Prior Research (prepared by 2026-09-25; current call and filed-period check above)

Summary

Bloom Energy makes the Bloom Energy Server, a solid-oxide fuel cell system that converts natural gas, biogas or hydrogen into electricity electrochemically, without combustion, at the customer's site. Its 10-K (filed February 9, 2026) frames the pitch as time to power: US generator-interconnection queues totaled 2,300 GW at the end of 2024, turbine makers quote multi-year delivery windows, and skid-mounted fuel cells install "in weeks and months rather than years, subject to permitting and site conditions."

That pitch found its buyer in AI data centers. June-quarter revenue was $1.065 billion, up 165.5% on the year, and operating margin reached 17.1% against −0.9% a year earlier (Massive financials, from the 10-Q filed July 28). On July 28 management raised 2026 revenue guidance to $3.9–$4.2 billion, from $3.6 billion, its second raise of the year.

The decisive uncertainty is conversion. The product runs on a gas line and needs an air permit, and the largest named project, Oracle's Project Jupiter in New Mexico, has had trouble with both: a rerouted pipeline in August, a pending fuel-cell air permit, and an Oracle force-majeure notice on September 24. The market has priced a great deal of success: about $85.0 billion of market value at the September 25 close.

Why This Moment Is Different

The customer changed. The 10-K describes three markets (data centers, commercial and industrial sites, utilities) and says nearly all product revenue so far has come from the Energy Server. In data centers Bloom sells across "hyperscalers, colocation operators, neocloud providers, developers, and infrastructure investors—each anchored by lighthouse customers."

The owner base changed. S&P Dow Jones Indices added Bloom to the S&P 500 in its September rebalance (Bloomberg, September 4), effective before the open on September 21. On the September 18 rebalance session the stock fell 5.39% on 5.55 times normal volume.

The first large project hit a formal notice. Project Jupiter's delay went from analyst warnings (SemiAnalysis, July 18) and an attributed loan-pricing report (about $18 billion of related loans at 89–91 cents on the dollar, relayed on X September 19) to an Oracle force-majeure notice on September 24. Morgan Stanley said the notice puts data-center loan and lease documents under fresh scrutiny (Bloomberg, September 25). That matters for Bloom's model: most of its end users buy power, not equipment, so a financier buys the server, and financing terms decide how fast orders ship.

Rates rose into it. Five-year Treasury yields crossed 5% on September 23 (Bloomberg) and the 30-year yield reached its highest level since 2004 on September 24 (CNBC and Bloomberg). Financed deployments carry that cost.

Where the Energy Comes In: Who Bloom Powers

Bloom's role in the AI build is behind-the-meter generation: power made on the data-center site and fed to the load without waiting for a utility interconnection. The table lists every counterparty on record; the stated sizes follow it, one source each, because most rest on reports rather than filings.

Oracle. @aakashgupta reported the April 14 deal at 2.8 GW in total, crediting Bloom's 55-day delivery of an earlier Oracle system; only part of it was contracted at signing, per the same post.

Project Jupiter. SemiAnalysis put the campus's April plan at up to 2.45 GW of Bloom solid-oxide cells, replacing gas turbines. A September 19 report relayed on X described the site's power plan as a gas-fired plant of similar size; Bloom's cells run on that same gas, without combustion.

American Electric Power. The 10-K says AEP "is expected to procure up to 1 GW" of Bloom's fuel cells under a November 2024 agreement, starting with a 100 MW order. The Motley Fool (February 3) reported a $2.65 billion AEP order in January 2026.

Nebius and Brookfield. Sherwood News (May 21) valued the Nebius partnership at up to $2.6 billion. Brookfield's framework was "up to $5.0 billion over five years" when signed (10-K) and $25 billion in total after the expansion described on the July 28 call.

Hyperscalers are claimed, not named. A summary of the second-quarter call in the news pull says all major US hyperscalers and more than a dozen other customers now use Bloom's on-site power; @saso_capital's account of the same call says they have "validated and approved" the technology and that chief executive KR Sridhar declined to separate active deployments from backlog and definitive agreements. Approval is not an order. Oracle is the only hyperscale cloud buyer named in the corpus.

The Meta–Amazon–Muse story does not run through Bloom. Meta launched its Muse agent on September 8 and Amazon blocked it from shopping on Amazon around September 20 (Bloomberg, September 21). The infrastructure constraint reported for Muse's WhatsApp rollout (September 21, relayed on X) was a shortage of CPUs, GPUs, DRAM and SSDs, not electricity. Agent-scale AI raises total compute and therefore total power demand, which is the general tailwind, but no Bloom contract with Meta or Amazon Web Services appears in filings, captures or news on file.

Meta buys power elsewhere. Meta's power arrangements on file are nuclear: 20-year agreements covering more than 2.6 GW across Vistra's nuclear fleet (per @amitisinvesting, August 25).

So does Amazon. AWS took warrants for up to $340 million of Generac stock in a backup-generator supply deal (CNBC, September 16), and Amazon is reported to be building an off-grid gas plant of 7.65 GW at its GW Ranch site in West Texas (relayed on X, August 7 and September 8). Neither names Bloom.

Where the power competition sits. SemiAnalysis estimated on July 21 that 140 GW of identifiable data-center projects have yet to contract power and will go behind the meter, and expects reciprocating engines, not fuel cells, to take the largest share. Management, per @saso_capital, put Bloom's share of fuel-cell data-center power in the "very high 90s." Both can be true: Bloom can dominate fuel cells while engines and turbines take most on-site megawatts.

How the Business Makes Money

Product plus a long service tail. Revenue comes mainly from Energy Server sales, with recurring revenue from operations and maintenance contracts of 5 to 20 years, cancellable annually (10-K). Most end users prefer to pay for the power they consume, so Bloom usually sells the server to a financier or strategic partner that owns it and sells power under a power purchase agreement. That is why Brookfield and AEP matter: they are the buyers of record for much of the equipment.

Concentrated revenue. In 2025, three customers and distributors accounted for about 43%, 13% and 12% of revenue; the largest is a related party (10-K). The systems run at about 1,100 sites in nine countries.

Manufacturing. Cells are printed in Fremont, California, and systems assembled in Newark, Delaware, where Bloom owns a 178,000 square-foot plant with 25 more acres for expansion; a joint venture with SK ecoplant assembles in Korea (10-K). The company plans "to double our factory capacity from 1 gigawatt to 2 gigawatts by the end of 2026" (10-K risk factors). Management said on the July call that capacity is not the constraint (per @saso_capital).

Tax support. The 2025 tax law kept a 30% investment tax credit that now names fuel cells explicitly, but excluded fuel cells from the domestic-content and energy-community bonus credits (10-K). Pennsylvania's House Bill 1834 reportedly classifies Bloom's cells as clean firm power even on natural gas (relayed on X, August 19).

Hydrogen is optional upside. The Bloom Electrolyzer shares the platform, but the 10-K says nearly all product revenue to date has come from power generation.

What the Company Says It Is Becoming

The 10-K states the goal plainly: "Our long-term objective is to establish our solid oxide fuel cell technology as the standard architecture for onsite power across data centers, advanced manufacturing, critical infrastructure, and other sectors globally." It calls power "a foundational layer of the AI and industrial ecosystems," and lists its edges as time to power against turbine makers and grid upgrades, emissions, reliability, modularity and load-following for AI racks.

In the July 28 results, as quoted in an investor Discord that day, Sridhar said "Bloom is now a standard for AI onsite power." The @saso_capital summary adds, attributed: service gross margin reached 22% in the quarter; four in five 2025 orders came from existing customers; Brookfield's extra $20 billion followed a review of how the first $5 billion performed; and units are fungible across sites under master agreements, so a delayed project need not strand inventory.

What the Street and the Public Conversation Say

Bull framing. Zacks (September 23) ranks Bloom a Strong Buy on earnings beats, the raised guide and the Brookfield expansion. @CKCapitalxx called the July call "one of the most bullish things I've heard for the entire power sector." @ChrisCamillo (July 20) bought the dip after the first Jupiter delay report, calling it "a classic overreaction."

Bear and cautious framing. The Motley Fool (September 24) notes a $20 billion backlog disclosed in February ($6 billion product, $14 billion service) and doubts the build rate, putting the stock at 56 times forward earnings on that date. Hunterbrook's July 8 report alleged undisclosed reliance on Chinese scandium; its later piece questioned whether the cells perform as claimed; and law firms filed class actions for buyers from February 27, 2025 to July 8, 2026 (GlobeNewswire, September 21–22). SemiAnalysis has been the most specific skeptic on Project Jupiter's timing since May and restated it on September 24.

Positioning. The name is crowded with fast money. CNBC (September 8) reported Bloom's options trading more actively than SpaceX's, and Leopold Aschenbrenner's Situational Awareness buying options positions (September 11); Nancy Pelosi disclosed a purchase on August 21. Short interest was 19.0 million shares, 1.33 days of volume, on September 15 (Massive). Insiders sold: five officers and directors reported sales of 54,625 shares, about $13.2 million, in Form 4s filed August 14–18, including directors Jeffrey Immelt and John Chambers, and four officers sold smaller lots on September 16 (SEC Form 4s); the Form 4s captured since August include no open-market purchase.

Today's social sweep was empty of substance: an old Hacker News item about ventilator repairs and two YouTube stock videos without transcripts.

The Conviction Chain

Weakest link: link 3. Demand and pricing evidence is strong; the proof that the largest named projects get gas and permits on time is not. Conviction stays medium until the October 27 report shows the order book converting through the Jupiter delay.

Financial Proof

Quarterly results, as reported (GAAP). Margins are computed by code from the reported lines.

The provider's per-share values for Q3 and Q4 2025 are unusable and are left out. SEC XBRL for the March 2026 quarter shows revenue of $746.4 million, slightly below Massive's $751.1 million; the June 2026 10-Q is not yet in the SEC data on file, so that row is single-source.

What the series says. Revenue climbed from $326.0 million in the March 2025 quarter to $1.065 billion in the June 2026 quarter, gross margin rose from 20.4% to 33.4% with two small dips along the way, and operating margin went from negative to 17.1%. That is operating leverage from volume, which is what a capacity-bound manufacturer should show when demand outruns supply. Non-GAAP earnings of $0.78 a share in the June quarter beat a $0.41 consensus estimate (Yahoo Finance), and management guided 2026 to about 34% non-GAAP gross margin and $2.55–$2.85 of adjusted earnings per share (company release, July 28).

Cash is lumpy. Quarterly operating cash flow in the table runs from negative through mid-2025 to strongly positive in the December 2025 and June 2026 quarters (Massive). The data on file does not break out how much came from customer prepayments or other working-capital timing, so quarterly cash flow is not a steady-state measure.

Full-year 2025 (10-K). SEC XBRL shows fiscal 2025 capital spending of $56.8 million and fiscal 2025 free cash flow of $57.2 million, with the same caveat on composition. Inventory stood at $758.2 million on June 30.

Balance sheet and dilution. Cash rose from $595.1 million on September 30, 2025 to $2.45 billion on December 31 and $2.49 billion on March 31, 2026 (SEC XBRL), after $1.47 billion of financing inflow in the December quarter (Massive). Long-term debt went from $1.13 billion to $2.62 billion over the same step and was $2.48 billion on June 30. The 10-K discloses a December 2025 credit agreement secured by a lien on substantially all of the company's assets, and warns that conversions of its outstanding convertible notes may be settled in shares. Weighted basic shares were 232.5 million in the June 2025 quarter, 234.9 million in September 2025, 281.7 million in March 2026 and 287.3 million in June 2026 (diluted: 323.3 million); 294.5 million shares were outstanding as of September 25. Total equity rose from $948.0 million on March 31, 2026 to $1.64 billion on June 30, 2026, more than the June 2026 quarter's net income in the table, while that quarter's financing cash flow was only $1.2 million; the filing text that would explain the difference, such as note conversions, is not in the data on file.

Risks

Gas supply. The servers run continuously and need "a constant source of fuel such as natural gas, biogas, or hydrogen," usually from local gas utilities; "adverse fuel supply constraints or fuel outside of our fuel specifications may delay or prevent the deployment" (10-K). Every large Bloom data-center deal is also a gas-delivery project. Project Jupiter's pipeline reroute (August 19 report) is the live example. Customer economics also depend on gas prices against grid prices.

Permitting. The 10-K warns that "delays in obtaining these approvals and permits could stall the installation process of our products and adversely affect our revenue." Fuel cells avoid combustion, but they still need air permits: Jupiter's fuel-cell air permit was pending with a public hearing ordered (SemiAnalysis, July 18) and officials delayed the permits to power the site (Semafor, September 25); Vineland drew a stop-construction order (August 10). In Texas, @KashRamki (September 21) noted that the state environmental agency's approvals are air permits and typically take about 90 days in an attainment zone; the post he quoted said Nebius signed its Vineland site before its air permits were final and then pivoted to Bloom. Air permits gate every on-site generator; cleaner emissions can shorten the process, not skip it.

Customer and project concentration. Three buyers made up about 43%, 13% and 12% of 2025 revenue, the largest a related party (10-K). A single campus, Project Jupiter, carries a plan of up to 2.45 GW, larger than the annual factory capacity the 10-K targets for the end of 2026 ("2 gigawatts").

Dilution and debt. The share count and debt both jumped over the past year; convertible notes may settle in shares; the credit agreement is secured by substantially all assets (10-K).

Supply chain and litigation. The 10-K names China as a source of rare-earth metals in tier-2 and tier-3 sub-assemblies and Taiwan as the principal source of some components. The scandium class actions have a lead-plaintiff deadline of September 28.

Crowding. Heavy options trading, a well-known AI-focused fund, a congressional disclosure and index buyers all arrived within two months. The drawdown reached 52.51% from the high on July 28; a crowded base can do that again.

Stock Setup and What to Watch

The live panel below updates each close. At the September 25 close the stock sat 12.6% above its 20-day average at $256.50 with RSI 62.8, neither overbought nor washed out. The nearest support shelf is $272.76; below it, the 20-day, then the $242.56 value-area high (the top of the heaviest-traded price range in the desk's volume profile) and the 50-day at $229.68. The June 18 close of $328.91 is the first overhead level; the stock is 17.81% below its 52-week high.

Dated items:

• September 28: lead-plaintiff deadline in the scandium class actions (scheduled).
• October 27: third-quarter report, scheduled after the close. The test is revenue pace against the $3.9–$4.2 billion guide, any change to Project Jupiter's status, and whether the 10-Q quantifies backlog or cancellations.
• Open: the Jupiter developer's response to Oracle's notice and New Mexico's permit hearing; the Nebius Vineland permits; any second named hyperscaler contract.

Cross-references

• Watchlists: AI power, electricity, energy, EV and clean energy, AI scan, fuel cells and alternative data-center power, natural-gas turbines for data-center power, war sensitivity, SpaceX S-1 supply chain, tech insider buys, insider universe.
• Momentum Board, September 25: the 40–60%-above-the-200-day tier; not among the multi-year outperformers in pullback or the steady climbers; not on the washout board.
• Campaigns: Bloom is not an instrument in "AI power: generation and grid equipment," whose stated failure condition includes a verified behind-the-meter commissioning that supplies usable load ahead of grid conversion — a Bloom success is evidence against that campaign's timing.
• Neocloud lane: Nebius is a named Bloom customer (Vineland). Related profiles: the September 23 Oracle profile (Project Jupiter), Vertiv, Nebius. The September 25 Amazon profile (Generac, its own gas generation) and the Meta deep dive carry no Bloom relationship.

Price Data (live)

Live values from the desk's settled-close summaries. They update with each close; the dated sections above keep the numbers of their own date.

Financial Proof

Filed-period check (as of 2026-09-28): The newest verified quarter ended 2026-06-30 and was filed 2026-07-28. Revenue was $1.07B; GAAP net income was $198.9M; diluted EPS was 0.62. Source: Massive income statements and the filing. Dated historical tables below retain their own period basis.

The recent filed revenue series shows the direction of the business; missing cells are left unavailable.

Research Log

Sources

• Tape (price, RSI, averages, returns, support shelf, value-area high, volume, market value): the desk's settled-close summaries as of 2026-09-25 — war-sensitivity (same row in ai-power.json, thebuildout-fuel-cells-alternative-dc-power.json); cohort and market comparison ai-power, market-pulse.json. September 24 close ($266.65) and tier: 2026-09-24-monster-scan; September 25 tier: 2026-09-25-monster-scan; rates context: 2026-09-25-market-pulse.
• Financials: Massive (Polygon) statements income-statements, cash-flow-statements.json, balance-sheets.json, short-interest.json, company.json; code-computed margins 2026-09-25; SEC XBRL key-financials (cash, capital spending, Q1 2026 cross-check); reconciliation run 2026-09-25 (revenue, operating margin, net margin and operating cash flow flagged disputed on a full-year-versus-trailing basis mismatch; capital spending and free cash flow single-source SEC).
• Filing narrative: 10-K filed 2026-02-09, business and risk-factor sections 10k-business, 10k-risk_factors.json — SEC 10-K. Q2 2026 filings (not yet in the text cache): 10-Q, 8-K results, 8-K July 9, Item 7.01; filing list submissions.
• Q2 print: 2026-07-28-be-earnings — Business Wire release, Bloomberg, July 28, MarketWatch, July 29. Scheduled Q3 report: 2026-10-27-be-earnings.
• Call summaries (attributed): @saso_capital via @ChrisCamillo, July 29; stockanalysis.com Q2 transcript page as summarized in articles (fetched 2026-09-18); @CKCapitalxx, July 30; release language 2026-07-28-dumb-money-trade-ideas-and-research-61365260-be-ceo-the-demand-for-bloom-energy-s-solutions-kee.
• Oracle and Project Jupiter: @aakashgupta, April 14 (2026-04-14-aakashgupta-leopold-aschenbrenner-13f); SemiAnalysis thread, July 18, air permit, 2.45 GW plan; SemiAnalysis, September 24; pipeline reroute relayed on X, August 19 (2026-08-19-thebuildout-oracle-and-openai-reroute-pipeline-for-project-jupiter-after-208932ba); loans and delay relayed on X, September 19; CNBC, September 24; Semafor, September 25; Bloomberg, Morgan Stanley on the notice, September 25; @ChrisCamillo, July 20; Yahoo quote-page line in latest (fetched 2026-09-25, ok, not stale; five quote pages, newest dated 2026-09-22).
• Nebius: Nebius Q2 2026 shareholder letter, SEC EX-99.2 (NBIS-5f05a8ef4e34ce27); Sherwood News, May 21; Hunterbrook on the Vineland stop order, August 13 via 2026-08-13-kawzinvests-nbis-vineland-got-hit-with-two-stop-construction-85267003; Nebius call language 2026-08-12-dumb-money-channel-478096-83122964-nbis-on-the-vineland-nj-site-and-the-switch-to-be-; Nebius at Goldman Sachs, relayed on X, September 9.
• AEP and Brookfield: 10-K (above); Motley Fool, February 3; Zacks, September 23; Brookfield expansion items in articles (fetched 2026-07-07).
• Permitting and policy: @KashRamki on Texas air permits, September 21 (2026-09-21-KashRamki-tceq-approvals-air-permits-remember-tcdc-is-in-a-41973655); Pennsylvania HB 1834, relayed on X, August 19; SemiAnalysis on behind-the-meter share, July 21.
• Meta, Amazon, Muse: Bloomberg, Amazon blocks Muse, September 21; Muse infrastructure constraints, relayed on X, September 21 (2026-09-21-thebuildout-meta-infrastructure-constraints-delay-whatsapp-muse-ai-launc-06f4b703). Corpus search for a Bloom–Meta or Bloom–Amazon relationship across captures, inputs, profiles and scans: none found.
• Meta and Amazon power: @amitisinvesting on Vistra, August 25 (2026-08-25-amitisinvesting-i-think-vst-looks-compelling-here-at-138-vistra--67061929); CNBC, Amazon and Generac, September 16; Amazon GW Ranch gas plant, relayed on X, August 7 and September 8; sibling profile 2026-09-25-AMZN.
• Index inclusion: Bloomberg, September 4.
• Street and positioning: Motley Fool, September 24 (news-latest); GlobeNewswire class action, September 21; GlobeNewswire, September 22; Hunterbrook performance piece; CNBC, options, September 8; CNBC, Situational Awareness, September 11; Pelosi disclosure, relayed on X; social sweep 2026-09-25-social-search-be-bloom-energy.
• Insider filings: Form 4, Immelt, August 17, Form 4, Chambers, August 13, Form 4, Joshi, September 16; cluster and rows filing-watch, 2026-08--be-4-, 2026-09--be-4-.
• Campaign context: the desk's own tooling (read 2026-09-25).

2026-09-28 refresh receipts

• 2026-09-28 settled-close receipt: *, BE row dated 2026-09-28; tape values above are bound to the current summary fields.

• 2026-09-28 financial-period check: income-statements and key-financials; newest verified 2026-06-30 period, filed 2026-07-28 (Massive statement / SEC period cross-check).

• 2026-09-28 news receipt check: latest, fetched 2026-09-25; receipt is not current; latest-session cause remains unassigned.

• 2026-09-28 stock-event scan: research/stock-events/, no BE row dated 2026-09-24 through 2026-09-28.

• 2026-09-28 source-page check: Bloom's September 16 800V DC report models $3.6B less non-compute capex for a hypothetical 1 GW AI data center versus traditional AC architecture. This is a vendor scenario, not a booked order or verified customer saving. Reuters reported September 24 that Oracle sent Project Jupiter's developer a force-majeure notice over possible power-supply delay; Oracle said the project remained on its planned schedule. Oracle's April announcement put Bloom's planned role at up to 2.45 GW. The notice raises execution and timing questions but is not evidence Bloom's contract was cancelled or that its full capacity is booked revenue today. No single-session price cause is established. The Chrome CDP source-page capture is under 2026-09-28-raw.

• 2026-09-28 captured-follows scan: tweets has 3 September 24–28 document(s) tagging BE.

• Retrospective social-search receipt (captured 2026-09-29): 2026-09-29-social-search-be; quality/attention check only, not a September 28 event or price source.

• 2026-09-29 browser discovery check for BE: news-discovery-google-sep29 (key BE); search links only, not evidence of an event or of absence of news.