Bloom Energy [BE] — Company Research & Unpublished Scoping

Phase Space Research | July 28, 2026 | Tasks 1–3 | Skill v1.4.2

Event: Q2 2026 reported after the close on 2026-07-28. Regular-session close $166.84 (−11.3%); aftermarket $183.90 (+10.23%). The −11.3% was pre-print derating, not a reaction to the print.

0. THE SETUP IN ONE PARAGRAPH

Bloom sells solid-oxide fuel cells as onsite power, and the AI data-center buildout has turned a 20-year-old ~20%-CAGR company into one growing 100%. Q2 was a genuine blowout — first billion-dollar quarter, GAAP operating income of $182.2m against a small loss a year ago, positive operating cash flow — and guidance went up for the third consecutive time. It is also, at $183.90, a 68x forward non-GAAP P/E on a business whose product backlog was last quantified in February and whose accounting was attacked by a short seller three weeks ago. Both of those things are true at once.

1. Q2 2026 RESULTS — TIED TO THE FILED STATEMENTS (C8 rule)

Source: 8-K filed 2026-07-28, accession 0001628280-26-050150, EX-99.1.

Tie-out: every line above is transcribed directly from the filed EX-99.1 condensed consolidated statement of operations, not derived. Revenue $1,065,365k, operating income $182,237k, net income to common $196,290k all agree to the dollar. Gross margin recomputes to 33.38% vs the 33.4% stated.

Full quarterly revenue history (press-release basis, total revenue):

* 2024 H1 quarters from XBRL ASC-606 revenue; note XBRL RevenueFromContractWithCustomerExcludingAssessedTax excludes lease/electricity revenue and therefore runs ~$5m/qtr below press-release total revenue. Do not mix the two bases.

2. THE GUIDANCE ARC — THIS IS THE THIRD RAISE

Midpoint moves: revenue $3.60bn → $4.05bn (+12.5%); operating income $675m → $850m (+25.9%); EPS $2.05 → $2.70 (+31.7%). Earnings guidance is being raised roughly 2.5x faster than revenue guidance — genuine operating leverage, not just a demand raise.

The H2 math — the bear's strongest arithmetic point

H1'26 actual revenue = $1,816.4m. FY guide $3.9–4.2bn ⇒ H2 implied $2,083.6–2,383.6m, i.e. $1,041.8–1,191.8m per quarter against Q2's $1,065.4m. The low end of guidance implies a sequential decline. Against H2'25 of $1,296.7m, however, H2'26 midpoint is still +72.3% YoY. So the deceleration is real in sequential terms and modest in year-over-year terms. Both framings are honest; the bear uses the first, the company the second.

3. UNPUBLISHED SCOPING — WHAT IS NOT IN THE NUMBERS

3.1 The Oracle warrant: $306.5m of future contra-revenue (the most underappreciated item)

Found by tracing an unexplained new balance-sheet line, not from any commentary.

• Oct 28, 2025: in connection with an onsite-power partnership for Oracle's AI data centers, BE agreed to issue Oracle a warrant over 3,531,073 shares at a $113.28 strike (that day's close).
• Accounted for under ASC 606/718 as consideration payable to a customer's customer, which means it is recognised as a reduction of revenue as the underlying Energy Servers are delivered.
• Fair-value march: $55.9m (Dec 31, 2025) → $183.6m (Mar 31, 2026) → $261.3m grant-date FV (Apr 9, 2026).
• $12.8m was already recognised as a reduction of revenue in Q1'26.
• June 30, 2026 balance sheet carries a "customer consideration asset" of $306.5m ($90,967k current + $215,533k long-term) — a line that did not exist at Dec 31, 2025.

Why it matters: $306.5m of contra-revenue is queued against future periods — ~7.6% of a full year of FY2026E revenue — and it lands precisely as the Oracle volumes (the growth story) are delivered. Reported revenue growth will be understated relative to units shipped, and gross margin optics will be pressured, for reasons that have nothing to do with pricing or cost.

Open item (flagged, not resolved): grant-date fair value is disclosed as ~$261.3m but the carried asset is $306.5m. The Q2 10-Q (not yet filed as of this writing) should reconcile the $45.2m difference. I have not assumed an explanation.

A dated catalyst falls out of this: the warrant is exercisable for six months from the April 9, 2026 grant date — i.e. expiring ~October 9, 2026. At $183.90 it is ~$70.62 in the money, ~$249m of intrinsic value; exercise is close to certain, delivering BE $400.0m of cash and 3.53m shares (~1.1% dilution).

3.2 Backlog: quantified exactly once, and not since

BE has quantified backlog once, at the FY2025 annual print, and has not updated it in either quarter since — including today, three weeks after a short seller publicly disputed it, and after the company's own July 9 rebuttal expressly invoked "our current fuel cell demand and backlog" as a defence. This is the same disclosure pattern flagged on TWST (which retired its orders disclosure for six quarters while promoting an adjusted metric).

And the definition inflates it. Per the company's own footnotes to that single disclosure: - Product backlog "reflects anticipated ITC and other tax incentives as applicable" — so the ~$6bn is not $6bn of Bloom product revenue; it includes tax credits accruing to the customer/financier. - Service backlog is contracted O&M "for installed Energy Servers as well as Energy Servers to be delivered" — i.e. the ~$14bn balance (the difference between the $20bn headline and the $6bn product figure) is long-dated O&M stretching over the multi-decade service life of the fleet.

So "~$20 billion of backlog" is roughly 70% long-duration service annuity and 30% product-including- tax-credits. It is not $20bn of orders, and — like Alphabet's $519.5bn RPO — it carries no disclosed weighted-average duration and no 12-month conversion percentage.

3.3 Transcript mention-frequency — hypothesis generation

Source: Alpha Vantage EARNINGS_CALL_TRANSCRIPT, pinned; normalised per 10,000 words. NOW COMPLETE — all 9 quarters, 2024Q1 → 2026Q1. The Q2'26 call transcript is not yet published by the vendor and is not included; that remains open.

⚠ A CLAIM MADE IN THE FIRST VERSION OF THIS MEMO WAS FALSIFIED BY THE COMPLETED SERIES

The first version ran on 4 of 9 quarters (the shared Alpha Vantage quota was exhausted mid-run) and stated: "Zero mentions of 'scandium' or 'china' across all four sampled calls. Management has never volunteered the input supply chain." The five missing quarters have now been retrieved and that claim is wrong:

term	24Q1	24Q2	24Q3	24Q4	25Q1	25Q2	25Q3	25Q4	26Q1
scandium (raw)	0	0	0	0	3	0	0	0	0
china (raw)	0	0	0	1	6	0	0	0	0

Management did discuss both — in 2025Q1, one of the quarters I could not retrieve. The absence was an artifact of the incomplete sample, exactly the failure mode the ISRG re-run taught (three claims falsified there for the same reason). I disclosed the limitation and made no "first-ever" claim, which is why this is a correction rather than a published error — but the underlying inference was still wrong, and it is recorded here rather than silently amended.

The corrected claim is narrower and, I think, stronger: scandium appears in exactly one quarter of nine, and in ZERO prepared remarks in any quarter — it has never once been volunteered by management, only surfaced in Q&A. china appears in prepared remarks exactly once (2025Q1). Both cluster in 2025Q1, which is also the peak quarter for supply chain (14.4 per 10k) and tariff (7.2) — i.e. management addressed the input supply chain once, under tariff pressure, in response to questions, and has not returned to it since. Including in the quarter after a short seller made scandium sourcing the centrepiece of an attack.

Per 10,000 words, full series:

A SECOND CORRECTION — the backlog finding is narrower than first stated. The first version implied backlog disclosure had been retired. The full series shows an annual rhythm: backlog peaks in the Q4 call (19.2 in 24Q4, 23.6 in 25Q4) and is discussed lightly in between, with 2026Q1 carrying 6 mentions, 5 of them unprompted. So management talks about backlog regularly. The precise, defensible statement is: BE has put a backlog NUMBER in a press release exactly once (Q4'25), has never put one in a 10-Q at all, and did not re-quantify it at Q2'26. That is still a real Gate 2B leg-3 failure — an undated, unquantified-in-filings backlog cannot be sized — but it is not the TWST/CAI "retired disclosure" pattern, and the calibration file has been corrected to downgrade BE as an instance of that technique.

Hypotheses that SURVIVE the completed series: 1. Capacity is the binding constraint, and 26Q1 is a genuine peak (42.8 per 10k vs a prior max of 26.6), with 18 of 30 mentions unprompted. Consistent with the H2 guidance shape: if throughput rather than demand is binding, revenue timing slips while backlog builds. 2. The hydrogen/electrolyzer franchise is dead — zero mentions for five consecutive quarters after being a core part of the 2024 equity story. Nothing in the financials flags an impairment. Open question: is there an unrecognised impairment risk? 3. Oracle was material from 2025Q2, five months before the warrant was agreed (Oct 28, 2025). The commercial relationship predates the equity consideration, which slightly weakens a "they bought the deal" reading and strengthens a "they paid to expand an existing deal" one.

Hypothesis KILLED by the completed series: "data center" is not in decline — the 4-quarter sample showed 16.4 → 10.0, but the full series has no trend.

4. THE SHORT REPORT — AND WHAT THE FILINGS ACTUALLY SAY

July 8, 2026: Hunterbrook Media published "Bloom's Big Lie," disclosing it may hold positions that profit from a decline. Core allegations: (a) BE sources scandium oxide — a critical dopant in its solid-oxide cells — from Chinese suppliers via intermediary countries despite the CEO's public claims of no China supply chain; (b) claims regarding financial results and accounting; (c) a backlog dispute.

July 9, 2026: BE filed an 8-K under Item 7.01 rejecting the report "categorically," stating it has sufficient scandium oxide for current demand and backlog, that supply "is not dependent on China," and that it has "clear visibility into our supply chain to support production of 25GW of fuel cells per year."

What I can and cannot verify from filings: - Scandium sourcing: not verifiable from public filings. BE does not name input suppliers, and the term appears zero times across the four transcripts sampled. Unresolved either way. - The 25GW/yr capacity claim: this is a supply-chain visibility claim, not a demand or capacity claim. At roughly $3,000–4,000/kW, 25GW would be $75–100bn of revenue against FY2026E of ~$4.05bn. It should not be read as a revenue-capacity statement, and BE did not present it as one. - Accounting: the two items I independently surfaced — the $306.5m Oracle warrant contra-revenue and the once-only backlog disclosure — are fully disclosed in the filings, not concealed. They are aggressive in presentation (a $20bn backlog headline that is ~70% multi-decade O&M) rather than improper in accounting. On the evidence I can check, the accounting allegations are not supported. - One genuine quality-of-earnings point in the company's favour: Q1'26 contained $373.3m of related-party revenue — 49.7% of the quarter — sold into the Brookfield JVs in which BE holds equity. Q2'26 contained $2.8m, i.e. 0.3%. Q2's growth is almost entirely third-party and arm's length. Q2 is materially higher-quality revenue than Q1, which cuts directly against the short thesis and is not something the bear case has credited.

5–7. FACTOR SCORECARD, VALUATION AND CONSENSUS BRIDGE

Moved to the companion document Bloom_Energy_Valuation_2026-07-28.md (the Valuation page of the published site) so that this document stays the business/scoping record and the valuation record stands on its own. Headline: probability-weighted value $72.43–$94.98 against $183.90; Gate 4 fails by 31.4 points annualised; the price requires BE to still trade at 53–80x earnings in 2029.

8. CALIBRATION NOTE — THE HORIZON MISMATCH (candidate new item)

BE is the sharpest test this framework has faced of the "too strict?" question, because the factor evidence and the valuation evidence point in opposite directions with unusual force:

• Empirically strongest long signals available — top-decile 12-1 momentum, a +29%/+90% revenue/EPS surprise, a +24.4% EPS-guidance revision, and −11.9% accruals — all of which are documented, replicated, positive-alpha signals over 1–12 month horizons.
• Gate 4 is a 3-year scenario DCF, which values the name 31% below spot on an undiscounted expected-value basis, failing the hurdle by 16.4pp.

These are not contradictory findings. They are answers to different questions on different horizons. Momentum and PEAD say "this stock likely outperforms over the next 1–2 quarters." The DCF says "the 3-year holding-period return from $183.90 is poor." Both can be — and here probably are — true.

The framework currently has no way to express that, so it defaults to no. That is a structural limitation worth logging: a name can be a good 3-month long and a bad 3-year hold, and Gate 4 is silently imposing a 3-year horizon on a system whose two strongest empirical signals are short-horizon.

Logged to CALIBRATION_WATCH.md as a new item. Not acted on — per the standing rule, gates are not loosened unilaterally on the strength of an entry there.