The question this piece answers: if a fuel cell's electricity is more expensive per kWh than a combined-cycle gas turbine, why are data centers signing gigawatts of it? Because the two products are not priced on the same axis. One sells energy; the other sells a start date. The evidence is in the seller's own income statement and in the customer announcements — both of which are primary sources, published by the companies involved, and both re-checked on 2026-11-17. Where a number is our own arithmetic it says so; where a claim rests on a single source, the piece says that too.

What is a fuel cell actually selling?

A solid-oxide fuel cell (SOFC) generates electricity by electrochemical reaction rather than combustion: oxygen ions cross a solid ceramic electrolyte at roughly 500–1,000 °C, and the resulting current is drawn off directly. There is no flame, no turbine, no steam cycle. That single architectural fact produces the commercial ones. It runs quietly, it produces very little local air pollution, it needs minimal water, it is modular — and, crucially for 2026, it can be delivered and commissioned in months rather than years, because it arrives as factory-built modules instead of a constructed power plant.

Bloom Energy's own product framing makes the same point. The company's November 2026 announcement with NTT DATA describes fuel cells as "ready-to-deploy" and "modular systems [that] can be deployed rapidly and expanded in phases, allowing power capacity to come online alongside data center construction". NTT Global Data Centers' chief executive Doug Adams states the buying logic in one sentence: "Our clients are building AI capabilities at a pace the grid alone cannot always match, and speed to power has become one of the defining factors in delivering capacity when and where it's needed." That is not a price argument. It is a schedule argument, made by the buyer, on the record.

So the honest definition for this piece's purposes: an on-site fuel cell is a schedule-compression product that happens to generate electricity. Everything below follows from taking that seriously — including why the revenue mix looks like a server vendor's, why the fuel arithmetic matters more than the efficiency headline, and why the cash flow has gone negative at record revenue.

Why is the revenue mix the most important number?

Because it tells you which business model is actually being financed. Bloom's Q3 2026 release breaks revenue four ways, and the split is lopsided in a way that no power-plant vendor's would be.

Q3 2026 revenue lineAmountShare of revenueWhat it is
Product$1,073.6M86.9%Fuel cell modules shipped — capital equipment
Installation$84.9M6.9%Site work, one-time per project
Service$67.6M5.5%Maintenance contracts on the installed base
Electricity$9.8M0.8%Power sold from owned installations
Total$1,235.8M100%+138.1% year over year

Now put next to it the mix at GE Vernova, whose gas turbines are the technology the fuel cell is supposedly displacing. GE Vernova's Q3 2026 disclosure splits a $197B backlog into $104B of equipment and $93B of services — 47% of the order book is a labour annuity, growing 14% year over year while equipment grows 92%. A turbine is sold as a machine plus two decades of maintenance, and the maintenance is where the durable margin lives. A fuel cell, at this stage, is sold as a machine. The service line exists and will compound as the installed base ages, but at 5.5% of revenue it is not yet the business. The business is shipping boxes.

That has three consequences a buyer should price in. First, the customer is the financier. Product revenue is recognised when equipment ships; the customer (or its landlord or its lessor) has already committed capital before a single kWh arrives. This is the opposite of a utility relationship, where the utility owns the asset and you own a bill. Second, the seller's margin is a hardware margin. GAAP gross margin of 33.8% and an 18.2% operating margin are respectable for capital equipment and unremarkable for an energy provider — and hardware margins compress when competitors arrive, which is exactly what a 5.5% service line cannot cushion. Third, the annuity is the option you are implicitly underwriting. If you are buying the equipment, you are also deciding who services it for the next decade; if the vendor's service economics never materialise, the installed base becomes somebody else's maintenance market.

What does a month of schedule actually cost?

More than the electricity, by an order of magnitude — and that is the whole case. The arithmetic (ours) needs three inputs: the capital tied up in compute, the life it is depreciated over, and the delay being avoided.

Take a 1 GW campus with roughly $40B of compute and supporting capital, depreciated over four years. That is $10B a year, or $27.4M a day, of asset-time. Pull the switch-on forward by twelve months and you have released about $10B of asset-time — before counting a single dollar of revenue or a single customer who did not churn.

Now price the power premium. A fuel cell's levelised electricity cost is higher than a combined cycle's; suppose the premium is +$40/MWh, a generous number for a first-of-a-kind on-site fleet. At 1 GW and an 80% capacity factor, that is 7.008 TWh a year — $280M of extra electricity spend. The ratio is 36:1 in favour of the earlier date. Even at a +$100/MWh premium the bill is $701M a year, and the date still wins by roughly 14:1. This is the same option-pricing logic that makes a $420/kW waiting cost rational in the gas-turbine slot market: when the alternative asset earns nothing while it waits, the price of the wait dominates the price of the input.

InputValueSource
Compute capital, 1 GW campus~$40BAssumption (ours), stated for arithmetic
Depreciation life4 yearsAssumption (ours)
Asset-time per day$27.4MOurs: $40B ÷ 4 ÷ 365
Value of a 12-month acceleration~$10BOurs
Power premium at +$40/MWh, 80% CF~$280M/yrOurs: 7.008 TWh × $40
Ratio, date vs premium~36:1Ours

The uncomfortable corollary: this arithmetic justifies any fast power, including the diesel and gas turbines that data centers have been criticised for. It is not an argument that fuel cells are good. It is an argument that in a market where the compute is worth $27.4M a day, the power source is judged on delivery date and dispatched in that order — which is why the same buyers who signed 20-year renewable PPAs are now signing on-site generation with a 12-month lead time.

How much gas does 1 GW of fuel cells burn?

This is where the "clean" framing needs care, and where a buyer should do the fuel bill before the press release. A fuel cell's electrical efficiency is around 60% in the non-combined-heat configuration, per the US Department of Energy figures cited in industry coverage; Bloom's own marketing material claims up to 65%. Take 60% and the fuel arithmetic is unforgiving: one MWh of electricity requires 1.667 MWh of thermal input, which is 5.69 MMBtu of natural gas.

At 1 GW and an 80% capacity factor that is 7.008 TWh of electricity a year, requiring about 39.9 million MMBtu of gas. At $3/MMBtu that is $120M a year; at $4 it is $160M. Per MWh sold, fuel alone is $17–23. A combined cycle at the same 60% efficiency burns the same gas per MWh — the fuel cell's advantage is not thermodynamic, it is that it needs no steam cycle, no water for cooling, no multi-year construction and no interconnection queue. Its disadvantage is that it captures no combined-cycle synergy, so its ceiling is that 60%, not 62–64%.

Two honest consequences. It still emits CO₂ — roughly 30–40% less than a simple-cycle turbine per MWh, but not zero, and the on-site location means the emissions are where the neighbours are. And the gas has to arrive, which is its own schedule risk: pipeline laterals require federal review, and a campus that substitutes fuel cells for turbines has traded an interconnection queue for a gas-supply queue. The one place the fuel cell is genuinely cleaner is water — it consumes very little on site, which matters because the water fight is local. But note the boundary: our earlier piece on data-center water found 92% of the footprint sits at the power plant, so a fenceline "low water" claim is a claim about 8% of the original problem.

Why did operating cash flow go negative at record revenue?

Because this is what a hardware ramp looks like, and the same signature appeared in the AI server builders' results this quarter. Bloom's Q3 2026 release reports $250.3M of net income attributable to common stockholders, GAAP operating income of $225.4M, and operating cash flow of −$27.1M — with $153.5M of capital expenditure in the same three months, against a cash balance of $2.49B. Profitable on the income statement, cash-negative in the quarter: receivables and inventory growing faster than collections, because the units being built have not been paid for yet, or have been paid for on terms that lag shipment.

That is not a red flag by itself; it is the normal shape of a company scaling a physical product, and it is the shape we documented at Dell, Supermicro and HPE in November. What it does mean is that the growth is being financed — by cash on hand, by the customer's down payment, and increasingly by third-party capital. Which is where Brookfield comes in: in June 2026 Bloom and Brookfield expanded their AI infrastructure power financing framework from $5B to $25B, a fivefold increase, explicitly to "build and finance rapid power for AI infrastructure". When the vendor's own balance sheet cannot carry the ramp, the financier becomes part of the product — and the cost of that capital lands in the price of the power, not in the price of the box.

What is the evidence that this is a fleet, not a pilot?

Three announcements, all primary, all dated, and each from a different kind of buyer — which is what distinguishes a category from a demo.

  • Oracle, April 2026: the two companies expanded their strategic partnership to deploy up to 2.8 GW of fuel cells to accelerate AI infrastructure. A hyperscale-adjacent buyer committing gigawatts is the demand signal that matters, because that buyer has the alternative of waiting for utility service.
  • NTT DATA, 5 November 2026: 900 MW across multiple data center projects globally, with initial shipments starting in 2027 — Bloom's largest international project to date. This is the first large commitment from a colocation operator rather than a compute buyer, and colocation is where the power constraint is usually passed to the landlord.
  • Brookfield, 30 June 2026: the financing framework expanded to $25B, fivefold. Capital committing at that scale is a statement that the projects are expected to be financeable assets, not research programmes.

One more technical thread worth noting because it changes the bill rather than the schedule: in September 2026 Bloom published the case for 800 V DC-native output from its systems, arguing that delivering DC directly can eliminate transformer and rectifier stages in an AI campus. We covered the arithmetic of 800 VDC rack distribution separately — the copper and loss savings are real, and the harder question is who owns the conversion stage you deleted. A fuel cell that natively produces DC is one of the few generation technologies that could plausibly skip that stage, which would make it an architectural choice and not just a stopgap. That claim is the vendor's, not an independent measurement; treat it as a direction, not a number.

What are the caveats?

These are one company's results. Bloom is the dominant Western SOFC vendor, so its income statement is the best available proxy for the category — but a category with one public pure-play is not a market with a price. The annuity is unproven. A 5.5% service line today says nothing about what it will be at a 20 GW installed base; it could compound into a GE-style annuity, or it could be competed away by third-party maintenance. The efficiency numbers are vendor and government figures, not independent tests. The 60% DOE figure and the 65% marketing figure are close enough to be the same engineering claim stated at different temperatures and loads; neither is a field measurement across a fleet. The schedule arithmetic is ours and rests on assumptions — $40B per GW of compute capital and a four-year life are round numbers chosen to be defensible, not audited; change the life to six years and the daily asset-time falls to $18.3M, which still dwarfs the power premium. And emissions remain. A gas-fed fuel cell is a lower-carbon generator than a simple-cycle turbine and a higher-carbon one than the grid average in many regions; the "clean" label is doing work that the chemistry does not support.