Deep|AIDC: Answering Six Key Investor Questions

After we published our AIDC report on 27 August, investors questioned our conversion of turbine nameplate capacity into data-center power. We had estimated that AIDC could claim roughly 18 GW of large gas-turbine nameplate in 2027, but investors felt our haircut to delivered power was too aggressive. The published version adjusted grid-connected units for derating, engineering readiness and PUE. For genuinely off-grid, behind-the-meter (BTM) units, we applied the engineering-readiness adjustment and then divided by an over-provisioning multiple of 1.9–3.0×. That multiple is the focus of this follow-up.

We conducted another round of channel checks and rebuilt the coefficients linking IT load to nameplate capacity for off-grid projects. The tables identify the source of each input so readers can reproduce the calculation.

We address six questions: why off-grid designs require 2.4–3.0× provisioning when nameplate supply appears ample; whether the derates are too severe; whether gas engines and batteries reduce the multiple; why a large backlog does not translate into near-term delivery; whether Chinese equipment can fill the gap; and whether Europe and Southeast Asia can absorb demand displaced from the US.

#1 Why does an off-grid data center need 2.4–3.0 MW of turbine nameplate per MW of IT when turbine supply looks ample?

The pushback: Investors raise three objections. First, natural gas supplied about 41% of US utility-scale generation in 2025, the largest single source. Combined-cycle plants are the mainstay; the EIA reference cited here is a roughly 620 MW new-build NGCC plant in a 2×1 configuration. Investors compare that plant’s nameplate with its output to the grid, without applying a campus-level redundancy multiple. Second, AI data centers have been built for several years without such a ratio being discussed. Why does a 3× multiple appear now? Third, investors’ own counts of OEM capacity and order books suggest substantial turbine nameplate deliveries in 2027. They therefore question whether our conversion assumptions overstate the shortfall.

Our view: Our off-grid BTM sizing multiple is based on inputs from gas-turbine OEM channel checks. Those sources put the multiple at 2.4–3.0×, with a central range of 2.5–2.7×. It applies to the off-grid designs discussed here — it cannot be used to discount the whole market’s turbine nameplate.

Why we think so: Grid-connected plants and off-grid campuses have different obligations for maintaining supply when equipment is unavailable.

A grid-connected CCGT can rely on other generation to help serve customers during its maintenance outage, subject to system reserve and transmission constraints. Its own output still falls to zero while it is offline. The cited roughly 15% reserve margin is a system-level planning measure, not evidence that each plant delivers its full nameplate at all times.

An off-grid campus must provide its own reserve. In a simplified example with 200 MW of site load and four 60 MW turbines, one unavailable unit leaves only 180 MW before site derating and auxiliary loads. Additional units are needed to maintain supply. They count toward installed nameplate and provide power when other units are unavailable, even if they run less often in normal operation.

The distinction is between a generator’s rated output and the capacity needed to support an uninterrupted IT load. Our 100 MW IT example below requires 2.4 MW of installed turbine nameplate per MW of IT under N+1. Grid-connected nameplate also needs adjustment for site conditions and net output, but plant-level redundancy is shared through the wider system.

“Ample supply” refers to total turbine nameplate, much of it destined for grid-connected plants. The 2.4–3.0× multiple refers to installed capacity per MW of IT in the off-grid designs discussed here. Those are different measures.

Meta’s Hyperion campus in Louisiana is planned to receive power from three Entergy combined-cycle plants totalling 2.26 GW, serving roughly 2 GW of site load through nearly 100 miles of new 500 kV transmission lines. The approximately 1.4× nameplate-to-IT ratio requires a separate site-load-to-IT conversion. The assets belong to the utility, reliability is supported by the grid system, and Meta is buying grid power. Our channel checks indicate that large frame machines are flowing predominantly to utility CCGT projects able to accommodate 2028–2030 delivery slots. Off-grid developers seeking earlier power are buying aeroderivatives, simple-cycle units and reciprocating engines. Our earlier report therefore treated the two paths separately and applied the over-provisioning multiple only to the 30–40% assumed to be off-grid. Aggregating both types of turbine without distinguishing their end use can overstate the capacity available to off-grid IT loads.

Most earlier AIDC developments relied on grid power, including electricity generated by gas plants. The specific sizing issue examined here arises when a campus must sustain its IT load with its own generation. BTM describes equipment behind the customer’s meter; it does not, by itself, mean that the campus is off-grid.

#2 Are the adjustments from nameplate capacity to IT load too severe?

The pushback: Investors argue that each input has a less conservative alternative, and that combining conservative choices inflates the multiple. A 60 MW machine should be counted at 60 MW if hot-day derating affects only a minority of hours; PUE could use the roughly 1.1–1.15 disclosed by Google and Amazon; and N+1 may be sufficient without N+2. On that reading, the 2.4–3.0× range reflects conservative assumptions rather than a general engineering requirement.

Our view: The coefficients come from vendors’ customer-sizing work, where output and availability carry contractual obligations. Our bottom-up calculation produces 2.4× under N+1 and 3.0× under N+2. Expert 1 also reports an internal reliability-tool result of approximately 2.4× for “N+1 plus six nines.” A second turbine OEM describes a developer provisioning 3 GW of installed capacity for 1 GW of IT with a 99.999%+ availability target. These channel observations provide separate checks on the order of magnitude; their assumptions need to be compared before treating them as equivalent designs.