Vantix Research

$INIO is one of the most mispriced names in the AI power trade. Posted a deep dive on them a few days ago. Here's everything I uncovered. Everyone knows the general thesis: AI needs power, the grid can't deliver it fast enough, so the money flooded into Bloom, GE Vernova, and the utilities. Meanwhile, the company actually powering some of the biggest AI campuses in America trades like a forgotten European industrial. INNIO makes gas engines (Jenbacher and Waukesha). Line up dozens of them next to a data center, and you have your own power plant. Modular, fast to ship, no waiting years in an interconnection queue. A data center developer doesn't care about the cheapest megawatt. They care about the date. What I found digging through permits, SEC exhibits, and city records: - Amazon's Pennsylvania air permit specifies 280 Jenbacher J620 engines at a single campus. - Oracle and OpenAI's Texas campus runs on Jenbachers through VoltaGrid. When the gas pipeline wasn't ready, Oracle trucked natural gas in to keep the engines running. That's how badly hyperscalers need power now, not in 2030. - Liberty Energy disclosed a $493M engine contract buried in an 8-K exhibit. - Waukesha city planning drawings show nine new engine test cells. That's a second US engine factory being built, and you won't find it in any sell-side note. - On top of that: a 1.25 GW deal with Rehlko, a 1.1 GW data center order, and Aggreko locking up capacity through 2031. All in, I traced roughly 8 GW of Jenbacher data center engines to named sites. INNIO shipped 3.4 GW last year and is building toward 10 GW a year. And every engine sold feeds a services business that runs for decades. Services were about half of revenue last year. Today's equipment sale is tomorrow's annuity. The comps are absurd. $BE Bloom: guiding $3.9-4.2B of revenue this year. Worth ~$74B. $INIO INNIO: guiding $3.8-3.9B of revenue this year. Worth ~$12B. Same revenue. One sixth the price. $FPS Forgent: another 2026 PE-backed AI power IPO. ~$10.4B market cap on ~$1.4B of trailing revenue and negative free cash flow. $INIO: more than double the revenue, ~$700M of trailing free cash flow, and barely a bigger market cap. INNIO is guiding to ~$730M of adjusted EBITDA this year. At ~$16, that's under 20x this year and roughly 15x next year if they deliver. If the stock just sits there, it gets cheaper every single quarter. At this price, the market is implying INNIO's capacity stalls out under 6 GW. They're building for 10. The grid can wait. The data center can't.

$FTAI $UAL FTAI has now named the “major US airline” from 2025, and it’s United. Back in Sep 2025, Adams said FTAI had sold a big batch of V2500s to a major US carrier at slightly dilutive margins, “in the context of looking for a bigger opportunity with that customer, which most probably would involve CFM56 engines.” Yesterday’s release confirms it: the relationship started with United’s V2500 fleet and now extends to designated CFM56-7Bs, United’s 737NG engines, under a multi-year exchange program built to grow over its term. How it works: FTAI delivers a serviceable engine ahead of each scheduled removal, then buys United’s worn engine and adds it to inventory. United skips the shop visit. FTAI gets the work plus the feedstock, and the CFM56-7B is the same engine its Mod-1 power turbine is built on. This is the margin story in action. Aerospace margins fell because FTAI is deliberately chasing Tier 1 airlines with heavier, lower-margin work to win share. United is what that share looks like.

Starting coverage on $FTAI. FTAI owns one of the largest independent pools of CFM56 engines (539 at YE25), the engine on the 737NG and A320ceo, plus the shops to rebuild them in Montreal, Miami, Portugal, and Rome. Instead of an airline waiting 4 to 7 months for an overhaul, FTAI swaps in a rebuilt module or engine in days and fixes the worn one on its own time. Montreal doubled output last year on ~60% more staff. In Dec 2025, it began converting the same engine into a 25 MW gas turbine for data centers. Seven months later, its JV with Jereh signed a $1.465B order, and Jereh has already received the first advance payment. The buyer is undisclosed. Jereh's first US AI customer bought its gensets in late 2025, those gensets later appear on a court-filed equipment list at xAI's Mississippi plant, and that same repeat customer placed this order. The timing is the edge. GE Vernova sits on ~116 GW of turbine backlog against ~20 GW a year of output, and Siemens booked 15 GW last quarter while shipping 6. Big turbines are a 2029+ product. FTAI ships in 2027, at roughly $7.6M of EBITDA per unit built from a core worth ~$1M as parts. The stock is ~48% off its high after Aerospace margins fell from ~35% to ~29%. That's part of management's plan. The old margins came from rebuilding lease engines that were already depreciated on the books, a pool that's running out. FTAI is now buying engines at market and going after the biggest airlines with full restorations, which earn less per job (~25% vs ~42% on light builds) but win share and turn into higher-margin module swaps later. Management called it "consciously going for higher market share." EBITDA dollars are still growing. A few things found along the way: - Engine maker CFM signed a parts, repairs, and thrust-upgrade deal with FTAI - High fuel prices are retiring 737-700s early. FTAI just bought 27 from WestJet, and Southwest has 288 to retire by 2031 - FAA registry filings flagged the WestJet deal 10 days before the press release - Investors in FTAI's aircraft fund are marking it ~10% above cost a year in - VietJet is exploring an FTAI shop in Vietnam plus LEAP and GTF support, the next-gen engines Deep Dive posted on @CVResearch_

$VICR raised Q3 guidance again today! Twice in nine days! Both times on the same license. July: guided to roughly 10% sequential growth. About $158M. Sept 21: raised to more than 20%. Sept 30: raised to more than 30%. More than $186M. Consensus was sitting around $158M to $165M. All of it is royalty. They signed a non-exclusive license to Vertical Power Delivery on Sept 16, and it has added something around $29 million to a single quarter, most of it in a partial one. If product tracks where it was, royalty goes from the $20M range to roughly $49M. Gross margin lands near 61% instead of 54%. That is what a business with no factory and no inventory attached to it looks like when it turns on. The mechanism, for anyone new to this name: Vicor does not sue its customers. It goes to the trade commission and gets the customer's supplier banned at the border. The customer then discovers that its supplier's indemnity clause reimburses legal fees and does not stop product from being seized. Then Vicor sells them a license that lets them keep buying from that same supplier, legally. Vicor collects either way. From the Sept 21 release, and this is the part worth reading twice: "Among OEMs and hyperscalers, four leading companies have secured licenses. Unlicensed hyperscalers first reached out after their computing systems were banned from importation. Since the first patent to VPD was only recently asserted, it is still possible to secure a license at an early stage of escalation with a low royalty rate." Four licensed. Patrizio has said the universe is about nine. The unlicensed hyperscalers are the ones that reached out, and he is publicly telling them the rate goes up from here. Four weeks before the Open Compute Summit, where all of them will be in the same room writing the 800-volt spec.

Cheap robots can't compute their own depth. That's the whole $NVDA angle on $CGNX buying RealSense. Look at the photo below. A robot needs depth, meaning how far away every object is. Either the camera computes it on its own chip, or the robot's main computer computes it on the GPU. NVIDIA's own reference stack, Isaac Perceptor, does it on the GPU with a neural network, fed by passive stereo cameras from Leopard Imaging. RealSense's CMO puts GPU stereo depth at about 100 TOPS per camera pair. On a Jetson Thor, that's 5% of the chip, basically free. On an AGX Orin, it's 36%. On the new Jetson Orin Nano 2, NVIDIA's entry-level robot computer shipping 1H27, it's 128%. The depth math alone is bigger than the whole computer, before the robot runs a single AI model. It gets worse with better cameras. Double the resolution and depth noise halves, but the matching work goes up about 8x: 4x the pixels, each searching twice as far. Better depth eats compute exponentially, and most mobile robots run 4 to 6 cameras. So on the cheap robots, which are the volume, depth has to be computed on the camera. RealSense has done exactly that since 2018, on its own chip. The new V5 chip also adds about 5 TOPS of on-camera AI. Now look at the timeline. Aug 25: NVIDIA names Cognex among the first adopters of Orin Nano 2. Moschner posts about combining Cognex's vision software with NVIDIA's edge AI platform. Sep 22: Cognex announces RealSense on day one of ROSCon, the biggest robot-software developer event of the year. That same morning, NVIDIA releases Isaac ROS 5.0 and credits RealSense twice: once for sponsoring AgenticROS, and once for optimizing the D585 Pro for Isaac ROS and Jetson Thor. 1H27: the D585 Pro and the Orin Nano 2 ship in the same window. On the deal call, Moschner said the D585 lets Cognex run its vision software directly on the camera, and called it a "huge leap" for the segment. So picture the stack. A RealSense camera computes depth for free. Cognex software handles inspection, barcodes, and OCR. The Jetson handles navigation, mapping, and language models. Orbbec has cameras but no industrial AI. NVIDIA has compute but no cameras and no factory-grade vision tools. Nobody else sells all three. A few more details make it tighter. Cognex's new In-Sight 3900 runs on Qualcomm, not NVIDIA, so Cognex is making its AI software portable across chips rather than betting on one vendor. RealSense's newest software features already assume an NVIDIA computer sitting next to the camera: its visual odometry is built on NVIDIA's cuVSLAM. RealSense has shown 8 of its cameras running on a single Jetson. And Leopard Imaging, the company that builds NVIDIA's own reference stereo cameras, also resells RealSense. Then there's Matic, one of NVIDIA's named Orin Nano 2 launch adopters. It navigates home robots with five plain color cameras and neural-network depth, no depth camera at all. NVIDIA's own navigation docs list where that approach breaks down: glass, featureless hallways, thin obstacles, vibration, changing light. Those are exactly the conditions an active infrared projector is built for. Consumer robots will lean on GPU depth. Industrial, safety-critical, and compute-limited robots will keep depth on the camera, and those are the robots Cognex already sells into. The other named adopter is Doosan Bobcat. It builds compact construction equipment, and it's reachable through RealSense's Korean channel partner, which has sold into Korean industrials since 2001. The one thing RealSense has to fix is speed. It moved one chip generation in eight years: the D4 in 2018, the V5 in 2026. Jetson roughly doubles AI throughput every 14 months. To keep the camera ahead of the GPU on cost-constrained robots, it needs a V6 in about 3 years, not 8. Intel never prioritized that. Cognex has $755M of cash and ~$270M of free cash flow a year to fund it. What's it worth? If entry-level Jetson robots hit 1.5M units a year by 2030 and RealSense takes 20% of that socket, at about $200 of camera plus $200 of Cognex software per robot, that's roughly $120M a year. That's more than RealSense's entire revenue today, and half of it is software at Cognex-level margins. Did NVIDIA know the deal was coming? No idea. What's clear is that the entry-level robot boom needs a camera that computes its own depth, and Cognex now owns the leading one.

Never posted about $RRX before. It's at $149 today, down 37% from June highs, trading around 9.6x EV/EBITDA. Finally cheap enough to start eyeing it. Regal Rexnord makes motion. Motors, bearings, gearboxes, couplings, brakes, actuators. Boring stuff until you see where it ends up. What they actually own: Thirteen plants certified to AS9100, which is the quality standard you need to sell into aerospace. Nine are in the US. That certification takes years to earn and locks you in once you have it. They're on all four moving parts of a guided missile. The fins that steer it, the seeker head that finds the target, the nozzle that vectors thrust, and the mechanism that deploys the wings. The fin actuator uses all seven of their product lines. Missile market compounds toward $77B by 2028, and every shot fired is a replacement order. Kollmorgen, their motor brand, sits on 711 approved-supplier lists with 1,080 federal awards, shipping gimbal motors into $RTX’s Tucson missile campus. Once you're qualified on a flight-critical program, swapping you out means the customer pays to re-certify. Aerospace bearings have been on backorder since 2023. They built a new certified plant in Chihuahua for $13.8M just to catch up on orders they couldn't fill. Marine propulsion approved by nine ship classification societies, the bodies that sign off on anything going into a hull. Twiflex, Stromag, Svendborg, Jaure. Offshore patrol vessels, icebreakers, LNG carriers. Surgical robot motion through Kollmorgen and Portescap. Management calls medical the highest-margin business in the segment. Radford, Cleveland, and Simi Valley are all running three shifts right now, and they're hiring a VP of continuous improvement. Plus, a billion in debt gets paid off by end-2027, which is roughly $15 a share before anything else happens. The knock is margins haven't converted yet. Their best motors need samarium-cobalt magnets, and China's export suspension expires November 10. That's the thing to watch. But you're buying an aerospace, defense, marine, and medical motion franchise at industrial multiples, and getting the data center business thrown in.

$CRDO Two weeks ago I named the one hole in this thesis. Credo was selling the light engine while someone else still sold the brain sitting next to it. Today they closed it. Why it matters: An optical module has two expensive parts. The photonics chip that turns electricity into light, and the DSP that cleans the signal up so it survives the fiber. Credo has always been strong at the second. It paid $770M for DustPhotonics in May to get the first. Until now, it sold them separately, and its two biggest photonics wins went to customers who bought someone else's DSP. That is roughly $70 of a module instead of roughly $200. The new 1.6T ZeroFlap transceiver puts both Credo parts in one box. Their words: the first product to combine them. They did not wait for a customer to buy the pair. They built the pair. Under four months from closing the deal to shipping the integrated part. And the same Kfir200 photonics chip carries straight over to the co-packaged optics roadmap in FY28, so one design is serving two markets. This is also the product that has to carry a $600M optical year that is almost entirely back-half weighted. I still want a named customer on it. But the gap I pointed at is gone.

$FPS guided FY27 to $2.4-$2.6 billion of revenue, up 76%. Adjusted EBITDA $575 to $625 million, up 86%. Adjusted EPS $1.26 to $1.40, up 95%. That is the second straight year of roughly doubling the company. Q4 revenue was $462M, up 94%. Adjusted EBITDA $113M at a 24.4% margin, up 640bps year over year. Adjusted net income $77M, up 275%. All three came in above the high end of guidance they had already raised in May. Orders were $1.5 billion in the quarter. That is more than the company's entire revenue for the full fiscal year. Backlog went to $3.0 billion, up 256%. Book-to-bill went 1.3x, then 2.3x, then 3.3x across three quarters while revenue doubled. The denominator got harder every quarter, and the ratio still climbed two full turns. Here is the math that matters: Orders = monthly demand × months of lead time At a 10.5-month lead time, Forgent was booking $381M a quarter, or $36M a month of underlying demand. At 13.5 months, they booked $815M, or $60M a month. Q4's $1.5 billion at 13.5 months is $111M a month. Stretch the assumption to an 18-month lead time, and it is still $83M. Underlying demand is up somewhere between 38% and 84% no matter which assumption you use. This is not customers ordering earlier. The book also got shorter while it got bigger. In March, 55-60% of backlog was scheduled to ship in FY27. Today, backlog covers more than 90% of a $2.5 billion guide, so at least $2.25B of the $3.0B ships inside twelve months. Conversion went from roughly 57% to roughly 74%. Customers are now paying to hold slots. Deferred revenue went from $133M to $264M in one quarter. That is $130M of new cash against $1.0B of new backlog, 13% down on the marginal order, against 2% over the prior nine months. Every dollar of that backlog is a firm purchase order. No LOIs, no MOUs. Pricing is flat with no erosion, and lead times are still inside industry-expected across every product line. Eighteen months ago, Forgent sold point products through EPCs and OEMs. Since then, they added regional colos, then national and international colos, then neoclouds. In Q4, they took their first direct order from a frontier AI lab and signed an MSA with a hyperscaler. Hyperscalers and frontier labs are roughly zero percent of revenue today. The frontier lab order is MV transformers and switchgear for a campus larger than one gigawatt, built at Waco, delivered on about a six-month lead time. The industry quotes 24-36 months for large transformers. It is the initial award on a program measured in multiple gigawatts. Management is in active negotiation with several customers on multi-year volume commitments rather than one-off POs. Powertrain Solutions revenue went $51M, $99M, $147M in consecutive quarters. Up 187% year over year and nearly tripled in six months. It is 32% of Q4 revenue and 40% of backlog. A data center buys six to seven dollars of low voltage gear for every dollar of medium voltage. Forgent owned the hard dollar and skipped the easy six. Powertrain is them going back for it: same customer, same project, several times the revenue. The margin on their own content inside an eHouse sits within a standard deviation of a point product, so wallet share multiplies without giving up rate. Modular was 10-20% of data center construction in the cloud era. It is about 40% today, and third-party work puts it at 60% by 2029. It compresses build schedules by 30-50%. Forgent just committed $35M to a dedicated 385,000-square-foot Powertrain plant, 80% the size of their largest facility, taking total revenue capacity to $5.8 billion. The share gain is measurable. Data center revenue up 161% against a market up 37%. Grid up 69% against a market up 10%. Customers they served a year or two ago are spending 30 to 100% more with them now. And the cash finally arrives. Operating cash flow went $45M to $109M, and is guided above $300M for FY27. Capex was 11% of sales, then 8%, and steps to about 3% next year, including the new plant. First real free cash flow year in the company's public life. First use of it is M&A. Two things to watch. Beginning in FY27, they report orders and backlog annually instead of quarterly, replaced by quarterly revenue and EBITDA guidance, so there is less to check them with between year-ends. And Q1 margin guides to roughly 21% because they are hiring ahead of the ramp again, which is exactly what they did last year. At ~$32, the enterprise value is about $10.2 billion, or 17x the midpoint of the EBITDA they just guided, on 76% revenue growth. Amazing quarter, much more growth to come!

$FPS “To put that in perspective, fourth quarter revenue growth exceeded our full year growth rate by approximately 500 basis points. We also booked more than $1.5 billion of orders in the quarter alone.”

$FPS reports tomorrow morning before market open! About 20% of the March-end backlog was already scheduled to ship in F4Q. That revenue is close to booked. Morgan Stanley put it plainly in its Sunday preview: the positive surprise, strong Q4 orders, has largely already been communicated, while the negative surprise would come through in the conversion pace and margins embedded in the FY27 guide. So the event is the first FY27 guide Forgent has ever given. On the quarter, consensus organic growth of 84.3% implies about $1,388M of FY26 revenue against a guide of $1,350-1,390M. Back out the nine months already reported, and the Street is modeling a Q4 of roughly $430M against a guide of $392-432M. On FY27, consensus sits near $2.09B of revenue and $490M of EBITDA. Morgan Stanley expects the company to guide roughly $2.1B with about 100bps of margin expansion to a 24% EBITDA margin, which lands in line. But MS also reports what its own clients are carrying: $2.4B of sales and roughly $600M of FY27 EBITDA. That whisper is about 22% above the published number. It is also above what the plants can make. Forgent's prospectus says the expansion enables it to more than triple fiscal 2025 production volume by the end of calendar 2026. Triple $753M is about $2.26B. The $5B figure everyone looks at is the eventual footprint ceiling, not a FY27 capability. Consensus at $2.09B already runs the system near 90% of stated capability. So why does a strong print matter here? Because the multiple already did the damage. TD Cowen called this expensive on June 22, at 37x blended forward EV/EBITDA against Vertiv at 30x, and defended it on a growth-adjusted basis at 0.88 versus Vertiv's 1.16. At today's price, the blended forward multiple is roughly 18x and the growth-adjusted figure is about 0.43. Less than half what it was, and less than half the next cheapest name in that same peer chart. Morgan Stanley's bear case is $25 on 17x a bear EBITDA of $455M. We are 17% above that. We are also 42% below the lowest published target on the Street, which is Morgan Stanley's own $51, and there is nothing between here and there. If management guides FY27 at or above $2.1B with a margin bridge that holds, this is cheap against its own peer group for the first time since it listed, with a lot of empty air overhead. That is where the steam comes from. So here is where I land going in. Under @CVResearch_, we wrote in July that the walls decide it. Tuesday decides something narrower. It decides whether the number is real. A clean Q4 and an FY27 guide at or above $2.1B, with margins near 24% and a conversion schedule that does not slip, and this reprices quickly because the compression already happened. I'll share my thoughts after the results and the earnings call!

New newsletter out now, completely free moving forward!

$KGS - the AI power trade nobody is calling an AI power trade Everyone's crowded into the obvious data center energy names. Haven't seen anyone speak on this one. Start with what you're actually buying today. The core compression business is one of the best-kept secrets in energy infrastructure: 98.2% fleet utilization 70% gross margins (all-time high), up from 64% two years ago 60% investment-grade counterparties ~5x build multiples, ~19% IRRs 84% of units have never been moved off their original site And the durability just stepped up in a way the market completely missed. In Q2 alone, backlog went from $1.63B to $2.64B. Weighted contract life doubled. The revenue contracted beyond 2030 went from $11M to $541M (50x) on two 10-year renewals with top-10 customers. Nobody even mentioned it on the most recent earnings call. Meanwhile, $CAT lead times have gone from 53 weeks to >180 weeks. Kodiak's competitors can't get engines. Their customers can't either. That's a supply-side moat, and it's why pricing keeps climbing. Now the part the market hasn't priced. Kodiak bought Distributed Power Solutions at 7.4x, secured up to 1.8 GW from Baker Hughes, and is targeting 2 GW of behind-the-meter power by 2030. Look at where they already are. Midland: new 22-acre, 140,000 sq ft campus Odessa: the DPS yard, 20 miles away Pecos: 9-acre facility opened in December Carlsbad, NM: satellite office Now look at the announced data center pipeline in those exact counties: Lea County, NM: 7,000 MW Pecos County: 2,000 MW Ector County: 600+ MW ~10 GW of announced load sitting directly inside a service corridor Kodiak already staffs, with 3.1M horsepower of compression, hundreds of Caterpillar-certified technicians, and existing relationships with every producer in the basin. Add the setup: ERCOT's interconnection queue is 474 GW against a grid that peaks near 90. Abbott froze it in August. Every project that can't wait 6-10 years for a wire goes behind the meter. Kodiak sells the compression that moves the gas, and now owns the generation that burns it. Stock is more than 20% off its May high, trading around 10x EBITDA. The Chairman bought 6,000 shares at $54.81 in March. Another director bought 5,000 at $54.75 (his first purchase ever). EQT's overhang is fully cleared. You're paying compression multiples for a compression business that's getting better, and getting the power optionality close to free.