702 MW. About $4 billion. Roughly 2,160 refrigerator-sized boxes. On October 14, AEP said it will build a 702 MW fuel cell plant near Cheyenne, Wyoming, using Bloom Energy solid oxide fuel cells (SOFC), with Google buying the output under a 20-year PPA and full operation targeted by the end of 2028 (Data Center Feed). The same month, Oracle reaffirmed up to 2.45 GW of Bloom capacity at Project Jupiter (Yahoo Finance), and Bloom said the first power block at Nebius's 300 MW New Jersey site was ready five months after signing (Blockspace).
Headlines quote one number. Engineers have to size four: how much to overbuild, how much gas it eats, how efficiency fades, and how to collect hundreds of megawatts from 480 V boxes. Here is that math.
1. SOFC overbuild: why 250 MW guaranteed needs 328 MW installed
A Nebius filing puts its Bloom contract at about 250 MW guaranteed and about 328 MW installed (Blockspace). That ratio is the first sizing rule of an islanded fuel cell plant:
Overbuild ratio = P_installed / P_guaranteed = 328 / 250 = 1.31 → about 31% extra
Why so much? With no grid behind it, the plant must still hit its guarantee while some power modules are out for service, while every stack is slowly degrading, and while a whole server trips. The 31% is the price of running behind the meter. For scale, at Bloom's 325 kW Energy Server rating (Bloom datasheet), 328 MW is about 1,009 servers, and AEP's 702 MW would be about 702,000 / 325 ≈ 2,160 servers if that figure is nameplate.
2. Heat rate math: how much gas does a 702 MW SOFC plant burn?
SOFCs don't combust the gas, but they still consume it. Bloom's datasheet gives a heat rate of 5,811 to 7,127 Btu/kWh (HHV) across stack life. Daily fuel at full output:
Fuel/day = P × 24 h × HR
= 702,000 kW × 24 h × 5,811 to 7,127 Btu/kWh
≈ 97,900 to 120,100 MMBtu/day ≈ 94 to 116 MMcf/day (at ~1,037 Btu/scf)
That is a dedicated transmission-pipeline customer, not a city gas tap. Two quick follow-ons fall out of the same number. Fuel cost per MWh is just heat rate times gas price: at an assumed $3.50/MMBtu, that's about $20 to $25/MWh before capital and service. And CO2 at Bloom's stated 308 to 378 kg/MWh, run flat out for a year (6.15 million MWh), is roughly 1.9 to 2.3 million tonnes, which is why AEP is talking about adding carbon capture.
3. SOFC degradation: 65% to 53% efficiency, and the 5-year module cycle
Bloom specifies cumulative electrical efficiency of 65% falling to 53% (LHV, net AC) over stack life. Fuel per kWh scales with 1/η:
Fuel increase at end of life = η_BOL / η_EOL = 65 / 53 ≈ 1.23 → ~23% more gas per kWh
The cause is resistance growth in cells and interconnects; Bloom's own SEC filing calls it area-specific resistance degradation and says modules are swapped on a schedule, not after sudden failure (SEC exhibit, 2019). Its 2024 impact report projects a fleet median time to refurbishment of 5 years (Bloom 2024 report). Over a 20-year PPA, that's about 20 / 5 = 4 module generations, or three full swap-outs. The same filing put one cell at ~25 W, so 702 MW is on the order of 28 million cells to keep in rotation.
A fuel cell plant isn't sized once. It's sized for its worst day: oldest stacks, highest heat rate, a few modules out for service. That's where the ~31% overbuild and the 7,127 Btu/kWh gas contract come from, not from the brochure's 65%.
4. From 480 V to 38 kV: the current math behind SOFC switchgear
Each Bloom server outputs 480/415/400 V three-phase AC. Per-server current at unity power factor:
I = P / (√3 × V) = 325,000 / (1.732 × 480) ≈ 391 A
Collecting 250 MW at 480 V would mean hundreds of thousands of amps. So the plant steps up. LS Electric's $138.5 million Bloom order is for 38 kV switchgear and distribution transformers at a Pennsylvania data center (DCD). At 38 kV, 250 MW is about 250e6 / (1.732 × 38,000) ≈ 3,800 A total, and 702 MW is about 10,700 A, split over many feeders. The payoff is in conductor loss, which scales with 1/V²:
Loss ratio (same conductor, same power) = (38,000 / 480)² ≈ 6,270× lower
That's why fuel cell farms look like substations with boxes attached. If you want the bigger picture on SOFC versus nuclear for AI loads, I covered it in this video: AI's Power Crisis: SOFC vs SMR Nuclear. More SOFC posts on this blog are here.
What to take away about SOFC plant sizing
When you see "X MW of fuel cells," ask four questions: what's guaranteed versus installed, what heat rate the gas contract assumes, how often modules get swapped, and what voltage the collection system runs at. Speed is the real selling point here, with five months from signing to first power at Nebius. But the engineering bill shows up as a 30% overbuild, a pipeline-scale gas supply, and a medium-voltage substation at every site.
