Company deep dive · personal research公司深度研究 · 個人研究
Bloom Energy (BE)
Onsite power that runs on a chemical reaction, not a flame — and it's become the fast way to power an AI data center while the grid waits. The reasoning is the model's; every number carries its source and date.靠化學反應、而非火焰來發電的「現場電廠」 — 在電網大排長龍時，它成了 AI 資料中心最快的供電方式。推論由模型產生，但每個數字都標註來源與日期。
How we analyze this company. A variant-perception deep dive — first understand the business and the technology (jargon decoded below), then ask what consensus may be misreading. Same nine steps as any stock: business → bull → bear → variant view → catalysts → management → moat → probability-weighted odds → what a double takes. The trust rule: no number from memory — each is grounded to a filing or live data with a date, and the share price is recomputed the moment this page is built.
Forward P/E = live price ÷ 2026 non-GAAP EPS guidance midpoint (~$2.05). 52-week range $24.04 → $351.25 — a ~10× run off the low, now ~34% below the high.預估本益比 = 即時股價 ÷ 2026 非 GAAP 每股盈餘財測中位數（約 $2.05）。52 週區間 $24.04 → $351.25 — 從低點漲約 10 倍，目前距高點約 34%。
headlineA real story — at a price that already believes it故事是真的 — 但股價已經大幅反應
The bull case is strong and probably right: AI data centers need power now, the grid takes years, and Bloom's fuel cells deploy in months. But BE is $232.34, +167% YTD, a ~10× run off the 52-week low (✓ IBKR · 07-10), trading at ~113× forward earnings (live price ÷ 2026 guidance EPS ~$2.05). At that multiple the market has still priced years of near-flawless execution.
The question isn't "is the power shortage real" — it is. It's "what's left at 113× earnings, when any execution slip resets the multiple." The asymmetry is less one-sided than it was at 160× — but a July short report moved the live question from the multiple to the backlog itself (scandium supply; $492.6M of binding RPO against a $20B marketed backlog). See the update log up top.
Update log — material news since publish更新紀錄 — 發布後的重要消息
Dated additions, newest first. The core thesis above stays as-is unless an entry says otherwise.
逐則標日期、最新在上。除非條目另外註明，上方核心論點維持不變。
✓ 2026-07-10Hunterbrook short report; stock round-trips ~29% of the run — valuation gap narrows, debate shifts to "is the backlog real"Hunterbrook 放空報告；股價回吐約 29% 漲幅 — 估值落差收斂，爭論轉向「在手訂單是不是真的」
Hedge-fund-backed short seller Hunterbrook Capital published "Bloom's Big Lie" (~07-08) and disclosed a live short. Three load-bearing claims: (1) scandium supply — scaling from ~1 GW to a 5 GW/yr goal would need ~220 tons of scandium oxide against ~240 tons of total global production (China-concentrated), which they call "physically and commercially unattainable"; (2) backlog quality — the ~$20B backlog Bloom markets is >40× the $492.6M binding remaining performance obligations (RPO) in its latest 10-Q, versus a ~2× max gap across the 10 peers they checked; (3) revenue concentration — 74% of Q4'25 revenue from the Brookfield JV. Bloom fired back the next day, calling the report "false and misleading" and saying audited financials, supply chain and scandium access fully support current operations and growth; the stock bounced ~6% on the rebuttal, then fell again. ✓ Hunterbrook / TipRanks / Investing.com · 07-08→09
Numbers refreshed: price $326.90 → $232.34 (−9.6% on the day, ≈−29% since the 07-01 update) ✓ IBKR · 07-10, +276% → +167% YTD, forward P/E ~160× → ~113× ($232.34 ÷ 2026 guide EPS ~$2.05), market cap ~$93B → ~$66B (~284M sh). 52-wk range now $24.04 → $351.25 (the old $21.53 low rolled out of the trailing window); price sits ~34% below the high.
Does it change the finding? Yes — it splits in two.Valuation: the drop cut the forward multiple 160× → ~113×, so the "priced for perfection" gap the headline flags has narrowed ~30% — the pure over-valuation objection is weaker than on 07-01. Thesis: but the report reframes the bear case from "too expensive" to "are the two pillars real" — the scale-up (scandium) and the backlog ($492.6M binding vs $20B marketed). Those aren't multiple questions, they're does-the-growth-exist questions — and the growth is what justifies any premium. So the decision is no longer mainly about price; it's thesis integrity: refute the scandium + RPO-gap claims and this is a ~29%-cheaper entry on an intact story; can't refute them, and the sell-off is the market pricing a real question, not an overreaction — where adding "because it's lower" is averaging down into an unresolved thesis. Narrative ≠ verified delivery. ◆ framing = judgment定位＝判斷
✓ 2026-07-01Brookfield expands the AI-power financing partnership 5× to $25B; stock jumps ~8%Brookfield 把 AI 供電融資合作擴大 5 倍至 $25B；股價跳漲約 8%
Brookfield and Bloom expanded their framework to build and finance onsite power for AI data centers from $5B to $25B — a 5× increase (the original $5B dates to Oct 2025). The money sits inside Brookfield's dedicated AI Infrastructure Fund (launched Nov 2025, ~$100B deploy target). ✓ Bloom IR / BusinessWire · 06-30 It is a financing framework — capacity to fund projects, not booked orders; no revenue guidance or backlog figure was attached.
Does it change the finding? Direction, no — the gap widened. The expansion is a real positive: it directly softens the bear case's financing leg (Brookfield supplies the capital customers would otherwise raise) and signals sustained hyperscaler demand. But it is optionality-to-fund, not orders — and the ~8% pop lifted the forward multiple to ~160×, so the price more than absorbed the news in a day. "A real story, priced for years of flawless execution" holds, and on valuation it's now more stretched, not less. Optionality ≠ catalyst booked. ◆ framing = judgment定位＝判斷
✓ 2026-06-30Numbers refreshed — little net change; still priced for perfection數字已更新 — 淨變化不大；仍是「完美定價」
Re-pulled live: $279.00, +221% YTD✓ IBKR · 06-30 — essentially flat vs the $274.86 build. But the 52-week range widened to $21.53 → $351.25: the stock spiked ~28% to a $351 high and round-tripped back, leaving it ~21% below the high. Forward P/E is ~136× ($279 ÷ 2026 guide EPS ~$2.05); market cap ≈$79B (~284M sh). Q1'26 was a beat — non-GAAP EPS $0.44 vs ~$0.13 consensus, revenue +130% ✓ Q1'26 release / Yahoo · ~06-26.
Does it change the finding? No. A net-flat price at the same ~136× keeps the headline intact: a real story, priced for years of flawless execution, where the edge is in execution surprises, not the idea. If anything the round-trip to $351 and back demonstrates the "any slip resets the multiple" fragility the bear case flags — a ~28% swing with no change in the long-term thesis. The Q1 EPS beat is a genuine positive (operating leverage showing up), but it's consistent with the price, not a reason to re-rate higher from 136×. ◆ framing = judgment定位＝判斷
Bloom's box is a solid oxide fuel cell (SOFC). It turns fuel into electricity through a chemical reaction — no burning, no flame — which is why it's efficient, quiet and clean. Here's the reaction, then a plain-language glossary.
Air feeds the cathode; the ceramic electrolyte ferries oxygen ions across to the anode, where fuel reacts and frees electrons; those electrons can't cross the electrolyte, so they detour through your building as electricity. Mechanism per Wikipedia / ScienceDirect.空氣進陰極；陶瓷電解質把氧離子送到陽極，燃料在那裡反應、釋出電子；電子無法穿過電解質，只好繞經你的建築 — 那就是電。原理依 Wikipedia／ScienceDirect。
The jargon, decoded術語白話解碼
SOFC固態氧化物燃料電池（SOFC，Solid Oxide Fuel Cell）
A fuel cell with a solid ceramic electrolyte that runs hot. Converts fuel to electricity by chemistry, not burning. Bloom's core tech.
用固態陶瓷電解質、高溫運作的燃料電池。靠化學、不靠燃燒把燃料變電。Bloom 的核心技術。
Bloom— the clear SOFC leader for stationary onsite power.— 定點現場電力的 SOFC 龍頭。
Baseload power基載電力（Baseload）
Steady 24/7 power, always on — what a data center needs. (Opposite: "peaking," only run at demand spikes.)
穩定、24 小時不停的電 — 資料中心要的就是這個。（相對是「尖載」，只在用電高峰才開。）
Bloom sells prime baseload, not just backup.Bloom 賣的是主力基載，不只是備援。
Behind-the-meter表後 / 現場發電（Behind-the-Meter, BTM）
Power made on-site, on your side of the utility meter — you skip the grid and its multi-year wait.
在你這側、電表「之後」就地發電 — 跳過電網與它好幾年的等待。
The whole speed advantage lives here.所有「快」的優勢都來自這裡。
Energy Server能源伺服器（Energy Server）
Bloom's product: a modular cabinet (~325 kW building block) you stack from hundreds of kW to GW.
AI compute is bottlenecked on power, not chips. New grid hookups take years; Bloom delivers reliable 24/7 baseload onsite in months, cleanly and quietly. That speed-to-power is the whole thesis.
AI 算力的瓶頸是電，不是晶片。接新電網要好幾年；Bloom 幾個月就能在現場供應可靠的 24/7 基載，乾淨又安靜。這個「供電速度」就是整個論點。
Oracle: up to 2.8 GW (1.2 GW contracted) — the validation that re-rated the stock.Oracle：上看 2.8 GW（已簽 1.2 GW）— 讓股價重估的背書。✓ PR
Bull case — what could go right多方論點 — 哪裡可能對
Multi-year AI power demand多年的 AI 電力需求
Grid bottlenecks last years. Bloom's speed, reliability (up to "five nines"), and fuel flexibility make it the prime onsite answer. Oracle (2.8 GW) is validation; more hyperscaler deals can follow fast.
Electrolyzers + fuel cells position Bloom for green-hydrogen production and use — upside not in the core product numbers. Plus Korea (SK) and broader C&I/microgrid markets.
~113× forward earnings after a ~10× run. Any growth wobble or margin miss compresses the multiple hard. The price still assumes years of near-perfect execution. live · 07-10
Rapid manufacturing ramp + long install cycles. A $20B backlog only matters if it converts — a July 2026 short report notes it is >40× the $492.6M of binding RPO on the balance sheet, plus a scandium-supply constraint on the 5 GW scale-up. A quality issue in a high-stakes AI deployment would also hurt the brand badly.
Gas turbines (established, long lead times), batteries/storage hybrids, grid upgrades, and longer-term nuclear/SMRs all compete for the same "power the data center" budget.
Natural-gas price swings, tightening carbon rules, reliance on tax credits, and customers needing to finance large systems. An AI-capex slowdown would hit demand directly.
Must hold for the thesis to fail:論點要被推翻，需同時成立：sustained hyperscaler demand, on-time backlog monetization at scale, continued margin expansion, no major competitive or regulatory setback. Break one and a ~113× multiple unwinds fast.雲端大廠需求持續、在手訂單能規模化準時變現、毛利持續擴張、沒有重大競爭或法規挫折。任一條斷掉，約 113 倍的倍數會很快回吐。
Variant perception變異觀點— the reason to read— 值得讀的地方
The variant thesis — a structural, decade-long power shortage and an under-modeled platform — is compelling. But even after the pullback the +167% run means much of it is already in the price. The edge now is in execution surprises, not in the idea.變異論點 — 結構性、長達十年的電力短缺，加上被低估的平台價值 — 很有說服力。但即使回落後，+167% 的漲幅仍意味著它大致已反映在股價。現在的超額來自執行面的驚喜，而非這個想法本身。
Is it like EOSE? And does it do HVDC?跟 EOSE 一樣嗎？有 HVDC 嗎？
Is it like EOSE?跟 EOSE（Eos Energy）一樣嗎？
No — same "clean/reliable power" universe, different jobs. BE generates power (fuel cells, 24/7 onsite). EOSE stores power (zinc batteries, 3–12h). They're complements (fuel cell + battery in a microgrid), not substitutes. BE is far more advanced commercially (~$2B+ revenue, profitable path, $20B backlog); EOSE is a smaller, higher-risk pure-play storage story.
No. HVDC (high-voltage DC) is long-distance grid transmission — the domain of Hitachi Energy, Siemens Energy, GE Vernova. Bloom is onsite generation, which actually reduces the need for new HVDC lines. Bloom does talk about facility-level DC architectures inside data centers (~400–800V DC busways for GPU racks) — but that's in-building distribution, not grid HVDC. This strengthens the bull case: data centers go onsite to bypass slow grid + transmission build-out.
Founder-CEO KR Sridhar✓ public — deep technical roots (NASA / space-tech origins), drove the pivot into the data-center/AI market. Insiders own ~2–3%; net sellers (~$83M net sold last 12 months), no major buying ✓ SWS — a mild negative near the highs.
Proprietary SOFC IP + system-integration know-how, a proven reliability record, and a growing installed base with sticky LTSA service. Modular design + fuel flexibility are hard to copy fast. The clear leader in behind-the-meter prime fuel-cell power.
judgmentThese are the source draft's weights — notably bull-skewed (45%). After the pullback to ~113× (from ~160×) the pure over-valuation argument for the bear weight is a touch weaker — but the July short report cuts the other way, putting a real question mark on the backlog and the scale-up. Net, my own read still leans base/bear over the draft's bull skew.這是原始草稿的權重 — 明顯偏多（45%）。回落到約 113 倍（自約 160 倍）後，「太貴」這條看空理由稍微變弱 — 但 7 月的放空報告從另一邊切入，對在手訂單與擴產打上真實問號。整體而言，我自己仍偏向基準／空方，而非草稿的偏多。
To double from $232.34 (~$132B market cap) needs revenue toward $8–12B+ by the late 2020s with 20%+ operating margins and the market keeping a premium growth multiple. A rising stock has two engines — earnings growth and multiple expansion; at ~113× the second engine is still well past redline, so a double rests almost entirely on delivering the backlog at scale.
