“Latency is a big issue, right? workloads like robotics, you need to be close to where the factory is.”

1

The Edge AI Trade Series is a collection of articles on individual companies that I believe could benefit from the growth of edge AI over the coming years. The goal is to look beyond the obvious mega-cap AI winners and focus on smaller, more specialized companies positioned around inference, distributed compute, power, connectivity, and edge infrastructure.

If you enjoy this kind of research, consider supporting the publication with a paid subscription. It helps me keep doing deeper dives into underfollowed companies, especially small-cap names that often sit outside mainstream AI coverage.

Latency

Revolutionary technologies rarely create value exactly where people expect. In the internet’s early years, few people would have predicted that a social network would become one of the world’s largest companies.

One thing, however, is becoming clearer: more AI inference is moving toward the edge.

Not all AI will move there. Training and many batch workloads will remain in giant centralized data centers. But applications that depend on real-time responses, bandwidth, data locality, or privacy can benefit from placing compute closer to users, devices, and data.

Latency is the time between sending a request and receiving a response. Faster chips can reduce processing time, but they cannot erase physical distance. Data still has to travel through fiber, and the speed of light remains undefeated.

That matters for real-time gaming, voice agents, video analytics, personalized media, robotics, autonomous fleet coordination, and multi-agent systems. Latency already matters far beyond large language models, and the list of use cases is growing. I believe we are still near the bottom of the S-curve for edge AI adoption.

Akamai

In my opinion, Akamai Technologies is undergoing a profound business-model transformation and is well positioned to play an important role in edge AI.

Today, Akamai is primarily a cybersecurity company. Security represented roughly 55% of revenue in the latest quarter. The company protects websites, applications, APIs, networks, and users from cyberattacks while also operating a large content-delivery network and a growing cloud infrastructure platform. Delivery and other cloud applications represented about 36% of revenue, while Cloud Infrastructure Services, or CIS, represented roughly 9%.

Over the last four reported quarters, total revenue grew in the mid-single digits. The mix improved underneath the surface: Security remained the largest and most dependable growth engine, Delivery continued to decline but stabilized, and CIS became the clear source of acceleration.

That transition is pressuring profitability. Akamai produced non-GAAP operating margins of roughly 29% to 31% through 2025, but margin fell to 26% in Q1 2026—you will soon find out why. Akamai ended the quarter with about $1.7 billion in cash, equivalents, and marketable securities, but it also carries roughly $4.1 billion of convertible debt. This is a profitable company with access to capital, not a pristine net-cash software business.

Where Is the Edge?

Enough beating around the bush. On the surface, AKAM looks boring: mid-single-digit growth, falling margins, and most of its revenue tied to cybersecurity and network delivery.

The real story is its smallest segment.

Cloud Infrastructure Services

Akamai’s Cloud Infrastructure Services is its distributed cloud platform. It includes compute, storage, networking, cloud-native services, serverless EdgeWorkers, managed containers, infrastructure as a service, and now AI inference capacity.

The idea is simple: do not send every workload back to a distant hyperscale data center. Move the necessary compute closer to the user, device, application, or data source.

Akamai’s advantage is that much of the foundation already exists. Its network spans more than 4,300 edge locations across roughly 700 cities and more than 130 countries. That platform already carries delivery and security traffic. Akamai is now layering distributed compute and AI inference onto it.

The architecture is tiered. Serverless functions can run across the broad edge footprint. Managed containers are active in well over 100 cities. Full infrastructure-as-a-service capabilities are available in several dozen cities, while GPU-equipped inference capacity is concentrated in a couple dozen larger locations.

The easiest way to picture this is as a hierarchy. Lightweight, inexpensive tasks can run on CPUs close to the user. More complex workloads can move to managed containers or larger compute regions. The most demanding inference jobs can be routed to GPU clusters. Akamai’s orchestration layer is designed to choose the most appropriate resource rather than wasting premium GPU capacity on every request.

That creates a wide range of use cases. A retailer can personalize a shopping experience in real time. A media company can process or generate video closer to viewers. A voice application can reduce the awkward delay between question and answer. An enterprise can place an agent, API, or inference service closer to employees and customers.

Akamai did not accidentally build an inference network. It deliberately built a global delivery and security platform over decades. What has changed is that the same distribution advantage is now valuable for AI compute.

The proof is starting to appear in the numbers. CIS growth accelerated from about 30% in Q2 2025 to 39% in Q3 and 45% in Q4, before growing 40% again in Q1 2026. Management now expects at least 50% constant-currency CIS growth for the full year.

The contracts are even more striking. In Q4 2025, Akamai announced a four-year minimum commitment of about $200 million from a major U.S. technology company, with most of the spend tied to Akamai Inference Cloud. In Q1 2026, it announced a seven-year, $1.8 billion CIS commitment from a leading frontier-model provider, the largest customer agreement in Akamai’s history.

For scale, CIS generated about $95 million in Q1 2026. If the $1.8 billion contract were spread evenly across seven years, it would average roughly $257 million annually, equal to about 68% of Q1’s annualized CIS revenue. That is only a simple comparison, not the actual revenue schedule, but it shows the magnitude of the win.

Management says the broader GPU pipeline exceeds both existing and projected inventory. That is bullish for demand, but it also means more large wins will probably require more capital.

Akamai Is Not a Conventional Neocloud

Akamai is not trying to build a handful of gigawatt-scale AI factories. Its model is distributed and modular.

Management has described its larger core compute deployments as roughly 5 to 10 megawatts (hey, I’ve seen this movie before), expandable to 20 to 30 megawatts. These are meaningful facilities, but they are much smaller than the giant centralized campuses pursued by hyperscalers and neoclouds.

That reflects a different job. Neoclouds concentrate enormous GPU clusters for training and large-scale inference. Akamai is optimized for workloads where latency, bandwidth, egress cost2, geographic distribution, reliability, and security matter.

The two models can compete, but they can also complement each other. Management has said the three largest U.S. cloud providers already use Akamai CIS for selected distributed workloads. They chose Akamai because its network can place application logic closer to users than their own centralized regions.

Beyond Raw Compute: Orchestration, Security, and Token Economics

Akamai’s opportunity extends beyond renting CPUs and GPUs.

Its AI Grid orchestration layer is designed to route model requests to the right resource, apply semantic caching3, and reserve premium GPUs for workloads that actually need them. Enterprises do not want every simple task sent to the largest and most expensive model. Better routing can reduce latency, infrastructure cost, and unnecessary token consumption.

Akamai can also wrap those workloads in security. Its platform includes API security, bot management, Firewall for AI, rate limits, quotas, and model-aware defenses. The same network that processes an agent request can help determine whether that request is legitimate, malicious, or outside policy.

Its partnership with Visa offers early validation of this role. Visa’s Trusted Agent Protocol is designed to authenticate AI shopping agents and communicate who they represent and what they intend to do. Akamai brings user recognition, bot detection, and threat intelligence at the edge, helping merchants distinguish legitimate agents from malicious automation before those requests reach checkout or backend systems.

This partnership is not yet proof of material revenue, but it points toward a plausible future. Consumers may eventually delegate travel booking, shopping, and payments to several agents working in parallel. Those agents will need identity, permissions, security, policy enforcement, and fast access to compute. The edge becomes valuable when decisions must be made quickly and close to the user.

Risks

The financial tension is obvious: Akamai’s fastest-growing business is also its most capital-intensive.

Cloud infrastructure requires colocation space, servers, GPUs, networking equipment, power, depreciation, and more sales and support investment. After the $1.8 billion contract, management raised expected 2026 capital expenditures to roughly 40% to 42% of revenue. Akamai expects to spend about $800 million to $825 million over 12 months to support that customer alone, with roughly $700 million falling in the second half of 2026.

The cash goes out before the revenue fully arrives. Management expects only about $20 million to $25 million of revenue from the new contract in Q4 2026, with a larger ramp in 2027. It also warned that additional GPU orders for the wider pipeline are not included in the current CapEx guide.

That creates several risks. Customer ramps could be delayed. GPU pricing could fall. Hardware could become obsolete faster than expected. Colocation and memory costs could rise. A handful of very large contracts could create customer concentration. Akamai must prove that the revenue and lifetime cash flow generated by these assets justify the upfront investment.

Margins will also remain under pressure. Management has said some large dedicated-capacity contracts may operate below Akamai’s historical 30% operating-margin level. On-demand GPU services can carry better economics, but they provide less guaranteed utilization. The trade-off is predictable contracted revenue versus potentially higher but less certain returns.

I do not see evidence that management is preparing for an immediate common-equity raise however. Akamai says it can fund the current buildout using its cash generation, existing liquidity, and $1 billion credit facility, while retaining access to debt markets.

Still, capital is not free. Akamai already has roughly $4.1 billion of convertible notes outstanding. More large deals could eventually require additional debt or convertible financing. Heavy CapEx may also leave less room for buybacks, which management uses to offset dilution from employee equity compensation.

Akamai’s share count is roughly flat to modestly lower over the past year because repurchases have offset most equity issuance. That is better than unchecked dilution, but the buyback cushion may become less dependable during the buildout.

Management Follow-Through

Management’s recent record has been solid.

Over the last four quarters, Akamai generally met or exceeded its near-term guidance. Q3 and Q4 revenue came in above the guided ranges, margins held up better than expected, and earnings per share beat. Q1 2026 landed within guidance. More importantly, the directional calls were right: CIS accelerated, Delivery stabilized, API Security remained strong, and AI inference demand converted into major contracts.

The next test is more demanding. Management now has to convert enormous capital spending into deployed capacity, recognized revenue, acceptable margins, and durable free cash flow. Signing the deals was validation. Earning an attractive return on them will be execution.

Valuation

There is no perfect public comparable for Akamai. The company combines a mature cybersecurity platform, a declining delivery business, conventional cloud services, and a fast-growing GPU-backed edge inference operation.

The closest strategic peers are Cloudflare and Fastly .

Cloudflare combines security, delivery, serverless compute, storage, AI Gateway, Workers AI, and agent-development tools on one global network. It is more software-like and developer-centric than Akamai, with less emphasis so far on large dedicated GPU clusters and customer-specific infrastructure commitments.

Fastly is the closest smaller-cap operating peer. It offers programmable edge compute, delivery, API security, bot and agent traffic management, and semantic caching. However, it is not yet an apples-to-apples GPU infrastructure peer. Fastly currently has more exposure to accelerating, caching, orchestrating, and securing AI traffic than to hosting large inference clusters directly.

I own both Akamai and Fastly. I like Fasly as the higher-torque small-cap expression of the theme. Akamai is my preferred risk-adjusted bet because its addressable opportunity is broader and its demand validation is stronger.

Akamai currently trades at about 19.7 times the midpoint of 2026 non-GAAP EPS guidance. Cloudflare traded at roughly 191 times its guidance midpoint, while Fastly traded near 62 times. These are non-GAAP forward multiples.

For this peer set, forward enterprise value to revenue is more useful because profitability differs so widely. Akamai traded at roughly 5 times 2026 revenue, compared with about 28 for Cloudflare and roughly 4 times for Fastly.

Cloudflare gets a substantial premium because it grew revenue 34% in Q1 2026, has a powerful developer ecosystem, and operates a more software-like model. But I believe the valuation gap becomes difficult to justify if Akamai’s CIS business continues to grow near 50%, turns its large commitments into revenue, and pushes consolidated growth into double digits.

The Fastly comparison cuts the other way. Fastly grew 20% in Q1 and guides to roughly 15% full-year growth, but its Compute and Observability category generated only $8 million in the quarter. Akamai’s CIS business generated about $95 million. Fastly also remains GAAP unprofitable and continues to issue meaningful equity compensation. Akamai is already substantially profitable and deserves a premium to Fastly’s revenue multiple.

Akamai’s screened PEG ratio is around 2.2, which does not scream cheap. Near-term earnings are being deliberately depressed by colocation, depreciation, infrastructure investment, and go-to-market spending. PEG can become artificially unattractive during an investment cycle, just as optimistic long-term growth assumptions can make it look artificially cheap.

The thesis is not that Akamai is cheap because one ratio is low. It is that the market may still be valuing the company mainly as a mature security and delivery provider while underestimating how large CIS could become.

The stock has already rerated, and current investors are paying for some success. The valuation becomes attractive if margin and free-cash-flow pressure prove temporary and the buildout creates durable contracted revenue rather than permanently lower returns on capital.

The Next Layer of the AI Trade

The AI trade has broadened in waves.

It began with GPUs. Then investors moved outward to power equipment, memory, storage, neoclouds, colocation, optical networking, and semiconductor inspection and test equipment makers. As AI infrastructure matures, the market searches for the next constraints or the next underappreciated layer.

Distributed inference, specialized edge infrastructure, agent orchestration, AI security, and token optimization could be part of the next layer.

The market has already begun to reward robotics, physical AI, and edge-silicon stories. The next step may be recognizing the companies that can run, route, secure, and monetize inference closer to the user.

That is where Akamai sits.

It already owns one of the world’s most distributed networks. It has enterprise customers, security, delivery, compute, orchestration, and now large AI infrastructure commitments. If edge inference becomes a major investment theme, Akamai is one of the few public companies that can quickly become an obvious, recognizable way to play it.

That does not guarantee multiple expansion. Execution still matters, and the CapEx burden is real. But if Akamai converts its current pipeline into revenue while preserving attractive long-term economics, the market may eventually stop seeing a mature CDN and start seeing a distributed AI infrastructure platform.

Disclosure: I own shares of AKAM and FSLY. This article reflects my personal opinion and is not financial advice, investment advice, or a recommendation to buy, sell, or hold any security. I am not a financial advisor. I may change my view or position at any time without notice. The estimates below rely on company disclosures, earnings-call commentary, conference transcripts, and my own assumptions, which may be wrong. Please do your own due diligence.

“I’ll go up against any of them (competitors) every day, and if you were part of my sales organization, I have one trick. Here’s my trick. When we go to sell somebody, like this is how we won our first hyper, I said, ‘Look, I’ll pay for you to go see their pod. I’ll fly you there. We’ll take a tour together, and then we’ll go see mine. If there’s no pod to see, you have to sign with me.’ I haven’t lost yet.”

1

The Edge AI Trade Series is a collection of articles on individual companies that I believe could benefit from the growth of edge AI over the coming years. The goal is to look beyond the obvious mega-cap AI winners and focus on smaller, more specialized companies positioned around inference, distributed compute, power, connectivity, and edge infrastructure.

If you enjoy this kind of research, consider supporting the publication with a paid subscription. It helps me keep doing deeper dives into underfollowed companies, especially small-cap names that often sit outside mainstream AI coverage.

The central thesis

My core thesis on is that the company has found an interesting niche inside the severely constrained AI infrastructure ecosystem. Based on management’s execution so far, CEO Doug Recker’s comments, and the company’s recent strategic shift, I believe Duos is making a credible transition from a legacy rail AI inspection company into a modular edge data center operator.

That transition came at a cost, especially dilution. But Recker’s background matters here. He has over 30 years of telecom and data-center experience, and he founded EdgePresence, which was essentially an earlier version of what Duos via Duos Edge AI is now trying to become: modular, purpose-built edge data centers with power, cooling, monitoring, physical security, and low-latency connectivity. So they have a proven, execution history, and deployment.

The opportunity is straightforward. Companies are spending billions on GPUs and compute, but grid infrastructure, labor, materials, permits, and traditional data-center construction are not moving fast enough. Duos’s Edge AI strategy aims to fill part of that gap by finding smaller markets where power is available or underused, then rapidly deploying modular edge data centers in those locations.

In simple terms, Duos is trying to become a powered-site optimizer for AI infrastructure. Instead of competing with giant hyperscale campuses, it is targeting regional and underserved markets where megawatts of capacity may be sitting idle or underutilized. These modular EDCs can support AI inference, distributed compute, and high-power colocation, while generating recurring revenue from infrastructure services.

I believe the high-power colocation model is attractive on its own. But as edge AI grows, Duos’ know-how could become more strategic. More inference workloads should move closer to where data is created and consumed: industrial automation, robotics, cloud gaming, traffic monitoring, surveillance, V2X, healthcare, and other latency-sensitive applications. If that happens, small modular data centers near power, fiber, and end users could become an important part of the AI infrastructure stack.

The Hydra Host GPUaaS partnership and Technology Solutions

Around this core strategy, Duos has several complementary businesses. The Hydra Host GPU-as-a-Service partnership gives Duos a near-term way to monetize high-density GPU infrastructure while proving it can execute on its broader Edge AI mission. The good news is that the financing side of this expensive project has been largely derisked by USD.AI’s $98.1M facility. Now the question is execution: can Duos deploy the infrastructure successfully, and can Hydra bring customers onto the GPUaaS platform?

Duos also has its Technology Solutions segment, which provides procurement, sourcing, logistics, and deployment support for data-center equipment. Together, these pieces make Duos a speculative but increasingly tangible AI infrastructure story. I like the Technology Solutions segment because it is bringing meaningful revenue this year and it also shows Duos prowess in executing a data-centre equipment project procurement and deployment. So the evidence here is clear to me, they have already demonstrated the ability to do what they are saying they are going to.

Financial picture

Small detour: New APR

Before getting into Duos’ numbers, New APR needs context. In late 2024, a Fortress affiliate formed New APR Energy to acquire roughly 850 MW of mobile gas-turbine power assets, essentially equipment used to provide fast temporary power. Duos helped manage those assets through an asset-management agreement and received a 5% non-voting stake in New APR’s parent company.

This created meaningful revenue in 2025, but it was never Duos’ long-term core business. After Q1 2026, New APR sold substantially all of its assets, and Duos received $50.4M, with another $9.9M held in escrow.

Sales

Duos’ financials are messy because the company is in the middle of a major transition. Over the last four quarters, revenue moved from $5.7M in Q2 2025, to $6.9M in Q3, to roughly $9.5M in Q4, before falling to $2.7M in Q1 2026.

That Q1 drop looks bad, but it was largely expected. New APR revenue is winding down, while the new AI infrastructure business has not fully hit the income statement yet.

The 2026 guide is what matters. Management still expects revenue to exceed $50M, driven mainly by roughly $26M from GPUaaS and roughly $26M from Technology Solutions, with additional upside from colocation, infrastructure services, new hosting deployments, and customer expansions.

In other words, Duos’ is guiding for a sharp mix shift in the second half of 2026, away from New APR and toward AI infrastructure revenue. I believe the guide may still be somewhat conservative if GPUaaS and Technology Solutions meet expectations, because management also said it has 10 MW contracted and another 15 MW planned for deployment in 2026. Higher-power EDCs should provide much higher monthly recurring revenue, with some deployments expected to begin contributing in the second half of 2026.

Gross margins

Gross margin has been volatile: about 26% in Q2, 37% in Q3, 27% in Q4, and 59% in Q1 2026. The Q1 margin needs context. Part of the improvement came from New APR-related revenue with no associated cost, and because Q1 revenue was small, that high-margin component had an outsized impact. So the margin trend is encouraging, but not yet clean.

Looking forward, management has indicated that GPUaaS and high-power colocation could carry roughly 80% EBITDA margins. That is very high, but the mix matters. Technology Solutions should have lower margins, while GPUaaS and high-power EDC colocation should carry much stronger profitability if they scale as expected.

Balance sheet

The balance sheet is now the more important part of the story. In mid-2025, Duos was financially fragile, with only $1.5M of cash at Q2 and a thin equity cushion. After equity raises in 2025 and early 2026, Q1 2026 looked very different: $33.0M cash, $41.2M of equipment deposits, $27.6M of property and equipment, only $16.0M of total liabilities, and $106.9M of stockholders’ equity.

The cost was dilution. Shares outstanding rose from about 12.3M at Q2 2025 to about 29.6M at Q1 2026. Painful, but Duos is much better funded today.

Since Q1, the balance sheet improved again. DUOT received $50.4M from the New APR asset sale, which would take pro forma cash from roughly $33M to more than $83M, before taxes, burn, and deployment spending.

Duos also secured a $98.1M three-year USD.AI facility to support the DUOT/Hydra deployment of 2,304 NVIDIA B300 GPUs. The facility is described as non-recourse, non-dilutive, and off-balance-sheet, secured by the GPU infrastructure and related offtake contracts.

That matters because one of the biggest fears around Duos was further dilution. On the Q1 call, CEO Doug Recker said he did not want to “go out for any more equity” and wanted to fund growth through debt financing or strategic backstops. The USD.AI facility does exactly that for the Hydra GPUaaS project. It does not eliminate future dilution risk, but it makes the hyper-dilution bear case less convincing in the short term.

In my view, Duos is no longer operating with its back against the wall financially. The story now comes down to execution: can they deploy the infrastructure and recognize the revenue?

Peer valuation

I have not found a publicly listed U.S. company with the same business model as Duos. There does not appear to be a clean public comp for small, modular, high-power edge data centers.

The closest comparison is not CoreWeave or Nebius. Those companies own or control GPU capacity and sell compute. Duos’ cleaner medium-term model looks more like high-density colocation: find power, deploy infrastructure, and let customers bring their GPUs.

On that basis, Core Scientific and Applied Digital are the better comps. Core Scientific is the cleanest match because it provides high-density infrastructure to host customer GPUs. Applied Digital is also relevant, but it is larger, more campus-scale, and more capital intensive. Duos is much smaller and earlier, but the model it is aiming for is closer to Core and Applied than to and .

Using a recent price of $11.76/share, about 29.3M shares, Q1 cash of $33M, and the $50.4M APR proceeds while excluding the $9.9M escrow, I get Duos at roughly $345M market cap and about $261M pro forma enterprise value.

On current TTM numbers, Duos is not “cheap” in absolute terms, but it is cheaper than Core and Applied today. Duos has about $24.8M of TTM revenue, so it trades around 10.5x EV/sales. Core recently traded around 26.4x EV/sales, while Applied traded around 38.9x EV/sales. Using Core Scientific as the closer comp, a CORZ-like current EV/sales multiple would imply roughly $25/share for DUOT. That said, the discount today is understandable. Duos’ current revenue mix still has very little contribution from GPUaaS or high-power EDC colocation. The market is not going to pay CORZ/APLD-type multiples until those segments show up in reported revenue and EBITDA. But when they do…

Inflection

The inflection should come if/when GPUaaS revenue begins to appear as expected in the second half of 2026. Management expects GPUaaS to represent roughly half of 2026 revenue and a meaningful portion of EBITDA. If EDC colocation and GPUaaS become a larger share of the business going into 2027, the valuation gap could become harder to justify.

So, to recap: Duos has light corporate debt, a much stronger cash position after the APR sale, project-level financing for the Hydra GPU deployment, potentially higher segment margins than peers, and exposure to a niche that could matter in the AI buildout. The risk is simple: the numbers still need to show up, Mr Recker and his team “just” need to execute. I am leaning toward yes given Mr Recker expertise in this segment, alongside Duos’ already proven Technology Services division. Additionally, they just financially derisked a meaningful portion of their GPUaaS project with Hydra. So it’s time to roll up the sleeves and deploy these modular EDCs.

The next few quarters should tell us whether this is a small company chasing the AI wave, or an early mover in a very specific and valuable corner of the AI infrastructure stack. I will be watching intently.

Disclosure: I own shares of DUOT. This article reflects my personal opinion and is not financial advice, investment advice, or a recommendation to buy, sell, or hold any security. I am not a financial advisor. I may change my view or position at any time without notice. The estimates below rely on company disclosures, earnings-call commentary, conference transcripts, and my own assumptions, which may be wrong. Please do your own due diligence.

However, let’s put things into perspective for a moment.

First of all, SPY currently sits at 8.16% YTD. Last year YTD return was 16.36%. SPY returned 23.30% in 2024 and 24.29% in 2023 —2022 was a dump (-19.48%).

We’ve been on a phenomenal run so far, but this year is not exceptional compared to the previous 3 years. All this talk about everything is overextended, and overbought —maybe. But we are on track to perform like 2025, which was a good year but not exceptional by any means. I am not in the business of predicting however. I will be watching SPY and VIX closely, that’s it. I am hoping for a 21EMA reclaim on the daily, that will be a good sign.

I am also hoping for a good CPI print this week. The key is to not elevate hike expectations. Job numbers were very strong. That is a good thing for the economy, but could be bad for the market. Markets are forward-looking. If they get a sniff that the economy is running hot, they will price in interest rate hikes in the near future. Multiples compress, stocks fall. The 2022 bear market was a function of interest rate hikes for instance. The Fed was careless with their rate policy, they thought inflation was transitory —it wasn’t. When they realized their mistake, it was too late. A series of aggressive hikes followed. This time, the issue may be oil prices. Maybe a bad print puts pressure on the WH to reach an agreement with Iran. I honestly have no clue though. These are just thoughts.

But looking at the market and individual names. I don’t find evidence of a pervasive euphoric environment yet, though there were signs that we were getting there. Like IPWR (30k revenue, 96mkt cap… 3x in 2 months.

What I am hoping is for SPY to eventually resume growth while stocks like IPWR, MRAM, AMBA come back to earth. Just like we saw with quantum stocks by the end of 2025.

Some names are getting increasingly attractive

Read more

“We're sitting on something that with power comes high value. It's not only that we're sitting in a very widely recognized, highly attractive area for these data centers and these businesses, and everyone that's listed here is already here. Not to mention the ones that we're talking to that want to come.”

1

Read more

Read more

“[this company] supplies low density aluminum sheet and plate for the tank barrel and domes of the Falcon 9 to help SpaceX meet its objective of providing reliable, efficient and cost-effective transport of cargo, and potentially human crew to low-earth-orbit destinations.”

“The first stage of the Falcon 9 launch system incorporates nine Merlin engines and aluminum-lithium alloy tanks containing liquid oxygen and rocket-grade kerosene (RP-1) propellant. Low-density Airware 2195-T8 plate used in the tank barrels and domes of the booster is supplied by [this company]”

1

I hope those quotes piqued your interest.

Most would think that a value added aluminum company does not require a lot of know-how and technology, that it can be easily replicated. This could not be further from the truth. The process to manufacture aluminum alloys that go on spacecrafts and rockets is usually extremely complex, and for specific applications, some of these alloys can only be produced by a handful of companies in the West. Most specifically, for aluminum-lithium (Al-Li) alloys used in rocket tank walls, domes, barrel panels, intertanks, thrust structures, crew modules, and others, there are only a handful of credible Western suppliers, with only two visible industrial-scale names that I found. One of these companies has already publicly disclosed that they supply these highly specialized alloys to NASA, Blue Origin, and Boeing for space missions purposes.

It is quite curious to me how no one is talking about them. Most importantly, there is strong evidence that they also supply these alloys to SpaceX. My research shows that their relationship goes back as far as 2010, with the last publicly documented mention in 2019. However, given SpaceX recent descriptions of Falcon’s tank structure and its composition, I have not found public evidence that SpaceX changed suppliers. This alloy plays a mission-critical role and these relationships are sticky.

I am talking about Constellium SE (CSTM).

What CSTM does

CSTM turns aluminum into higher specification rolled and extruded products used in packaging, aerospace, defense, automotive, industrial, and transportation markets. Its largest businesses include beverage can sheet, automotive body sheet, structural automotive components, and aerospace/defense materials.

Aerospace & Transportation (A&T): The Growth Engine

The more interesting part of the story is its Aerospace & Transportation (A&T) segment. On revenue/shipments, Q1 2026 A&T revenue was $609M vs $468M in Q1 2025, and shipments were 60 kt vs 51 kt. Most importantly, A&T segment has higher margins than the other operating segments, at 17% vs 7% and 5% in the remaining segments.

The star of the show in this segment is Constellium’s space related aluminum lithium technology called Airware. Airware is Constellium’s family of proprietary aluminum lithium alloys designed for applications where weight, strength, stiffness, weldability, corrosion resistance, and extreme temperature performance matter.

In space, these materials can be used in rocket tank walls, domes, barrel panels, intertanks, thrust structures, crew modules, and other mission critical structures. Rockets or spacecrafts need materials that are extremely light but still strong enough to survive launch loads, pressure, vibration, and cryogenic fuel conditions. Airware is built for that kind of job.

A&T Relationships

First, the ones that we know.

Constellium has flight heritage in programs such as NASA’s Space Shuttle external tank, Orion, SLS, and Blue Origin’s New Glenn, and management says demand for space and military aviation is strengthening.

But what about SpaceX?

SpaceX is highly vertically integrated, security-sensitive, and procurement-sensitive. Publicizing exact suppliers for critical rocket structures can expose supply-chain dependencies, weaken purchasing leverage, help competitors understand material choices, or create unnecessary security/ITAR-style sensitivity around launch hardware. Most importantly, SpaceX’s most S-1 also acknowledges that despite vertical integration, SpaceX still depends on third parties for specialized materials, components, and services, and that supply disruptions or supplier qualification issues could affect operations.

Naturally, we need stronger evidence, not just those suppositions.

In this 2010 article, Constellium’s predecessor business, Alcan Global ATI, was named as a Falcon 9 supplier:

Alcan Global ATI supplies low density aluminum sheet and plate for the tank barrel and domes of the Falcon 9 to help SpaceX meet its objective of providing reliable, efficient and cost-effective transport of cargo, and potentially human crew to low-earth-orbit destinations.

Alcan Global ATI was rebranded a few years later as Constellium.

Then in 2019, this article acknowledges Constellium’s role in providing a specialized alloy to Falcon 9’s rocket:

The first stage of the Falcon 9 launch system incorporates nine Merlin engines and aluminum-lithium alloy tanks containing liquid oxygen and rocket-grade kerosene (RP-1) propellant. Low-density Airware 2195-T8 plate used in the tank barrels and domes of the booster is supplied by Constellium (formerly Alcan Global ATI)

This is important because a lot change since 2010. Still, 9 years later, Constellium was still a key cog in SpaceX’s rocket supply chain.

In SpaceX’s most recent Falcon Payload User’s Guide, dated May 9, 2025, it states that the “first stage propellant tank walls” of Falcon vehicles are made from aluminum-lithium alloy, and that the tanks are manufactured using friction stir welding. It also describes a common dome separating the liquid oxygen (LOX) and kerosene rocket fuel RP-1 tanks, which confirms the tank architecture context that fits Airware’s specs. Especially the language around CSTM’s alloy being suitable for, and optimized around, friction stir welding.

So, all in all, Constellium’s role appears to be supplying the specialized Airware-family aluminum-lithium plate/sheet that can be formed and friction-stir-welded into Falcon’s tanks.

Arconic, the only visible competitor

The only clearly visible Western industrial-scale competitor with a comparable aluminum-lithium capability is Arconic. However, the best public evidence for Falcon-like cryogenic tank plate/sheet points more specifically to Constellium’s Airware 2195/2050 family, not to Arconic.

Why? And it bears repeating, I am hammering this point again: SpaceX says Falcon tank walls are made from aluminum-lithium alloy and manufactured using friction stir welding. That tells us the material must be suitable for large welded rocket tanks, not just any aircraft Al-Li part. NASA’s materials literature says the relevant launch-vehicle plate alloys are mainly 2195 and 2050 alloys: 2195 is the incumbent plate alloy for cryogenic tank/dry-bay structures, while 2050 may be better for plate gauges above 2 inches.

Constellium publicly maps directly onto that use case. Its Airware brochure says Airware 2195 plate delivers improved performance for cryogenic tankage through optimized friction stir welding, and that Airware 2050 can replace legacy 2219 for thicker space components between 2.0 and 6.5 inches. Constellium also says fuel-tank material needs low density, formability, machinability, weldability, corrosion resistance and crack resistance, and specifically positions Airware 2195/2050 for space launchers, crew modules, Orion and SLS-type applications.

Arconic is real competition, but its public marketing is broader and more aircraft-focused. Arconic says it has the world’s largest aluminum-lithium plant, can produce more than 20,000 metric tons per year, and offers a broad Al-Li portfolio including extruded, forged and rolled parts. It also highlights wing skins, fuselage skins, wing stringers, floor beams and seat tracks. I do not see Arconic publicly saying, in the same direct way, something like “our 2195/2050 plate is optimized for cryogenic rocket tankage and friction stir welding.”

Conclusion

For the reasons explained above, I believe CSTM supplies the Al-Li material used in the manufacturing of SpaceX’s Falcon rocket fuel tanks.

The most exciting part is that the market has not given the proper rerating to this story. Relative to peers, CSTM still looks cheap. At roughly 11x P/E, CSTM trades at a large discount to Kaiser Aluminum at about 20x, and at an even wider discount to higher-quality aerospace/materials peers like ATI, Carpenter, and Howmet, which trade around 50x to 60x earnings. That discount might make sense if CSTM were just a cyclical aluminum converter with no differentiated technology, but it becomes harder to justify when you factor in Airware, its space-grade aluminum-lithium platform, its NASA/Blue Origin/Boeing exposure, and the credible historical evidence tying Constellium/Alcan material to Falcon tank structures. The market does not appear to be assigning CSTM any meaningful “space materials” premium, let alone a SpaceX-adjacent premium. Compared with peers that already receive rich aerospace/defense multiples, CSTM is still being valued more like a basic aluminum processor than a scarce supplier of qualified, mission-critical Al-Li materials for rockets, crew modules, defense aircraft, and next-generation aerospace structures.

Then we have LiteWave800, ADTN’s new 800G linear pluggable optics module for short-reach intra-data-center links. It is a low-power optical transceiver designed to connect AI racks and switches while consuming far less power than conventional 800G optics. ADTN claims LiteWave800 operates at 0.8W and 1 pJ/bit, with power dissipation roughly 12 to 18 times lower than DSP-based optics and 6 to 10 times lower than first-generation LPOs. If those numbers hold up in real deployments, the economics are not trivial. Lower module power means lower electricity use, less heat, reduced cooling burden, and potentially denser AI interconnects. That is a real hyperscaler problem that people are not scared to throw a lot of money into. Management was emphatic in their last earnings call on how “market reaction to it has been fantastic.” Then they said that they’ve had “some very large, very well-known customers that have been very encouraging(…)to get the product out as quickly as possible.”

There is a lot to like. ADTN is improving operationally, optical networking is becoming a larger part of the mix, and LiteWave800 gives the company a credible wedge into AI data-center connectivity. The problem is that ADTN also has a structural overhang that makes the stock difficult to trade cleanly.

Read more

“Based on these conversations stand today, we expect to be mostly sold out of our 2027 target production in the near term with a meaningful portion of 2028 spoken for.”

“we indicated that the margins would be as good or better than (current business)(…)people have been using a reference point of $1 million per megawatt hour, so that's $25 million. So roughly about $7 million to $8 million of EBITDA per unit.”

1

Start with the simple math.

Management has now given enough detail to model its new Power segment on a rough 2027 annualized basis. Their target is 100 Mod-1 units in 2027. Each unit is 25 MW. Management framed the economics at roughly $25 million of revenue-equivalent per unit and $7 million to $8 million of Adjusted EBITDA per unit.

So:

$7 million to $8 million of Adjusted EBITDA per unit × 100 units = $700 million to $800 million of annualized Adjusted EBITDA.

That would be a large number for this company. Against FY2025 Adjusted EBITDA, it implies roughly 58% to 67% growth. Against Q1 2026 annualized Adjusted EBITDA, it implies roughly 54% to 61% growth. Against management’s FY2026 business-segment EBITDA guide, it implies roughly 43% to 49% growth.

There may be more beyond that. Management said the long-term service agreement revenue, or LTSA revenue, is not included in that math. Customers also want service contracts, and FTAI expects to be paid based on usage, similar to its aviation exchange model. So I treat the $700 million to $800 million as the baseline Power contribution, not the full lifetime economics of the installed fleet.

The company I am talking about is FTAI Aviation, ticker FTAI.

The Current Business

FTAI’s core business is already attractive. At its heart, FTAI is an independent engine maintenance and exchange platform focused on CFM56-5B, CFM56-7B, and V2500 engines, the older narrowbody engines that power 737NG and A320ceo aircraft.

They buy or source engines, modules, and used material, repair or refurbish them through their own facilities and joint ventures, then sell or exchange serviceable engines, modules, and parts to airlines, lessors, asset owners, and strategic partnerships under their Maintenance, Repair, and Exchange model, or MRE.

The economics are driven by scarcity and speed. Airlines need cheaper, faster alternatives to full OEM shop visits, and FTAI captures margin by controlling feedstock, repair capability, PMA and used material sourcing, and module exchange capacity.

This core business is growing fast. In Q1 2026, Aerospace Products generated $743.8 million of revenue and $222.6 million of Adjusted EBITDA. That was 104% revenue growth and 70% Adjusted EBITDA growth year over year. Annualized, Q1 2026 Aerospace Products is running at almost $3.0 billion of revenue and about $890 million of Adjusted EBITDA.

The nuance is margin. FTAI is chasing larger airline programs and heavier work scopes, which means EBITDA dollars are growing quickly but percentage margins are lower than prior peak levels. Management has been clear that 2026 is more about capturing market share than maximizing near-term margin percentage.

The 2025 MRE Partnership

A quick pause on the 2025 Partnership, because I think this part of the story is underrated.

The structure is simple: FTAI uses third-party capital to own aircraft, while FTAI keeps the engine-maintenance and exchange work. Outside investors help finance aircraft ownership. FTAI contributes a minority stake, acts as the manager or servicer, and keeps the engine relationship.

That matters because it can turn aircraft ownership into a captive demand source for FTAI’s core engine business. The partnership owns the planes, airlines lease the planes, and when the engines need work, FTAI is positioned as the engine solution provider.

Financially, this is already material. FTAI recorded $335.8 million of MRE Contract revenue in 2025 from engine and module sales or exchanges with the 2025 Partnership. In Q1 2026 alone, MRE Contract revenue was $221.2 million. That represented about 26.6% of total Q1 2026 revenue, calculated as $221.2 million divided by $830.7 million. It also means Q1 2026 MRE Contract revenue already covered roughly two-thirds of the full-year 2025 MRE Contract revenue generated from the same structure.

This is why I like the model. FTAI contributes some capital and owns a minority stake, but it also creates a recurring channel for engine work.

Management’s track record

For starters, the CEO, COO, and CFO bought shares back in November and haven’t sold them since.

Execution matters a lot here, especially because FTAI Power is still ramping toward commercial delivery. Management’s recent follow-through gives me more comfort than I usually have with a new business launch.

Across the Q1 to Q4 2025 calls, the pattern was consistent: management gave operational targets, then generally met them or raised them later.

The main exception was adjusted free cash flow. FTAI ended 2025 around $724 million of adjusted free cash flow versus the revised $750 million target. I do not view that as a core execution miss. Management chose to reinvest more aggressively in growth inventory, Strategic Capital, and Power.

The Q4 to Q1 2026 follow-through was also solid. In Q4, management raised FY2026 business segment EBITDA guidance to $1.625 billion, raised the module target to 1,050 CFM56 modules, and kept the Power timeline intact. In Q1, they reaffirmed the $1.625 billion EBITDA guide, reaffirmed the $915 million adjusted free cash flow guide, produced 270 CFM56 modules, and gave stronger evidence that FTAI Power is moving toward commercialization.

The only caveat is Aerospace Products margin. Margins stepped down to around 30% as management prioritized market share and larger airline programs. I do not love lower margins, but I understand the strategy. They are trying to win the customer relationship now and expand the profit pool over time.

I would not say there is no execution risk. There is always execution risk. But management has earned some credibility.

FTAI Power

FTAI Power is the new segment that could change the company’s earnings power.

FTAI officially launched FTAI Power as a platform focused on converting CFM56 aircraft engines into aeroderivative power turbines. The opportunity is tied directly to data-center power demand, especially demand driven by AI hyperscalers. The first product, Mod-1, is expected to be delivered in Q4 2026, with a planned production target of 100 units in 2027.

This has moved beyond a slide-deck idea. On the Q1 2026 call, management said prototype testing was ahead of schedule, all major mechanical milestones were complete, final testing should wrap in Q3, and customers were visiting the site to see the prototype. Management also said it was in advanced negotiations and expected to be sold out of 2027 volumes imminently, with discussions already extending into 2028 and beyond.

That last point matters, but I want to be precise: management has not disclosed signed 2027 backlog in detail. The language was “advanced negotiations” and “sold out imminently,” not a fully published order book.

The Jereh Group JV

In Q1 2026, FTAI announced a JV with Jereh Group, which meaningfully de-risks the path to commercial scale.

FTAI will provide the converted turbine. Jereh will handle much of the surrounding power package: the trailer, generator, gearbox, controls, and overall packaging. Jereh already packages turbines for major players like GE Vernova, Baker Hughes, and Siemens, and has manufacturing capacity across the U.S., Canada, the UAE, and China.

That is important because FTAI is outsourcing a major execution bottleneck to a scaled specialist.

The tradeoff is accounting complexity, not necessarily worse economics. Management said the overall unit economics should be preserved. Because part of the value now sits inside a JV, reported revenue may understate the full economic value of each unit, with some profit flowing through JV earnings instead. That is why I focus more on Adjusted EBITDA than revenue in my model.

If FTAI treats the Power JV like its other unconsolidated entities and includes its pro-rata JV EBITDA in Adjusted EBITDA, the JV structure should not materially reduce the economic EBITDA contribution, even if reported revenue looks lower. That assumption still needs to be confirmed once Power becomes a reporting segment, but it is consistent with how FTAI has treated other unconsolidated entities in its non-GAAP reporting.

Working The Valuation

Before I work on my preferred valuation lens for this business. I want to bring up two things that happened since their Q1 call. Management has been aggressive more recently.

Recent Balance Sheet Moves

Before getting to valuation, two recent capital-structure moves are worth mentioning.

Redeeming Series C Preferred Shares

FTAI announced that they are redeeming all of its Series C preferred shares, ticker FTAIN, by paying holders $25 per share in cash on June 15, 2026. This costs about $105M, but it eliminates an 8.25% preferred dividend, saving roughly $8.7M per year. FTAI is using liquidity to retire an expensive layer of capital. In my opinion, long term, this is modestly positive for common shareholders because fewer cash flows go to preferred holders, but near term it slightly reduces cash on the balance sheet. Keep this in mind moving forward, especially when we move through quarters.

Entry into ABS Capital Markets

The ABS is FTAI’s Strategic Capital vehicle borrowing $612M against a pool of 48 A320ceo and 737NG aircraft leased to 23 airlines. They have aircraft-backed financing at the Strategic Capital vehicle level, and the notes are expected to be investment grade. Put it simply, outside capital is helping finance aircraft ownership while FTAI earns servicing fees, MRE revenue, and profit participation. Long term, this is positive if it lowers capital intensity and scales the asset-light model, but it also increases dependence on ABS/credit markets and the performance of aircraft collateral.

There is no free lunch. The Strategic Capital model adds accounting complexity and financing-market dependence. But I like the direction because it supports the same flywheel: more aircraft relationships, more engine demand, more MRE revenue.

Valuation

For FTAI, I prefer EV/FY2026 guided EBITDA as the main valuation metric.

At a stock price of about $260, and pro forma for the Series C preferred redemption, I estimate enterprise value around $30.35 billion. FTAI reaffirmed $1.625 billion of FY2026 business segment EBITDA. That puts FTAI at roughly 18.7x EV/FY2026 guided EBITDA.

The question is what happens if FTAI Power delivers?

My baseline scenario assumes the market does not give FTAI any extra multiple for Power. Multiples stay the same. If Adjusted EBITDA rises by 40% to 50%, using today’s multiple, the implied stock price is roughly $377 to $4062.

GE Vernova (GEV) is the most obvious public-market comparison for the power-turbine theme, but the comparison needs to be adjusted carefully. GEV’s FY2026 guide is $44.5 billion to $45.5 billion of revenue and 12% to 14% adjusted EBITDA margin, implying about $5.85 billion of midpoint adjusted EBITDA. Based on current market values, that implies roughly 43.7x EV/EBITDA.

But GEV is far more exposed to power than FTAI would be under the 100-unit FTAI Power case. GEV’s Power segment likely represents around 70% of GEV EBITDA on a full-year estimate, while FTAI Power would represent only about 30% to 33% of FTAI’s pro forma EBITDA. GEV’s Power segment also includes more than gas turbines, so I do not treat GEV as a perfect comp.

Using an exposure-adjusted method, I discount the GEV multiple by relative Power exposure:

GEV multiple × FTAI Power EBITDA share / GEV Power EBITDA share

That framework gives FTAI a stock-price range of roughly $385 to $416.

So both valuation methods point to a similar conclusion. If FTAI Power delivers the 100-unit 2027 case, the stock can justify a meaningfully higher price than today, even without assuming growth in the core business or any extra LTSA economics.

What Could Go Wrong

The core engine platform is growing and already profitable. The biggest risk is expectation risk.

FTAI Power still needs signed orders, completed testing, commercial delivery, working units in the field, and proof that the Jereh structure preserves the economics management has described. The stock has already moved a lot. If contracts are delayed, margins disappoint, or the 100-unit ramp slips, the multiple could compress quickly.

There is also balance-sheet risk. FTAI is improving liquidity and reducing expensive preferred capital, but it remains a leveraged growth company. Inventory, Strategic Capital commitments, Power development, and acquisition activity all require capital.

However, management’s recent execution history gives me comfort that they should keep executing on all fronts.

Bottom Line

FTAI’s current business is strong enough to matter on its own. Aerospace Products is growing fast, the MRE model has real strategic value, and the 2025 Partnership shows how FTAI can use outside capital to create more engine demand.

FTAI Power is the upside case. Management’s own math implies $700 million to $800 million of incremental annualized Adjusted EBITDA if 100 units are delivered in 2027. That would be a major step-up versus FTAI’s current EBITDA base.

If FTAI Power executes, today’s valuation could be reasonably cheap. My base valuation work points to something around $377 to $416, before giving credit for core-business growth in 2027 or LTSA revenue.

That is why I own the stock. The core business is working, management has shown decent follow-through, and FTAI Power gives the company a credible path to a much larger earnings base.

Disclosure: I own shares of FTAI. This article reflects my personal opinion and is not financial advice, investment advice, or a recommendation to buy, sell, or hold any security. I am not a financial advisor. I may change my view or position at any time without notice. The estimates below rely on company disclosures, earnings-call commentary, conference transcripts, and my own assumptions, which may be wrong. Please do your own due diligence.

Leave a comment

Ampco-Pittsburgh (AP) is a small industrial manufacturer with two main businesses. Its Forged and Cast Engineered Products (FCEP) segment makes large forged and cast metal rolls used by steel and aluminum mills to flatten, shape, and process metal. This is the older, more cyclical part of the company because demand depends on steel and metals production. Its Air and Liquid Processing (ALP) makes specialized pumps, heat exchangers, and custom air-handling systems used in markets such as power generation, nuclear, U.S. Navy ships, pharmaceuticals, industrial facilities, and data-center-related power infrastructure. My thesis on this stock revolves around its ALP segment overtaking FCEP as the main engine inside AP.

Read more

Hey Everyone,

Welcome to Tropical Penguin, my new home for sharing trading and investing research in a transparent, education-first way.

I am moving everything here to Substack so I can deliver more in-depth analysis, better organization, and a cleaner experience for everyone. Whether you’re a beginner looking to learn or an experienced trader wanting thoughtful discussion, this is the place for it.

What You’ll Get Inside Tropical Penguin

Here’s what I’ll be sharing regularly:

• Market commentary and detailed research breakdowns

• Valuation and thesis discussions

• Watchlists and ongoing idea tracking

• Risk management and portfolio construction education

• Educational trade journal updates shared for discussion and learning purposes

Everything here is designed to help you think critically about markets and improve your own process.

Important — Please Read

This community is for education and research discussion only.

It does not provide personalized financial advice, investment recommendations, or trade instructions. All content is general information and reflects my personal opinions and research for educational purposes only.

You are responsible for your own investment decisions and should always do your own due diligence. Past performance is not indicative of future results.

Disclosures: I may discuss securities that I currently own or am actively watching. Any positions or interests will be disclosed in the relevant posts.

How to Get the Most Out of This Substack

• Subscribe so you never miss new research and updates

• Paid subscribers will get access to deeper analysis, exclusive watchlist updates, and additional commentary

• Jump into the comments — I will read and reply whenever possible

• Treat this as a learning community, not a signal service

Thanks for being here. Let’s learn and grow together!

— Tropical Penguin