I published my $SIMO deep-dive with a base case that had the company hitting a 28% operating margin in 2028. I was higher than anyone on the Street. Silicon Motion just told the market it will exit 2026 above 30%. That’s the whole story of this quarter in one line. The thesis is intact, the ramp is faster than I modeled, and I’m raising my numbers.

The quarter

Revenue of $451 million, up 32% sequentially and 127% year over year, blowing through the top end of the $393-411M guide by a full 10%. Third straight record quarter. Gross margin of 50.2%, above the 48.5-49.5% guide and through the psychological 50% line Kou promised earlier this year. Operating margin of 23.1%, past the 21-22% guide. EPS of $2.43, versus the ~$2.11 consensus.

The mix shift underneath the beat is what matters most.

Read more

I sat down 1x1 with this company a few weeks back, and we walked through each business opportunity in detail. What I found: two brand-new business lines, enterprise boot drives and enterprise controllers, that could each become double-digit percentages of revenue within two years, stacked on top of a mobile controller franchise where the competitive moat is being poured in concrete right now.

The full report below covers the business model, what they told me in my sit-down, the boot drive opportunity nobody is modeling correctly, the enterprise controller line’s compatibility with NVIDIA’s CMX architecture, my full 2027-2030 base, bull, and bear model, and where I’d be wrong.

Read more

So how do I actually handle a day like this. I go back to my research. I reread my own reports and remind myself where my conviction sits and why it’s there. Price action moves fast and it’s easy to let it drag your perspective around with it. My job is to keep the two separate, price is what the market is doing right now, conviction is what I know about the business.

That conviction comes from a specific checklist. Do I want to own leading stocks? Yes. Do I want them trending above a rising 50 day moving average? Yes. Do I want them building tight flags and breaking out clean? Yes. That’s the setup I’m always hunting for. The market doesn’t always make it that easy. Trading isn’t easy, and days like today are part of the game.

I run heavy size and I hold leading names for as long as they let me. Part of that process is letting cost basis work in my favor. I own $VPG at $45 and $PENG around $30, both up well over 100% from cost. Both businesses are executing on all cylinders. Stocks like that have earned the right to keep working even when they cool off with the rest of the momentum trade.

Drawdowns come with that territory. Chasing real outperformance means accepting more volatility, that’s the trade you’re signing up for the moment you decide to build concentrated conviction instead of hugging an index.

The other piece of today’s work was scanning monthly and weekly charts across 50+ names over the last few hours, checking which higher timeframe setups held up and which fell apart. $PENG on the higher timeframe still looks fantastic. It’s off its high of $89, but that’s the sell everything tape doing what it does, not the business or the high timeframe setup deteriorating. $VPG on the weekly is the same story, well off its highs but doing nothing wrong from a business or high timeframe setup. The thesis on both has only gotten stronger since I bought them, and my targets sit well above where they’re trading today.

None of that means I get sloppy on new risk. New buys get managed tight, especially with any real size behind them. I’m fine running smaller size and letting a new position breathe, but the moment I’m putting larger capital to work, the risk management gets tighter even when conviction is high. This environment will punish you if not. I also have advised to not press the gas and probe. I have come out liking the risk to reward on put sales, but still don’t think new, size risk in this environment is worth it. Can that change in a moment’s notice? Yes. That is the beauty of the market.

Today stung. Honestly one of the worse days of my trading career, but it’s been a great year, and I own names I believe in deeply. Go back to why you bought the stock. If the thesis has only gotten stronger since you did and you have a cost basis advantage, you’re in a good spot. And if you’re sitting on cash, days like today are the gift to probe and be ready to press the gas when the environment becomes ripe for more risk-on.

Have a good night guys!

Humanoid robotics is running through the same arc autonomous vehicles ran through a decade ago, except faster, and public markets have not caught up to how fast. For years the story was almost entirely software: could a machine perceive, reason, and act quickly enough to be trusted near people. That question is still not fully answered, but it has stopped being the binding constraint. The harder problem now sitting in front of every humanoid program on earth is a hardware problem, and hardware problems get solved by supply chains, not by demo videos.

That hardware problem looks a lot like a problem the automotive industry already spent the last decade solving. A humanoid robot needs to see, hear, and process its surroundings in real time, on a tight power and thermal budget, using components that have already been proven reliable at scale, because nobody is trusting a machine walking through a warehouse or a living room to run on unqualified silicon. Vision processors, radar, LiDAR, and sensor fusion chips built for advanced driver assistance systems solve almost exactly this problem, just wrapped around a different chassis. The companies that spent ten years and hundreds of millions of shipped units proving out that silicon for cars are sitting on a sensing stack that transfers to robots at close to zero incremental engineering cost, while the humanoid manufacturers themselves are largely starting from a blank sheet on everything else.

Layer a second, less obvious constraint on top of that. Global memory markets broke in 2026, with AI datacenter demand pulling in the overwhelming majority of DRAM and HBM output and leaving everything else, laptops, cars, and now robots, fighting over what’s left or paying multiples higher for it. Architectures that need little or no external memory at all, once a minor cost optimization buried in a spec sheet, have quietly become one of the more important competitive advantages in hardware right now, and one of the few genuine moats available to a component supplier in this cycle.

Put those two dynamics together and the mispricing becomes obvious. The public companies sitting at the overlap of automotive sensing and humanoid-ready silicon are still being valued almost entirely on their car business, because the market has not yet connected a decade of unglamorous automotive design wins to the robotics wave sitting on top of them for free. One name fits that description more directly than almost any other public company we can find.

Three days ago, China’s securities regulator approved Unitree Robotics’ IPO registration on the Shanghai STAR Market. The approval took 104 days from application to registration, the fastest review the STAR Market has ever processed. Unitree is targeting a raise of roughly $618-620M, implying a valuation near $6.18B, with a listing that could land as early as late July 2026. This is not a someday catalyst. It’s weeks away.

indie Semiconductor ($INDI) has a confirmed, multi-quarter design-in at Unitree, and a confirmed design-in at Figure AI, the other company most people would name if asked who’s building the future of humanoid robots. indie is, as far as we can find, the only U.S.-listed Western semiconductor supplier with disclosed content at Unitree. Today, the overwhelming majority of indie’s revenue still comes from cars, priced not unlike an industrial auto supplier such as Schaeffler. You aren’t being asked to pay up for the humanoid or quantum story yet. The market is pricing indie almost entirely on its automotive base, and the robotics and photonics exposure is riding along for free until something forces a re-rating. The Unitree IPO is that something.

At ~$4.10 a share, with 227M shares outstanding, market cap runs ~$931M. Net debt sits at roughly $230M ($415M total debt against $184.7M of cash exiting Q1 2026), putting enterprise value at ~$1.16B. Against consensus revenue of $266M in 2026 and $521M in 2028, that's ~4.4x 2026 sales, compressing to ~2.2x 2028 sales. Small-cap hyper-growth semiconductor names posting comparable growth rates typically trade at 10x-15x forward sales. INDI is priced at a third of that for a business about to nearly double revenue in two years, before you even get to the robotics and quantum optionality riding underneath it for free.

At today’s price, you’re getting the entire Physical AI business for nothing

This is the same setup as the VPG thesis: a solid, understandable core business already implying more value on its own than the current stock price, with a funded, ambitious humanoid optionality thrown in for free on top of that.

Rather than build a separate discounted estimate, this uses the same Automotive multiples already established in the SOTP work below, applied to the nearest-term years rather than reaching out to 2030. In 2027, using the 7.0x sales multiple this report assigns to Auto, the core Automotive business alone is worth $1.58B. In 2028, using the 6.0x sales multiple established for that year, Auto alone is worth $2.01B. Zero humanoids included, zero photonics, in either figure. That Auto figure properly excludes Wuxi indie Microelectronics, the Chinese subsidiary indie is fully divesting for roughly $135M in cash, since Wuxi represents close to “40%” of current revenue (Note, they disclose Wuxi entire revenue on income statement, even though they own ~35%) and deconsolidates entirely once the sale closes.

At today’s ~$1.16B enterprise value, the 2027 Auto figure alone is already 1.4x today’s entire company, and the 2028 figure is 1.7x. You aren’t paying a fair price for the core and receiving Physical AI for free, you’re paying a fraction of what the core business alone already implies one year out, using multiples this report has already defended on their own terms, and receiving the entire Physical AI and Photonics opportunity as pure upside on top of that gap.

You starting to see the opportunity here?

That gap, between today’s $1.16B and the $1.58B-$2.01B the core alone already implies over the next one to two years, is the real setup. Physical AI and Photonics don’t need to close that gap on their own, they’re free regardless. What they add is a second, separate reason for the stock to re-rate even after the core catches up to its own near-term value, since by the SOTP’s own base case those two segments are worth $1.12B in 2027 and $1.31B combined by 2028, growing from there.

The ADAS base: growing content, widening design wins

The automotive semiconductor TAM is growing from $77B in 2025 to $133B by 2030. Average semiconductor content per vehicle climbs from $892 in 2023 to $1,514 by 2030.

indie’s own silicon content per vehicle is growing faster than that industry average, from roughly $25 today to an estimated $150+ by 2030, a 6x expansion driven by stacking front camera, rear camera, in-cabin camera, eMirror, ultrasonic, front and rear radar, corner radar, in-cabin radar, and incremental LiDAR onto a single platform.

Proof this isn’t a slide deck fantasy: indie has shipped 550M+ SoCs cumulatively through 2025, up from 400M+ in 2024 and just 29M in 2016. Design win breadth keeps widening too. The Tier 1 radar partner’s $25M production order booked in Q1 2026, representing “the first 4TX 8RX radar available in the industry” per management, spans mainstream and commercial vehicles across Europe, Asia, and the U.S., and management confirmed more customers are layering in behind the initial two OEMs.

On the Q4 2025 call, management sized the underlying opportunity directly, describing annual demand as “well above 50 million units” once past the ramp-up phase.

That’s the base case. Now here’s the part of the story that actually changes the multiple.

Why the DRAM-less architecture matters more than anything else, specifically in auto

Every semiconductor company on earth is currently explaining DRAM cost inflation as a headwind on its earnings call. indie is one of the only companies that gets to present the shortage itself as a selling point.

The iND880 vision processor eliminates the external DRAM chip from the image signal processing pipeline entirely, a deeper change than a cost optimization on a spec sheet: a part that physically doesn’t need to exist on the board. No external memory means no memory allocation risk, no board space for it, no power budget for it, and no separate supplier relationship to manage for it.

Removing that chip matters far more today than it would have two years ago. TrendForce started 2026 forecasting DRAM contract prices up 50-55% quarter over quarter, then revised that to 90-95% QoQ for conventional DRAM by February, with PC DRAM alone projected to more than double in a single quarter. By Q2, LPDDR5X was up 89% QoQ and DDR4 up 51%. AMD told the industry not to expect DDR5 pricing to normalize until 2028. IDC estimates AI data centers will consume roughly 70% of global memory output in 2026, and every wafer that becomes an HBM stack for an Nvidia accelerator is a wafer that doesn’t become conventional DRAM. Global DRAM supply growth is running at just 16% year over year, and new fab capacity from Micron and SK Hynix doesn’t reach volume production for years.

Don McClymont’s own words on the Q1 2026 call capture it better than any outside data point. He called the DRAM-less design “a defining advantage of the iND880 and increasingly a focal point in our customer engagements,” and put the customer problem plainly:

“In many cases, our customers are unable to source memory at all.”'

Not expensive. Unavailable. When memory can be sourced at all, he added, it comes at a price premium measured in multiples, not percentages. The language across the last three earnings calls has escalated too, and that progression tells you this wasn’t a one-quarter talking point. On the Q4 2025 call, management framed the architecture as creating new opportunities tied to BOM savings, a competitive-differentiator pitch. By Q1 2026, it had become a focal point in customer engagements in management’s own words, expanding from Chinese OEMs specifically to U.S. customers on compressed timelines.

There’s a structural reason this holds up beyond just good timing. Automotive vision systems run on legacy DDR3 and LPDDR4-class parts, older nodes that Samsung, SK Hynix, and Micron are actively winding down as they redeploy fab capacity toward HBM and DDR5. So even if the headline AI memory supercycle eventually cools, automotive-grade commodity DRAM could stay scarce for a reason that has nothing to do with AI demand at all: the suppliers are exiting that category, and automotive qualification timelines run roughly seven years, so a supplier that exits doesn’t come back on any schedule that helps a car program next year. That advantage doesn’t mean-revert with spot pricing.

There’s also precedent for how far this kind of architecture can travel. Silicon Motion built its business on DRAM-less SSD controllers, a design that started as a niche, cost-down alternative and became the default architecture across client SSDs. Eliminating a dependency has a track record of going from stopgap to category standard once the reliability case proves out.

The buyer signing off on this design win matters as much as the design itself. It’s automotive procurement and that function lived through the 2021-22 chip shortage and the line-down events that came with it. “This part cannot strand your line on memory allocation” reads to that buyer as an insurance policy against a repeat of 2021-22, and insurance sells on a different urgency curve than a cost pitch ever does.

It’s already converting to real revenue too. The iND880 went from design win to volume production at NIO for eMirror functionality in roughly six months. A camera mirror system with the largest Chinese OEM is entering volume production. Beijing Auto Show design wins include the Buick GL8, AITO M9, NIO ES9, and Cadillac VISTIQ. Management has also disclosed design wins in eMirror and camera mirror systems across passenger vehicles and trucks with leading Tier 1s, production beginning in late 2026.

Management is already citing “tens of millions of dollars per year in annual revenue” for this single line, growing fast enough that it could exceed radar revenue in 2026, the same radar program that just booked a $25M order.

One more data point worth knowing: indie itself had to navigate a package substrate shortage in 2025, the same AI-driven demand spike squeezing its own manufacturing supply chain, before qualifying second-source substrate and packaging vendors. A company that has had to solve its own component scarcity problem in real time can credibly sell a scarcity-proofing architecture to its own customers, having lived inside the same dynamics it’s now monetizing downstream.

I am unsure if this iND880 vision processor will be used for humanoids or the iND881, but this DRAM-less architecture is a real commercial advantage for smart / autonomous vehicles and any prospective clients/applications adopting it.

iND881: the newest chip, and the one aimed straight at the Physical AI thesis

indie published a new product page for the iND881 series in June 2026.. The iND881 builds directly on the iND880’s ISP core, the same low-latency, multi-camera image processing engine already in eMirror and camera mirror production, and adds a genuine on-device AI compute engine on top of it: a 2 TOPS neural processing unit paired with a 500 GOPS DSP, tuned for real-time object detection and scene understanding rather than just image cleanup.

Two variants ship, iND88140 at the higher end (700 MP/s, 2 TOPS, quad-core application processor) and iND88120 at the lower end (350 MP/s, 1 TOPS, dual-core), a deliberate two-tier structure built to win a range of customer price points rather than one (Think the lower price and higher price Unitree models, for example).

The applications list is the part that matters most here. Alongside the expected automotive uses, ADAS, eMirror, DMS/OMS, parking and back-up cameras, indie’s own product page lists “smart cameras for AMRs and Humanoids” directly under Industrial and Consumer applications. This comes straight from the company’s own product marketing naming humanoid robotics as a target market for a specific, shipping part.

It’s also a more natural fit for what a humanoid actually needs than the ISP-only iND880. A robot navigating a warehouse or a living room has to make decisions from its video feed in real time, not just clean the image up, object detection and scene understanding, the exact workload indie built this chip’s NPU and DSP to handle. If that’s the direction indie’s humanoid content moves over time, and a product roadmap this specific suggests it is, iND881 rather than iND880 could end up carrying more dollars of content per unit, not less, even though it gives up the DRAM-less advantage to get there.

iND881 needs a 32-bit LPDDR4(x) external DRAM interface, confirmed in four separate places on its own spec sheet, so it doesn’t inherit the memory-scarcity moat built around the iND880. Whether that matters commercially depends on how the disclosed Unitree and Figure AI design-ins evolve from here, both are currently described as running on iND880-powered cameras, the ISP-only tier. If those relationships stay on iND880, the DRAM-less thesis holds as written above. If they migrate toward iND881-class compute as robot autonomy gets more sophisticated, which looks like the more likely direction over time, the content opportunity probably gets bigger, but the memory-scarcity argument stops applying to it specifically. Both things can be true at once, and management hasn’t yet disclosed which tier future design wins will run on.

Physical AI: four straight quarters of confirmation

In August 2025, on the Q2 call, McClymont disclosed indie’s first commercial success with iND880-powered cameras shipping into humanoid robots. In November 2025, on the Q3 call, he confirmed the relationship in the clearest terms indie has given, describing new content at “leading robotics providers, figure.ai and Unitree,” who use indie’s automotive products directly for their own applications. Humanoid robotics was added to indie’s $7.4B strategic backlog for the first time, and he called the interest an “unexpected positive.” By February 2026, indie’s ADAS technology was framed as the robot’s eyes and ears. By May 2026, the language shifted to sizing the opportunity, citing Yole Group’s forecast of the global humanoid market growing from $600M in 2025 to $6B by 2030 and then $51B by 2035.

Normally a stock would soar on confirmed content at just one of these two companies. indie has disclosed deals with both, confirmed by the CEO directly. On an earlier call, McClymont went further on why the fit works, saying indie’s automotive products are “basically 100% compatible with the needs that these guys have for these applications,” the same silicon, not a redesigned product line. Figure AI alone has stated plans to produce 100,000 humanoid robots over the next few years out of its dedicated BotQ manufacturing facility.

By Q1 2026, management described “tremendous growth in interest and activity in quantum and robotics” as a defining feature of the quarter.

One speculative data point worth flagging as exactly that, speculation, not a confirmed relationship: indie’s LinkedIn recap of its presence at the Optica Quantum 2.0 Conference, covering a presentation delivered by Philipp Vorreau, drew likes and comments from multiple Intel employees, including Run Levinger, whose title is Principal Engineer for quantum computing control at Intel Corporation, and Roee Ben-Yishay, an mmWave and RFIC designer at Intel, who commented directly on the quality of the papers presented. Engagement from name-matched employees at a specific company, especially one working on quantum computing control specifically, the exact discipline indie’s DFB laser platform sells into, is the kind of pattern worth noting rather than ignoring. It is not evidence of a partnership, a design win, or even a conversation, and should be read as nothing more than an incremental data point until something firmer is disclosed.

Why humanoid scale keeps compounding well past 2030

Everything modeled below through 2030 should be read as early-innings scale, not the final destination. The realistic debate in humanoid manufacturing isn’t whether the leading OEMs eventually reach multi-million and then double-digit-million annual unit volumes, it’s when. Unitree shipped 5,500 units in 2025 and guided to 20,000 for 2026, a 3.6x jump in one year. Figure AI’s BotQ facility is explicitly designed to keep adding production lines toward its own multi-hundred-thousand-unit ambitions. Tesla has publicly floated an eventual 10-million-unit annual capacity target at Gigafactory Texas. These companies are sizing their factories for the demand curve they expect in the early 2030s, not for where volumes sit today.

The binding constraint on how fast that curve bends isn’t primarily capital anymore, and it isn’t primarily the AI stack either, which has improved faster than almost anyone modeled eighteen months ago. It’s the upstream precision component supply chain: harmonic drives, NdFeB magnets, and dexterous-hand tactile sensors, categories where few qualified global suppliers currently exist. That’s a real bottleneck, closer to standing up new semiconductor fab capacity than adding a stamping line, and it’s the actual variable that determines whether the industry crosses into the multi-million range by 2030 or by 2032/2033.

There’s a real case that this bottleneck breaks faster than the McKinsey-style supply chain reports currently assume, because this exact pattern, “impossible” component scale-up on a compressed timeline, has already played out in adjacent Chinese industrial supply chains. Green Harmonic, now likely Tesla’s primary harmonic reducer supplier, was a small-cap gearing company in 2022. Ningbo Tuopu was an obscure auto suspension parts maker in 2016 and now likely builds actuators for Optimus. China’s own playbook in solar, batteries, and EVs has consistently been to flood capital into a component category ahead of confirmed demand and let capacity growth pull the market forward, exactly the BYD precedent underneath this whole argument: BYD went from roughly 200,000 NEVs to over 4 million in about five years once that playbook engaged. If that same state-backed, capital-flush approach gets pointed at harmonic drives and NdFeB magnets specifically, and there’s real reason to think it will given China’s explicit “hard tech” policy language around embodied intelligence, the five-supplier bottleneck could loosen considerably faster than conservative sell-side models currently assume.

That playbook comes with a catch worth remembering: it tends to arrive with brutal price destruction and consolidation along the way, the same thing that happened to solar module makers and is now happening to EV assemblers competing on price in China. That matters more for Unitree’s own hardware margins than it does for indie’s content-per-unit economics, since indie sits one layer up the stack selling sensing and processing silicon into whichever OEMs win the volume war, not competing for share of the robot itself. A supply-flood scenario that multiplies units while compressing OEM margins is, if anything, a better outcome for indie’s silicon content than a slower, higher-margin, lower-volume path would be.

That’s why the multiples in the valuation work below don’t collapse by 2030 the way a normal maturing business’ would. The market has to price in some of the 2031-2035 curve today, and increasingly over the next several years, precisely because by the time the hockey stick is obvious in the reported numbers, the re-rating will already be behind you.

Blending Unitree and Figure AI into one Physical AI shipment base

Unitree is the headline name because of the IPO timing, but indie’s disclosed exposure runs through both companies, and Figure AI’s own production plans are a meaningful second engine. Figure’s BotQ facility launched with an initial line capable of up to 12,000 units annually, building toward a stated goal of 100,000 cumulative robots over four years. Our base case uses a Figure AI ramp reaching 250,000 units by 2030, blended with a Unitree base case reaching 500,000 units by 2030.

That Unitree figure runs well above Morgan Stanley’s 446,000-unit forecast for the entire China market, not just Unitree, by 2030. We’re making that call deliberately rather than defaulting to the bank number. Sell-side humanoid estimates get revised every few months right now, and the pace of price collapse, capacity announcements, and design-win disclosures over just the past two quarters argues that the consensus curve is lagging reality rather than leading it. Treat this as a house view that we’re willing to be wrong on, not a hedge.

Content per unit is $150 across all three scenarios, bear included, so bear’s lower dollar contribution comes entirely from fewer units shipping. Bear assumes both companies undershoot current guidance. Base reflects my own above-consensus view of both companies’ trajectories rather than a straight read of disclosed plans. Bull compounds Unitree’s own demonstrated 2025-2026 growth rate forward and gives Figure AI a similarly steep, if less proven, ramp. Even the bull case’s 2.18M combined units by 2030 sits well below the industry-wide multi-million-per-OEM outcome discussed above, an early stage of that curve rather than its endpoint. I still sit in the camp that we see many million deployments of humanoids by OEMs, including Optimus.

There’s also an evidence trail worth naming. Automotive supplier pivots into humanoid actuators have another concrete precedent now beyond Green Harmonic and Ningbo Tuopu: Shuanglin Group, an existing gear and motor manufacturer retooling toward planetary roller screws for robotics. Over half of humanoid component categories already overlap with the EV and consumer electronics supply chains, which is exactly why Tesla, Mercedes-Benz, Changan, SAIC, XPeng, and GAC are all entering embodied intelligence through internal programs, joint ventures, or incubation, several of which are already indie automotive customers or prospects in their own right.

Worth addressing directly: does Unitree’s own aggressive price collapse threaten indie’s content-per-unit assumptions? The evidence says no, and the reason is informative. Unitree’s own production cost per unit fell far less over the same period, from 73,200 yuan (~$10,800) in 2023 to 62,200 yuan (~$9,175) by late 2025, only a 15% decline against a 72% drop in selling price. That gap says the ASP collapse is coming from Unitree compressing its own margin to chase volume and scale, sourcing over 90% of core components in-house, not from gutting the functional sensor and compute stack a robot needs to actually work. A camera, an ISP, and a perception chip are specialized components a humanoid cannot function without, closer to the electronics in a car than to a trim level that gets value-engineered down first when a product chases a lower price point.

GGII’s fresh numbers reinforce the scale case: China’s humanoid shipments hit 14,400 units in 2025 and are projected to reach 100,000 to 200,000 units in 2026 industry-wide, well above Unitree’s own 20,000-unit guide alone, meaning the broader China market indie’s OEM relationships sit inside is scaling faster than a single-company view suggests. It also means that Unitree is likely scaling faster than their stale guide.

What’s on a Unitree robot, and what indie’s share is worth

Unitree’s G1 sells for roughly $13,500-$16,000. The R1 entry model starts around $5,600-$5,900. Against that pricing backdrop, indie’s addressable content sits in vision (camera ISP via iND880), potentially LiDAR (iND83301 SoC), and perception software royalty (emotion3D). Content per unit is $150 across bear, base, and bull cases alike. The logic: a humanoid processing visual and auditory input at the low latency embodied AI requires isn’t meaningfully different in sensing and compute demands from what indie already charges for in an autonomous car, so there’s no reason the per-unit content should sit meaningfully below where the ADAS content curve is already headed, and no reason it should vary by scenario. What differs between bear, base, and bull is entirely shipment volume, not content per unit.

This is, transparently, an assumption that I am still working on. It does adjust the potential valuation model if it goes up or down.

The image sensor acquisition: completing the stack

The June 2026 investor deck disclosed indie’s agreement to acquire AMS OSRAM’s CMOS image sensor group: €40M total consideration, €35M cash at closing, expected close Q3 2026, immediately accretive per management. It completes indie’s full sensor stack and is explicitly described as a “Physical AI Beachhead,” sitting inside a market indie sizes at $4.21B by 2030 for CMOS image sensors specifically. Before this deal, indie built ISPs downstream of someone else’s camera sensor, the single most expensive and IP-rich component in a camera node, and never captured that dollar. Now it does, which is the direct mechanical enabler of the jump from indie’s current ~$25 content per vehicle toward the $150 endpoint, and the missing layer in the Physical AI pitch specifically.

Separately, the June deck disclosed a new software royalty line: per-vehicle licensing on DMS/OMS and forward vision perception software, the first meaningful software licensing stream in indie’s history. The Mahindra win is already in production, near-100% incremental margin once it scales.

The third leg: photonics and quantum

indie’s DFB laser platform now includes the world’s first commercially available ultraviolet DFB laser at 399nm, matched to the atomic cooling transition of ytterbium. The company shipped roughly $1M of optical products into quantum applications in 2025, expecting that to roughly triple in 2026. Small in dollars today, a pure call option on a market growing independent of anything happening in Detroit or Shenzhen.

The financial model

Everything below is an illustrative modeling exercise built on stated assumptions, not company guidance.

Wuxi is a Chinese subsidiary indie controls but only holds a 34.38% equity stake in, a lighting and motor-control chip business, structurally separate from the vision, radar, and ADAS silicon this report is actually about. Because indie controls it, US GAAP requires full consolidation, 100% of Wuxi's revenue shows up in indie's reported top line today, not a 34.38% pull-through. indie agreed in October 2025 to sell its entire stake for roughly $135M in cash, and Wuxi fully deconsolidates once the deal closes, expected sometime in 2026. Wuxi has run flat to slightly down for two years, it's roughly 40% of current revenue and has been consistently so, not a growth contributor, and management has said directly that it's meaningfully lower-margin than the core business, so its exit should improve blended gross margin, not hurt it. Since Wuxi fully deconsolidates on close rather than fading out gradually, the honest way to model it is a clean cutover: Wuxi contributes through 2026, then drops out entirely, with core Automotive growth picking up the pace management itself has described as accelerating, roughly 20% year-over-year most recently, before layering in the radar, eMirror, image sensor, and software royalty ramps this report already covers in detail.

The 60% bull-case gross margin ceiling comes straight from management’s own long-term target, not something invented for this model.

Asked directly on the Q3 2025 call what margin and breakeven look like excluding Wuxi, Don McClymont didn’t hedge, reaffirming the company remains “committed to getting to the 60% gross margin level of the target model” it set for itself.

Base case 2030 deliberately stops three points short of that, at 57%, on the view that base reflects a slightly slower mix shift than bull’s more optimistic content and volume ramp. Both are real, disclosed ceilings management has pointed to, not invented multiples.

Segment-level figures shown in the SOTP below are that same consolidated margin applied to each segment’s own revenue, not a separately assumed margin per segment. Segment revenue splits into three lines: Automotive (core ADAS/UX, Wuxi through divestiture, software royalty), Physical AI (blended Unitree and Figure AI content plus, in bear and base, a supporting Other Physical AI line), and Photonics/Quantum.

Automotive carries the overwhelming majority of revenue throughout. In the base case it grows from $264.4M in 2026 to $696.0M by 2030, with Physical AI scaling from a rounding error to $214.5M and Photonics from $3.0M to $28.0M over the same span. The bull case pushes Automotive to $1,300.0M and Physical AI to $327.0M by 2030, while the bear case holds Automotive to $495.5M and Physical AI to $101.5M. Total revenue, gross margin, and the resulting EBITDA across all three scenarios sit below.

The 2026-to-2027 step is a one-time optics effect, not a sign of weakness, roughly $94M of flat, lower-margin Wuxi revenue exits entirely while the core business keeps growing underneath it. Base and bull show only a modest net dip, -9.0% and -3.1% respectively, since accelerating core growth offsets most of the loss. Bear shows a sharper -28.3% step, since its slower core growth can't offset the exit as well, arguably appropriate for a bear case, but worth naming plainly rather than leaving an unexplained decline sitting in a headline number.

Non-GAAP EPS: bear stays negative through 2027, building from ($0.22) in 2026 to ($0.19) in 2027, crossing into thin positive territory at $0.02 in 2029 and reaching $0.21 by 2030. Base turns profitable already in 2027, essentially breakeven at $0.01, building to $0.81 by 2030. Bull turns profitable in 2027 too, at $0.04, the strongest of the three, reaching $1.72 by 2030.

2026 is a shared, roughly breakeven-adjacent year across all three scenarios, base and bull both inflect in 2027, helped by gross margin doing most of the early work, and bear stays a slower, later story. Every one of these figures sits on a revenue base that conservatively excludes a stagnant, lower-margin, soon-to-be-divested Chinese subsidiary.

One more piece worth calling out explicitly: the Wuxi cash proceeds are a real balance sheet event, separate from the revenue story. The sale brings in roughly $135M gross, or about $108M net of an assumed 20% tax, arriving roughly in step with the divestiture closing. That cash strengthens the balance sheet precisely as indie is scaling into radar, eMirror, and image sensor production ramps that need working capital, taking pro forma net debt down to roughly $122M, and is explicitly reflected in the net debt figures used throughout the SOTP below. Losing Wuxi’s revenue and gaining Wuxi’s cash are two different, both real, effects, worth keeping separate rather than letting the first obscure the second.

I might be understating the auto revenue growth in 2027 & 2028, but want to be pragmatic with the Wuxi divestiture. Something I can continue to iterate upon. I might be understating the photonics / quantum growth, too. I just haven’t spent enough time here, as it is likely the most far out of the three segments.

I AM CONSERVATIVELY SITTING BELOW CONSENSUS SALES FOR 2027 & 2028 AND WE STILL GET ~200% UPSIDE IN THE BASE CASE.

Sum-of-the-parts valuation

A single blended multiple understates what’s inside this company, since it would cap a hypergrowth robotics and quantum business at the same multiple as a maturing auto supplier.

Bear stays on EV/Sales for all four years, since consolidated margin only reaches 19.5% by 2030 and the segment-level EBITDA base is still thin. Base and bull both keep the same sales-basis multiples used in 2027 and 2028, since the underlying revenue in those years hasn’t changed, before shifting to EV/EBITDA from 2029 onward once real, disclosed-scale profitability is established.

Segment multiples, all three scenarios:

The base case’s Automotive multiple stepping down from 19x (2029) to 14x (2030) rather than climbing further reflects the segment’s own EBITDA scaling quickly enough that a stable, more moderate multiple already produces a rising enterprise value, the same logic as before, a business late in its growth curve doesn’t need an ever-richer multiple once the earnings base itself is compounding.

The multiples compress hard across the sales-basis years in every scenario, which is exactly the pattern you’d expect: richest when the segment is most dependent on the story, more modest as revenue scale grows and real EBITDA becomes the better yardstick.

Implied share price, all three scenarios:

The bear case is a grinding, unexciting climb rather than a collapse, a gradual move from $2.61 to $4.44 as revenue keeps growing even though the business never really turns the corner on profit within this window. Auto’s own value stays modest throughout, a direct consequence of the Wuxi-excluded revenue base combined with bear’s weaker core growth assumption, the two effects compounding on top of each other in the scenario least equipped to absorb them. The full segment-level build behind it: Auto enterprise value climbs from $535M to $779M, Physical AI from $118M to $152M, Photonics from $80M to $98M, net of a balance sheet that swings from $122M net debt to a net cash position over the window.

The base case runs from $11.02 to $19.28 across the four years, with Auto EV growing from $1,583M to $2,789M and Physical AI from $999M to $1,566M. The bull case is where the divergence gets dramatic, $15.74 to $44.28, with Auto alone reaching $6,912M of enterprise value by 2030 and Physical AI $3,617M.

Auto stays the largest single segment in the bull case throughout this window, with Physical AI a substantial but consistently smaller second, one that carries relatively more weight in this corrected build than a simple eyeball of the segment sizes might suggest, since Auto’s own base is now properly scaled to exclude Wuxi. Treat the bull case for what it is, the scenario where the content ramp, the shipment compounding, and a sustained premium multiple all show up together across every segment at once. It’s a live possibility given how fast this market is moving, not a base rate, and the bear case is the reminder of what a stalled version of this story implies instead.

12-month price target

Pulling this together into a single number rather than a four-year table: our base case 12-month target is $11.02, and our bull case 12-month target is $15.74. Both are the 2027 output of the SOTP build above, not a separate estimate, so the reasoning behind them is everything already walked through: Auto's own content and design-win trajectory on a revenue base that properly excludes Wuxi, and Physical AI carrying a genuine premium multiple that reflects the hyper-growth expected across humanoid robotics through the 2030s rather than a discount for being early. The base case implies roughly 2.7x from today's $4.10, the bull case roughly 3.8x, over the next year. Both depend on the Unitree IPO actually landing in the next few weeks and the market starting to ask who else sits in that supply chain, the base and bull cases differ mainly in how much of Auto's own near-term content ramp converts on schedule.

Why the IPO is the trigger

Every thread above has been sitting in plain sight in indie’s own transcripts and decks for months without moving the stock. IPOs change that, they put a name and a valuation in front of generalist money that has never had to think about who supplies the sensors inside a humanoid robot. Once Unitree prices, likely within weeks, the natural next question from anyone allocating into that theme is who else is exposed to this. Right now, with a confirmed multi-quarter design-in and next to no market attention on the position, the answer is $INDI.

Risks worth sitting with

Four things could break this thesis independent of anything modeled above.

Slower humanoid adoption or development. If Unitree, Figure AI, or the broader humanoid ecosystem takes materially longer to reach scale than even the bear case assumes, Physical AI could stay a rounding error on indie’s income statement well past 2030, not just through the near-term years modeled here.

ADAS market slowdown. Auto still carries most of enterprise value in every scenario in this report. A broader auto production downturn, a pause in ADAS regulatory mandates, or an OEM capex pullback hits the segment everything else here is built on top of.

A valuation discount for China exposure. indie’s revenue runs through Chinese OEMs (NIO, AITO, the largest Chinese OEM’s camera mirror win) and the pending Wuxi divestiture. The market has shown a willingness to apply a blanket discount to US-listed companies with meaningful China revenue exposure on tariff, export control, and geopolitical grounds, independent of how the underlying business actually performs.

ADR-adjacent risk. indie trades as ordinary Nasdaq common stock, not an ADR, but its Wuxi equity stake and China-linked revenue invite some of the same scrutiny US-listed Chinese ADRs face, PCAOB audit access questions, VIE-style structural concerns, and headline risk from US-China regulatory or delisting actions, any of which could compress the multiple the market assigns even if operations are unaffected.

What has to go right, and what could go wrong

The bull case shipment path sits above institutional consensus, treat it as the bull case it is. indie hasn’t disclosed a per-unit content figure for Unitree or Figure AI, so all the BOM math here is a bottoms-up estimate. The SOTP multiples are stated assumptions built for illustration, swap in your own view freely. The real industry-wide bottleneck on reaching multi-million unit scale sits in the precision component supply chain, harmonic drives and magnets specifically, more than in capital availability, where fewer than five qualified global suppliers currently exist and a multi-year buildout stands ahead even with unlimited capital behind it. DRAM pricing could normalize faster than AMD’s 2028 timeline if hyperscaler capex cools. Wuxi still needs to clear Chinese regulatory approval before that ~$135M shows up. The core radar and vision ramps still need to execute cleanly, since none of the adjacent-market optionality matters if the base business stumbles.

None of that changes the shape of the trade. indie built a chip to solve its own bill-of-materials problem in cars. The memory market handed it a much bigger problem to solve, for free, in robots, and the company is still priced at ~4x sales and 2x 2028 consensus sales. Unitree’s IPO is about to test how long that lasts.

Not investment advice. Do your own diligence.

$PENG printed the best quarter in company history and raised guidance for the second straight quarter. Management dropped a preliminary FY27 view that I think undersells what this business can actually do. Here’s the review, and the updated bottoms-up model through FY30.

Q3 was a clean beat across every line

Record net sales of $479 million, up 48% year over year and 40% sequentially. Non-GAAP diluted EPS of $0.84, up 79% year over year and 62% sequentially, crushing the FactSet estimate of $0.54. Non-GAAP operating income hit a third quarter record of $64 million, up 67% year over year, and operating margin expanded to 13.4%.

Segment breakdown:

• Integrated Memory: $275 million, up 111% year over year and 60% sequentially. Both volume and pricing drove this. Memory is now 57% of total company revenue.

• Advanced Computing: $138 million, up 4% year over year and 19% sequentially. The headline growth number looks soft, but the mix underneath is the real story. Non-hyperscaler AI infrastructure within this segment grew 81% year over year and is now 58% of segment revenue, up from 33% a year ago. Hyperscaler and Penguin Edge wind-down are dragging the reported number down while the actual AI infrastructure business scales hard.

• Optimized LED: $66 million, up 7% year over year. Steady, cash-generative, not the story.

AI-driven businesses (Memory plus non-hyperscaler AI infrastructure) were 74% of total company revenue and grew 104% year over year. Management said demand is outpacing net sales growth in both of these businesses, which is building backlog into Q4. That’s the kind of statement you want to hear from a company that just beat by 55%.

Penguin added four new AI Infrastructure logos in Q3, bringing the trailing four quarter total to 13 new logos, with 7 of those logos already increasing business with Penguin in the timeframe. One new data point worth flagging: management said they picked up a Hyperscaler customer in the memory business this quarter, contributing to volume growth. On the call, Kash Shaikh was explicit that this is not a large contributor yet and is consistent with the existing FY26 run rate, not some immediate multi-hundred-million dollar customer.

Opex leverage is already showing up. Q3 non-GAAP opex was $70.4 million on $478.7 million revenue, or 14.7% of revenue, up $8.7 million sequentially from Q2’s $61.7 million. Backing into Q4 off the full year opex guide of $260 million ± $5 million, nine month opex of $193.4 million implies roughly $66.6 million for Q4, which against the ~$505 million implied Q4 revenue works out to about 13.2% of revenue. Opex is growing in dollar terms every quarter, but slower than revenue. That’s the leverage story in one data point.

CFO transition: Nate Olmstead steps down July 8th, and Aaron Johnson, VP of Finance and Accounting, steps in as interim CFO while a formal search runs. Management framed it as a smooth handoff with no change to operating philosophy. Worth tracking, but not a thesis-changing event on its own.

Read more

The electrical infrastructure stack for AI data centers is being rebuilt from the ground up. Not the cooling. Not the networking. The wiring itself. For decades, large-scale computing facilities have run on alternating current distribution, typically at 415 VAC in Europe and 480 VAC in North America, with power cascading through multiple conversion stages from the utility meter to the server rack. Each stage is a loss event. Each transformer bleeds efficiency, generates heat, and adds latency to power delivery. That architecture worked because it had to, and because rack power demands were manageable enough that the inefficiency was tolerable. That tolerable threshold has now been crossed. The Vera Rubin generation of compute racks runs at up to 370 kilowatts per rack, against roughly 120 kilowatts two years ago on Hopper. At that power density, high-voltage direct current stops being an engineering preference and becomes a hard physics requirement. Running 800 VDC distribution means that at equivalent power delivery, current drops proportionally, conductors get smaller, resistive losses in the distribution network collapse, and PSUs operating on DC input run materially more efficiently than the same hardware on AC input. The industry is eliminating conversion stages it tolerated for a century because it can no longer afford them.

Nvidia formalized this transition at Computex 2026. In a joint architecture roadmap developed with Siemens, Nvidia published the blueprint for what they are calling next-generation AI factories: 800 VDC distribution fed through a medium-voltage rectifier or solid-state transformer, replacing the old chain of step-down transformer, main switchboard, AC UPS, and power distribution unit that has been standard for fifty years. The document is explicit. The 2025-era data center diagram shows battery storage with a dotted line, labeled optional. The 2027-era AI factory diagram shows battery storage as a solid block, labeled in the same weight as the medium-voltage rectifier and the 800V DC bus. No dotted line. No optional. Battery Energy Storage is now a standard, mandatory component of the reference architecture that Nvidia and Siemens are putting in front of every hyperscaler, neocloud, and sovereign AI program building to this specification. When Nvidia draws a mandatory box in an official blueprint, the procurement cycle for that box begins immediately.

The reason BESS is mandatory rather than optional in this architecture comes down to three engineering problems that all surface simultaneously as rack power crosses the 100+ kilowatt threshold. First is grid interconnection. The US interconnection queue across major ISOs and RTOs now stretches five to ten years in most territories, a direct product of the volume of large load connection requests filed by data centers, EV charging networks, and industrial electrification projects. A developer who needs two gigawatts of capacity to run a full AI campus can contract for 500 megawatts of grid connection today and deploy battery storage to buffer peak compute demand within those limits, effectively compressing a decade-long infrastructure bottleneck into an engineering workaround. Speed to power is the single most commercially urgent problem in the hyperscaler buildout right now, and BESS is the only economically viable solution at scale. Second is power quality. AI workloads generate demand fluctuations at millisecond timescales that passive grid infrastructure cannot absorb. When a GPU cluster cycles through a training run or inference burst, the instantaneous power delta destabilizes voltage on the 800V DC bus, tripping server PSUs across the cluster. BESS systems with sub-10 millisecond response capability inject or absorb power fast enough to maintain bus stability. Traditional AC UPS systems using lead-acid or VRLA chemistry operate at response times an order of magnitude too slow for this application. Third is backup power. Lithium-based BESS has reached total cost of ownership parity with diesel generator backup at multi-hour durations, while eliminating refueling logistics, emissions liability, and the maintenance overhead of large rotating machinery. For facilities drawing 8 megawatts per compute cluster, the operational and ESG case for replacing diesel backup with battery backup has closed completely.

The market that emerges from this transition is large and accelerating. Global BESS market projections from Wood Mackenzie, BloombergNEF, and S&P Capital IQ converge on $105-150 billion by 2030, built primarily on utility-scale and front-of-meter storage demand. Those models were constructed before the Hyperscaler data center demand vector was visible in the order intake data of the leading BESS manufacturers. US installed battery storage capacity is projected to exceed 170 gigawatts by 2030, from 2.4 gigawatts in 2020. The Siemens/Nvidia 800V reference architecture is now the certified specification for sovereign AI programs in the Middle East, Southeast Asia, and Europe, markets where national AI infrastructure programs are deploying capital at speed and have neither the time nor the institutional knowledge to engineer power architecture from scratch. The 800V future data centers’ procurement cycles are active today.

Read more

Attendees

• Eric Blashford, President & CEO, Broadwind, Inc.

• Thomas Ciccone, VP & CFO, Broadwind, Inc.

• Brian Hawthorne, Investor Relations

• 9 Ventures, just a guy who likes his job

Overview

I got time with Eric Blashford and Thomas Ciccone earlier this week. The conversation covered backlog dynamics, competitive positioning, margin targets, and the long-term Supercycle thesis. The key takeaway: management is not hedging and are laser focused on delivering for customers, driving both growth and margin. They are explicitly targeting historical best margins in both Gearing and Industrial Solutions in the future, and the demand picture they are describing is unlike anything they have seen before.

Read more

When we broke down the FCEL CEPA, the central question was: who exactly is Fit Energy, and can they actually identify 380 MW of U.S. data center sites?

The report below walks through that.

Two data points change how we frame this entirely.

1. Fit Energy was incorporated in October 2025, six months after a major merger inside Fit Ventures, and eight months before the CEPA was signed. It was purpose-built to be the energy counterparty in this deal.

2. Hypertec, the infrastructure operator inside Fit Ventures, is NVIDIA’s Canadian Partner of the Year and Intel’s Strategic OEM Partner of the Year in Canada. NVIDIA and Intel don’t hand those awards to fringe operators. One of North America’s most decorated AI infrastructure companies, 40 years in business, sits inside the same family office that just signed a fuel cell supply agreement with FCEL.

The Fit Ventures Empire

The Fit Ventures portfolio page, confirmed directly from fitventures.com, lists the following holdings:

• 5C Data Centers, next-gen AI data center campuses across the U.S.

• Hypertec, HPC & AI infrastructure, founded 1984, NVIDIA and Intel Partner of the Year

• UVEYE, AI-powered vehicle inspection (Series C-stage)

• BusPatrol, AI safety technology for school buses

• Shielda, cybersecurity

• Thrio, AI-powered contact center software

• DBFF, David Baazov Family Foundation

And now: Fit Energy, the entity that signed the FCEL CEPA.

The ownership chain is straightforward: David Baazov → Fit Ventures → Fit Energy, 5C Data Centers, and Hypertec, all under the same roof.

Hypertec Cloud acquired 5C Data Centers in April 2025, forming 5C Group, a combined entity that unites Hypertec’s cloud computing capabilities with 5C’s data center footprint. Baazov’s family office seeded both sides of that merger. Fit Energy was incorporated six months later. The CEPA followed eight months after that.

Fit Ventures runs a vertically integrated AI infrastructure operation. They build the data centers (5C), supply the servers and compute clusters (Hypertec), and with Fit Energy now in place, they have the dedicated arm to provision baseload power for all of it.

The FCEL CEPA is the final piece of a supply chain Fit Ventures has been quietly assembling.

Hypertec: 40 Years, Award-Winning, Blackwell Already Deployed

Founded in 1984. Headquartered in Montreal. Serves clients in over 80 countries. A partner network of 800+ technology providers globally.

Awards:

• NVIDIA Canadian Partner of the Year, 2025 (announced at GTC 2025, March 19, 2025). Recognized for advancing HPC and generative AI across Canada using the full-stack NVIDIA platform across financial services, higher education, and research.

• Intel Strategic OEM Partner of the Year in Canada, 2024 (and 2022, and 2020). Recognized at Intel Vision in Phoenix for pioneering sustainable immersion computing via Hypertec’s TRIDENT platform.

• Best Innovative Manufacturer for Large Enterprise, 2023

• EcoVadis Best Performer in the Americas, 2023

NVIDIA’s own blog called Hypertec out by name at GTC 2025, the most important AI infrastructure conference in the world. Peer validation, direct from the most important chip company on the planet.

The Blackwell connection is where it gets interesting. At GTC 2025, Hypertec Cloud deployed its first NVIDIA Blackwell HGX B200-accelerated cluster with Together AI, the same Together AI co-building the Memphis MEM01 Frontier AI Factory with 5C Data Centers. Both companies also announced plans to co-build an NVIDIA GB200 NVL72 cluster together.

Read that once more: Hypertec (Fit Ventures portco) builds NVIDIA Blackwell clusters for Together AI. Together AI deploys those clusters at MEM01, built by 5C Data Centers (Fit Ventures portco). Fit Energy (Fit Ventures energy arm) procures fuel cell baseload power from FCEL to supply those same facilities.

Closed loop.

Hypertec’s Five Divisions and Why Each One Matters

Hypertec runs five divisions that map almost perfectly onto the FCEL deployment thesis:

HPC & AI: CIARA-branded servers across AI & GPU, immersion-born, high-frequency, and high-density form factors. The compute infrastructure that goes into 5C’s AI factories.

Solutions Partner: IT products, cloud solutions, and integration. The services layer on top of the hardware.

Construction: High-density data center design and build, backed by 15+ years of HPC data center expertise. The division that physically builds the campuses where FCEL fuel cells would sit.

Custom Manufacturing: 30+ years of precision manufacturing. Hypertec builds its own servers via the CIARA brand, immersion-cooled and GB200-ready, designed for the AI factory use case.

Health: Telehealth and healthcare IT. Separate from the FCEL thesis.

The Construction division is the one to focus on. Hypertec has the in-house capability to design, build, and deploy the physical data center infrastructure at 5C’s expansion sites. Fit Ventures has no need to hire an external EPC contractor. The builder is already in the family.

One of Hypertec’s recent blog posts: “Planning a Data Center in the Next 4 to 12 Months? Prefab First. Power First. Or Expect Delays.”

Power First. That’s the product FCEL sells.

The October 2025 Incorporation: Fit Energy Was Built for This

The 5C/Hypertec merger closed April 2025. That transaction created a combined entity with a 2 GW U.S. data center roadmap, $835M in Brookfield and Deutsche Bank capital, and a team capable of designing, building, supplying, and operating AI factories at hyperscale.

Fit Energy was incorporated October 2025, six months after that merger.

The sequence: merge the data center and compute infrastructure businesses, create the dedicated energy procurement entity, sign the fuel cell power agreement. Fit Energy didn’t exist before this trade. A functional SPV inside an ecosystem with $835M in institutional capital, a 2 GW site pipeline, 40 years of HPC manufacturing history, and awards from NVIDIA and Intel.

The “fresh company” counterparty risk flagged in our original report looks very different with this context.

What 5C Data Centers Actually Is

5C describes itself as the full AI digital infrastructure provider, combining Hypertec’s cloud computing capabilities with a purpose-built data center footprint across the U.S.

The client roster is serious. 5C counts Together AI, Zyphra, NVIDIA, Intel, AMD, VAST Data, WEKA, DDN, and SambaNova as partners. Together AI’s CEO described the partnership as enabling “projects of vast scale with a high level of confidence in their speed, reliability, and success.”

The capital behind 5C is institutional. In July 2025, 5C secured $835 million, equity financing led by Brookfield Asset Management and debt financing led by Deutsche Bank, to accelerate development of its data centers and AI infrastructure across North America.

Brookfield. Deutsche Bank. $835M.

5C’s U.S. data center roadmap is 2 GW. The FCEL CEPA covers up to 380 MW, roughly 19% of that roadmap from a single supply agreement.

Note, that Brookfield also has a partnership with T5, who has signaled to be using FuelCell Energy’s fuel cells in their Fairfax County data center. Brookfield has a multi-GW partnership with Bloom Energy also.

The Five Sites: Where 380 MW Actually Goes

5C’s U.S. site table, scraped directly from 5c.ai, maps a current footprint of 515+ MW across five locations, all with 2027 next-phase delivery timelines. Four of five sites are fully committed today. Expansion capacity across the portfolio totals 1,535+ MW.

CMH01, Columbus, OH

Current: 75 MW | Status: Fully Committed | Expansion: 400+ MW | Next delivery: 2027

The flagship site. 5C’s most recent deal was a Corporate Asset Purchase covering 320,000 square feet in Columbus. Already full. The 400+ MW expansion is where fuel cell baseload becomes highly relevant. Grid interconnection timelines in Ohio are long, and combustion-free on-site generation that bypasses the queue is exactly the product FCEL’s 12.5 MW Energy Blocks are designed to deliver.

MEM01, Memphis, TN

Current: 20 MW | Status: Fully Committed | Expansion: 35 MW | Next delivery: 2028

Watch this one. 5C is building a Frontier AI Factory in Memphis jointly with Together AI, featuring NVIDIA GB200 NVL72 systems. The MEM01 expansion target is 35 MW. FCEL’s Phase 0 is 30 MW. The near-perfect sizing is deliberate. MEM01 is the most obvious Phase 0 deployment target in the portfolio. If Phase 0 delivers clean baseload power to Memphis while the Together AI factory ramps GB200 clusters, this becomes the reference site for the entire 380 MW buildout.

U.S. South Region

Current: 200+ MW | Status: Fully Committed | Expansion: 300+ MW | Next delivery: 2027

Full on current capacity. The 300+ MW expansion is a direct candidate for Phase 1’s 100 MW allocation. At 300+ MW of planned build, exactly the scale where FCEL’s modular block architecture (2.5 MW per block, up to 12.5 MW Energy Block clusters) works best.

U.S. Southwest Region

Current: 20+ MW | Status: Fully Committed | Expansion: None listed | Next delivery: 2027

Smaller footprint, no expansion listed. Phoenix was previously disclosed as a 5C location. Lower priority for FCEL deployment given the lack of expansion runway.

U.S. Southeast Region

Current: 200+ MW | Status: Selective Capacity Available | Expansion: 800+ MW | Next delivery: 2027

The biggest number on the board. 800+ MW of planned expansion at a site where 5C still has selective capacity available. The endgame location for Phases 2 and 3 of the FCEL CEPA, representing a combined 250 MW if both phases elect. The “Selective Capacity Available” status means 5C is still actively signing customers here, which tends to precede the large capacity commitments that would justify a Phase 2 or Phase 3 deposit.

The Full Stack

The Fit Ventures infrastructure build-out as it exists today:

Build the data center, Hypertec Construction

Deploy the servers, Hypertec HPC & AI (CIARA servers, GB200 clusters)

Operate the facility, 5C Data Centers

Power the facility, Fit Energy / FCEL

Every piece of that stack is linked to one family office. The FCEL CEPA is a potential power utility layer of an already-operating AI factory platform with real sites, real customers, real institutional backing, and real hardware partnerships.

Updated Counterparty Risk Assessment

Fit still has to elect each phase and post non-refundable deposits. The warrants expire June 22, 2028 if nothing moves. The structural protections for FCEL remain in place.

This is not financial advice. 9 Ventures holds a position in $FCEL. Always do your own due diligence.

On the surface, this is the kind of contract announcement that moves the tape. There is nuance here that must be discussed. Let me walk through both the bull case and the due diligence notes so you can frame this properly.

Read more

Read more

Every AI chip on the market today is running into the same wall. Raw compute power is no longer the scarce resource. Moving data between chips fast enough to keep the processor fed is.

To understand why this matters, it helps to know how chips are physically assembled. For most of semiconductor history, individual chips were connected using tiny metal wires or solder bumps, miniaturized versions of the connections inside any consumer electronic device. These connections are measured in microns, a millionth of a meter. As AI workloads have grown exponentially in complexity, the number of signals traveling between a processor and its memory has grown with them. Conventional interconnect methods cannot carry enough data fast enough without consuming enormous amounts of power. The gap between what AI silicon demands and what traditional packaging can deliver has become the central engineering problem of this decade.

Thermo-compression bonding, or TCB, is the industry’s current answer. Instead of attaching chips using relatively large solder bumps, TCB uses heat and precisely controlled pressure to join chips through much finer copper pillar structures. The result is a denser, more reliable connection that carries significantly more bandwidth at lower power. TCB is what makes today’s most advanced chip packages possible, including the heterogeneous logic designs inside data center AI accelerators. It is in high-volume production today at the world’s leading foundries, IDMs, and increasingly at the outsourced assembly providers handling packaging at massive scale. TCB is the foundation the industry is building from to make the next generation conceivable.

That next generation is hybrid bonding, and the market has largely decided it is coming soon. Where TCB still uses physical bumps as intermediaries between chip surfaces, hybrid bonding eliminates the bump entirely. Two surfaces are polished to an almost perfectly flat finish and brought into direct contact, bonding copper to copper and oxide to oxide at the atomic level. Interconnect pitches compress from the 40 to 50 micron range achievable with TCB down to below 10 microns.

The performance case is real. Performance and production readiness are two different things, and the gap between them is wider than most investors realize. I sat down with one of the companies at the center of this transition. What they told me changes the timeline and my perspective. The full interview is below.

Read more

The solution the industry is converging on is 800V DC distribution. Instead of fighting the physics of low-voltage high-current delivery, you raise the voltage and lower the current for the long journey from the grid to the rack. Less cable thickness, less energy waste, more compute density per square foot. NVIDIA has basically mandated an 800V DC architecture for next-generation AI factories, and the Open Compute Project (co-authored by Google, Meta, and Microsoft) published the Mt. Diablo (Diablo 400) specification defining ±400VDC bipolar and 800VDC power distribution standards for high-density AI racks scaling from 100kW to 1MW.

This inflection has begun already. The NVIDIA Rubin Ultra NVL576 Kyber rack, arriving mid-2027, will be the first production 800VDC architecture at >600kW per rack. The Feynman generation in 2028 is expected to top 1MW per rack. The infrastructure spend per rack follows the same trajectory: from roughly $36,000 at the GB200 generation to north of $360,000 for Rubin Ultra Kyber.

A 10x expansion in power infrastructure cost per unit of compute deployed.

The semiconductor content inside every 800V power system is where durable margin and some of the highest BOM share increase lives.

Subscribe to 9 Ventures to access the full single-stock analysis, including: the complete the 2026-2030 bear/base/bull revenue and EBITDA model, current datacenter revenue estimates, gross margin assumptions by product line, and our position sizing framework.

Read more

The next trillion-dollar technology deployment cycle is all about intelligent machines that can interact and bring value to the physical world. Humanoid robots assembling battery packs. Autonomous trucks hauling freight across the Southwest without a driver. Industrial vision systems rejecting defective parts at speeds no human inspector could match. NVIDIA’s Jensen Huang called it the “ChatGPT moment for Physical AI” at CES 2026, and for once the hyperbole is underselling it. We are at the transition point where decades of robotics research, AI model training, and precision manufacturing engineering converge into commercial deployments. The first production-scale humanoid robots are being shipped now. The first driverless commercial trucking routes are live and generating recurring revenue now. Tesla Optimus and Figure are gearing to ramp production before the end of 2026. The market is pricing this, but not fully, and not in all the right places.

The hardware backbone of Physical AI is more complex and more defensible than people realize. Physical AI applications need eyes: LiDAR sensors, 3D cameras, and radar chips that can map environments, detect velocity in real time, and resolve the difference between a parked car and a moving pedestrian. They need brains at the edge: ultra-low-power SoCs that run perception and inference without a cloud connection, often on battery power. They need muscles: precision motors, servo drives, strain wave gears, and cycloidal reducers that translate electrical signals into controlled, repeatable joint motion across dozens of degrees of freedom. They need the ability to touch and feel: force and torque sensors that let a robot feel what it’s touching, enabling the dexterous manipulation that makes key tasks in the factory, and more importantly, within the home, commercially viable and SAFE. They need to learn: massive, carefully engineered datasets of real-world physical interactions, egocentric robot perspectives, and affordance-rich scenes that train the world models powering all of it. The supply chains for each of these verticals are long, specialized, and defensible.

Multiple secular trends are compounding at the same time right now. Humanoid robot manufacturers are locking in component supply chains. Autonomous vehicle companies are expanding routes and securing OEM deals that will run through the 2030s. Machine vision is getting layered with AI that turns cameras into intelligent quality inspectors capable of catching defects that have evaded production lines for years. Industrial edge compute is becoming critical infrastructure inside defense platforms, mining equipment, and autonomous construction machinery. The most interesting equity opportunities sit in the components, the sensing stacks, the precision motion systems, and the data infrastructure that powers all of it, not in the consumer-facing robots grabbing the headlines.

The full breakdown of the Physical AI favorites is below.

The Full Stack: 9 Ventures Physical AI Favorites

Read more

The market continues to price FuelCell as a subscale power generation company, even though we know that business is inflecting soon with a 5.5 GW data center pipeline and a 500 MW capacity expansion underway. The carbon capture business, which just shipped its first commercial modules to ExxonMobil’s Rotterdam refinery this month, is priced at roughly zero. The serviceable addressable market for this technology is $380-440 billion in installed equipment. At base case penetration, the annual revenue run-rate by 2040 is in the billions. FuelCell Energy did $158 million in total revenue last year.

The full analysis, including a decade of study results, the competitive picture, the Exxon commercial arrangement, all four revenue levers, and the capacity-constrained model we built from the ground up, is below for paid subscribers.

Read more

Full call notes, full interactive valuation model (Assumptions adjustable), an LCOE analyzer, and my takeaways are paywalled below. If you’ve been following the $FCEL work, this is required reading before the market figures out what’s sitting here.

Read more

The McKinsey humanoid supply chain mapping report is making the rounds. Edge AI & robotics are trending on every FinX account. Thematic investors are scrambling to figure out where the real money is in Physical AI, the part of the AI revolution moving through the physical world in robots, autonomous vehicles, and intelligent machinery, not running on a GPU rack in a data center.

I have worked tirelessly to find where the durable, defensible, asymmetric upside actually lives inside the Physical AI supply chain. Everyone can find the robot manufacturers and the autonomous leaders. The harder question is: who do those leaders depend on that the market hasn’t figured out yet?

I’ve spent the last 12 months studying this trade with the kind of obsessive depth that most people won’t. That means breaking down individual company earnings reports line by line. Pulling and reading full earnings call transcripts, not summaries, the actual transcripts. Listening to live earnings calls. Mapping the humanoid bill of materials component by component to understand where the bottlenecks and the margin pools actually sit. Studying Physical AI applications beyond humanoids: autonomous logistics, precision agriculture, robotic surgery, industrial automation. Reading every industry market report I could find. Building financial models from the ground up and pressure-testing every key assumption.

The conclusion I keep arriving at is the same one. There is one company that sits at the intersection of every meaningful Physical AI application.

In this report, available to paid subscribers, I lay out the full thesis for why this is my highest conviction Physical AI play. That includes a complete valuation analysis of the core business, a revenue model with base and bull case scenarios, a breakdown of the strategic moat and why switching costs are essentially prohibitive, and a critical analysis of the key inputs, catalysts, and risks that every serious investor in this name needs to understand.

If you’ve been looking for the Physical AI play that isn’t already on everyone’s radar, this is the one.

The company is named below the paywall.

If you want the full thesis on this name, and every Physical AI leader I identify before the market figures it out, subscribe to 9 Ventures.

Subscribe now

I’m building out a concentrated Physical AI research portfolio, and this is the first deep-dive of several. Each one gets the same treatment: months of prep, institutional-quality modeling, IR-level diligence, and a clear-eyed view of where the real money sits in the supply chain. No hype. Just high-conviction original research on names the crowd hasn’t found yet.

The Physical AI Supercycle is just getting started. The window to get positioned ahead of the consensus is still open. It won’t be for long.

Read more

In the last 24 months, I have left my fair share of multi-baggers on the table by being cute.

I sold AEHR at $25 from $10 and again at $50 from $33. I sold LITE at $400 from $240 and SNDK at $325 from $200. In every instance, I was early, and in every instance, I missed the meat of the move. These were top conviction positions in my portfolio, yet I let short-term noise dictate my exit.

The biggest lesson I have learned is that patience and deep-rooted conviction is where the alpha is made.

Get the Edge Today

The Art of Sitting Tight

Trading is not easy; the market is designed to shake you out. I remember being a hyper-bull on DOCN at $54 in January/February. I watched it rip to $70, only to see it pull back to $50. That volatility shook me out of the name. It is now trading at $180.

You will not bat 1.000 in this game. I aim to bat over .600. The secret is recognizing that when you hit, you need to let those positions work over multiple quarters. You must know what you own. If you don’t understand the business, the supply chain, the trends ahead, and the inflection points, then volatility will destroy you.

When my winners draw down, I do not worry and obsess over the price action. I re-read my thesis reports. If the fundamental story holds and price is still above a rising 50 SMA, the drawdown is merely noise. It is an opportunity to remind myself why I have conviction in the name.

My Rules of Engagement

I keep my process simple. Complexity creates noise, and noise leads to self-sabotage. My framework rests on four pillars:

1. The 50-Day Rule: I do not own stocks living below their 50-day moving average. When a stock consistently closes below this level, the primary trend is broken. Opportunity cost is paramount, and I would rather rotate into a new leader than hold dead money. While I may continue to track a name after a breakdown, it has no place in my active portfolio.

2. Concentration as a Strength: I allocate to a concentrated portfolio of 5–10 stocks, sized between 5% and 30%. This range allows me to know my companies inside and out. It provides enough diversification for my personal mandate while ensuring that when I have a winner, it is large enough to move the needle.

3. Risk Management at the Entry: My goal is to keep risk tight on every entry. If that means trying 4-5 times to get positioned in a high-conviction idea, that is exactly what I will do. The positioning battle is often difficult, but when the thesis works, the effort becomes well worth it. Managing risk at the point of entry is one of the most important parts of the game.

4. The SMID Cap Edge: My universe tends to live in the $250M - $5B market cap space. Identifying early, major inflections in these businesses can yield special return profiles. I look for companies with the most leverage to emerging trends at strong valuation profiles. These are often businesses that institutions or the sell-side misunderstand, allowing me to position myself before the market catches on.

Get the Edge Today

From Research to Execution

My research process is deep. I map supply chains and identify companies at the intersection of critical growth trends. I want players who can fundamentally reshape the market’s perception of their earnings growth.

When I have conviction, I am not afraid to average up. I have recently done this with PENG and FCEL. As I gather more information, my thesis moves from speculative to high-probability, and I press the gas.

In uptrends, I press the gas on my highest-conviction research. In downtrends, I move to cash or target high-quality shorts that meet resistance against key moving averages. I needed to be more aggressive in identifying that the trend was down earlier this year, but once I recognized the environment, I successfully pivoted to high-quality shorts as they met resistance against key horizontal and moving average levels.

Join the Community

If you want to see how this process looks in real time, consider upgrading to a paid subscription. I walk through my research, my thoughts, and my actual execution in real time for my members. This is the exact same process that has me returning 100% YTD with only shares, even after the drawdown we saw this past Friday.

Get the Edge Today

Stop getting cute. Trust your research. Trust the management and the business to execute. You will have landmines. That is the cost of doing business. If you hold your monster winners through the volatility, those landmines will look minuscule in comparison.

Let price tell you the story, and when the trend is in your favor, have the courage to stay put.

Power demand in the US has hit a level nobody in this industry was prepared for. The culprit is obvious: AI data centers are pulling electricity at a scale that has completely overwhelmed grid planning models built for a different era. Hyperscalers are signing PPAs for gigawatts they need delivered yesterday, and utilities are scrambling to source generation fast enough to keep up.

The solution the market landed on is gas turbines. Not because they’re perfect, but because they’re the only dispatchable generation technology that can be permitted, financed, and built fast enough to matter. Solar and wind are great, but they’re not firm power. Batteries help at the margin. Gas turbines are what you reach for when you need guaranteed megawatts on a timeline measured in months.

The OEMs are overwhelmed. GE Vernova, the dominant player in large gas turbines, entered Q1 2026 with 100GW in combined backlog and slot reservations, up from 83GW just the quarter before. $13B of new backlog added in a single quarter. CEO Strazik has said reservations will be sold out through 2030 by year-end 2026. They’re currently shipping at roughly 20GW/year. They cannot deliver 110GW in two years.

Siemens Energy is running the same playbook at greater scale. Their backlog just hit a new all-time high of €154B with a book-to-bill of 1.72x. In the first half of their fiscal 2026, they sold 179 gas turbines, nearly replicating all of fiscal 2025 in six months.

This is a supply chain event that spans most of the decade.

The Opportunity One Layer Below

Here’s what most people miss. Neither GE Vernova nor Siemens manufactures everything in-house. They depend on a network of qualified precision manufacturers embedded in their production processes, suppliers with the equipment, certifications, and institutional relationships that took decades to build and cannot be replicated quickly. These suppliers are the invisible infrastructure of the gas turbine supercycle.

The full deep-dive, including the specific name, ticker, segment-level analysis, margin progression, capacity utilization data, OEM relationship history, re-rating catalysts, valuation model, and price targets, is below.

NASDAQ: BWEN · Share Price ~$3.8 · Market Cap ~$88M

For informational purposes only. Not financial advice. Do your own due diligence. The author is long this stock.⚠️ LIQUIDITY DISCLOSURE — PLEASE READ BEFORE ACTING

This report covers a small cap security (~$88M market cap) with limited daily liquidity. If you read this research and determine $BWEN warrants a position in your portfolio, please be disciplined in your execution. Do not market buy. Use limit orders. Do not deploy a full position in a single session. Keep individual orders well below 10% of average daily volume. Aggressive buying in a thin market drives your own cost basis up. A patient entry matters as much as the thesis itself.

Broadwind ($BWEN) is a precision manufacturer with a six consecutive record quarters in its highest-margin segment, and 90%+ order growth YoY in its best quarter since 2022. It supplies the gas turbine supply chains of 4 of the top 10 global OEMs. It holds hardware that effectively does not exist anywhere else in the world. It has irreplaceable equipment and nearly a century of engineering expertise.

Broadwind has spent the last two years deliberately exiting a low-margin legacy business to concentrate entirely on the two segments now printing 40%-60% YoY growth.

Management owns a 14% stake in the company. Their variable compensation is 75% tied to earnings growth. They are incentivized to make exactly the decisions they’ve been making.

00 — The Thesis

Power demand in the US has hit a level nobody in this industry was ready for. Gas turbines are by far the most popular behind-the-meter power solution right now, and $GEV and Siemens are already reporting 5-7 year lead times. OEMs are in full frenzy mode, prioritizing timelines and simplicity over everything else.

Broadwind is the overlooked precision manufacturer sitting in the middle of all this. They supply component kits directly embedded in the $GEV/$CAT/$ENR supply chain: precision gearing for gas turbines and outsourced kitting and assembly for natural gas turbine installations. They make rare, highly capable gears that enable higher rotational speeds, and they’re harvesting a weakening Wind Tower segment, cutting operating expenses and redeploying capital into the two segments that have real pricing power.

Combined backlog just hit $73M and growing. Industrial Solutions has set a record six consecutive quarters. Orders were up 90% YoY in their best quarter since 2022. Market cap is $90M.

A $90M micro-cap with a growing backlog, contractual OEM relationships, and irreplaceable manufacturing equipment, attached to one of the most powerful secular themes in public markets.

“While our exit from the Wind market will result in a smaller company over the near term, our remaining businesses are higher-growth, more predictable, and more profitable, with a meaningfully improved quality of earnings profile.”

— Eric Blashford, President & CEO, Q1 2026 Earnings

Check out this monthly base and this ignition volume last month!

01 — Industrial Solutions: Six Straight Record Backlogs

Revenue in Q1 2026 came in at $9.2M, up 64% year-over-year. $14.6M in new orders that quarter alone. Then April booked another $10M+ on top of that. Backlog hit a new record of $43.3M, more than $5M above the prior record set just the quarter before.

To understand why the IS position is defensible, you need one distinction: the “hot gas path” versus the “not hot gas path.” The hot gas path refers to components in direct contact with combustion gases at temperatures exceeding 1,000°C. Turbine blades, rotors, the core. These require exotic nickel superalloys, enormous R&D budgets, and are dominated by a handful of specialists.

Broadwind operates entirely outside this zone. CEO Blashford stated explicitly on the Q1 2026 call: “We tend to support those gas turbine installations on what’s called not hot gas path, but surrounding the hot gas path.” Valves, pumps, piping, heat exchangers, lubrication systems, generators, transformers, wiring harnesses, compressors, instrumentation, controls. Everything the turbine needs to actually run, manufactured from conventional materials at far better economics. A missing fuel valve stops the turbine just as effectively as a missing blade.

What makes IS specifically sticky is the service model wrapped around those components.

Pre-Testing. Broadwind pre-tests sub-assemblies before they leave the factory. For an OEM managing a complex installation timeline in an exponential demand environment, ready-to-install kits aren’t a convenience, they’re a requirement. Pre-testing integrates Broadwind deeper into the OEM workflow with every shipment.

Engineered Packaging and Sequencing. These kits are co-engineered with OEM teams: cable harnesses coiled for point-of-use installation, assemblies labeled for individual project specifications. Switching suppliers means rebuilding that co-engineered system from scratch. Nobody does that in a rush.

Just-In-Time Delivery. By aggregating components from multiple suppliers and holding them centrally, Broadwind lets OEM customers cut their own inventory carrying costs and eliminate stock-out risk. One partner, one purchase order.

“A leading global energy corporation sought a streamlined approach to procurement and supply chain management. It outsourced ALL aspects of its kitting container program to Broadwind Industrial Solutions, reducing storage needs and administrative costs. A single partner with a single purchase order.”

“Quote activity continues to increase in both Gearing and Industrial Solutions, generated by our ability to solve the complex precision manufacturing and sourcing challenges faced by customers in this growing market.”

The margin progression tells the real story:

Margins went from 7.6% to 15.8% to 19.1% in two quarters while revenue was simultaneously growing. Operating leverage alone doesn’t do that. Broadwind cited increased labor costs in the mid-2025 dip, specifically flagging they were hiring ahead of the demand wave. Those costs have since been fully passed through to customers, which is the proof of pricing power, not just scale.

The Q1 2024 spike to 23.7% was attributed directly to “a major increase in demand for new and aftermarket gas turbines.” That’s the potential of this business when the cycle is fully firing. The current $43M backlog is concentrated in this segment. 20%+ IS EBITDA margins are a base case from here, not an optimistic scenario.

“Precision machining market remains very tight from a capacity standpoint as we look into the next several years.” — Eric Blashford, Q4 2025

02 — Gearing: The Technical Moat Nobody Is Talking About

Gearing is where the real moat lives. Brad Foote Gear Works engineers custom gearboxes and precision-machined components for aeroderivative gas turbines, high-torque power generation equipment, oil and gas processing, and aerospace and defense. Applications where failure has no tolerance.

Q3 2025 Gearing orders: $16M, up 260% YoY, the highest quarterly total in years. Inside that number was a $6M follow-on order from a leading gas turbine OEM, described as the minimum first-year commitment on a two-year agreement with potential for additional volume.

Why do 4 of the top 10 global gas turbine OEMs contract a micro-cap in Cicero, Illinois? The answer is irreplaceable hardware.

Brad Foote’s facility houses Gleason 675 Spiral Bevel Machines, two of four that exist worldwide, capable of machining AGMA 2000 Q10 gears up to 2.6 meters (104 inches). Gleason operates in a duopoly in the spiral bevel gear market, and these specific machines are no longer manufactured. A competitor cannot replicate this capability with capital alone. If an OEM needs a spiral bevel gear machined to AGMA Q10 at 104 inches, Broadwind is one of their only options, full stop.

It goes further. Broadwind built the largest climate-controlled grinding room in the U.S., housing 13 form tooth grinders, capable of producing gears to Q15 precision grade, the highest possible standard. Q15 requires climate control to minimize thermal expansion effects during grinding. Broadwind is a physical prerequisite for turbines running at the efficiency levels OEMs are promising their customers.

The GE relationship predates the current power hype by nearly two decades. In April 2008, Brad Foote signed a supply agreement with GE Transportation for wind turbine gearboxes. GE’s VP publicly stated that Brad Foote had “grown into a world class supplier to GE Transportation.” That institutional trust, built over more than 2,700 gear sets produced, is now being deployed on the gas turbine side. Broadwind also co-develops gearboxes with customers, optimizing design, manufacturing components, and performing internal testing before delivery. Structural integration, not transactional supply.

The capacity utilization lever. Management flagged on the Q3 2025 call that the Gearing facility was running at roughly 45% utilization. The business is booking orders at a pace unseen since 2022, and the factory is barely half-full. As utilization pushes toward 70-80%, fixed cost absorption improves dramatically. That lever hasn’t been pulled yet.

The proof is already in the numbers. Gearing went from -$0.3M EBITDA in Q4 2025 to +$0.6M EBITDA in Q1 2026, on a revenue increase of only $1.5M. That inflection on minimal revenue growth illustrates exactly how the fixed-cost math plays out from here. Management has guided for further operating leverage as utilization continues to climb, and a 100-year-old precision machining business with 4 of the top 10 global OEMs under contract is not a business that runs at 45% for long.

03 — Reading the OEM Scoreboard: $GEV and Siemens Are Booked Through 2030

Most people playing the power generation supercycle are buying $GEV or Siemens directly. That’s valid. BWEN is a different animal, a precision manufacturing supplier embedded in the OEM production process itself, one layer below the names everyone knows, at a fraction of the valuation.

GE Vernova. $GEV entered Q1 2026 with combined gas turbine backlog and slot reservations of 100GW, up from 83GW just the prior quarter. $13B of new backlog added in a single quarter. Targeting at least 110GW by year-end 2026. CEO Strazik has said reservations will be sold out through 2030 by end of this year. Currently shipping at roughly 20GW/year. They cannot deliver 110GW in two years. This is a supply chain event spanning most of the decade.

Siemens Energy. Similar story at greater scale. Backlog just hit a new all-time high of €154B with a book-to-bill of 1.72x. In fiscal 2025, Siemens sold 194 gas turbines total. In just the first half of fiscal 2026, they sold 179, nearly the full prior year’s volume in six months.

Neither GEV nor Siemens manufactures everything in-house. They depend on qualified precision manufacturers embedded in their supply chains. BWEN is one of those partners. When GEV adds 17GW to its backlog in a single quarter, BWEN starts receiving orders 6-18 months later. GEV’s backlog growth is a leading indicator for BWEN’s order flow, and that backlog is growing by $13B per quarter.

Three sequential orders illustrate the compounding relationship: Q1 2025 (largest gearing order to date), July 2025 (year one of a two-year LTSA with minimum volume commitment), and March 2026 (year two follow-on, same OEM). That’s a loyal customer under contract, not a transactional vendor relationship.

“Recent sizable orders we received from the power generation sector are the beginning of a multiyear cycle for which we are prepared.” — Eric Blashford, Q4 2025

04 — Book-to-Bill: The Cleanest Leading Indicator in Manufacturing

When a company consistently books more than it ships, backlog grows, forward revenue becomes visible, and revenues accelerate as capacity unlocks. BWEN’s two core segments have been running above 1.5x book-to-bill for multiple consecutive quarters.

Then April happened. $16M+ in new orders in a single month. $10M+ in Industrial Solutions. $6M+ in Gearing. This segment base is currently running at ~$75M annualized revenue. Converting that order rate comes down to manufacturing throughput, which is exactly why the North Carolina expansion and new Gearing equipment are the most important capital allocation decisions BWEN is making right now.

Clean precision manufacturers with similar profiles, sustained 1.5x+ book-to-bill, multi-year OEM agreements, 15-20%+ EBITDA margins, domestic manufacturing, trade at 14-18x EBITDA. BWEN’s implied multiple on core segment run-rate EBITDA is well below that band. When Q3 2026 is the first quarter reporting zero wind revenue, the street has no choice but to re-anchor the multiple to what this business actually is.

05 — The Wind Exit: Surgery, Not Triage

Blashford has been telling the same story consistently for two years. The actions match the words.

Manitowoc, Wisconsin — Sold September 2025. BWEN sold the Manitowoc industrial fabrication facility for $13M total consideration, generating an $8.2M net gain. The facility had generated over $25M in revenue in 2024 at roughly 8-9% EBITDA margin. Decent economics, but not compelling for a precision manufacturer moving upmarket. Management followed the sale with a $3M share repurchase authorization.

Abilene, Texas — Sold April 2026. The decisive move. BWEN sold the Abilene wind tower production facility to IES Infrastructure for up to $19.5M. In 2025, Abilene generated $56.3M in revenue and $9.7M in EBITDA. They walked away from nearly $10M of EBITDA to eliminate a business carrying single-customer concentration risk, federal tax credit dependency with a hard expiration at end of 2027, excess industry capacity creating price erosion, and the structural volatility of the domestic onshore wind market.

“Strategic exit from Wind market reduces exposure to regulatory, legislative, and business risk. In recent years, Broadwind’s wind-related revenue has been largely concentrated with one large wind OEM customer, while wind project economics have been largely dependent on federal tax incentives that will expire by year-end 2027.” — BWEN Press Release, May 2026

Both facilities were purchased by IES Holdings subsidiaries: Manitowoc in September 2025, Abilene in April 2026. A negotiated, structured exit to a single strategic counterparty who evaluated both assets deliberately. Clean.

The Abilene decision is counterintuitive on the surface. You’re selling a profitable business. But when Gearing and Industrial Solutions are growing at 40-60% per year and the remaining business has a century of precision machining expertise now clearing certification pathways into aerospace and defense, holding a fixed-cost wind factory serving one customer at the mercy of federal tax policy isn’t worth owning. Management’s compensation, 75% tied to earnings growth, makes their conviction on this clear.

“Despite the volatile trade policy environment, our 100% domestic manufacturing base remains a key competitive advantage positioning us to partner with tier one OEMs who value our deep technical expertise, commitment to quality, and on-time service.” — Eric Blashford, Q3 2025

06 — The Hidden Tax Asset: $300M in NOL Carryforwards

Almost nothing written about Broadwind mentions this. Broadwind carries nearly $300 million in net operating loss carryforwards, accumulated during years of wind business underperformance. At a 21% corporate tax rate, that’s approximately $63 million in future tax savings. For an $90M market cap company, that’s a material hidden asset sitting right on the balance sheet.

Management’s own press release stated: “Domestic acquisition strategy will seek to capitalize on nearly $300 million of net operating loss carryforwards.”

A company approaching profitability with $300M in NOL protection pays effectively zero federal income tax for the better part of a decade. Every dollar of EBITDA that converts to pre-tax income flows directly to the bottom line without a federal tax haircut. A rough NPV of that tax savings (spread out over a decade), discounted at 10%, adds $25-40M of value to the equity, or roughly $1.10 to $1.75 per share on a fully diluted basis. Most analysts won’t model it separately. It’s real regardless.

07 — Management Alignment: Skin in the Game

The incentive structure is what matters. 75% of variable compensation is tied to earnings growth, not revenue growth, not backlog growth. That alignment explains every capital allocation decision in this thesis: exit low-margin wind, concentrate resources in high-margin Gearing and IS, pursue LTSAs with earnings visibility, and clear certifications into higher-value markets. The remaining 25% tied to working capital management ensures the balance sheet discipline to execute without over-leveraging.

14% combined insider and employee ownership in a $90M micro-cap means the people making every decision in this thesis are personally invested in getting it right.

08 — The Model

Key Assumptions (FY2027): Industrial Solutions expands to $65-78M annualized revenue at 24-28% EBITDA margins. Gearing expands to $48-55M annualized revenue as utilization climbs from 45% to 70%+, producing 16-20% EBITDA margins from the fixed-cost base.

Base Case — FY2027

~$10.8/share at 14x EBITDA + $1.10 to $1.75 per share of PV tax savings = $11.9 - $12.55/share · ~$280M EV · ~3.3x from current levels

Bull Case — FY2027

~$21.4/share at 18x EBITDA + $1.10 to $1.75 per share of PV tax savings = $22.5 - $23.15 · ~$524M EV · ~6.1x from current levels

The 14-18x range is grounded in peer comps for clean pure-play precision manufacturers with 15-20%+ EBITDA margins, double-digit growth, multi-year OEM agreements, and direct exposure to the power generation supercycle. Post the Abilene sale, BWEN reports zero wind revenue. The street will eventually value the business on what remains, and what remains is very good.

Most would rather keep the NOL part separate, but for those interested, this math is how it (could) play out.

Additional Upside Case

A potential defense contract is likely given the DoD has only spent 17% of its newly awarded budget. Much of the budget is to be allocated to vehicles, weaponry, and other infrastructure, all of which require gears and have complex supply chains which could benefit from supply chain simplifiers. A multi-year contract at the least would fruit another $15-25M. These would likely be higher margin, at 30%+ EBITDA margins.

I expect BWEN to continue sustained 20%+ growth through FY30 given the backlog their customers face, implying a higher multiple than I have quoted above.

09 — Risks

Customer Concentration. Growth is driven by a small number of large OEM relationships. Switching costs on both sides provide protection, but the concentration risk is real and worth sizing around.

Execution During Transition. Running two divestitures while simultaneously scaling two fast-growing segments is operationally complex. Management has executed cleanly so far, but complexity compounds.

Gas Power Policy Risk. Any material policy shift slowing gas turbine permitting or penalizing natural gas generation is a headwind to both the OEM customers and BWEN’s order flow.

Micro-Cap Liquidity. $90M market cap means thin daily volume and wide bid-ask spreads. This is a position-sizing and patience game.

Guidance Withdrawal. BWEN pulled 2026 guidance when they announced the Abilene sale. Q2 2026 will be the first truly clean look at the core business without wind noise.

10 — Conclusion

$GEV is sold out through 2030. Siemens is booking 179 gas turbines in six months. The delivery ramp hasn’t even started. And sitting at $90M, with a $73M backlog growing every month, is the precision manufacturer kitting their turbines and machining their gears, with irreplaceable equipment, a 17-year OEM relationship, contractual minimum volume commitments, a defense certification stack being built in parallel, and a 30% expansion of North Carolina floor space coming in 2026.

$300M in tax-free earnings potential. A management team with 14% of the company and 75% of their variable pay tied to earnings growth.

This monthly base might just be the mother of all bases.

For informational purposes only. Not financial advice. Do your own due diligence. May 2026.

What I found gives me even deeper conviction on the holistic thesis: broadband tailwinds, LPO optionality, hyperscaler relationships, deep-rooted European DC ties.

LiteWave800 technology works. The efficiency advantage over DSP incumbents is 12-18x. The advantage over first-generation LPOs is 6-10x. Management committed publicly, while still in prototype, that it will not be dilutive to gross margins. The CEO is telling you the IP is defensible and the cost structure is understood.

In reality, LiteWave800 is only part of the story.

Another key part the market is missing is what sits underneath it. Adtran’s FSP 3000 platform is foundational infrastructure for the European data center interconnect layer that Microsoft, Google, and Amazon depend on. euNetworks, which directly connects 600+ data centers across 17 countries and runs cloud interconnect services for every major hyperscaler in Europe, runs entirely on Adtran’s FSP 3000. Colt Technology, a $2.2B revenue operator connecting 32,000 buildings across 230 cities, has been an ADVA/Adtran customer for over 15 years.

The CTO overseeing LiteWave800 development, Christoph Glingener, is the same person who built those hyperscaler relationships at the optical layer over the last two decades, with named Microsoft Azure architects on record endorsing his work.

There’s also a product nobody’s talking about. The Quattro. A 400G companion module in the same SM-VCSEL family as LiteWave800, launching end of 2026, before LiteWave800 even hits volume production. The Quattro is effectively a production qualification run for the technology that powers everything.

Behind the paywall: the full manufacturing readiness breakdown, the Meiningen Terafactory analysis, the VCSEL supply chain risk, what the gross margin signal actually means, the FSP 3000 hyperscaler flywheel, and the open questions that will drive the next leg of this trade.

This is a top opportunity in the market right now.

Full report below for paid subscribers.

Read more

operates at the critical intersection of high-performance accelerated computing and next-generation memory architecture. That alone is interesting. What makes it a real investment thesis is the strategic pivot management executed away from low-margin hyperscaler hardware contracts.

Here’s the thing about hyperscalers, their procurement scale is a weapon. They use it to grind down margins on hardware integrators. Most importantly, they don’t need the high-touch optimization services that generate real margin. So, Penguin walked away from that business on purpose.

The new focus: enterprise, neocloud, and sovereign AI clients. These customers lack specialized internal engineering departments. They need solution design, end-to-end cluster optimization, and robust cluster management software. That unlocks strong margins that you simply can’t get selling to Amazon or Microsoft.

CEO Kash Shaikhpreviously led Dell’s high-performance computing and AI business. He knows this market better than anyone in the world. Speaking of…

$DELL Earnings: Reading the Tea Leaves for What Comes Next

The following section is free. The full $PENG bull case, financial model, and price target are behind the paywall. If DELL just told you something important — and it did — you’ll want to read what comes next.

On March 16th, when $DELL was trading at $150, I published a post on Substack laying out exactly why management’s guidance for FY27 and FY28 was a sandbag. I said they’d shatter consensus and deliver EPS well above even the most bullish analyst on the street.

They didn’t shatter it. They pulled my FY28 revenue estimates forward into FY27. One year early.

Read the original post →

The Numbers First

📊 $DELL FY27 Q1 AI Server Highlights

AI-Optimized Server Revenue: $16.1B (+757% YoY)

AI Orders Booked: $24.4B | AI Backlog: $51.3B

AI Customer Count: 5,000+ (Up 50%+ in the last six months)

Let that revenue number land for a second. $16.1 billion in AI-optimized server revenue in a single quarter. Up 757% year over year. And management exited Q1 with a record $51.3B in AI backlog with a pipeline that is described as multiples of that backlog, growing across every single vertical: Neocloud, Sovereign, and Enterprise individually.

This is not a hyperscaler-driven story anymore.

The Three Quotes That Matter

Management said a lot on this call. These are the three lines you need to read carefully.

On supply:

“Demand continues to exceed supply with memory as the primary constraint and we expect to exit the year with meaningful backlog.”

Memory. Primary constraint. Not GPUs. Not power. Memory. Penguin Solutions has a strategic supply relationship with SK Hynix.

On what customers actually want:

“Customers are not just buying components, they are looking for integrated solutions they can put into production quickly on infrastructure they control with the performance, security and data foundation their workloads require.”

Components are a commodity. Integrated solutions are a business. There is a company that has built exactly that — a full-stack, ready-to-deploy AI factory platform with proprietary orchestration software and a certified memory solution. Pay attention.

On the pipeline:

“As we look at our pipeline over the next five quarters, that is multiples of our backlog and it is growing across each individual vertical — Neo Clouds individually, sovereign individually and the enterprise space. So it is broad based and it is prevalent either geography wise or vertical wise.”

Five quarters of visibility. Every vertical expanding. If you’re building a position in anything that touches AI infrastructure at the enterprise, neocloud, or sovereign layer — this is your confirmation signal.

What $DELL Is Trading At Right Now

$DELL is now trading at 25x FY27 EPS. That’s the new benchmark. That’s what the market is willing to pay for a company with AI server exposure and supply chain discipline.

Write that multiple down. You’ll need it in the next section.

The Operating Leverage Setup

DELL is conducting a masterclass. Revenue is compounding at record levels. Backlog keeps building. Margins are expanding. And they’re just getting started on the enterprise and sovereign layers.

If you think DELL’s operating leverage story is impressive, you are in for an awakening with what comes next.

The company I’m going to walk you through in the next section is smaller, earlier in its inflection, and positioned at the exact constraint DELL just identified as the primary bottleneck across their entire AI business. It has a direct supply allocation from the world’s leading HBM manufacturer, who happens to be an equity investor. It has a certified, deployed product that solves the memory problem DELL flagged. And it’s trading at a fraction of the multiple DELL just earned.

DELL just told you where the puck is heading. The next section breaks the Penguin opportunity down in detail.

Subscribe to unlock the full $PENG report: financial model, FY27–FY30 EPS scenarios, price target, and the complete bull case.

Read more