2026
4,600 boxes
Range 3,700 – 5,500
≈ 1.50 GW at 325 kW/box (1,500 MW ÷ 0.325 MW = 4,615)
Independent estimate · mill-cut · 12 Sep 2026
How many Bloom boxes can Bloom Energy produce in the next three years?
A “Bloom box” here is one Energy Server cabinet — the commercial building block Bloom sells, not a fuel-cell stack and not a whole energy farm. New production is treated as the current 325 kW Energy Server (five ~65 kW power modules). Years are calendar 2026, 2027, and 2028. The numbers are factory output the plants can physically build, not last year’s product acceptances rolled forward.
2027
6,800 boxes
Range 5,500 – 8,000
≈ 2.20 GW at 325 kW/box (2,200 MW ÷ 0.325 MW = 6,769)
2028
8,600 boxes
Range 6,800 – 10,800
≈ 2.80 GW at 325 kW/box (2,800 MW ÷ 0.325 MW = 8,615)
Average kW per Energy Server (box size mix)
Forecast units are reconcilable from MW by dividing annual factory MW by 0.325 MW (325 kW) per box. That is the current commercial Energy Server Bloom shows in investor decks and the 2025 product brochure: ~65 kW power module × 5 = 325 kW Energy Server; ten servers make a 3.25 MW “power block.”
- New-build mix (used for 2026–2028 units): 325 kW average — essentially all current-generation ES cabinets.
- Legacy fleet mix (not used for new production): older ES-5000 ~100 kW and ES-5700 ~200 kW cabinets still sit in the installed base, so fleet-average kW per historical box is lower than 325 kW. Cumulative ~1.5 GW across ~1,200 sites does not imply 325 kW per site; sites are multi-box farms.
- One Motley Fool Q2 2025 recap cited “more than 22,000 energy servers.” At 1.4–1.5 GW that would be ~64–68 kW each — that is the power-module size, not the cabinet. We do not use that figure as box size.
Worked conversion used on this page: boxes = MW shipped ÷ 0.325. Example: 1,500 MW in 2026 → 4,615 boxes.
Acceptances vs factory production (channel lag)
The three yearly box counts on this page are factory production (nameplate × utilization at Fremont / Copy Exactly lines). They are not last year’s product acceptances grown by a percentage. Bloom’s own filings treat those as different clocks.
What “product accepted” is
Through FY2023 Bloom published a key operating metric: product accepted — “the number of customer acceptances of our Energy Servers in any period. We recognize revenue when an acceptance is achieved.” Each server was counted in 100 kW equivalents. FY2023 10-K table:
| Year | Product accepted (100 kW eq.) | MW accepted, net | Product revenue |
|---|---|---|---|
| 2022 | 2,281 | 228 MW | $881 million |
| 2023 | 2,682 (+17.6%) | 268 MW | $975 million |
Acceptance is not the factory gate. The 10-K says control (and therefore product revenue) transfers at customer acceptance, which depending on the contract is (1) shipped and delivered, (2) mechanical completion, or (3) commencement of operations (power on). Installation is a separate line, often later, percent-complete. The FY2024 10-K then stopped reporting product-accepted kW because solutions (islanded microgrids, BOP) made a simple kW count “no longer as relevant.” CFO on the Q4 2024 call: in earlier years “timing of product revenue recognition was somewhat divorced from timing of product shipments”; now product revenue is “primarily recognized on shipments,” but still not on factory completion.
Where production sits before it is accepted
- Finished-goods inventory (GuruFocus / 10-Q): $149M (Dec 2024) → $282M (Jun 2025) → $241M (Sep 2025) → $167M (Dec 2025). A $100M+ swing in finished servers is months of channel lag, not a change in nameplate.
- Total inventory: $545M (YE2024) → $643M (YE2025) → $758M (30 Jun 2026). Q2 2026 10-Q: inventory +$115M because “we built additional units to support future customer demand”; deferred cost of revenue +$36.5M as systems hit acceptance; AR + contract assets +$277M on “customer acceptance cycles.”
- Site / channel delays: FY2024 10-K — 73 MW sold to a distributor, end customer did not deploy, alternative site “expected to be operational in 2027.” Those boxes were produced and sold years before they generate power. FY2025 10-K still cites permitting / facility-access delays on some 2025 installs.
Why 2026–2028 units are not 2025 acceptances rolled forward
2025 product revenue was $1.531 billion (FY2025 10-K). At ~$3.0–$3.4/W that is roughly 450–510 MW accepted, or about 1,400–1,570 current 325 kW boxes. Rolling that forward at 20–40% would give ~1,700–2,200 boxes in 2026 — far below the factory forecast. We do not do that.
Factory constraint is the other clock: management said ~1 GW/year now, 2 GW/year by end of 2026. A 1.5 GW produced year is ~4,600 boxes at 325 kW. Q2 2026 product revenue of $935M already implies an acceptance run-rate that only exists if Fremont is building well above 2025 accepted MW. Channel lag can still leave factory MW ahead of accepted MW (finished goods, SK timing, site readiness) or, in a catch-up quarter, behind it.
| 2025 (actual, accepted) | 2026 (this page: produced) | 2027 (produced) | 2028 (produced) | |
|---|---|---|---|---|
| MW | ~0.45–0.51 GW accepted | 1.50 GW factory | 2.20 GW factory | 2.80 GW factory |
| Boxes @ 325 kW | ~1,400–1,570 accepted | 4,600 produced | 6,800 produced | 8,600 produced |
Assumption used in the three yearly numbers: production can exceed acceptances by ~0–30% in a ramp year (inventory + delayed sites). 2026 factory 1.50 GW vs ~1.0–1.2 GW implied by the $3.9–$4.2B revenue guide’s product slice is that lag, not a second count of the same boxes.
Factory capacity vs utilization for box output
Bloom’s disclosed nameplate is annual production capacity run rate at the Fremont, California cell-print plant (164,000 sq ft lease at 44408 Pacific Commons Blvd., grand opening 2022). Stacks printed in Fremont are assembled into Energy Servers at the Newark, Delaware factory (owned 178,000 sq ft plus leased R&O/warehousing) and a Republic of Korea full-assembly plant (SK ecoplant JV). Company filings treat Fremont GW as the bottleneck; Delaware and Korea are Copy Exactly assembly, not a second GW of cells.
Plant nameplate (as of FY2025 10-K / Q2 2025–Q2 2026 calls)
| Plant | Role | Nameplate used here | Notes from filings |
|---|---|---|---|
| Fremont, CA (164k sq ft + 89k R&D/mfg) | Cell print / stack capacity (the GW number) | 1 GW/year now; 2 GW/year by end of 2026; site can go to ~5 GW | FY2025 10-K: “expanding our annual production capacity run rate at our Fremont facility from 1 gigawatt to 2 gigawatts … complete … by the end of 2026.” Each extra 1 GW ≈ 6–9 months and $100–$150M. 2022 press: Fremont “annual output of more than one gigawatt.” |
| Newark, DE (owned 178k + leased R&O ~133k) | Energy Server assembly / Copy Exactly | Sized to take Fremont stacks; 25 acres for more lines | FY2025 10-K: “same manufacturing lines … Fremont, California cell print factory and Delaware assembly facility.” Not a separate GW of cells; utilization tracks Fremont stack output. |
| Korea assembly (SK ecoplant JV) | Full assembly for Korea offtake | Local assembly of U.S.-printed stacks; not additive GW | FY2025 10-K: “full-assembly facility in the Republic of Korea.” Offtake ~60–80 MW/year of the factory total, not extra nameplate. |
Utilization and implied annual box output (the three yearly forecasts)
Bloom does not publish a utilization %. We apply a factory-utilization assumption to Fremont nameplate (Delaware/Korea assemble what Fremont prints). Formula: implied boxes = Fremont nameplate (GW) × utilization × 1,000 MW/GW ÷ 0.325 MW/box.
| Year | Fremont nameplate (annual run-rate, time-weighted) | Utilization used | Implied factory MW | Implied boxes @ 325 kW |
|---|---|---|---|---|
| 2026 | Ramps ~1.0 GW → 2.0 GW YE; mid-year ~1.75 GW run-rate | ~85% of that mid-year run-rate (ramp, staffing, yield) | 1.75 × 0.85 ≈ 1.50 GW (range 1.2–1.8) | 4,600 (3,700–5,500) = 1,500 / 0.325 |
| 2027 | Full-year ~2.0 GW plus ~0.2–0.4 GW Copy Exactly if they add a line | ~90–95% (first full year at 2 GW) | ≈ 2.20 GW (1.8–2.6) | 6,800 (5,500–8,000) = 2,200 / 0.325 |
| 2028 | ~2.6–3.4 GW if they keep adding 1 GW lines “ahead of committed orders” | ~85–95% (new lines start under-utilized) | ≈ 2.80 GW (2.2–3.5) | 8,600 (6,800–10,800) = 2,800 / 0.325 |
Ceiling check: a fully utilized 2 GW Fremont nameplate is 2,000 / 0.325 ≈ 6,150 boxes/year. 2026 sits below that because the 2 GW line is not available all year and we use 85% utilization. 2027–28 exceed 6,150 only if Newark/Fremont Copy Exactly adds more than 2 GW of stacks. Delaware and Korea do not add extra boxes beyond Fremont stack output in this model.
SK ecoplant / Korea offtake in boxes
FY2025 10-K: Korea deployed base ~682 MW. PDA with SK eternix runs through 31 Dec 2027. SK ecoplant take-or-pay was at least 500 MW for 2022–2024; December 2023 added 250 MW of purchase commitments through 2027 and stretched remaining original take-or-pay. That 250 MW over 2024–2027 is roughly 60–80 MW/year if even, or ~185–250 boxes/year at 325 kW — a floor in the Korea channel, not the whole factory. U.S. AI / AEP / Oracle volumes dominate 2026–28 output.
Revenue cross-check (not the primary method)
If product ASP is about $3.0–$3.4 per watt (2024 product $1.09B; 2025 product ~$1.5B; Q2 2026 product $935M at ~88% of revenue), then:
- 2026 product ~$3.4–$3.7B (88% of $3.9–$4.2B guide) → ~1.0–1.2 GW accepted.
- Factory 1.5 GW in 2026 can exceed acceptances by channel lag (finished goods, SK timing, site readiness).
- 2027–28 have no official MW guide; the factory table is the constraint we use.
Sources
- Bloom Energy Form 10-K for year ended 31 Dec 2023 (filed Feb 2024): product accepted 2,682 systems / 268 MW (vs 2,281 / 228 MW in 2022), measured in 100 kW equivalents; revenue at customer acceptance (ship, mechanical completion, or COO).
- Bloom Energy Form 10-K for year ended 31 Dec 2024 (filed 27 Feb 2025), SEC: ~1.4 GW deployed; Korea ~600 MW at that date; product revenue $1.09B; discontinued the product-accepted kW metric; 73 MW distributor shipment later re-sited to 2027.
- Bloom Energy Form 10-K for year ended 31 Dec 2025 (filed 9 Feb 2026): product revenue $1.531B; ~1,100 sites; Korea ~682 MW; Fremont expansion to ~2 GW by end of 2026; SK PDA through 2027 and +250 MW commitments; AEP up-to-1 GW supply agreement; remaining performance obligations $394.4M product/install.
- Q2 2026 Form 10-Q: inventory $758M (+$115M “built additional units to support future customer demand”); deferred cost of revenue and contract assets moved on customer acceptance cycles — factory build ≠ same-period acceptances.
- Finished-goods inventory (10-Q / GuruFocus): $149M YE2024, $282M Jun 2025, $167M YE2025 — the visible channel buffer between production and acceptance.
- Q2 2025 earnings (31 Jul 2025): 1 GW factory now, double to 2 GW by end of 2026; 2025 revenue guide $1.65–$1.85B (later ~$2.0B actual).
- Q2 2026 earnings (28 Jul 2026): revenue $1.065B, product $935.4M; FY2026 revenue raised to $3.9–$4.2B; Copy Exactly capacity adds; Oracle MSA up to 2.8 GW with 1.2 GW contracted.
- Investor presentation (Nov 2025): stack / power module ~65 kW, Energy Server 325 kW, 3.25 MW block; ~1.5 GW deployed as of Q3 2025.
- Bloom Energy Server brochure (2025): “Figure 1… 325 kW Energy Server”; electrical output 325 kW; 1 GW manufacturing capacity cited at that printing.
- AEP supply agreement press release, 14 Nov 2024: up to 1 GW, initial 100 MW order.
- Fremont plant: Bloom 20 Jul 2022 grand-opening release (164,000 sq ft, >1 GW annual output); 8 Apr 2024 48C tax-credit release (same 1 GW Fremont nameplate); FY2025 10-K Manufacturing Facilities (Fremont 164k + 89k, Newark 178k owned, Korea full-assembly); Q2 2025 call (double 1 GW → 2 GW by YE2026).
Not company guidance. Not investment advice. Point estimates will be wrong; ranges are the honest output. If Bloom’s next Energy Server rating moves off 325 kW, divide the same MW by the new kW/box.