7 years in software IB and private equity investing & 5 years (so far) in L/S equity investing in TMT. I like to opine & share some personal wisdom I’ve learned
San Francisco, CA
Joined March 2025
- Tweets337
- Following179
- Followers312
- Likes140
$BE to the moon 🚀
- MGMT signalled strong demand from various markets (utilities, international, and C&I), positive signals of carbon capture commercialization and continued margin expansion
-strong demand from utilities, int markets, C&I customers -> reinforces value prop of Fcells beyond hyperscalers; in industrials, BE's CHP offering has seen adoption curve
-Positive catalysts next few quarters: midterm elections, Texas data center clarity in Dec'26, large fuel cell project announcements, progress at Project Jupiter, 4Q earnings (annual backlog update), and upcoming analyst day in 1H27
Carbon capture is untapped:
-AEP and Google's 700 MW Wyoming site announced that site's fcells will be carbon capture sequestration (CCS)
-$1Tn TAM
Path to >5GW:
-signed 12.5yr lease on 160K sq ft facility
Strong deployment forecasts:
- I think can hit 2.3MW by 2027 and 3.8 MW by 2028 (26% & 20% growth)
Key themes to play out :
- time to power for large load customers
-growing value prop of distributed energy
-rising grid instability
-30% fuel cell tax credit thru 2030's
$BE’s CEO just made some HUGE predictions about the future of power during the company's earnings call.
One of the biggest takeaways was how confident management is about the future of on site power generation. The CEO essentially argued that the traditional electrical grid was never designed to handle the enormous, concentrated power demands we’re seeing from AI data centers. Hyperscalers are increasingly realizing that waiting years for grid connections isn’t realistic, which is why they’re turning to companies like Bloom to generate electricity directly at their facilities. What I found particularly interesting is that management believes this shift will eventually extend far beyond AI data centers. Factories, hospitals, military installations, commercial buildings and even entire international markets could increasingly move toward generating their own electricity. The CEO compared this transition to developing countries skipping landline telephones and moving directly to mobile phones.
Management also made it clear that they’re already preparing for demand beyond 5 GW of annual manufacturing capacity. They aren’t committing to a specific timeline, but the CEO said he doesn’t see them dialing down production expansion anytime soon. The goal is to increase capacity alongside customer demand without becoming the bottleneck preventing projects from moving forward. Speaking of bottlenecks, there was an interesting update on Nebius $NBIS. Bloom confirmed that its equipment is already ready for the first phase of Nebius’s 300 MW New Jersey data center. Management emphasized that they will have power available as Nebius brings additional capacity online, meaning Bloom does not expect to be the reason the project gets delayed.
The CEO also referenced Jensen Huang’s comments at GTC Berlin, where Jensen reportedly said that every quarter of delay for a 1 GW AI data center could translate into $10 billion in missed revenue. Think about what that means. When hyperscalers are potentially losing billions because they cannot bring compute online quickly enough, the speed at which a company can deliver electricity becomes incredibly valuable. Another important development was Bloom’s growing relationship with utilities. Historically, the company has largely sold on site power directly to customers. Now, utilities are beginning to recognize that installing Bloom’s fuel cells throughout their own networks could be faster and more economical than waiting for massive transmission and generation projects. Management believes this could become a major growth opportunity because utilities could purchase systems in anticipation of future electricity demand rather than waiting for individual customers. Carbon capture was another major topic. Bloom highlighted its 700 MW project involving Google and AEP, which will use fuel cells alongside carbon capture to provide near zero carbon electricity. Management explained that its technology produces a more concentrated stream of CO2 than conventional combustion systems, potentially making carbon capture easier and more economical. The CEO believes this could eventually become a trillion dollar market opportunity, and said hyperscalers are already asking about making their systems carbon-capture ready. There was also a discussion around 800V DC power delivery, which I think is worth paying attention to. Today, data centers lose efficiency converting electricity multiple times before it reaches the actual AI servers. Bloom believes its technology could eliminate some of those conversions, potentially saving billions in operating costs. Management is already having conversations with hyperscalers and expects adoption to accelerate as newer server architectures become available.
Overall, my biggest takeaway is that Bloom is positioning itself as much more than a fuel cell manufacturer. Management sees the company becoming a major part of how electricity is generated and delivered, especially as AI, automation, and electrification continue pushing power demand higher. I thought the call was very encouraging, particularly the comments around manufacturing expansion, Nebius, utilities, and the growing demand for on site power. The big question now is how effectively Bloom can execute as the scale of these projects continues increasing. Yes, $BE is down nearly 7% after hours despite reporting an extremely strong quarter, but I think a lot of this has to do with the current market environment we’re in. High growth stocks are getting punished across the board, regardless of how well their underlying businesses are performing. I’m not going to pretend I know where the bottom is, but my long term conviction hasn’t changed. Once the broader market stabilizes and investors become more comfortable taking risks again, I believe $BE will recover as well. The fundamentals are moving in the right direction, and that’s what matters most to me.
$BE
Backlog monetization is accelerating a lot
-Product revenue has increased well beyond the FY25 delivery pace
-FY25 product revenue: $1.53B
-9M26 product revenue: $2.66B
-Q3'26 product revenue: $1.07B
-Q3'26 annualized product revenue: $4.29B
At Q3's run rate, Bloom is recognizing product revenue at 2.8x the FY25 annual pace The relationship to the disclosed starting backlog is also key: 9M26 product revenue ÷ FY25-ending product backlog $2.662B ÷ $6.0B = 44.4%
In 9 months, Bloom recognized product revenue equivalent to ~ 44% of its Dec'25 product backlog
Bloom expanded product backlog by $3.5B during FY25 despite recognizing $1.53B of product revenue. Subsequently, 9M26 product revenue already reached $2.66B, with Q3's annualized pace of $4.29B representing 2.8x FY25 product revenue. The key question for sustaining the bull case into FY27 is whether new bookings continue replenishing backlog at a rate comparable to or faster than the accelerating shipment pace.
BLOOM ENERGY $BE Q3’26 EARNINGS HIGHLIGHTS
🔹 Revenue: $1.24B (Est. $1.05B) 🟢; +138.1% YoY
🔹 Adj. EPS: $0.94 (Est. $0.68) 🟢
🔹 GAAP EPS: $0.78
🔹 Adj. Gross Margin: 34.5%; +412 bps YoY
Raises FY26 Guide:
🔹 Revenue: $4.40B-$4.60B (Est. $4.1B) 🟢
🔹 Adj. EPS: $3.10-$3.40 (Est. $2.73) 🟢
🔹 Adj. Gross Margin: ~34%
🔹 Adj. Oper Income: $960M-$1.03B
Announcements:
🔹 Added to the S&P 500
Other Q3 Metrics:
🔹 Adj. Oper Income: $285.7M (Est. $214M) 🟢
🔹 GAAP Oper Income: $225.4M
🔹 Operating Cash Flow: -$27.1M
🔹 GAAP Gross Margin: 33.8%; +453 bps YoY
Comments:
🔸 “Bloom is leading that transformation.”
Semicon West and OCP Global Takeaways
1. Wafer Fab Equipment: Supply Chain Preparing for Sustained Hypergrowth
-WFE growth expectations bullish: expectations for 30%+ annual WFE growth for the foreseeable future with few equipment suppliers directing their supply chains to prepare for as much as $500B in annual WFE demand
-DRAM a major demand driver: Industry DRAM capacity could expand to ~3M wafer starts / month by 2029 and 4M by 2031, with certain equipment suppliers already filling production slots into 2030.
-Terafab emerging as significant incremental catalyst: aggressive capacity plans and customized fab architectures could generate substantial equipment demand, including specialized tools beyond conventional configurations.
-Intel could become an incremental WFE catalyst: Rising demand for EMIB-T advanced packaging offering could prompt TSMC to restrict wafer allocations to Intel beginning in 2027 potentially driving Intel's internal capacity investments and equipment purchases
Takeaway: The magnitude and duration of the WFE upcycle may exceed conventional semiconductor capital spending expectations, with DRAM expansion, Terafab and Intel providing multiple sources of demand
2. Semi Testing: Rising Complexity Could Drive Outsized TAM Expansion
-$TER and Advantest conveyed bullish demand outlooks with supply chain indications that combined tool production could ~double between 2H26 and 2H28
-Testing intensity per AI accelerator is structurally increasing: Higher transistor counts, greater HBM integration, more extensive mission-mode testing and relatively immature design-for-test capabilities mean each new generation of AI silicon requires substantially more testing
$AMZN remains a growth catalyst for Teradyne: Trainium related test demand could increase substantially again in 2027 following strong growth in 2026
-Google TPU represents another potential growth engine wher Teradyne's Google opportunity could eventually exceed the size of its Amazon relationship
Takeaway: Semi test market benefits from both increasing chip volumes and rising test requirements / chip. Test revenue growth could materially outpace semiconductor unit growth, potentially making consensus Test TAM assumptions too conservative
3. Semi Materials: Emerging Bottleneck as Wafer Starts Recover
-Wafer production is approaching historical highs: Industry silicon wafer shipments reached an annualized 14,292 million sq in in 2Q26, passing 2021 levels and sitting ~3% below the 2022 peak
-Equipment shortages began emerging in Oct'2026 potentially signaling broader shift in supply constraints from manufacturing equipment toward semiconductor materials as utilization and wafer starts increase
-Materials demand is growing faster than wafer volumes alone suggest: ENTG estimates ~25% more materials content with successive process-node transitions, creating incremental demand independent of wafer-start growth
-Supply constraints could intensify: ~55% of industry wafer shipments are associated with trailing-edge foundry/logic production, where volumes are only now recovering toward prior peaks.
Takeaway: Headline wafer shipment growth may materially understate underlying materials demand given advanced manfact requires more consumables / wafer. Results in favorable setup for Entegris and other semi materials suppliers while wafer producers like SUMCO, GlobalWafers and Siltronic benefit from volume recovery
$MRVL: Upside from conservative interconnect and custom silicon assump
-Interconnect estimates may have upside as scale-out networking remains the primary growth driver with mgmt indicating ~50% growth in scale-out alone would satisfy most of its growth expectations. Scale-up optics represents additional optionality, given limited assumed adoption today
-Custom XPU opportunity may be understated where $30B oppty looks conservative with mgmt assuming < two attached companion devices per XPU
- NVIDIA's system level co-design establishes a competitive benchmark: NVIDIA's integration across 6 chips suggests hyperscalers will increasingly require similar rack-level integration. Marvell is positioned to supply multiple supporting chips around custom accelerators
-Microsoft Maia is optionality where large scale deployments could drive incremental custom silicon demand
$ALAB: Broadening beyond PCIe retimers but execution is key
-Orion introduces a heterogeneous rack-scale opportunity. Targeted for Q1'27, the compute agnostic architecture lets enterprises and neoclouds to integrate different asics through Astera's reference designs, Scorpio fabric and management software. mgmt cited ~50% higher performance and 25% memory offload
-Scale-up represents a potential new revenue stream: A bridge chip could connect XPUs to NVLink Fusion or UALink fabrics, enabling 1 chip / accelerator
-Optics broadens addressable content. Aries 7 is Astera's expansion into optical drivers and TIAs. Management estimates ~$200 of potential NPO content per rack, increasing toward $1,000 with CPO
-CXL applications are expanding: KV-cache and HBM offload introduce additional AI-specific opportunities beyond traditional memory expansion
$INTC: Advanced packaging gives near-term upside; GPU recovery is speculative
-EMIB-T packaging could drive incremental equipment spending: Potential restrictions on TSMC wafer allocations could increase Intel's need for internally controlled manufacturing capacity
-Repositioning its accelerator roadmap toward inference: Its renamed D-series GPU emphasizes memory capacity and power efficiency -> 1.3 PFLOPS FP8 performance, 1.5TB/s bandwidth, up to 480GB LPDDR5X and 350W power consumption
-Customer sampling is in process but mgmt hasn't disclosed a broad availability date or demonstrated meaningful production adoption.
Wanted to share this chart I came across recently..
a) AI infra spending remains exceptionally strong with leases masking scale of investment with $NVDA capturing ~67% of AI capex in 2027
b) Network, optics & power are crucial secondary oppties - beneficiaries beaten down are $ANET , $CIEN , $VRT , $ETN
c) Key investment risk shifting from demand visiblity to monetization, deployment contraints & capex sustainability. I.e. rev capture doesn't translate to equivalent incremental earnings or shareholder value.
Generally, I see most attractive exposures are those capable of delivering earnings growth w/o relying on perpetual upward capex revisions..
$NBIS essentially running airline seat yield management pricing. Capacity is fixed & perishable bc unused GPU hour today can’t be sold tomorrow so committed customers get discounts while urgent on demand pay a premium. So if utilization holds, Nebius can optimize revenue / GPU-hour similar to airlines can charge more for the same seat.
🚨 $NBIS just announced ~20% price increases for on-demand H100, H200, B200 and B300 GPUs, effective October 1.
This comes just four months after the company raised prices by ~30%.
Demand remains insatiable.
Some fun probabilities to think about 😂
The >10% extinction estimate is a chain of conditional probabilities:
P(AI automates AI R&D)
P(rapid recursive improvement | AI automates R&D)
P(humans lose control | rapid recursive improvement)
P(extinction | humans lose control)
So: P(extinction) = P(A) × P(B|A) × P(C|B) × P(D|C)
If I assume:
50% chance AI automates AI R&D
× 60% chance that causes rapid recursive improvement
× 50% chance humans will lose control
× 70% chance loss of control leads to extinction
=~10%
While Q2 print for SaaS didn't establish a case for extreme re-acceleration, I think it shows that SaaS still alive driven by moats such as a) less than expected seat dependence b) app layer control c) inference leverage d) value expansion e) embedded distribution
a) As agents do more work, traditional usage metrics can decline as overall platform activity spikes. Agent workflows can generate many API/MCP calls data requests, tool actions, downstream transacts.. w/o user interaction -> increase demand for SW, data, governance systems that agents rely on. Vendors can monetize thru premiums, consumption pricing, API access, agent actions -> pricing shift is key risk
- $CRM: sales, service & Slack seats grew YoY while agentic usage greew 6x thru MCP and lab agent calls (+95% Agentic Work Units in Q2)
- $NOW 50% of net new biz is non-seat based
- $TEAM Cloud upside from paid-seat expansion
b) Bull case is frontier models provide probabilistic reasoning & apps provide governed data, permissions, biz logic, workflows to translate reasoning to action. Agents can abstract the UI but still need SoR to establish workflow layer that includes harness + scaffolding.
- $CRM: 9/10 top AI companies use CRM and Slack with cohort spend +435% in Q2
- $NOW AI Control towe rhas >500 live customers in 6M
- $TEAM Teamwork Graph MCP calls up 400% Q2
- $WDAY 5.5K customers using 1+ agents up 35% QoQ
c) open weight & frontier model competition shifts value to app layer. SaaS serves as orchestration layer across various models, selecting the right model with prop. data, permissions, semantic understanding and execution environments as the value differentiation.
Lower inference costs can reduce AI COGS limiting GM pressure or funding lower prices / richer products.
d) Historically, SW has been prcied around product. gains. As apps begin to automate or do work, vendors can price against labor capacity crated, costs avoided, outcomes delivered. This can expand TAM beyond traditional IT budget.
e) Distribution is not only GTM moat but an input into faster product iteration & more effective monetization. Existing SaaS install base provides lower cost channel to drive AI features, x-sell, expanding usage, etc.
-Salesforce Agentforce ARR grew 240% YoY
- WDAY -> Deployment Agent is live with 4.6K customers with query volumes up 500% in Q2. >50% of customer base migrated to UMSA giving them on-ramp to introduce AI products across install base
Been thinking about sovereign AI implications lately… world is shifting to country specific AI cluster where each cluster needs its own power, land, DCs, networking, Gpus, cloud, etc. Fundamentally pretty economically inefficient but overall bullish to AI infrastructure buildout.
One leg of beneficiaries I see are
A) neoclouds like Nebius where they own sovereign assets with int partnerships. Their model maps to my mental web of sovereign AI requirements
B) colocation players like $EQIX and $DLR as I’m most bullish on Equinox. They have direct route thru Fabric Geo Zones and have larger direct China exposure vs comps
C) $MSFT is clearest. They have Sovereign Public /Private Cloud, National Partner Cloud and their entire stack allows customers to run models within customer fenced loads. Along with their identity, security, governance overlays
I see $NOK as a mix shift and operating leverage story (not growth). The long bet is that its business will be ~30% AI Infra by 2030 with owned fab capacity in a supply constrained market. 35% discount to $CIEN is defensible today given Nokia's smaller cloud base + mobile drag. Let's see tomorrow:
1) Mix shift is entire story
-Infinera acq. took Nokia from RAN duopoly loser to optical competitor (Non-Ran now ~50% of group) and Optical now 19% of revenue. AI & Cloud is 8% of rev and grew 49% YoY in Q1 and I see this growing 30%+
2) Nokia priced as optical laggard but its a systems vendor
- Nokia sells optical + IP Routing + RAN into same customer base as Ciena.
-The 25% long-haul DCI share understates Nokia's addressable wallet per customer. If attach economics work at even one hyperscaler, revenue per win is materially higher than the DCI share numbers imply.
3) Fab ownership
-Ciena has chosen share over price for years and is fabless. If Nokia's 25x San Jose capacity lands into a still-constrained 2027-28, Nokia is not the price taker forcing Ciena's margins down. They are the one with volume to sell into a tight market at a good price, which is a gross margin story that's not priced in.
Keep an Eye on:
1) Execution risk on DSPs
-Nokia is targeting 4 DSPs by H2 2027 against a Nokia/Infinera track record of 1 DSP every 4 years.
-Ciena runs a 2-3 year cadence
- Miss the 2027 delivery and the multi-rail/800ZR window closes
2) Share position is lopsided
-Long-haul DCI: Ciena 63%, Nokia 25%. Metro DCI: Nokia 23%, tied with Cisco, Huawei 21%.
-Nokia's strength is in the fragmented, more competitive, lower-margin segment.
3) Mobile infra is 54% of revenue and structurally declining
Should we switch to DCF vs multiple valuation given near-term profitability?
J.P. Morgan: Nokia $NOK
> Q2 Preview & Guidance Potential: J.P. Morgan expects strong Cloud and AI orders for 2Q26, with the core question for management being whether supply availability will allow them to raise FY26 guidance.
> Revenue and Guidance Upside: If supply is accessible, Network Infrastructure (NI) revenue growth could be revised upward to as high as 28% (compared to the current guided midpoint of 19%).
> Profitability and Seasonality: Q2 EBIT is projected to beat consensus slightly by 5.7% (€393m). However, due to normal seasonality, the strongest sales and EBIT improvements are anticipated in the second half of the year (2H).
Segment Dynamics & AI Drivers
> IP Networks Momentum: IP Networks is expected to become a major revenue and order contributor via data center switching wins with international neo-clouds and hyperscalers. Growth is further reinforced by higher front-end DCI intensity for AI and inference/agentic architectures.
> AI & Cloud Order Momentum: Following €1bn in AI & Cloud orders in Q1 (largely driven by optical), Q2 orders are expected to climb even higher, supported by optical products and IP network demand. New optical products launched at OFC are also gaining traction despite long lead times.
> Supply and Margin Realities: Component tightness (such as lasers) and elongated product lead times mean the revenue ramp will take longer. Margin expansion in NI will be partially offset by ongoing business investments.
Estimate Revisions (2Q26 and Forward)
> 2Q26 Adjustments: J.P. Morgan raised 2Q26 revenue estimates by 1.2% (to €4,778m, up 5.0% YoY) and EBIT estimates by 16.9% (to €393m) due to accelerating IP networks revenue and operating leverage.
> Multi-Year Estimates: Minor upward adjustments were applied to FY26–FY28 EBIT estimates (by 2.5%, 0.9%, and 0.5% respectively) and EPS estimates (by 2.3%, 0.8%, and 0.5%).
Valuation and Investment Thesis
> The Multi-Year Compounder Narrative: With 30% of revenue coming from AI-exposed Optical and IP Networks (guided to grow 10–12%) and earnings compounding potential via operating leverage and Infinera deal synergies, J.P. Morgan believes Nokia can re-rate to a high-teens earnings multiple
I expect beat & raise for $INTC driven by stronger demand and pricing on data center biz.
1) I believe DCAI performing better than expected
Channel checks bring up:
- Server CPU demand remains very strong with ability to increasingly able to raise CPU prices.
- Both server CPU prices and unit shipments could increase by a high-single-digit to low-double-digit % QoQ in June.
- Additional price increases could support further growth in September.
Also benefiting from improving manufacturing yields on its Intel 3 process and older manufacturing nodes.
2) PC Biz improving
I think CCG should return to sequential growth in the Q2 driven by Panther Lake, their upcoming notebook processor manufactured on its 18A process.
I hear Panther Lake ramp is progressing well. Desktop growth will likely remain modest, in the low-single-digit range.
3. Foundry Remains the Biggest Question
Key areas to watch include:
-Manufacturing progress on 18A, 18A-P and 14A
-The timing of Intel’s 14A 1.0 process design kit
-New customer engagements for 14A and 18A-P
-Progress in advanced semiconductor packaging
4. MediaTek Packaging Oppty
I think they're making progress with MediaTek on EMIB-T advanced packaging. Production shipments for MediaTek’s future v9 generation could begin in the second half of 2027 or during 2028.
Intel’s current packaging yields remain well below normal industry norms but can still win customers because it could offer the first available panel-level packaging solution capable of supporting four-die configurations.
5. Gross Margin Should Beat Guidance
Intel guided to a GM of ~39%, implying a ~2% sequential decel. However, they tend to outperform its GM guidance. Positive factors include higher server CPU pricing, better manufacturing yields. stronger data-center demand
The main offset is Panther Lake as margins are expected to remain below Intel’s avg during the early ramp and prob won't be margin-accretive until 2H'27
And we have Bernstein estimates on Intel $INTC
Easy enough!
We model Q226 at $14.3B/ $0.20, a touch below consensus at $14.4B/$0.21.
For Q326, we model $14.9B/$0.28, below to in-line compared to consensus at $15.1B/$0.28.
For full-year 2026, we model $57.6B/$1.07, below consensus at $58.6B/$1.11.
For full-year 2027, we model $63.3B/$1.50, a bit below consensus at $66.0B/$1.57.
$BE is the fastest path to a megawatt in a country that's 38 GW short 🚀
U.S. data centers need ~68 GW of new power during 2026-2028. Only around 15 GW is covered by data centers already under construction and 15 GW is covered by available or contracted grid capacity.
That leaves a potential 38 GW power deficit.
Bloom fuel cells are expected to supply ~5-8 GW through 2028 -> large assumption relative to Bloom’s current manufacturing footprint.
1. Need for Speed
Bloom can deliver equipment in ~50 days, with certain projects reaching commercial operation within approximately six months of the initial customer discussions.
That compares with ~3-7 years for a traditional grid interconnection and > four years in some cases to obtain newly ordered large gas turbines. Accelerating a data center by 1 year may be worth ~$45-$70 per MWh to the developer
2. Time is Money
An AI operator that bought expensive compute but cannot obtain power delays revenue, loses useful GPU life, risks hardware becoming obsolete + gives competitors additional time to deploy new infra
For a data center that can receive power 2 years earlier, a hyperscaler could rationally pay a 101% premium on power costs, assuming a 6Y GPU economic life and a base electricity cost of $100 per MWh.
When comparing a) a site available 2 years earlier, versus B) a slower site with electricity priced 30% lower. The financial value of the earlier site is estimated to exceed the extra electricity cost by :
a) 3x with a six-year GPU life.
b) 2x with a ten-year GPU life.
3. Bloom vs gas turbines - Bloom has several advantages.
1) Deployment timing - turbines may have multi-year procurement queues. Bloom can be operational in less than 6 months.
2) Reliability and redundancy - A data center microgrid requires extremely high availability. A turbine installation may require excess generation capacity, backup units, batteries, additional microgrid equipment, greater balance-of-plant infrastructure.
Bloom’s modular architecture reduces the impact of an individual unit going offline and can lower the amount of excess capacity required for reliability and narrow the apparent cost gap with turbines.
3) Permitting - Gas turbines face more extensive local air permitting because of conventional combustion emissions such as NOx. Bloom still consumes natural gas and produces carbon, but its electrochemical process can be easier to permit in dense or environmentally sensitive locations.
4) Power-density and land use - Bloom’s newer stacked configuration can deliver ~100 MW on one acre, compared with ~25 MW per acre under its historical single-layer configuration.
5) Decarbonization- Both turbines and Bloom systems may eventually use hydrogen or carbon capture. Bloom may have an advantage because its carbon stream is more concentrated potentially making capture easier and less costly.
$BE | JPMorgan 𝐦𝐚𝐢𝐧𝐭𝐚𝐢𝐧𝐬 𝐎𝐯𝐞𝐫𝐰𝐞𝐢𝐠𝐡𝐭 on 𝐁𝐥𝐨𝐨𝐦 𝐄𝐧𝐞𝐫𝐠𝐲, 𝐫𝐚𝐢𝐬𝐞𝐬 𝐏𝐓 𝐭𝐨 $𝟑𝟒𝟔 𝐟𝐫𝐨𝐦 $𝟐𝟔𝟕
Analyst sees significant growth to FY30, supported by existing partnerships, with new contract announcements expected to be the primary near-term catalyst.
4. Manufacturing capacity is not a constraint
Bloom exited 2025 with ~1 GW of annual manufacturing capacity at its Fremont facility.
Estimated utilization was only around 60%-70%, which means immediate growth can be absorbed without a large expansion. Plus, the facility can reportedly scale from 1 GW to as much as 5 GW:
a) In less than two years.
b) For ~$150 million of incremental capex
c) While supporting both new equipment demand and replacement or servicing requirements for the installed base
This is unusually capital efficient relative to the potential revenue opportunity.
5. $BE is not solely dependent on data centers
Mgmt indicated growth above 30% among existing commercial and industrial customers facing similar grid constraints.
Quanta Computer is the main example as it increased planned Bloom capacity at its California manufacturing facility by ~1.5x, less than a year after the original agreement.
The same problem affecting data centers also affects: semiconductor facilities, electronics manufacturers, advanced manufacturing, etc
6. Project economics
Bloom power at ~$100-$130 per MWh:
a) Landed equipment cost of ~$4,700 per kW.
b) ITC of 30%-40%.
c) Natural gas at $4.75 per MMBtu.
d) Heat rate of ~6,500 Btu per kWh.
e) Annual non-fuel OPEX of $212.50 per kW.
f) Unlevered project return of 7.5%.
Actual economics vary by project. A grid-connected customer using Bloom for steady baseload generation will generally have lower costs than a fully islanded data center requiring sophisticated load following, backup equipment and complete microgrid infrastructure.
Duration matching for AI infrastructure -> fixed liabilities vs. falling hardware values
$CRWV commits to 10K servers at $100K each ($1 bn) funded at 80% LTV. Let's assume 30% is memory so $300M. A 40% memory price decline makes system ~$120M cheaper so market value f alls to $880M while Coreweave still owese $1 bn. So LTV rises from 80% -> 91% ($800/$880M)
It hedges only $250M of the $300M exposure. The theoretical payoff is $250M × 40% = $100M, but with 90% hedge effectiveness/basis correlation, it receives $90M. After a $10M premium, the net benefit is $80M, reducing the economic loss from $120M to $40M. That remaining loss reflects $20M of unhedged exposure ($50M × 40%), imperfect index matching, and the hedge premium. Including the $80M hedge cash, effective collateral support is $960M, bringing adjusted LTV back to 83.3%. If prices stay flat or rise, the hedge pays nothing and CoreWeave loses the $10M premium though its fixed purchase price looks attractive if prices rise.
Yes, $IBM Q2 miss caused by collapse in enterprise IT demand.
Mgmt thinks this was a short-term quarter-end phenomenon involving a few large deals that slipped during end of quarter and it did not see customers migrating away from mainframes or evidence that demand had simply been pulled forward into 2025.
However, if high memory prices and enterprise compute shortages remain multi-year issues, customers may repeatedly prioritize servers and storage over IBM software and mainframe commitments. That would turn a few delayed deals into a persistent share of wallet problem extending through 2H26 and potentially 2027
Important positive read through outside IBM
IBM’s Distributed Infrastructure business grew 37% year over year, supporting the thesis that enterprise server and storage demand remains very strong and is taking a larger portion of IT budgets.
This as positive for:
Direct beneficiaries: $DELL , $HPE , $P , $NTAP , $CDW , $INGM
Indirect storage-component beneficiaries $STX $WDC
SAASOCALYPSE 5.0!?
$IBM CEO Arvind Krishna:
“In the last few weeks of June, we saw clients shift their quarterly capex spend toward servers, storage, and memory purchases to secure supply-constrained infrastructure ahead of expected price increases.”
IBM “did not anticipate the magnitude of the capex reprioritization,” and said clients were also distracted by “rapidly-evolving, industry-wide cybersecurity concerns.”
Krishna also said: “We did not adapt and move quickly enough,” with large deals failing to close on expected timelines.
Look, I think the upside case for $MU rests on SCAs (Strategic Customer Agreements) which transitions them from cyclical to real multiple. A few additional thoughts:
They signed 16 SCAs, mostly 5-year (C2026-2030, autos at 3yr), covering ~20% of DRAM bits and ~33% of NAND bits, with a stated target of >50% of revenue under contract. Four "very large" customers are in; with the gap as a few remaining hyperscalers, potentially NVDA. 70-75%, backing into a DDR floor price of ~$6/GB vs ~$13/GB implied in the Aug Q guide, and rising. So pricing would need to collapse 70-80% within the contract term for a customer to even contemplate breaking. Supply math says enough capacity can't come online to push spot below floor until C2029, more likely C2030 -> i.e., the floor is essentially unreachable during the deal window.
Margin normalization and re-rating. If you assume 1/2 of business not under SCA reverts to the C2015-2025 average of ~35% GM, blended normalized margin lands mid-50s, above anything $MU ever printed through-cycle. Higher structural RoE justifies a higher book multiple; TBV approaches ~$300/share exiting C2028.
Capital return . ~$380B cumulative FCF in C2027-2028 (FCF was already 44% of revenue in the May Q). Buybakcs is capped by CHIPS Act restrictions until 12/9/26. After that, I expect buybacks to mirror FCF, ~$380B cumulative in C2027-2028, enough to retire ~20-30% of the company at the current price.
HBM pull-forward. HBM4 already past $1B, with 12-high ramping twice as fast as HBM3E 12-high. They pulled its >$100B HBM TAM call into C2027, a year early.
Valuation. I think worth ~15x NTM P/E, in line with the 3yr average and peer semis on ~$121 C2029E EPS, discounted back one year at ~11.8% COE. If anchoring on C2029, this already embeds a modeled downcycle (revenue -37% Y/Y, DRAM ASPs -40%) with LTAs holding the earnings floor. The upside case isn't peak-multiple-on-peak-earnings but rather it's normal multiple-on-trough'ish earnings that are still $121 because half the book can't reprice below $6/GB.
$MU Holy Shit.
Morgan Stanley projects Micron will generate ~$400B in combined operating income across 2026 & 2027. That's nearly 40% of the entire market cap. That's fucking ridiculous.
2026 → 2027
Revenue: $172.9B → $285.9B (+65%)
Operating Income: $140.7B → $248.7B (+77%)
So many easy holes to address but will only address a few.. $BE
1. Scandium math is weak sauce. Reproducing Hunterbrook's own inputs yields ~25 t Sc₂O₃/GW "sealed" into product (~33–40 t/GW purchased after yield loss and FRU add-ons) but only if per-cell power is frozen at the 2021–2022 figure of 25 W. Bloom's disclosures show ~7.5× cumulative power-density improvement through 2021 and significantly fewer layers since. The entire deficit rests on per-cell power NOT having improved since 2021; t/GW scales inversely with watts/cell. A 2020 Bloom blog states fuel cells fielded since 2016 had "power density five times greater than its first-generation products," with 2020–2021 next-generation cells adding "an additional 50 percent increase in power density". The July 7, 2026 Bloom blog by COO restates that optimizing the doped electrolyte yields "a stack that requires significantly fewer layers of fuel cells to generate the same amount of electrical power."
Sensitivity table below (electrolyte thickness vs. watts per cell): scandium deficit is not fundamentally about geology but rather about an assumed cell-power spec. If that spec is stale, the conclusion weakens materially
2. Supply Elasticity where a Byproduct Market Is Not Fixed
Historical USGS series (Sc₂O₃ equivalent): 2019 ≈ 15–20 t supply/consumption; 2020–2021 ≈ 15–25 t; 2022 production ~40 t; 2024 production ~40 t, consumption 30–40 t, capacity ~80 t (MCS 2025); 2025 production ~80 t, consumption ~60 t, capacity >90 t (MCS 2026). Production doubled 40→80 t and consumption rose ~50% in one year, with capacity already exceeding production. CAGR 2022→2025 ≈ 26%/yr; 2020→2025 ≈ 32%/yr. Hunterbrook's 79→240 t by 2030 implies ~25%/yr a continuation of the observed trend not far fetched at all...
The binding historical constraint has been demand, not supply. With demand now present and a price signal, byproduct capacity is being built. Here's a quick table on ex-China project pipeline...
Hunterbrook adds ~35 t/yr of field-replaceable-unit demand on top of 185 t (new deployments) to reach ~220 t, treating FRU scandium as net-new. But scandium in a replaced stack is recoverable. Bloom operates repair-and-overhaul facilities in Newark, DE (~56,000 sq ft per the 2024 10-K). Patents (US10651479 "Method of recovering metal compounds from solid oxide fuel cell scrap"; US10036082 zirconium/scandium extraction from used SOFCs) and academic work confirm ScSZ recovery is established (acid dissolution + solvent extraction). If Bloom recovers 30–50% of FRU scandium in a closed loop, the ~35 t add-on falls ~11–18 t/yr. The July 7, 2026 blog claims proprietary recovery from waste streams; closed-loop recovery on returned stacks (concentrated, known composition) is materially easier than from dilute ore.
3. Other claims - shows a lack of basic knowledge of project finance
Related-party revenue: this was already flagged by auditor (so not news)... figure is disclosed, audited, GAAP-compliant on transfer of control (not fraud)
Backlog vs. RPO: BLOOM IS NOT A SaaS COMPANY!! RPO is a narrow ASC 606 snapshot that explicitly excludes short-duration and right-to-invoice obligations, and Bloom’s Q1 revenue exceeded beginning RPO by almost 80%, proving RPO is not the order book. Bloom's backlog is LONG DURATION with terms ranging from 5-20 years
- Peer comp is apples to oranges: GE defines backlog=RPO. Software has committed subs / capacity, etc.
-Bloom’s historical system backlog has converted into product / installation / electricity revenue at 40–80% of beginning backlog annually, while service backlog naturally converts at only 2–4% per year because it is a 5–20 year O&M stream. If the system backlog were just not real demand, you would not see 40–80% of beginning system backlog turning into annual revenue
-If RPO were the true order book, Bloom should not be able to generate $751M of quarterly revenue against only $493M of period-end RPO but it did. The reason is accounting mechanics where revenue can be booked, delivered, invoiced, and recognized without ever sitting in period-end RPO.
Summary on points 1-2:
1. Stale input. 25 W/cell is 2021–2022; Bloom has ~7.5× cumulative power-density gains through 2021 and "significantly fewer layers" since. Freezing per-cell power maximizes the deficit and is mislabeled conservative
2. False independence of Methods 1 and 2. Method 1 (patent physics → ~25 t/GW) and Method 2 is a market-share triangulation built from opaque estimates of global scandium demand, SOFC share, Bloom share, and 2025 deployed GW including FRUs. If any of those market inputs are off, the 37–43 t/GW confirmation moves materially. Can be a sanity check, but not an independent proof that Bloom’s current cell spec is still 25 W/cell.
3. Byproduct fixed error. Probability-weighting discrete projects like a fixed mine schedule ignores that byproduct capacity (>90 t vs. 80 t produced) expands with demand and price as it just did (40→80 t in one year).
4. FRU double count. Field replacement scandium is recyclable from the very units replaced; counting it as net-new overstates 2030 needs by up to ~35 t.
5. Single-source supplier quotes. Key claims lean on a Hunan Oriental sales rep in a chat an interested party with incentive to overstate its role. WTF is this institutional grade evidence...shit looks like my uncle's WeChat
NEW: The CEO of an AI giant said again and again: "There is no China supply chain for us."
Trade records, satellites, Chinese filings and even its own suppliers in China say otherwise.
The truth raises an existential threat to Bloom Energy’s story.
Our $BE investigation:
Here's some more color on $BE RPO vs. Backlog argument I hope clears things up
1. What RPO legally cannot contain
RPO under ASC 606 includes only enforceable, non-cancellable performance obligations and Bloom explicitly excludes "(i) contracts with an original expected length of one year or less and (ii) contracts for which we recognize revenue at the amount to which we have the right to invoice"'. Three structural features of Bloom's model then empty RPO of most economic content:
1) Service contracts are one-year renewable by design. Per the 10-K: service revenue is "recognized ratably over the term of the first or renewed one-year service period," with renewal options treated as material rights recognized "over a period of up to 21 years." Only the currently binding year can ever sit in RPO.
2) Product flows through inside the disclosure window. Book-to-bill runs 9–18 months, so most product contracts are excluded as ≤1-year or are right-to-invoice. Look at the deferred-revenue: FY24 additions $1,142.6M vs. revenue recognized $1,148.6M, ending balance only $66.3M (2024 ). ~billion dollars a year moves through a pipe whose standing water is ~$66M
3) Framework/financier commitments are not ASC 606 contracts until an SPV issues a binding PO for specific units.
GEV (inconsistent comp) defines backlog as RPO because it sells turbines under long-dated binding EPC/CSA contracts to utilities. Bloom sells into financing pipelines with annually-renewing O&M.
2. The service book validated by cash
The 2–4% service backlog conversion rate is the arithmetic nature of a 5–15 year contract book. Hunterbrook's coverage is misleading as the numerator contains the service tails of servers not yet deployed. $14B service backlog over ~15-year average terms implies ~$930M/yr steady-state service revenue, ~4x today's run-rate which is consistent with a fleet scaling ~4x as the $6B product backlog deploys.
3. Product Backlog
41–44% conversion for 3 years means beginning product backlog × 0.42 is proxy for the year's revenue within rounding (2024: $3.0B × 42% = $1.26B vs. actual $1,085M product + $122M install + ~$53M electricity = $1.26B). A fabricated order book does not produce a stable, verifiable conversion. Second, the implied duration: ~42% annual conversion ≈ 2.4-year average delivery tail, says most bookings deliver within two years, some at three
4. Project finance 101
Bloom does not sell appliances to end users ok? It sells 15–20 year generation assets, and generation assets are overwhelmingly bought by capital. PPA structures where an SPV owns the servers and sells electricity, Managed Services sale-leasebacks with financiers, portfolio financing entities. Five consequences follow:
1) The buyer of record is a financing vehicle. The end customer signs a long-dated PPA for electrons; the SPV/JV buys the boxes. Brookfield's Fund JVs (Bolt US JVCo, ORC HoldCo; Bloom retaining 9.9%/15% minority stakes) are the 2025–26 instantiation of what SK ecoplant and the PowerSecure/utility structures were earlier. Related-party revenue concentration (74% of Q4'25, ~50% of Q1'26) is the accounting shadow of capital formation moving upstream of deployment. Infrastructure capital pre-buying manufacturing capacity ahead of PPA origination. Solar built out the same way (developer dropdowns to yieldcos).
2) Project finance mechanically widens backlog vs RPO. A financier's framework commitment to fund X MW is economically real committed capital, deposits ($220.1M customer deposits at 12/31/24, with the Q1'25 10-Q noting deposits "becoming non-refundable" as they season; cash that walks away if orders don't). But it is not an ASC 606 performance obligation until specific units go under binding PO. So a project-financed order book systematically shows large backlog and small RPO, and converts in lumps as financings reach FID.
3) The service backlog is the bankability instrument. Lenders and tax equity will not finance a 15Y PPA on an asset with 5-year stacks unless O&M with performance guarantees spans the financing term. That is why the O&M book exists at 5–15+ year expected terms while remaining annually renewable in legal form: the annual renewal is a customer-friendly option, but the economic substance is locked by the financing documents and by Bloom's own guarantee exposure. I.e. you see performance guaranties capped at $588.1M aggregate ($460.8M related to portfolio financing entities), $476.6M remaining potential, with actual payouts of only $17.0M in 9M25 vs. $18.8M in 9M24. Bloom carries half a billion dollars of contingent liability keyed to portfolio financing entities across the service term!!
The important part of Thinking Machines Lab fine-tuning Qwen3-235B is not really Qwen but where the AI stack is going. Pre-training is becoming less differentiated at the frontier of enterprise value capture.
The value is shifting toward what you do after the base model (increasingly available and commoditized) post-training, RL, proprietary data, and workflow-specific productization and SW/business logic built around the model.
That changes the map:
OpenAI still has the consumer distribution layer through ChatGPT. Anthropic has a strong software/product wedge through Claude Code and CoWork.
But in the API market, plain base-model access probably compresses toward token cost plus hosting. The model itself becomes closer to pass-through compute.
The real enterprise value pool is taking proprietary company data, or specialized human/trained datasets, and using RL/post-training to make models actually useful inside workflows. That is where Mercor, Handshake, Turing, Scale and similar companies all start colliding.
The competitive set gets messy fast:
1. LLM labs: They move into domain-specific models, like biotech, finance, healthcare, etc
2. Inference supply: Nebius, coreweave, etc. As inference becomes more cloud-like, they will be forced to compete in this layer too.
3. CSPs: Microsoft especially has the distribution, compute, and enterprise control points to be a major winner.
4. Data companies: They can package their datasets with training services, and likely buy RL or neolab capability over time.
5. Snowflake: They already sit next to the proprietary enterprise data, so they can become the system of record plus services layer
6. App platforms: Salesforce and similar platforms can embed this directly into enterprise workflows.
7. RL labs: This is exactly their core business.
8. Neolabs: Their only real angle is specialization. Training another generic base model increasingly looks like a dead-end use of capital.
So the enterprise AI market is not one clean category. It is base model companies, RL labs, data vendors, app platforms, warehouses, clouds, and inference providers all trying to monetize the same budget.