Energy · Deep dive

Bloom Energy

Twenty-five years of unprofitable fuel-cell R&D collided with an AI power shortage — and Bloom went from science project to the fastest way to put a gigawatt on a greenfield site.

well positioned

Sold out, profitable, and holding the only product that puts a gigawatt on a greenfield site in months rather than years — though the position is time-boxed by GE Vernova's turbine capacity and the $20B backlog sits against far less audited RPO.

Snapshot

Bloom Energy builds solid oxide fuel cells — the Bloom Energy Server — that convert natural gas or hydrogen into electricity electrochemically, on the customer’s site, without combustion. For two decades that was an expensive niche product sold to companies that wanted resilient, clean-ish power and were willing to pay for it. Then AI data centers created a power shortage the grid cannot answer on schedule, and Bloom’s one durable attribute — you can be generating in months, not years — became the scarcest good in energy. FY2025 revenue was $2.02B, up 37.3% (company release, February 2026). Q1 2026 revenue was $751.1M, up 130.4%, with GAAP net income of $70.7M against a loss a year earlier. Management guides FY2026 to $3.4-3.8B. Backlog is described as roughly $20B. The market cap sits near $79B — which is the part of this story that has to be argued, not asserted.

Founding story

KR Sridhar was not an energy entrepreneur. He was a mechanical and nuclear engineer running the Space Technologies Laboratory at the University of Arizona, where his team won NASA contracts for the Mars Surveyor 2001 mission. The specific job: build a device that could take Martian carbon dioxide and produce oxygen for life support and rocket propellant. It was a solid oxide electrolyzer — a ceramic cell that, fed power, splits gas into oxygen and fuel.

NASA cancelled the mission. Sridhar had a working ceramic stack and no rocket to put it on, so he ran the reaction backwards: feed it fuel and air, get electricity. He incorporated Ion America in January 2001, moved the company to the NASA Ames research campus in Silicon Valley in 2002, renamed it Bloom Energy in 2006, and unveiled the Bloom Box in February 2010 with a level of Silicon Valley theatre — a 60 Minutes segment, Colin Powell on the board, John Doerr as evangelist — that the product’s economics did not yet support.

What followed was the least glamorous stretch in cleantech: $1.1B+ of venture capital consumed across Kleiner Perkins, NEA and DAG, a reported private mark near $2.9B in 2011, a 2018 IPO that priced at $15 and raised only $270M, then years of losses, dilution, a restatement, and a stock that spent most of 2019-2024 below its first-day close. The point for underwriting Bloom today is that it spent 24 years building manufacturing capacity and field data for a demand curve that did not exist until 2024. Nobody else was that patient — which is why nobody can copy the position quickly.

How it works

A solid oxide fuel cell is not an engine. There is no flame, no piston, no turbine blade. A ceramic electrolyte wafer runs at roughly 800°C; oxygen ions migrate through the ceramic lattice from the air side to the fuel side, where they react with reformed natural gas or hydrogen. The electrons take the external circuit instead of the reaction path, and that current is the product. Bloom quotes electrical efficiency of 50-60%, and ~60% on pure hydrogen (company announcement, 2024). Because nothing combusts, the exhaust carries essentially no NOx, SOx or particulate matter.

That last sentence is the entire business. A combustion source — a gas turbine, a recip genset — triggers an air permit, which triggers a public process, which in permit-constrained markets takes years and sometimes fails. A fuel cell in most jurisdictions does not. And because Bloom sits behind the meter, the customer skips the utility interconnection queue entirely. Bloom does not sell the cheapest electron. It sells the electron that arrives first.

Physically, four to six power modules combine into a 200-300 kW Energy Server occupying roughly half a 30-foot shipping container’s footprint. Capacity is added in increments, not 500 MW lumps. Oracle reportedly took delivery of its first system 55 days after order against a 90-day target — that number, not efficiency, closed the follow-on.

The catch is the ceramic. Stacks degrade — DOE’s program target is under 0.2% performance loss per 1,000 hours, and real-world stacks have historically run closer to 0.5% — so power modules must be swapped every 5-10 years depending on duty cycle and fuel quality, at an estimated 20-40% of the original system price. Bulls call that an annuity. Bears call it a liability Bloom has historically been optimistic about reserving for.

Product and business overview

Four things get sold. Energy Servers — the hardware, sold outright to customers, to utilities like AEP, or to financing partners. Long-term service agreements covering monitoring, maintenance and the periodic power-module swap, on essentially the entire installed base. Electricity, under PPAs where Bloom or a partner owns the asset. Electrolyzers, the same ceramic stack run in reverse to make hydrogen — strategically important, commercially immaterial in 2026.

The installed base is around 1.5 GW deployed globally, more than 400 MW of it data centers. Anchor customers include Equinix (100+ MW across 19 sites), AEP (up to 1 GW announced November 2024, the largest commercial fuel-cell procurement to that date), Oracle (up to 2.8 GW under a master services agreement), and SK ecoplant, which committed to a minimum 500 MW / ~$4.5B purchase in 2023.

Business model and pricing

Equipment revenue dominates — product revenue was $653.3M of the $751.1M Q1 2026 total and grew 208% year over year. Service is the recurring layer, and the power-module swap is roughly 65% of the cost inside a service contract, which tells you what those contracts really are: a hardware replacement subscription with a monitoring wrapper.

Bloom publishes no rate card, but third-party figures cluster. Under a PPA, delivered electricity has been quoted near 9.9 cents/kWh (~$99/MWh) on shorter structures. A customer who buys and operates the asset carries an all-in levelized cost — capex, gas, O&M, stack swaps, cost of capital — that independent analysts put around $120-130/MWh, with service agreements estimated at 8-12 cents per kWh generated. Against a Northern Virginia industrial tariff plus a five-year wait, that works. Against a $40-60/MWh combined-cycle turbine, it does not — which is why Bloom’s sales motion is entirely about sites where turbines are not available.

Gross margin is the number that decides the equity. GAAP gross margin went 14.8% (2023) to 27.5% (2024) to 29.0% (2025) to 30.0% in Q1 2026, and management guides to ~34% non-GAAP for FY2026. That improvement is almost entirely throughput: Fremont is being taken from 1 GW to 2 GW of annual capacity for about $100M, with headroom to ~5 GW at roughly $100-150M and 6-9 months per incremental gigawatt. The whole bull case is a factory utilization curve.

Traction over time

The step-change is unmistakable, and it happened in one year. Product backlog reached roughly $6B at FY2025 results (February 2026), up about 2.5x year over year, inside a total backlog Bloom describes as ~$20B. Order flow in early 2026 included roughly $7.65B of data-center contracts booked in a 90-day window, the Oracle expansion to as much as 2.8 GW, and a Brookfield project-financing partnership that was expanded from $5B (October 2025) to $25B on June 30, 2026 — meaning Bloom can put fleet in the ground without carrying it on its own balance sheet.

Read the backlog carefully. “Backlog” here bundles firm purchase orders, master services agreements with capacity ceilings, and framework commitments. Hunterbrook Media, publishing a short thesis on July 8, 2026, pointed out that Bloom’s audited remaining performance obligations were on the order of $492M — three orders of magnitude below the headline. Bloom disputes the framing. Both things can be true: the demand is real, and the $20B number is a marketing construct, not a GAAP one.

Market analysis

The demand side is not seriously contested. DOE and Lawrence Berkeley National Laboratory (December 2024) put US data-center consumption at 4.4% of national electricity in 2023 (176 TWh), rising to 6.7-12% by 2028 (325-580 TWh). Hyperscaler capex is tracked near $750B for 2026, up roughly 67% year over year (CreditSights, early 2026), and is expected to cross $1T in 2027. PJM’s 2026/27 capacity auction cleared at the price cap of $329.17/MW-day, up 22% on the prior record. There is not enough power, and the shortage has a rising price.

Bloom’s addressable slice is narrower than the headline: the behind-the-meter, permit-constrained, need-it-inside-a-year segment, where a $120/MWh electron beats a $50/MWh electron that arrives in 2031. Plausibly tens of gigawatts — but not the whole data-center power market, and it shrinks as turbine and grid capacity catch up.

Competitive intel

The named set is in the frontmatter; the shape of the fight is this. GE Vernova is the actual adversary, and right now it is not competing on merit — it is competing on availability, and it has none. Its ~83 GW turbine backlog stretching into 2029-2030 is Bloom’s entire moat. Caterpillar and Cummins own the fast, cheap, dirty end and lose on permits. FuelCell Energy and Plug are narrative competitors with a fraction of Bloom’s manufacturing reality. SMRs are a 2030s story that only affects terminal value. The quiet competitor is the grid itself: every month of interconnection reform erodes Bloom’s premium.

History and evolution

• January 2001 — Sridhar founds Ion America after NASA cancels Mars Surveyor 2001.
• 2002 — Company relocates to the NASA Ames campus in Silicon Valley.
• 2006 — Renamed Bloom Energy; Kleiner Perkins and NEA fund the long build.
• February 2010 — Bloom Box public unveiling; Google and eBay named as early customers.
• 2011 — Reported private valuation near $2.9B; profitability still years away.
• July 2018 — IPO on NYSE at $15/share, raising $270M. Closes first day near $25, then grinds down.
• July 2019 — Hindenburg Research publishes a short report alleging understated service costs, subsidy dependence, and emissions worse than the grid in key states.
• February 2020 — Bloom restates financials over Managed Services Agreement revenue recognition; net losses increase by an estimated $55-75M across 2016-2019. Securities class actions follow.
• 2021-2023 — SK ecoplant invests roughly $566M across two tranches and commits to a 500 MW / ~$4.5B purchase agreement.
• November 2024 — AEP agreement for up to 1 GW; the fuel-cell industry’s largest commercial order to date.
• October 2025 — $2.2B of 0% convertible notes due 2030; initial $5B Brookfield framework.
• February 2026 — FY2025 results: $2.02B revenue, first full year of GAAP profitability, ~$20B backlog claimed.
• June 30, 2026 — Brookfield partnership expanded fivefold to $25B; Oracle expanded to up to 2.8 GW.
• July 8-9, 2026 — Hunterbrook short report on scandium supply and backlog quality; stock falls ~6%, Bloom files an 8-K rebuttal, stock recovers.

What people say

The case for. The sell side is broadly constructive: as of mid-July 2026 the consensus rating across 21-36 covering analysts sits at Buy, and Jefferies raised its target in April 2026 on the strength of the order book. Morningstar’s Brett Castelli has framed the AEP-scale deals as evidence that utilities and hyperscalers are genuinely scrambling and will pay for optionality. The strongest testimonial is not an analyst note but a customer action: Oracle got a system energized in 55 days, then signed for up to 2.8 gigawatts. Brookfield, whose job is underwriting 20-year asset lives, quintupled its commitment to $25B in June 2026. Both are capital, not press releases. Employees rate the mission and the engineering highly; Glassdoor’s 3.3/5 across 457 reviews (mid-2026), with 61% recommending, is unremarkable rather than damning.

The complaints. They are numerous and mostly not new. Hindenburg’s 2019 thesis argued that Bloom systematically understated the cost of servicing its fleet, that it had absorbed over $1.1B in public subsidies without which a former employee said it would not exist, and that its servers emitted more CO2 than the grid in some of its core states — the “clean energy” branding sitting awkwardly on top of a machine that runs on natural gas. Bloom rebutted the emissions math on methodology grounds (marginal versus average grid emissions factors), which is a fair argument but not a knockout. Then the company restated its financials in February 2020 over exactly the kind of revenue-recognition aggressiveness the bears had flagged, which permanently costs a company the benefit of the doubt.

The current bear case, from Hunterbrook’s July 2026 report, is sharper: Bloom’s stacks require scandium, a genuinely obscure input, and a 5 GW production run would need roughly 220 tons of scandium oxide annually against a projected global supply near 240 tons — with China dominant in the chain. Hunterbrook also flagged that a chunk of recent revenue routed through Bloom’s own joint ventures, and that the $20B backlog is unaudited against ~$492M of booked performance obligations. Bloom denied the claims and the stock recovered within days. But the valuation is doing a lot of work here: BE has traded at multiples in the hundreds of times forward earnings, and several published price targets — including a consensus in the $186-$246 range depending on the compiler — sat below the market price through mid-2026. When the sell side’s bullish consensus target implies downside, that is the analysts declining to say the quiet part out loud. Internally, Glassdoor’s recurring negative themes are top-down decision-making, an arrogant senior team, repeated layoffs with poor communication, and favoritism — the profile of a company that has never had to be a good employer because it has always been in survival mode.

Outlook: well positioned or at risk?

Well positioned — but the position is time-boxed, and the dispute is about permanence, not whether the product works. Bloom is executing. Revenue is nearly doubling, GAAP profitability arrived, the gross margin curve is bending the right way, and Oracle and Brookfield have put real money behind the technology at hyperscale. The operating story has no obvious hole in it.

The exposure is structural. Bloom’s advantage is a temporary supply bottleneck in someone else’s product. Every gigawatt GE Vernova adds toward its 24 GW/year 2028 turbine target, every month FERC shaves off an interconnection queue, and every co-located gas plant a hyperscaler negotiates directly narrows the window in which speed is worth a 2-3x premium per megawatt-hour. Meanwhile Bloom must ramp manufacturing from 1 GW to 2 GW and beyond, on a supply chain with at least one concentrated rare input, at a valuation with no room for a bad quarter.

Three things break the thesis: a gross margin miss during the ramp; a visible gap between the $20B headline and revenue actually recognized; or a hyperscaler publicly choosing turbines or SMRs at multi-gigawatt scale for its 2029 fleet. Three things make it: stack replacement proving out as a real high-margin annuity, permitting friction getting worse rather than better, and Bloom holding 30%+ gross margins at 2 GW. Watch gross margin, watch remaining performance obligations in the 10-Q rather than the press release, and watch whether the scandium answer survives an audit cycle. The company would rather you watched the backlog number.

Sources and further reading

• Bloom Energy — Fourth Quarter and Full Year 2025 Financial Results (Bloom Energy, February 2026)
• Bloom Energy Q1 2026 supplemental disclosure — Form 8-K (SEC EDGAR, 2026)
• Bloom Energy and Oracle expand partnership to deploy up to 2.8 GW (Bloom Energy, 2026)
• Bloom Energy rebuts short-seller report, defends scandium supply and financials (TradingView / Bloomberg wire, July 2026)
• Bloom’s Big Lie (Hunterbrook Media, July 8, 2026)
• Bloom Energy — A “Clean” Energy Darling Wilting to its Demise (Hindenburg Research, July 2019)
• Fuel cell builder Bloom Energy continues streak of business math errors and lack of profits (pv magazine USA, February 2020)
• GE Vernova expects to end 2025 with an 80-GW gas turbine backlog that stretches into 2029 (Utility Dive, December 2025)
• DOE report — data centers consumed 4.4% of US power in 2023, could hit 12% by 2028 (DataCenterDynamics, December 2024)
• How the AI data center bubble story is playing out inside one booming energy stock (CNBC, January 2026)