. Set against a lease worth $3m a month for a 20 MW campus, energising a year early pays for that premium several times over.

The risks to underwrite

On-site power solves the timing problem but introduces others:

• Gas supply: pipeline capacity, pressure and contract terms. A generator without firm gas is not firm power.
• Permitting: air permits and emissions limits, which have become a gating item in several US states.
• Delivery and commissioning: slot dates, factory testing and the EPC schedule around them.
• Resilience: redundancy design (N+1 or more), maintenance windows and battery support for load swings.
• Long-term role: whether generation becomes backup once the grid arrives, or remains the primary supply.

What this means for investors

• Treat secured generation slots as an asset: ask for order confirmations, delivery dates and factory test plans.
• Confirm gas supply terms and capacity, not only the generator order.
• Check the air permit status and the emissions strategy.
• Model the power cost premium against the value of earlier energisation, using realistic slip scenarios.
• Look for sponsors who own the power plan, rather than depending on a third-party schedule they cannot influence.
• Nodus secures on-site gas generation for data-centre programmes and screens projects for power readiness before investors see them.