A fuel cell is a box that turns a fuel — either hydrogen or natural gas — straight into electricity through a chemical reaction, with no burning and no engine. The product these companies sell is power-generation hardware sold by the megawatt (MW = a million watts; roughly enough to run a few hundred homes, or one small slice of a data hall), plus a long-running service contract to keep the boxes running. Four public names anchor the group: Bloom Energy (BE), which sells solid-oxide cells that run on natural gas to make round-the-clock on-site power; Plug Power (PLUG) and Ballard Power (BLDP), which build hydrogen cells aimed more at forklifts, trucks, and backup power; and FuelCell Energy (FCEL). The reason this group connects to an AI/AGI build-out: data centers need large amounts of electricity now, and the electric grid in many places cannot connect them for years (grid interconnection queues in some US markets run multiple years). A fuel cell can be installed on the customer's own land in months, sidestepping that queue.
Demand for on-site power is rising because data centers face multi-year grid hookups, and AI compute growth adds to that need. On the supply side, Bloom states it is "not capacity constrained" — its factory ceiling (a management claim of ~5 GW/yr) sits well above current shipments, and its capex per megawatt is low. By Bloom's own account, what gates how fast its boxes get installed is customer site readiness and competing on-site options (gas turbines and reciprocating engines), not the assembly line. So within this group the binding constraint, by management's framing, is on the demand-fulfillment side rather than the factory. On price: Bloom (the one name with revenue growing toward profitability) carried a large-cap valuation against ~$2.0B of 2025 sales; the three hydrogen-focused names have historically run losses. The arithmetic below lets the reader weigh today's price against forecasts that are not yet contracted. No buy/sell view is expressed.
The physical product is an electrochemical "stack" packaged into a unit (Bloom calls its unit an "Energy Server"). Fuel goes in one side, electricity comes out the other, with heat as a byproduct. There is no combustion and no rotating engine, which is why it can sit quietly on a customer's property. Different chemistries serve different jobs:
The money comes in two streams. First, the hardware sale: the customer buys the boxes by the megawatt. The prior Bloom report models its blended system price at roughly $3M per MW of hardware before service and installation. est. (modeled) Second, a multi-year service contract to monitor, maintain, and eventually replace the stacks as they wear out — a recurring cash stream that, in principle, smooths the lumpy hardware sales. For an owner deciding whether to put capital in, the two facts that drive the cash math are whether the hardware is sold at a profit and how large and durable the service tail is.
Source: Bloom Energy Manufacturing Capacity Deep Dive (May 18, 2026), $3M/MW hardware is the report's modeled figure; 500-stocks Energy & Power scan, Fuel Cells section.
What is contracted and real today. The clearest hard demand signal in the group is Bloom's order book. Oracle signed a master agreement for up to 2.8 GW, of which an initial 1.2 GW is contracted with deployment underway (Apr 13, 2026 release); on the Q1 2026 call management also described an Oracle "Project Jupiter" power block of up to 2.45 GW. Separately, Bloom has a 1 GW master supply agreement with utility AEP (Nov 2024). Bloom's revenue grew from $1.2B (2022) to $2.0B (2025) — real cash from real shipments. Full-year 2026 revenue guidance is roughly $3.4-3.8B (a forecast, not yet shipped). contracted (order book, revenue)
The drivers (AGI lens). Reasoning from the premise that AGI is arriving and compute demand is compounding, a binding constraint on the AI build-out has shifted toward electricity supply. The 500-stocks scan notes that behind-the-meter generation (power made on-site, bypassing the grid) is growing as hyperscalers route around grid interconnection queues that can run multiple years. A fuel cell can be deployed in months versus years for a grid connection — Bloom's Apr 2026 Oracle release cites a prior deployment completed in 55 days versus an anticipated 90-day schedule. So the demand premise is not "fuel cells beat the grid on cost"; it is "fuel cells deliver power on a timeline the grid cannot match, and time-to-power is what AI capex is short of." That is the demand pull, amplified under the AGI premise. forecast / reasoning
Who the buyers are. Hyperscalers and the cloud/AI infrastructure firms serving them (Oracle is the named example), utilities buying capacity to serve those loads (AEP), industrial sites, and — for the hydrogen players — logistics/material-handling operators and, eventually, heavy transport. The scan also notes Microsoft has tested hydrogen fuel cells for data-center backup, an early, non-committed signal rather than a contract.
The counterweight to demand. The same scan states that fuel cells compete with cheaper gas turbines and reciprocating engines for on-site power, and that hydrogen infrastructure is "nascent" (early-stage, not yet built out at scale). So forward demand for fuel cells specifically is a function of how much of the on-site-power market they win versus turbines — not just how big the on-site-power market gets. Per the provided files, that share-of-on-site-power split is the largest single uncertainty on the demand side.
✓ VERIFIED — the following figures were confirmed from primary sources after initial publication:
Current capacity. On its Q1 2026 call (Apr 28, 2026) Bloom management stated that its current manufacturing footprint can deliver about 5 GW of product per year and that the company is "not order constrained and not capacity constrained." That is a management nameplate claim made on an earnings call, not an audited or SEC-filed figure; the firmer SEC-filed baseline (2025 10-K) is that the Fremont, California plant is being doubled from roughly 1 GW to 2 GW by the end of 2026. The prior report applies a discount, treating achievable shipping capacity as closer to 3-3.5 GW/yr until quarterly shipment data confirms more. 2 GW = 10-K baseline / 5 GW = management claim
What the bottleneck is — and is not. The 500-stocks scan states this group is "not significantly constrained" on supply: fuel-cell companies have excess manufacturing capacity and are scaling up, and the stated challenge is cost competitiveness, not making enough units. Bloom's reported numbers are consistent with that: capex has fallen as a share of revenue (9.7% in 2022 → 6.3% in 2023 → 4.0% in 2024 → 2.8% in 2025; 3.5% in Q1 2026). The prior report infers from this cash-flow pattern roughly $75-100M of capital per incremental GW of capacity (an inference, not a disclosed per-GW budget). On those figures, the limits Bloom and the scan cite are not the factory itself; they are:
Who controls supply / share structure. Within the data-center-relevant slice (natural-gas baseload), the scan describes Bloom as the standout among the scaled public players; the hydrogen-cell names (Plug, Ballard) address different end markets (forklifts, transport, backup) and are not competing for the same data-center sockets. Unlike high-bandwidth memory or advanced chip packaging — where one or few parties control a scarce physical input — the provided files identify no single physical input here that one party controls and others lack. The differentiator the files cite is know-how plus installed base and service relationships, not control of a scarce raw material. est. (share structure)
Source: Bloom Energy Manufacturing Capacity Deep Dive (capacity timeline, capex table, bottleneck section, Q1 2026 earnings call quote); 500-stocks Energy & Power scan, Fuel Cells "Supply" note.
Putting the two sides together: in this group the demand pull is strong and (under the AGI premise) growing, while supply is, by the makers' own framing, not the scarce thing. That is the opposite of products like GPUs or grid transformers, where the maker's output is the binding constraint and sold-out capacity drives pricing power. Here Bloom states it can make more than it currently ships — 2026 revenue guidance of $3.4-3.8B at the modeled ~$3M/MW (plus service and installation) implies roughly 1-1.2 GW of hardware shipments, i.e. only about 20-25% utilization against the stated 5 GW nameplate claim. On those numbers, the shortage is of finished, energized customer sites, not of factory output.
| Question | This group's answer (per the provided files) |
|---|---|
| Is the product physically short (sold-out factories, rising prices)? | No. Excess capacity; Bloom states "not capacity constrained." |
| What is the binding constraint instead? | Time-to-power. Customer site readiness + competing on-site options gate installs. |
| Evidence of demand strength | Oracle 2.8 GW (1.2 GW contracted), AEP 1 GW, revenue $1.2B→$2.0B (2022-25) |
| Evidence supply is loose | Capex 2.8% of revenue (2025); ~$75-100M/GW inferred; scan: "not significantly constrained" |
| When could it flip to oversupply? | By the factory measure it is already supply-long; the stated risks are grid catch-up and cheaper gas turbines taking share, not a fuel-cell glut. |
So the shape of the bet here differs from most AI-power sub-sectors: it is not "supply is scarce, so the maker has pricing power." It is "demand is large, and Bloom captures some share of on-site power on a faster timeline than the grid recovers and before turbines undercut it." Because supply is loose and turbines compete, the files frame fuel-cell economics as depending on holding price while driving cost down the curve, rather than on a scarcity-driven price spike. Stated neutrally: strong demand, loose supply; per the provided files, margin comes from execution and cost-down rather than from a structural shortage. The reader judges what that is worth.
| Company | What it makes | Exposure to this product | Rough size | Position (per the files) |
|---|---|---|---|---|
| Bloom Energy (BE) | Solid-oxide cells (natural gas) for on-site baseload power; "Energy Server" units | Pure-play, ~100% of revenue | ~$2.0B 2025 revenue (reported); large-cap (prior report referenced stock ~$258) est. cap | Scaled, data-center-relevant; Oracle + AEP contracts; revenue growing toward profitability; ~5 GW stated capacity claim |
| Plug Power (PLUG) | PEM hydrogen cells, electrolyzers, hydrogen ecosystem (forklifts, backup) | Pure-play hydrogen, ~100% | Smaller-cap; historically loss-making est. | Hydrogen-focused; scan notes it "has struggled financially"; less data-center relevant |
| Ballard Power (BLDP) | PEM hydrogen fuel cells for buses, trucks, marine, rail | Pure-play transport fuel cells, ~100% | Small-cap; loss-making est. | Transportation focus; scan says "less relevant to data centers" |
| FuelCell Energy (FCEL) | Molten-carbonate / solid-oxide platforms, carbon capture | Pure-play, ~100% | Small-cap; loss-making est. | Hydrogen/utility focus; scan groups it with PLUG as financially struggling |
Note: every name in this group is a fuel-cell pure-play — there is no diversified conglomerate here whose fuel-cell line is a small slice. That makes the group a direct (and concentrated) way to take exposure to the product, with Bloom carrying the large majority of the group's revenue. The size descriptors for PLUG, BLDP, and FCEL are general-knowledge estimates, not live-verified.
Source: 500-stocks Energy & Power scan, Fuel Cells "Key Companies" and verdict box; Bloom Energy Manufacturing Capacity Deep Dive (Bloom revenue, contracts, ~$258 stock reference). Market caps for PLUG/BLDP/FCEL are general-knowledge est., not live-verified.
The arithmetic the reader can do. The prior report referenced Bloom's stock at roughly $258 and noted that at that level the market was, in the report's words, pricing the "bull-to-lottery" end of its own scenario set. Translated to money-in/money-out terms: against ~$2.0B of 2025 revenue, a large-cap valuation means an owner pays several dollars of market value for each $1 of this year's sales — i.e., the price embeds years of forecast growth, not just current cash. Against the prior report's 2029 base-case revenue (~$10B), the same price is a smaller number per dollar of future sales — but that 2029 revenue is a forecast, not contracted. The distance between "price per dollar of today's sales" and "price per dollar of 2029 forecast sales" is exactly the AGI-demand forecast the buyer is relying on. Live quotes were unavailable here, so any precise current multiple would be a guess; the framing is neutral arithmetic, not a valuation verdict. est. multiple framing
The money-in / money-out shape. On the reported figures, Bloom is the capital-light, scaling name of the four: low capex per megawatt means each new MW shipped does not require a proportional fixed-cost outlay, so operating leverage can build as volume grows — that is the mechanical reason its path to generating owner cash is clearer than the other three on the numbers provided. The other three (PLUG, BLDP, FCEL) have historically been cash-consuming: they raise money and spend it faster than they earn it, so an owner there funds losses today against a hydrogen future the scan itself calls nascent. Two drags specific to Bloom to keep in view, both from the prior report: its Korea joint venture generates equity-method losses (cited at $17-21M per quarter) that reduce reported net income even while the JV helps fulfill demand; and it carries convertible debt ($2.5B due 2030) tied to the stock staying above roughly $195 (below that conversion level the obligation behaves more like straight debt). Stated neutrally: one name with revenue growing toward profitability and low capex, priced as a multiple of today's sales that embeds forecast growth; three loss-makers priced on a hydrogen option. The reader decides whether the forecast justifies the price — no buy/sell view is given.
Source: Bloom Energy Manufacturing Capacity Deep Dive (capex/revenue table, ~$3M/MW, ~$75-100M/GW inferred, 2029 scenario revenue, ~$258 stock and "bull-to-lottery" scenario language, Korea JV losses, convertible-debt note). Multiple-of-sales framing is neutral arithmetic; live quotes were not available.
Source: 500-stocks scan verdict box (BE standout; PLUG/FCEL struggled; BLDP transport-focused); Bloom Energy Manufacturing Capacity Deep Dive.
What was used.
/Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/500-stocks/01-energy-power.html) — for the product description, the supply/demand qualitative notes, the player list (BE, PLUG, FCEL, BLDP), and the scan statements that Bloom is the standout and supply is not significantly constrained./Users/ravf/projects/work/.claude/worktrees/sector-hub/research/investments/be-manufacturing-capacity.html, treated as V1) — for Bloom's hard numbers: revenue 2022-2025, capex table and intensity, the ~5 GW management capacity claim vs the 2 GW SEC baseline, the modeled $3M/MW pricing, the inferred $75-100M/GW capex, the Oracle (2.8 GW / 1.2 GW contracted) and AEP (1 GW) contracts, 2026 guidance, 2029 scenario revenue, Korea JV losses, the convertible-debt note, and the ~$258 stock reference.Which figures are hard vs approximate. Hard (grounded in the provided files / filings): Bloom's revenue ($1.2B 2022 → $2.0B 2025), the capex figures and intensity (2.8% of revenue in 2025), the Oracle and AEP contract sizes, the 2 GW SEC capacity baseline, and the qualitative scan statements. Modeled or inferred from the prior report (not disclosed line items): the ~$3M/MW hardware price and the ~$75-100M/GW capex per incremental GW. Management-stated, not audited: Bloom's ~5 GW capacity claim and the 2026 revenue guidance — labelled as a claim/forecast above. Approximate and NOT live-verified (general knowledge, flagged with est.): the ~$5-8B global stationary fuel-cell market size, the market-growth rate, all market-share percentages, and the current market caps of PLUG, BLDP, and FCEL. The single unverified-note in the Demand section states this caveat in full. No buy/sell view, price target, or valuation verdict is expressed anywhere in this sheet — the arithmetic is provided for the reader to judge.