In the case of Nu, and the reason I’m opening this breakdown like this, I listened to one of the best and most engaging stories of a founding I’ve ever come across with the Crucible Moments episode featuring Founder & CEO David Vélez. I’d recommend anyone reading this also listen to the episode (hey, your dog would love the walk too).

This, at the time of writing, $70B company is a bank. It generally wouldn’t show up on a tech website, but here we are. I think it’s fair that it shows up here given that Sequoia were an early investor, and given the tech-forward nature of what we’re looking at.

Nu, though, is a bank. It’s a bank that started in Brazil and has since expanded to Mexico and Colombia in the hopes of running the same playbook, and that playbook appears to be working.

So why are we here? Nu is not your standard bank. It’s actually an AWS-native software platform with a $1 cost to serve that happens to hold a banking license. Let’s dig in.

The Revolving Door

In the podcast I linked above, Vélez tells the story of trying to open a bank account in São Paulo back in 2012. He had just moved to the country for Sequoia, but they ended up shutting down their LatAm office almost immediately due to a lack of investment opportunities. He was there, in a new country, without a job.

I’ve never banked in Brazil, but it sounded like a nightmare. You didn’t just walk in and make your way to the teller, you had to pass through bulletproof glass, metal detectors, the works. If you set off those metal detectors while inside that bulletproof glass, you were locked in while the situation was figured out.

Once Vélez made it inside the bank it took him six trips over four months and stacks of notarized paper to get an account approved. And the reward at the end? $30 in monthly fees for his bank account and a credit card charging 300% interest.

At the time, Brazil was run by a five-bank cartel. Itau, Bradesco, Santander, Banco do Brasil, and Caixa. Between them, they controlled around 80% of all the banking assets in the country and made 20%+ ROE every year while providing what sounds like the worst service on Earth. They got away with it because customers had nowhere else to go.

Anyway, the story ends up coming full circle. Fresh off the nightmare of opening a bank account, Vélez decided to start a bank with no branches, all run by an app. A tech forward bank. He was laughed out of many rooms, but Doug Leone of Sequoia wrote the seed check.

There’s a lot more to the story, but we’re trying to be efficient with words here. The bank and its credit card did eventually become a reality. A purple card entered the Brazilian market on April 1, 2014. It wasn’t a hit out the gate, but once it caught its stride it ended up taking the market by storm.

The $1 Cost to Serve

The thing that really sets Nu apart is the $1 monthly cost to serve. I’ve worked in fintech. Trust me, that’s low.

The obvious savings come from lack of branches. Nu doesn’t have to carry all that baggage. But the tech question is more interesting than simply moving banking to an app.

I work in fintech, so I’m not automatically impressed by the words “AWS” and “microservices.” What interests me is whether the setup makes useful changes easy to deliver and the business cheaper to operate.

For example, let’s assume customers keep getting stuck partway through an application. Most would give up, some would call support to get it finished.

The useful promise of splitting software into smaller, independently changeable services is that a team can improve that part of the experience without having to change the entire banking platform at the same time. Of course you’ve still got to test and deal with controls... smaller pieces don’t mean fewer problems in all cases.

But suppose the team can test a simpler application. They can see where people get stuck, fix, test, iterate, and improve again.

Once they get to a place where more people finish the application, they’ve built a winning process. The structure that Nu has allows them to get to that winning process much quicker than at a legacy bank.

That’s what I’m looking for at Nu. Not more software releases for their own sake, but evidence that the company can improve the product while keeping the work required to serve each customer under control.

The architecture is a possible enabler of that advantage. It isn’t proof of it. A low cost-to-serve figure tells us something about the outcome; it doesn’t, on its own, tell us how much credit belongs to a particular technology choice.

When your cost is just a dollar, you can bank anyone. You can hand out a credit card that carries no annual fee. You can provide savings accounts with zero minimum balance requirements. And you can run the whole operation in a way that’s friendly to the humans on the other end.

In short, the $1 cost-to-serve leads to a company that’s able to provide a fantastic product in a part of the world where fantastic generally hasn’t been available.

“Fantastic” made customer acquisition practically free too. The bright purple card kept customers coming. It was a badge of honor that made friends want to get involved too. Soon, they captured the market.

The Flywheel

The purple card got millions of users in the door, but popularity alone doesn’t pay the bills. Nu still needed to build a profitable business.

If you apply for a credit card right now, assuming you’re in the U.S. and have decent credit you’re likely to get a silly limit... $20k? $50k? The Amex Platinum “limit” seems to be endless. It’s easy to do that now in the U.S. because the playbooks have been run through time and time again.

It’s not easy to extend credit to what has historically been an underbanked population. And, while those banks have been there for a while now, maturity of information and outcomes just isn’t as strong as it is in some parts of the world.

Nu’s playbook combats this by giving out those purple cards to first time credit holders with tiny limits... sometimes as low as $20. They get to see at this essentially riskless scale that a customer will pay their bills. When they do, the credit limit grows over time.

To fund that credit limit, Nu needed a source of low cost funds. They got there in 2017 by creating NuConta, a zero fee high interest savings account. No more borrowing expensive wholesale debt in the open market, they could now fund their own operations entirely.

Year one, the customers in this flywheel of credit and banking generally pull in around $3-4 per month, primarily from swipe fees on their cards.

Nu then encourages them to direct deposit their paycheck. Offers them personal loans if they need it. Allows them to buy term life insurance or trade crypto from the app. All of these slowly add up and monthly revenue, in those older cohorts can grow upwards of $12-15/month.

Best part? The cost to serve remains right around $1.

That’s a ridiculous operating leverage for a bank. The company closed Q2 2026 with $1.1B in net income, a 50% jump from the year before. Return on equity? 28%. Phenomenal for a bank.

One important distinction: cost to serve is not the entire cost of banking with Nu. It measures a specific set of servicing and transaction expenses. Funding costs, credit losses, and other expenses still sit between customer revenue and shareholder profit. The advantage is operating efficiency, not an exemption from banking economics.

Nu can make the operating side of banking cheaper. The investment question is how much of that advantage survives the rest of the business.

Expansion into Mexico & Colombia

Saturation was reached in Brazil with 118 million customers, well over half of the adult population. So, the next natural step was to take the playbook elsewhere.

Mexico and Colombia were the next two targets. Mexico, arguably, could be an even better target than Brazil initially was. More than half the population have never had a bank account and a lot of daily commerce still happens in cash. The problem: people are skeptical of banks.

Going into Mexico, Nu did tweak the strategy a bit. Instead of going credit first, they led with a high-yield savings account. A beast of a high-yield that paid 15% interest! Billions in deposits showed up pretty quickly.

Hitting 16 million customers in Mexico came pretty quick. Time will tell, as Nu starts scaling back those interest rates, if the people came solely for the initial offer, or if it’s a quick exit once that money doesn’t yield fantastic returns anymore.

On the Colombia side of things, the playbook looks the same. 5 million new customers in that country are helping Nu continue to find higher growth even after saturating home.

The best part of this expansion strategy though? Well Brazil prints so much cash that Nu is able to fund the entire rollout out of its own pocket, no need to ask Wall Street for another dime.

Then What’s Next?

The U.S. market looks to be in play next. I noodled on this for a bit because it’s largely where I start to worry that the story could break down. There’s a 0% chance that Nu can keep customer costs around $1 per month in the U.S., it’s just simply too expensive.

You’re also not going to have much success with even an entry level credit card in the U.S. against the behemoths Chase and American Express that have every single nook and cranny covered.

So, then, what? Well there’s a good amount of workers in the U.S. that send funds back to Mexico. Moving that money can be a little tricky, and sometimes costly. Nu, should they make an appearance in the U.S. could have a very quick in-road right there.

Take funds, make it easy, free, and take zero time to move money to your family in Mexico / Colombia / Brazil who also have a Nu account? They’d see some quick account openings here stateside with that.

Where This Can Break

So far there’s a lot to like about Nu. We have a lower-cost way to serve customers, more products to offer them, and countries where the playbook could be re-deployed. But there are places where I’d be careful about assuming the good story automatically.

First: credit.

Starting someone off on tiny credit is great. Increasing that limit and offering them a personal loan is another. What we’d want to see in an investment scenario is that, as lending grows, the repayment story stays the same. Put differently, additional lending should still remain attractive after credit losses.

Second breaking point: cost of deposits.

I raised the question earlier: did Mexican customers come for Nu, or did they come for the artificially high savings rate?

Ideally, the rate gets someone through the door, but the experience has to give them a reason to stay. Nu has not been around in Mexico for long enough to know with certainty the answer to this question, but it’s something we’d need to watch.

Third: assuming Brazil’s success travels automatically.

Reusing the technology is definitely appealing. Recreating customer relationships is much harder. The world has shifted since Nu first got its start. Many other countries have pleasant financial entities now, so Nu is playing in a new (ha!) world everywhere it goes.

An investor should want to see Nu develop attractive economics in new markets, not just impressive account totals. If it’s simply the number of accounts, we end up with a big business that’s less profitable.

A big bank doesn’t trade at a Nu multiple. And that would be a massive problem for any would-be investors to reckon with when they’re faced with a re-rate.

Thank you for reading this far! The rest of the article is reserved for paid subscribers. In it we will cover:

• What the economics look like for shareholders

• What the valuation tells us today

• Our bull/bear/base case

• What decision we’re making and whether NU will be added to the portfolio

All the woes got me thinking a little bit about GitLab (NASDAQ: GTLB) and what potential it has in the world of enterprise software. It’s a company with a great story to tell, and lots to break down.

The Invisible Domain of Software Development

Developing and selling software is a lot more involved than simply typing syntax. That’s only one part of the puzzle and GitLab covers half the pieces.

You’re going to want to plan things out. GitLab isn’t really playing in that space so we can safely disregard it.

Once you’ve planned things out you’ll write the code. With potentially hundreds of software engineers working at any one time on the codebase there needs to be a system for organizing all that code and making sure no one steps on another’s toes. That system is Git.

Engineers can write their code, open a “pull request” and then merge that code into the main branch of the software they’re developing. GitHub, GitLab and others like Atlassian’s Bitbucket run services that help make working with Git easier overall.

After that comes a whole bunch of “invisible” stuff. There’s the CI/CD pipelines (continuous integration / continuous delivery). These are the pipelines that automatically run tests, deploy software, configure services, and keep things running. There’s security to handle too, making sure the code is clean and packages are always kept up to date. If all those are handled well you’re going to need to prove it out which is where compliance logs come in.

Every one of those steps above, GitLab handles. GitHub handles them too (with Actions, Advanced Security). There are also a number of third parties that handle these steps individually too.

GitLab’s Edge - An Upper-Market Monopoly

While there are many competitors in the space, there is a distinct advantage that GitLab carries: its ability to work with customers who are under strict national security controls.

Could GitHub work with the DoD? Maybe. But the Department of Defense, many of the contractors that work with the DoD, and lots of other regulated entities side with GitLab more than GitHub.

GitHub offers Enterprise Server, a self-hosted instance but it is packaged as a closed-source, black-box virtual appliance. GitLab is a Kubernetes-native open-core piece of software. Under initiatives like Platform One and Iron Bank, the military can harden GitLab. They can’t do so with GitHub.

GitLab also comes as one package. You’re not having to piece things together, add in plugins, integrate vendors. If you’re in a mission critical area and need one vendor to handle everything, GitLab is a solid choice.

If you need open-source and the ability to verify / validate every touchpoint your code goes through, GitLab has you covered. If you’re in Europe and subject to the DORA framework (like the banks are) you might already struggle using GitHub due to vendor concentration risks. Using GitLab means you’re still able to be on Azure.

The US CLOUD Act also presents some challenges as well for GitHub when it comes to large international corporations who want to control their data. If a US company has “possession, custody, or control” of the data or encryption keys, they have to hand it over. GitLab’s self-managed instance is essentially a sovereign fortress.

The Switching Costs

One domain where GitLab will struggle is on getting existing enterprises to switch. To switch your git repos takes time, to change your CI/CD pipelines is risky.

No one really wants to take on these kinds of risks and tell the CEO, “hey, we’re slowing down development on new features for a bit.” So how does GitLab compete?

Well, first has to be the consolidating sale. GitLab might not have to convince a CTO to go beg for the change, they might be able to convince the CEO or CFO themselves with a look at potential savings.

GitHub, Circle, Snyk, some other tool to monitor compliance, they can easily run up to $120-150 per engineer. GitLab’s top offering is $99 per engineer. Their sales team could make the pitch of “one vendor, not several for one lower price.”

A team of 1000 saving $600k/year and only having to deal with a single vendor can be an easy sell in a time of cost cutting, a finding backed up by Forrester’s Total Economic Impact Study showing a 483% 3-year ROI for Ultimate consolidation.

Another way they potentially make those switching costs come a little easier is by offering the tooling in stages. A company can integrate GitLab’s CI/CD tooling into their existing GitHub workflows with ease and then change to GitLab’s code repositories over time.

The SMB Dilemma

Turning now to the business side of things a little more, I’d like to start out with a concern. A little while back, GitLab made the decision to step away from the SMB market and the hobbyists. Those markets are effectively free for GitHub to land.

Wall Street, of course, rewarded the decision. By focusing on Fortune 2000 and larger foreign entities, GitLab is putting itself in a position to benefit from a massive operational margin tailwind.

The costs of servicing SMB is higher, the costs of hobbyists is high too. Thousands of small companies at $9/month can be replaced by one or two large entities paying $99/mo for their engineers.

The dilemma I think GitLab will face though is losing out on that next generation of seed startups (the next Stripe, the next OpenAI).

Startups will not pay the $29 GitLab price point out the gate and, as we discussed above, will have a hard time switching barring a lot of push from people within the org. A lot of these companies may just stick with the status quo as long as it’s not actively harming things.

So the tailwind that GitLab receives now may well turn into a structural challenge three to five years from now. GitHub may solve its tech issues, it may broaden its offerings in some way to capture other markets, all while GitLab is sitting back and focusing on the Fortune 2000.

In fairness, if you’re going to focus somewhere, the Fortune 2000 is pretty good.

Back To Positive: The AI Paradox

AI is supposed to make us all efficient and it is succeeding in the software engineering world. It used to be that writing code was the bottleneck to progress but those days are now behind us.

Bottlenecks now exist downstream from the code writing. There’s a blockage at the review phase, the CI/CD phases, and on infosec teams keeping on top of all that’s getting delivered.

AI has allowed engineers to create more, much more. It’s partly the reason why GitHub is on the struggle bus as much as it is GitHub Incident History. GitLab is building for that future. Their goal is to build the platform to handle 100x the demand it sees today.

Growth in the shipping of code means that companies will require bigger CI/CD workflows. They’ll require better infosec tooling too, tied right into their codebases and GitLab can handle that.

The company is also down the path of building tooling for agentic development. Agent based code review guidance already exists. In times of old (two years ago, ha) engineers would have to comb through lines of code to make sure what was being shipped was coded effectively.

Now? Well you can have agents do a significant first pass which should save time and, in a virtuous cycle, give engineers time back to ship even more, backed by a Forrester Study showing a 400% 3-year ROI on the Duo Agent Platform.

The expansion and building out of this agentic platform, Duo Agent, means that GitLab has the potential to rake in incremental income from customers by charging usage credits for these AI workloads.

The Money Troubles

Seems a bit weird to write “money troubles” about a company with $1.36B in cash on the balance sheet, right? Well, it is. I’m actually talking about stock based comp at the company.

Stock-based compensation eats up effectively the entirety of free cash-flows. According to GitLab’s SEC Form 10-Q, Free Cash Flow comes in at $263.4M, a fantastic 26.2% margin over the last twelve months. Stock based compensation of $209.2M means that actual owner FCF ends up settling at $54.2M. A true owner FCF margin of 5.4%.

The good news is that things are looking a little better in Q1 of 2027 than they were in Q1 of 2026 so the ship may very well be turning, although it is a slight turn.

The CEO is fond of remote work (hey, me too), and also fond of keeping engineers around which requires, at least for some of the higher levels of talent, a nice pay package.

Getting SBC under control was actually the thing that drew me towards the company. I was using our cannibal and reinvestor screeners and thought to myself, “are there opportunities for companies to move into these buckets?” GTLB came up as a top contender.

Others? ADSK which is already undergoing activist involvement, and DocuSign. Two other cash cows that could do a better job of capital allocation and it would seem GitLab is no different.

Could the company succeed on its current path? Sure. Is there a chance for fantastic returns with corrections? Absolutely.

Because GitLab already generates ~26% standard cash margins, every 500 bps reduction in SBC drops straight to the bottom line. Taking SBC from 20% down to 10% of revenue would almost triple owner FCF and compress the EV / Owner FCF multiple down from 86x to 32x.

In the spirit of our cannibal portfolio, a GTLB at 10% SBC when combined with GitLab’s existing cash hoard could very quickly transition from a share diluter to a compounder. In fact, they could run that playbook right away with the cash on hand, unless they have better things to spend that cash on.

The M&A Angle

At ~5.9x enterprise value to sales, 87% gross margins, and holding $1.36B in cash, GitLab is one of the premier strategic takeover targets in the software space. Rumors have circulated of interest from many parties including Alphabet (which owns a chunk already), Datadog, and even some massive private equity firms.

So why hasn’t an acquisition happened yet?

There are a few roadblocks in the way...

The Founder Vote

Founder Sid Sijbrandij controls ~45% of the voting power via Class B shares. A hostile takeover cannot happen. He has to approve and personally sign off on the price and terms or the deal has to be so good to somehow convince every one of those other shares to vote in favor.

Antitrust

Regulators blocked Adobe’s acquisition of Figma for $20B. That rightfully makes other big tech question if they can get away with an acquisition.

Google is a solid suitor for GitLab. They could integrate it into their Google Cloud Platform. But is it worth running the risk of an 18-to-24 month regulatory review? Maybe not.

PE Dilemma

Private Equity buying GitLab makes sense. It’s a company that, at the surface level, is not making the best use of its resources. In the eyes of private equity the best use of all that cash is in returns to their shareholders.

But if PE was to buy GitLab and immediately slash the SBC load the best engineers will head for the door right away. With an open-source core, some of those engineers might even build competitors by forking a codebase they’re already familiar with.

There’s more to this article for our premium subscribers. You’ll get an in depth look at the valuation including our bear, base, and bull cases along with the actions we’re taking in the TechBreakdowns portfolio.

But spreadsheets assume smooth mathematical trends. So, I did what we do here on TechBreakdowns and got right into the name. What I found was a business that is doing a fantastic job of abstracting away the complexities of dealing with emerging market payment rails. As I dug in more, however, I saw some flags worth discussing deeper.

dLocal Explained

In the United States or Europe, charging a customer money is trivial. You can setup on Stripe, Adyen, or one of the other countless mega acquirers available and the user enters a 16-digit Visa for money to transact in seconds.

In Brazil, Nigeria, Indonesia, and many other countries, the path is not the same. The vast majority of consumers in emerging markets don’t have credit cards. There are a multitude of payment options, APIs, rails, and pain points that make moving money a big chore.

So, put yourself in the shoes of a global product manager at Spotify, Uber, or Amazon. You want to be able to accept payments in these emerging markets, but you don’t want to set up a team to build to each and every rail out there. What do you do? You hire dLocal.

When Brazilian customers choose to pay instantly with Pix, it’s covered. If a Mexican consumer wants to use an OXXO voucher, no problem. In Kenya you might find a customer wanting to pay with M-Pesa, a mobile SIM balance, and you can accept that without having to build anything.

The list goes on. dLocal covers 40 emerging markets and is integrated with 900+ bespoke local banking APIs. The company has the local entities required to transact, the banking licenses each nation needs and they’re able to help navigate foreign exchange capital controls along with filing local tax returns. It is, for all intents and purposes, a payment ecosystem in a box that covers emerging markets.

This can all sound like a lot. If I’m Spotify, how do I work with dLocal? Well, it’s actually quite easy. DLO takes the geopolitical payments maze and breaks it down into three simple promises:

1. 1 direct API integration

2. 1 legal contract

3. 1 consolidated settlement (and payouts in USD / EUR to a company’s local bank account)

Any new nations or payment methods added become instant value adds to dLocal’s customers. Oh! And there’s more value too... businesses that have pay-outs (disbursements) like Uber or DiDi can be handled by DLO as well in the end user’s local currency in real-time.

The Economics of dLocal

The economics of acquirers are always fun to try and figure out. Much like the maze dLocal is trying to navigate when it comes to international payments, analysts (like myself) have to navigate a “who owns what” when it comes to revenues and cash flowing through the companies. Thankfully I got way too deep into the weeds breaking down Adyen a few years back and I already knew my way around a little bit.

Let’s assume a $100 purchase in Brazil via Pix. dLocal may collect $2.38 off that and book it as gross revenue. A big chunk of that though is what’s called “pass-through interchange” which includes your local bank fees and FX processing costs. In our case we can call that $1.66 which leaves dLocal $0.72.

That $0.72 is what you’ll find on the DLO income statement as gross profit, and it’s the number we actually care about. If we divide that gross profit into the “Total Processed Volume” of $100 we get 0.72% which is the “Take Rate.”

OK, those are the terms we need: Total Processed Volume (TPV) and Take Rate. These tell the story of any payments processor.

And the story we’re seeing here with dLocal is an interesting one. If you look at Q2 2025 the Take Rate was 1.07%. Q1 2026 it’s 0.84%. Q2? It’s that 0.72% that we just outlined.

When asked about this significant slip, CEO Pedro Arnt (formerly the long-time CFO of MercadoLibre) said a single massive ride-hailing client (we can assume Uber or DiDi) doubled volume which triggered pricing step-downs. Both a good and a bad.

Asked further, Arnt told analysts on the call “We don’t manage to a take-rate floor; we manage to incremental gross profit dollars.”

The Customer Dilemma

Looking at the “logos” that sit on the slide decks of dLocal you’d be blown away that it’s a name you’ve never heard of. They’ve got 4 of the 5 largest ride-hailing giants, 5 of the 10 largest e-commerce platforms, and the top 5 streaming platforms.

The company is also posting a massive 188% TPV retention meaning that those existing clients are growing at a significant clip on dLocal’s platform. The math here, by the way, for every $1 equivalent last year existing customers are bringing in $1.88 this year. That’s unthinkable same customer growth.

From my research though, I believe this falls apart in a few critical ways:

1. Customers demand lower costs

2. Customers have cheaper options

3. Payments will consolidate

Customers Demand Lower Costs

We see this already playing out. Customers expand and they already have negotiated contracts to provide lower rates with that expansion. This in itself is not a negative thing but when combined with the second bullet point customers having cheaper options dLocal has no other alternatives except to comply.

Customers Have Cheaper Options

A company like Uber might choose to use dLocal to enter a market but unless dlocal continues to provide lower pricing for a company like uber once the company reaches a mature scale it may choose to go direct and build its own payment rails to save on the margin.

If a company like Uber is doing two, three, or four billion dollars of revenue in a country, it makes more sense for Uber to build its own payment rails to start capturing back the 0.1% or 0.2% it may save by doing so. That 0.1 or 0.2% on billions of dollars works out to a meaningful sum of money in the market.

Sure it might not be able to tackle all the payment methods that dLocal does but by operating in the market it will see what payment methods customers prefer and then be able target majority of those.

Better yet is bullet point three where payments will consolidate over time.

Payments Will Consolidate

Over the coming years these emerging markets will consolidate onto a handful of payment methods that work for the majority. When they do consolidate onto that handful of payments it makes market entry easier for the acquirers that large multinational companies are already working with like Adyen to get a better foothold in the market.

I see all three of the reasons above as key things working against DLO over the long-term. Sure the economics look great now And you cannot argue that dLocal is growing at a phenomenal clip But when it comes down to long term growth and how this company will succeed over the coming five or ten years I believe we’re more likely to see struggles than easy success.

The DLO Moat

I always like to make sure we cover the moat and talk about what sets this business apart for dLocal it is dealing with regulatory bureaucracy not necessarily the technical implementation.

And we’ve mentioned Adyen a few times in this article but comparing it to Adyen, Adyen is perfect from a technical implementation standpoint deals with regulatory stuff too But where dLocal stands out is on making sure that it can operate in all the markets where nobody else wants to deal with the red tape.

dLocal holds 40 direct central bank licenses across 26 markets, and operates directly in 39 markets in total with 217 integrated APMs (different payment technologies).

The moat is all the accumulated regulatory bureaucracy, local banking clearing accounts, the merchant of record compliance in all these different nations. It would take a company like Uber or Spotify 18-24 months and millions of dollars in collateral to get a central bank license. They don’t want to do that, so they need a partner, and dLocal is a partner they can pick across those 39 different markets.

As mentioned above though, that terminal vulnerability is the standardization of payments. Pix in Brazil, UPI in India, Bre-B in Colombia. These rails make the difficulty of collecting local payments go down, and an easier market erodes dLocal’s complexity premium.

The Financial Snafu

Finally before we dive into the premium portion of the article I did just want to call out what i consider to be a fantastic financial snafu.

dLocal is a fast growing company that has a 3% market share in the areas it operates. The company also believes it only has somewhere around 15% of its existing customer TAM. Sounds like a great position to be in, and a position where you’d want to, as a business, re-invest your cash to capture that market.

Instead of reinvesting in the business, dLocal has a dividend and share buyback program that’s returning 30% of FCF to shareholders.

Returning cash is better than funding low-return projects, but by all accounts there should not be low-return projects in this business. A true reinvestor or moonshot would be focused on building a massive business and then focus on returning capital once they get there.

I hope you enjoyed the article. Up next, if you’re interested in more, our premium subscribers will get a look at the owner economics, a float reality check, the valuation, and our verdict on whether to add this to our portfolio.

Disclaimer: Not personalized investment advice.

Adobe as an investment makes my palms sweat. The reason my palms sweat is the same reason that the stock is more than 60% down from its all time high of ~$688 back in late 2021. That reason: AI.

At roughly eleven times the cash it throws off, Adobe is priced for a business that has already stopped growing. Which makes it one of two things: a generational bargain because AI fear is overblown, or a classic value trap because AI really does hollow it out. Our quantitative screening cannot answer that question, only a deep dive qualitatively can.

The AI Question

If we look at Adobe through the Christensen lens we’d say that incumbents generally win sustaining transitions and lose disruptive ones. Bulls lean on generative AI being a sustaining change for Adobe while the bears call it completely disruptive.

Me, I say it’s both, and it splits roughly down Adobe’s customer base. Adobe doesn’t hand us a clean consumer-versus-enterprise line anymore (more on that reporting mess below), but think of it as two ends: the casual consumer, prosumer, and student on one side, and the professional and the enterprise on the other.

On the “cheap” consumer and prosumer side of the market, one might consider Adobe dead on arrival. Speaking from personal anecdotes I can say that whenever I need to make quick tweaks to an image, I’m on Canva. Whenever I need a flyer or a poster, or even something a bit unique it’s more likely that Gemini and ChatGPT are helping me out.

I’m not alone in this either. Just look around at the event posters for your local establishments. Nearly every one of them is using ChatGPT to get that made, it’s incredibly easy to tell.

This casual end is the slice most exposed to free AI tools. While I don’t expect it’d drop to absolute zero, I do think it’s on the safer side to assume that Adobe isn’t going to be the tool families reach for when editing their photos. That slice erodes. (Hold that thought, though, because when we get to the actual numbers, the group carrying most of those consumer dollars is growing faster than the professional one, not shrinking.)

On the opposite end of the market, the enterprise, they need the likes of Adobe. An enterprise cannot open up ChatGPT and say “make me a flyer.” They can’t open Midjourney and say “make me a commercial.” They can, however, open Adobe’s suite of products and utilize AI to speed up their overall delivery.

The enterprise moat that Adobe offers also comes with IP indemnification, a real benefit to the enterprise. Firefly is trained on licensed and public-domain content, Adobe Stock and openly licensed work and material whose copyright has expired, and Adobe offers enterprise customers an IP indemnity against copyright claims on the output of eligible Firefly workflows.

Other enterprise tools exist to go along with it. Adobe’s GenStudio is built to be the enterprise content supply chain. Its ending ARR grew more than 25% year over year, with marquee customers like Coca-Cola, and it now touts Brand Intelligence, contextual brand rules that AI agents can read to keep every asset on brand, a lock-in that could make it difficult to move away from.

There’s also the Experience Platform, plus Marketo, Workfront, and the recently acquired Semrush in the arsenal. Looking across the vast array of offerings, and looking at their growth, it leads me to think that a vast amount of the dislike for Adobe centers around Photoshop. I 100% believe that Photoshop will never be as great as it once was, but I do not think that Photoshop is a reflection on the future of Adobe.

What’s Actually Growing At Adobe?

When faced with a company that’s on shaky disruption ground I usually like to turn to the segment reporting. You can, quite easily, spot if that shaky ground is shaking them.

You’d think that’d be easy. Pull up Adobe’s segments, see which are growing, which are dying, tie it back to the bigger picture. Well, Adobe made that hard this year.

For a decade you could see Digital Media ARR, Creative Cloud, Document Cloud, the Experience business, each on its own line. Now, management has collapsed the whole thing. The Creative Cloud and Document Cloud split is gone, the three old segments are folded into one, and we get a single headline number: “Total Adobe ARR,” which grew about 11.5% to $25.2 billion exiting fiscal 2025, and $27.1 billion by Q2 of this year.

You can look at this in a couple of ways. There’s the cynical read: they stopped showing the numbers at the exact moment the market decided AI was going to kill Creative Cloud. Or there’s the more charitable read: the business is so cross-woven that the old lines stopped meaning so much. It’s probably a little of both, with the timing making it seem even more suspect.

What Adobe does still give us is two customer groups, both of which continue to show growth. In Q2 of fiscal 2026, their subscription revenue grew:

1. Creative & Marketing Professionals - about 13% Y/Y

2. Business Professionals & Consumers - about 16% Y/Y

(Those are customer groups, not segments. Adobe now reports a single reportable segment.)

Is there then anything shrinking at Adobe? Yes, there is, but it’s a legacy Publishing and Advertising business that fell 7% to $256 million, about 1% of revenue and freshly written down, plus a few low-margin services. All told, not worth much and effectively a rounding error in the grand scheme.

So we’re looking at a company that’s being priced for death, growing at 13% per quarter, where the only shrinkage we’ve seen has come from scraps they wrote off, and where the more consumer-weighted of the two customer groups is actually the faster-growing one.

The Bear Case: AI

The bear case that we’ve been dancing around throughout is the rise of AI. This would make sense if Adobe was standing still, but it’s not, it is actively selling the AI.

Adobe has a book of AI products, standalone tools and add-ons like Firefly and the Acrobat AI Assistant that customers pay extra for (paid Acrobat AI users more than doubled year over year).

Firefly has generated more than 22 billion assets, and the money is starting to show: Adobe’s AI-first ARR tripled year over year to more than $500 million as of Q2. People are paying for this, be it by credits, subscriptions, or an enterprise seat.

I’ve come to think of Adobe as the Switzerland of generative AI. It doesn’t matter if Google, Anthropic, OpenAI, or any other model is best, Adobe puts them in the app and charges for the workflow wrapped around them. Models are a commodity racing to zero, but the place and way that work gets done is not. Adobe currently owns that place.

Now let’s not mince things here, Adobe does face a lot of competition across its key areas. In the design space there’s Figma, the company Adobe tried and failed to buy for ~$20 billion before regulators killed the deal at the end of 2023. In the graphic design space, OpenAI can largely get quick jobs done as can Canva. Video editing has CapCut. Sending of emails and marketing tooling has hundreds of competitors touching each piece of the stack.

So, it’s not all smooth sailing, but Adobe is showing that the company is able to grow despite what seem like incredibly insurmountable headwinds.

The $25B Machine

Alongside everything we’ve spoken about, there’s the part that made Adobe the top stock on our Cannibals screener: buybacks.

In April, Adobe’s board authorized a new buyback: up to $25B of its own stock, running through April 2030. Stacked on the program already running, that leaves about $26.8B of buying power authorized.

Adobe is worth about $105B today. A buyback of $25B is close to a quarter of the whole company. At around $265 a share, $25B retires roughly 94 million shares, and Adobe has about 397M outstanding.

Adobe needs to generate this cash though to be able to buyback the stock. Can it? Well, the company threw off about $10B in free cash flow last year. $25B is about 2.5 years of cash flows spread across a four year program. If things remain flat, Adobe generates $42B and gives back just $25B of it. No debt or financial engineering here, it’s purely funded by the printer.

Shrinking the count of your shares by close to a quarter ends up growing EPS by about a third, all without lifting a finger. Spread that over four years, that’s ~6-7% of EPS growth before Adobe sells a single new subscription.

The paradox of this whole thing... the cheaper the market makes Adobe now out of the AI fear, the more of itself each of those dollars can buy. The market’s hatred, while Adobe continues to grow, is doing the compounding for you.

Paid subscribers get a valuation breakdown and what we’re going to do when it comes to our portfolio.

*Everything here is research and a model portfolio we track in the open, not personalized investment advice. Do your own work before you act.*

We’re here, though, to craft a portfolio of tech that helps us sleep soundly at night while offering the potential for outsized growth. One of the first names that came to mind, and the first company we’ll be breaking down is MercadoLibre (MELI).

MELI meets the quantitative criteria for a Reinvestor. We found that if we can get ourselves buying the top half of the Reinvestor cohort, we could be onto 30%+ gains per year. So, here’s the first swing!

The Rails No One Would Build

It’s a name most Americans don’t know. A quick search would very quickly lead you down a path of learning it’s “the Amazon of Latin America.” But, just like Amazon, the company isn’t stopping at being an e-commerce platform.

Before we get to all that fancy stuff though, let’s dive into that comparison with Amazon a little more. Amazon is a fantastic company, I’m not going to doubt that even one bit (I’ve been a small holder since 2009, and it’s grown massively).

Amazon though had some benefits in getting started. It was plugged into a country that already had FedEx, UPS, Visa, and a banked population. MercadoLibre started in a continent that had none of that, so MELI built it itself.

There was no logistics network to deliver goods, MELI built one. There were no payment rails for the largely unbanked, MELI built them (and now is its own bank). Along the way it survived the dot-com bust that wiped out the vast majority of its Latin American peers, it turned eBay from a rival into a shareholder that later cashed out, it held Amazon off at number two, and it kept growing through Argentina’s currency crises and triple-digit inflation. It has been a 100-bagger since its 2007 IPO.

MELI Today

Today, the company operates in 18 countries offering various parts of its massive ecosystem in each one. That ecosystem is made up of three main businesses: E-Commerce, Acquiring, and Fintech Services.

The company is LatAm’s E-Commerce leader, continuing to grow at a significant clip and showing no real signs of slowing down.

MELI’s ad business is growing at 73% year-over-year, and digital is still only about half of Latin America’s ad spending versus three-quarters in the US.

And they’re building a massive bank as a sort of side gig. 83 million monthly active users. $14.6B credit portfolio from lending and $20B of savings and investments through the platform.

Oh! And they’re one of the largest acquirers in the continent. While not quite South America, I did recently go to Mexico and every payment terminal I encountered had the company’s Pago logo displayed. Taxis, vendors, restaurants. MercadoPago has done a fantastic job at getting their solutions into the wild.

The vast ecosystem is the company’s competitive advantage. Merchants can use payment systems, take loans, bank with Mercado, and even ship what they’re selling via the logistics network.

The Founder Who Chooses to Invest

Marcos Galperin is the cofounder and former CEO of MercadoLibre. The stories above? He’s the one that made them happen.

He wrote the business plan for MELI in a class at Stanford in the late 90s, after studying eBay of all things. As the story goes, a professor asked him to drive a visiting investor to the airport, a private equity guy named John Muse. Galperin pitched him the whole way there, and Muse wrote one of the first checks. He pulled in a few classmates, Hernán Kazah among them, and they went and built it. That was 1999. Galperin ran the thing for the next 26 years.

On January 1, 2026 Galperin stepped down as CEO, handing the reins to Ariel Szarfsztejn, the insider who had been running MELI’s commerce business. Galperin remains with the business as its Executive Chairman, which in founder terms usually means he kept the one job that matters to us: deciding where the money goes.

And that’s the whole reason MELI shows up on our Reinvestor list. Given the choice between showing Wall Street a fat profit today, or feeding that money back into the machine to make the business harder to kill tomorrow, Galperin picks the second option every single time.

He did it with logistics. For years MELI ran its own shipping network at a loss while everyone waited to see if it would ever pay off. He did it again in 2016 with free shipping, ate the hit to margins, and let it compound into a moat. And he’s doing it right now with the credit business. That one we’ll get to in a minute, because it’s the whole reason the stock is down, and the whole reason it’s interesting.

The Number to Watch

If MELI stock is as good as it looks, why has the stock been on such a roller coaster and down ~20% over the last year?

Well, we can sum it up with one number, and a number any buyer of MELI should be paying close attention to when earnings are released on August 5th: operating margins.

Last year, MELI was keeping about 13 cents for every dollar it took in. Last quarter it was down to 7 cents. Wall Street has seen margins cut in half, and it did what Wall Street does: it sold.

Did margins collapse because customers are leaving? Did they collapse because a competitor is undercutting them? Nope. Nobody undercut them. Nobody left. The business didn’t get weaker.

The margin fell because Galperin spent it. When a loan gets booked, so does the expected loss the moment that loan is written, even before it earns a single cent back. So, when your loan book nearly doubles in a year, that up-front charge explodes and eats today’s margin. There was also some big spending on shipping and marketing to defend their Brazilian territory.

So, the margin hasn’t broken. It is being reinvested.

A Risk To Watch

Those operating margins will be something to watch on August 5th, but another thing is going to be that credit book as a whole. That same credit book that’s ballooning and eating margins could be a burden on the company as a whole.

It’s a $14.6B pile of consumer loans, in Latin America, growing at 87% per year. Sure, MELI has already reserved for its bad loans, but those are only management’s presumed losses. If we see signs of early delinquencies rising, things might not look too rosy on the other side of earnings.

So that, to me, is likely the biggest risk the company faces near term. It’s a fantastically built business with lots of arms, and lots of “outs.” But the future of the business is being staked on fintech, and it’s an arm of the business that has seen solid growth.

Verdict

Our paid subscribers get to go a little deeper, but I do just want to call out here for all readers my thoughts. I believe that MercadoLibre is a fantastic business offering investors diversification away from U.S. tech.

Still sitting within the leadership structure is Marcos Galperin who has proven to be an elite-level capital allocator over the last nearly three decades of this company’s existence.

That said. This wonderful business is not just a buy-it-and-forget-it. It’s one that you keep your eye on, particularly the loan book.

Cannibals shrink to win. Today’s topic is Tech Reinvestors and they do the opposite. They feed every dollar back into the machine. Both compound, they just take different paths towards doing it.

Tech Reinvestors Explained

I think the reinvesting analogy outside of tech helps make this a little more understandable, and there’s no better company to point to than Costco.

Costco has a fantastic product. They build a warehouse, gain a ton of members, keep them happy and then parlay that cash into another warehouse. Soon, assuming Mike Judge was right, Costco will be our everything.

That reinvestment is what has set Costco stock apart over the years. It has a template, it takes the cash it earns and reapplies that template somewhere else in the world. Well, it’s a little harder for tech companies, but we found a way to screen for them all the same.

How They Perform

Much like the Cannibals, we have back tested this thing. 2006 through to 2026, buy and hold for 3 years any stock that meets the criteria (which are below).

Overall annualized return came out to 15.7%, max drawdown was a whopping 44% but the doubling in 3 year number was 22%.

When compared to the entire tech universe that we could have selected from, the total return there was 11.4%, same max drawdown, and a 16% doubling rate.

Overall, the screen performs better than the cohort of tech that we’re selecting from.

Interestingly though, unlike our cannibals, these stocks do not serve us well if we buy and hold “forever.” Extending out the hold to forever leads to a 12.9%/yr return.

Why the split? Reinvestors tend to carry a premium (just look at Costco in our example). They win for a few years, marry them for a decade though and the market chooses to re-rate. This portion of our portfolio then ends up being, largely, a rotational portion.

In short, reinvestors beat the tech market, 22% of them double. But, buy-and-hold them forever is not a path to success. We own them for diversification, rotate when the timing is right.

We’re Qualitative Geniuses!

It did get me wondering... tech has been phenomenal, this screen averages ~30 names per year (2 on low end 51 on the high). What if we were actually qualitative geniuses (hey, we’re going to try and be), how would we have performed?

The results are inspiring, and honestly are going to make me show up and actually put in the work right here to be correct. But, if we were to split the stocks right down the middle, and we were able to select the best half every time, holding again for 3yr our return becomes 30.7%. And here’s the real kicker: that top half had a 0% blow-up rate, while the bottom half lost more than half its value one time in five. The edge isn’t picking the winners, it’s not stepping on the landmines.

Just for fun, the absolute ceiling: if you could somehow nail the top 3 stocks in the cohort every single year, you’re up to a 51.5% annualized return. Nobody picks the top 3 in advance, though, so file that one under fantasy.

We don’t need the fantasy, though. Nail the achievable 30% from today onwards, starting with $100k, and you land north of $20M by 2046. (The perfect-foresight version would clear $300M, but I’m not going to promise you something no human can actually do.) Sign me up for the ride either way.

The Criteria We’re Looking For

Super simple set of criteria, all gates need to be true to pass. We scan the universe of U.S. listed $2B+ market cap tech companies. From there we’re looking for:

1. Gross margin > 40%

2. 3-yr revenue CAGR > 15%

3. Reinvestment rate / gross profit > 40%

4. Rule of 40 > 25%

5. Growth efficiency > 0.15

6. Share count growth -2% up to 6% (nice and controlled)

7. Altman Z > 1.0

8. Beneish M < -1.78

In plain English: high, durable margin, real growth, and genuinely plowing money back into the business.

What’s the list look like?

Like with the cannibals, here’s a teaser of what meets the criteria. If you’re a premium user the full listing should show up on this page, and we’ll maintain a semi consistent updating version in the datasets.

In total 35 names hit the criteria today. Here’s ten of them for those reading for free:

1. MercadoLibre

2. PDD Holdings

3. Klaviyo

4. Global-e

5. Trip.com

6. GitLab

7. Zscaler

8. Taiwan Semiconductor

9. Rambus

10. Kanzhun

Next Up

From here I want to go ahead and start diving in. There is a third pillar, the “Activists,” that we’d originally planned as a full piece. Honestly though, it’s the minor one (the numbers say it only pays off when a human spots a real catalyst), so it’s better folded into a breakdown when the setup shows up than given its own post.

If you’re interested in following the journey, consider signing up for a premium subscription. The goal from here on out is to push toward that 30% (the top-half number, if our selection is good) in as sleep-well-at-night a manner as we can. We’re not chasing moonshots; we’re breaking down genuine companies and the resilient businesses they’ve built.

From time to time, sure, we will dive into some interesting high-growth tech, but that’s a very speculative slice we’d never let exceed 5% of the portfolio, and only if we ever dabble at all. Sleep well, get big returns is the goal and it all starts fresh in August 2026!

*Everything here is research and a model portfolio we track in the open, not personalized investment advice. Do your own work before you act.*

“It?” Yeah, we should probably start with what “it” actually is.

I’ve long been in love with the investing world, and the tech world hasn’t been far behind it. Throughout my time on the internet I have, at various points, attempted to build a number of investing projects, and TechBreakdowns is another iteration of that.

The rationale going back and forth has been market inspired. I see tech soaring I say “I work in tech, I know tech, I can write about that.” But, when tech takes a turn, my mentality shifts back to what I’ve always been fond of: high-quality.

High-quality can be boring, right? It’s the names that have a system down and they rinse & repeat for many years grinding out little-by-little. It’s that little-by-little that adds up though.

So, what is “it?”

“It” in this case is my clash of two worlds. We’re here on TechBreakdowns to talk about tech, that’s true. There is, however, no reason that it can’t be high quality tech.

I have adapted a couple of methodologies that I’ve long used in my regular portfolios to become tech centric. I’ll use them to guide content and discussion here, as well as utilize them to build a TechBreakdowns portfolio that is built for growth, and lets us sleep soundly at night.

Sure, we may dabble in those high-growth names from time to time, but there will be no chasing the beta.

Three Components

There are three components to the screeners / rulesets. We’ll cover Cannibals today, then move onto Reinvestors in our next piece and end on “Activist.” All three are playbooks. All three were built a long time ago and have been backtested (we’ll talk about that), and all three have undergone a little revitalization to focus on technology.

The Cannibal

In a paragraph, the Cannibal is a mature, dominant, high-margin, cash-gushing tech company that buys back stock at a fair price.

These names are “defensive” by nature. One could easily consider this the ruleset Buffett would quantitatively follow if he were solely interested in tech.

Emphasis, of course, on quantitative. We’re focusing here on finding companies first and foremost, then we’ve got to do the leg work qualitatively.

The question you’re most likely dying to know though is “how does this thing perform?” Well, you’re in luck, because I backtest literally everything.

To backtest this, a “high quality” screener for the tech world, I took the last 20 years of data, accounted for survivorship bias in the data, and then set the rule that any stock meeting what we look for is bought and held for three years.

The benchmark is the entire universe the cannibals are drawn from where market cap is >$1B. Every “Technology” and “Communication-Services” stock is up for grabs. In all, those stocks returned 11.8% annualized.

The cannibals: 16%.

Better yet... average max drawdown was 31% for the cannibals versus 43% for the “tech market.” Better than that? Sure... the “blow-up rate” was just 2% versus 11% for the benchmark (this is the amount of times a buy was down >50%). We also see the doubler rate (stocks 2x’ing in a 3 year period) of 18% versus 16% for the market.

So, overall, fantastic. Now let me be honest about *why*, because it matters: most of that extra return is the “quality premium” — a well-documented factor anyone can tilt toward, not some secret sauce I cooked up. The real prize is the *defense*. You compound right alongside the tech market, but you blow up a fifth as often. That’s a version of buy-and-hold that actually lets you sleep.

In short:

How Cannibals Looked Through The Years

A quick look at who was topping the list in some random past years. These companies made the cut, and without qualitative intervention would have been bought and held for three years.

2006

The oldest data we have, but it’s an incredible list that would have returned 1215% over the last 2 decades versus 648% for the S&P 500. A 14.1% CAGR versus 10.8%.

Again, just to call out, buying blindly on quantitative measures...

AND this was on the eve of the GFC. Beating the market, even through that, is considerable. $10,000 would have been $132k, not the $75k invested in the S&P index.

The list included just 9 names: Microsoft, Cisco, Intel, Oracle, Texas Instruments, Applied Materials, Maxim Integrated, Intuit — and, amusingly, Coca-Cola, which snuck into the *tech* screen purely on its fat gross margin (and even it roughly matched the market at ~10%). One footnote: Maxim was acquired by Analog Devices back in 2021, so that slot simply rolls into ADI — a buyout is just another way a Cannibal hands value back to its owners.

2015

11 names made the list in 2015, many familiar: Apple, Microsoft, Oracle, Intel, Cisco, Expedia, Gartner, F5, Jack Henry, Manhattan Associates, Marvell.

Even back then it was boring, but if you’d have bought that portfolio in 2015, you’d be happy in 2026. There are some roller coaster rides in the mix though.

2024

Even going 2 years back there’s a list of very familiar names in the bucket.

Apple, Alphabet, Meta, Adobe, Qualcomm, Applied Materials, ADP, KLA, Cadence, FICO, Verisk, EA, Gartner, Manhattan Associates (again), Dropbox, InterDigital.

Alphabet and Meta were fresh to the list. They’d hoarded cash through COVID and found themselves returning that cash en-masse as they matured. This screen captures the exact moment a company “grows up” and becomes mature.

What Are The Criteria?

We’ve made it this far but haven’t hit on the criteria, let’s do that. A stock meets our Tech Cannibal screen only if all 10 of the following criteria are true:

1. Gross margin > 40%

2. ROIC > 15%

3. Owner-FCF margin > 12%

4. Share count shrinking > 0.5%/yr

5. Net buyback yield > 0.5%

6. Revenue growing (look for 3yr CAGR > 0%)

7. Rule of 40 > 20

8. Net debt / EBITDA < 3.5x

9. Altman Z > 1.8

10. No Beneish earnings-manipulation flag

The stock must also have a market cap of >$2B, so I guess that makes it 11 criteria.

At the time of writing there are 25 names that meet what we’re looking for and, going forward, we’ll be covering these companies along with our other investment models to make a trio we can track for the longer term.

Interestingly, two names you might *expect* — Microsoft and Alphabet — didn’t make today’s cut. Both were Cannibals as recently as 2021–24, but their enormous AI and cloud capex has shrunk their net buybacks enough that they’ve quietly slipped back into “Reinvestor” mode. The screen catches companies growing up — and, occasionally, un-growing-up.

Top 10 Tech Cannibals

I’ll leave you with ten names that meet the above criteria. I’ve ranked them using a bit of a weird methodology, but largely focusing on capital returns.

1. Adobe

2. Expedia

3. GoDaddy

4. Verisk

5. Fair Isaac

6. AppLovin

7. Qualys

8. eBay

9. Corpay

10. Cirrus Logic

The question now turns to the qualitative side for me. We have a lot more names because a lot more tech companies have matured over time, and I’m not sure that buying blindly is going to be beneficial. That’s the point of this publication though, to track that and figure it out.

The complete list of stocks will be provided for paid users below. We’ll also provide the non-tech version of this for those that are curious in our premium data sets. As always, happy to answer questions on the whole thing.

*One last thing, and an important one: none of this is personalized investment advice. It’s research and education — the output of a quantitative screen plus my own opinions. Do your own homework before you buy anything.*

This week we’re breaking down an autonomous trucking company, Aurora. The central question here is not whether autonomous trucking happens because it does. The question that we need to answer is: who wins and are the winner’s economics as good as everyone hopes?

The reason we’re looking at autonomous trucking is because the total addressable market is absolutely massive. We know that the TAM has to occur because nobody’s just going to sit on their hands and not let this happen. We want to figure out who’s going to win and place our chips on the table.

And in looking at autonomous trucking, we will end up covering other companies in this space over the coming weeks. Companies like Kodiak and maybe taking a look at some private competitors in the space too.

Now for me this is a very interesting topic as well. I drive a Tesla Model Y and I say “drive” but realistically it drives me. The car uses a completely different approach to what Aurora is proposing for trucks and it gets the job done for me. So part of this article will be looking at what Aurora does differently from a company like Tesla and seeing if that is an edge in the long term.

What is Aurora?

In the fewest possible words, Aurora is a company that is trying to solve autonomous trucking.

The company today is in the midst of a capital-heavy pilot where it runs autonomous trucks between Dallas and Houston up the I-45 corridor. This is an insanely straight road, very little going on. It’s a highly common route for a lot of freight.

In addition to that route, Aurora is also running between Fort Worth and El Paso and has a 1,000-mile lane between Fort Worth and Phoenix.

As mentioned this is a capital-heavy pilot today where Aurora is building these trucks and then leasing them out as though they were drivers for hire. So anybody looking to move freight between Houston and Dallas can contract with Aurora to move that freight.

Long-term though, Aurora does not want to be in a capital-heavy business. They want to transition into a capital-light driver-as-a-service provider.

In this second phase when they are a Driver-as-a-Service provider, Aurora will provide the software and necessary hardware components for fleets owned by the carriers themselves. Aurora would then charge a per-mile subscription fee for the Aurora driver system.

And that right there is essentially the investment case for Aurora. Can this company get from an expensive phase one pilot, where they are responsible for all the hardware, all the trucks, all the fleet management aspects, to a high margin primarily software business before cash runs out or before competitors do it?

To help them get there they do have a lot of good partners. A lot of big names and logos on the slide decks. OEM names like Volvo, PACCAR, International. Nvidia is a partner for Compute. And they also have customers like Hirschbach, Werner, Schneider, FedEx, Uber Freight, Ryder, and McLane.

How the Aurora Driver Actually Works

I mentioned in the intro to this piece that I own a Tesla Model Y and I use FSD. Conveniently FSD version 14.3.3 adds this nice new stats screen so I can actually show you how much I don’t drive this car.

As you can see 95% of my miles are not driven by me; they are driven by the vehicle itself and I rarely have to intervene. Most of that 5% is when driving through parking lots or on sections of road where I’d prefer to go a little bit quicker than the car would like.

I’m not just saying this to gloat. I am saying this because Tesla has a very different approach to automation versus what Aurora is doing.

Tesla has placed a bet on vision. The car sees the world through cameras alone and a stack of neural networks turns what they see into driving decisions. There is no LiDAR and that’s on purpose. Elon Musk has previously called LiDAR a crutch. The logic behind that is simple: humans drive with two eyes and a brain so a good enough camera and a good enough brain should be able to do the same. And then you train that system on billions of miles of fleet video and let it learn to drive the way people do. When it works for me and it mostly does, it feels less like a robot following rules and more like a competent driver who happens to live in my dashboard.

Aurora has gone the opposite way. The Aurora driver does not lean on cameras alone; it carries more than two dozen sensors. Cameras, yes, but also radar and lidar, all working at once, each one covering the blind spots of the others.

The part of the Aurora stack that the company is proudest of is its LIDAR called FirstLight. Most LIDAR fires a laser and times how long the light takes to bounce back, which tells you how far away something is. FirstLight uses a different method called FMCW and it adds one trick that matters a lot. In the same instant that it measures distance, it can also measure how fast every point is moving. A camera has to watch an object across several frames to estimate its speed, but FirstLight just knows. It also sees a long way, with the generation currently being released designed to see around 1,000 meters. Bonus: it works in the dark.

That is just the hardware. On the software side Tesla and Aurora’s philosophies split even harder. Tesla lets one large network learn the whole task end to end while Aurora calls its approach “Verifiable AI.”

Verifiable AI blends machine learning with hard-coded rules of the road built so that the company can test them, check them, and prove them. Something like stopping at a red light is not a behavior Aurora hopes the system has picked up from enough video. It is a rule it must follow.

Chris Urmson, Aurora’s CEO, helped build Google’s self-driving program and he has been very blunt that solving driving with pure end-to-end systems leaves you playing whack-a-mole with edge cases forever.

So why does any of this matter if cameras already work well enough in my car?

Because a truck is not a Model Y, a loaded semi weighs up to 80,000 pounds and needs around 500 ft to stop from highway speed. A car needs less than half of that. Spotting a stalled car a thousand feet away instead of 150 turns a hard problem into an easy slow down. Knowing the car’s speed the instant you see it instead of a beat later buys you distance that you cannot win back. If sun glare or a smear on the road blinds one sensor, you badly want others that can still see.

And that is the case for Aurora’s heavier and more expensive setup. Redundancy and range count for more when the thing driving itself weighs 40 tons and there’s no human around to handle it.

My own car does so much with so little and the LIDAR-and-radar-and-everything approach looks like over-engineering. The sort of thing that you’d build when you don’t fully trust the software. After digging in, I still think both things are true at the same time. Tesla has gone a long way to prove that vision alone can drive a car very well in most conditions. Aurora is chasing a harder target where “very well in most conditions” is not the bar because the failure mode is a driverless 40-ton truck.

I’m not ready to call the extra hardware a permanent edge. Sensors themselves keep getting cheaper and software keeps swallowing problems people swore it never would. But for the exact job at hand, taking the driver out of a heavy truck on a highway today, the case for seeing far, measuring speed directly, and never staking the whole system on a single sensor is a strong one.

A quick truck driving primer

I realize we’re going down this path of talking about trucking but haven’t actually spoken about how trucking works today. I mean it looks pretty simple from the outside, right and I guess in a sense it is pretty simple. Somebody picks up a load and they have an agreed-upon rate to deliver that load somewhere else.

Those agreements are usually priced per mile. If you’re delivering a load 1,000 mi, you might charge $2.25 per mile or $2,250 for that load.

You then, the truck driver, would deliver that load, submit all the necessary paperwork, and collect your $2,250 on a net 30 basis, 30 days later.

After you’ve delivered that load and submitted your paperwork, you are now looking for your next load. There’s a near certainty that that next load is not set exactly where you just dropped a load off. Which means you have to drive something called deadhead miles.

Deadhead miles are miles that you, as a truck driver, are not being paid for. Instead it’s driving from one location to another to pick up your next load, which you are being paid for. Deadhead miles are roughly 16% of all miles driven by a truck driver.

Human drivers also have to deal with the intricacies of being human. Humans are only legally allowed to drive for about 11 hours a day give or take. There are a lot of rules that come into play but you cannot drive as a human for 15 hours. Do you know what can drive for 15 hours? That’s right, an automated truck.

Do you know what’s also not wasting hours on driving deadhead? Again an automated truck.

So when you consider trucking at this very high level, there are a ton of benefits to automated trucks happening, which is why I said earlier in the article this will happen. It’s not a case of if, it’s just a case of when and who.

And so talking about price for just a minute, Aurora’s DaaS target is about $0.85 per mile. The company believes that its truck will be able to drive roughly 250,000 miles per truck per year versus a human that is really maxing out about 125,000 miles. So these things end up being insanely cost-effective (if delivered).

Because this is a hub-to-hub or terminal model that Aurora is building out, there’s likely to be a lot of other efficiencies gained there. I’m sure there will still be the trouble of figuring out last mile and how those intricacies work but just doing over-the-road trucking at a rate of 85 cents per mile will change the economy massively.

The TAM, Guaranteed

The total addressable market, which we’ve alluded to throughout, is without a doubt the most interesting story. It is rare that a market will come to fruition. It is rare that you know the size of a market before it can even exist but in the case of automated trucking we have exceptionally good ideas.

And yes this total addressable market is pretty much guaranteed. The winner though is not and Aurora could absolutely be it but it’s also equally likely that one of the competitors gets there first.

So how big is this total addressable market? Well, Goldman Sachs puts the global autonomous trucking market near $560 billion by 2035 ($105B in the U.S.), which is larger than the robotaxi number that the company has put out there.

They believe that the cost per mile crossover is going to be happening around 2028. That crossover is when it is more cost efficient to hire a robot than it is to hire a human.

Other forecasts include McKinsey, who is more bullish on the US. It forecasts that autonomous heavy-duty trucking could reach $178 billion by 2035. McKinsey’s global figure lands around $600 billion, which is closer to Goldman’s.

Overall US trucking generated $906 billion in gross freight revenues in 2024, just down from about $1 trillion in 2023. It employs 3.58 million professional drivers and 8.4 million people in trucking-related jobs.

Driver wages are estimated to be somewhere around 32-40% of trucking operating costs, which is the single largest cost bucket.

This is a lot of big numbers so let’s bring it back to the one number that really matters: how many miles are being driven each year by truck drivers over the road on highways?

The number I was able to come up with is 196 billion combination truck (where Aurora is targeting) miles in a year. If we roughly guesstimate that 80% of these miles are driven over the road or on highways, that gives us 156 billion miles with which we can build our Aurora model.

156 billion miles at a cost of 85 cents per mile is $132B. Aurora, of course, won’t capture every single highway mile with its DaaS fee but capturing 10% of those highway miles at that 85 cents gives the company potential $13.2 billion in revenue.

Who else is here?

Aurora is not running this race alone. There’s a crowd. The one that matters most is Kodiak, and it’s close enough that it gets its own breakdown next week. There’s a field of private names too, like Waabi and Torc, that I’ll dig into when I cover them. For now just know the competition is real.

What’s more telling is how short that list keeps getting. This is a brutal business. The tech is hard, the timelines are long, and the cash burn is enormous. Companies don’t limp along in this space. They die.

Look at who already tried. UPS, the largest carrier in the country, bet on TuSimple and on Waymo’s trucking arm. TuSimple delisted and moved to China. Waymo shut its trucking program in 2023. Knight-Swift, the largest truckload carrier, bet on Embark. Embark collapsed.

So the two biggest, most sophisticated buyers in trucking placed early bets and both got burned. Neither has committed to anyone since. Picking the winner early has been a great way to lose money. It also means the largest buyers in the country are still on the sidelines, uncommitted, waiting for someone to prove this at scale. That business is up for grabs, and whoever wins it could see their deployments jump overnight.

It might just come down to last man standing.

Product Noodling

This section will likely be a staple going forward in our breakdowns. A chance to talk about what a company could become, about the paths the product could take. Product Noodling.

Aurora is a product in its infancy. We know the company wants to go down an asset-light path and build a Driver-as-a-Service offering that carriers of all kinds subscribe to. That’s fantastic. But I want to focus on the part of that dream nobody is really looking at: the logistics.

Today every truck Aurora runs drives over the road only. It runs highway to highway, terminal to terminal. Those terminals are where the over-the-road leg ends and the last-mile leg begins. Until a truck can do the messy last mile by itself, that terminal layer has to exist. The question is who builds it, and who captures the value.

Here’s the tension. Aurora’s asset-light dream has a physical-world dependency. Software trucks still need a national grid of places to stop, and that grid does not exist yet. Someone has to pour the concrete. So Aurora either compromises the asset-light model and builds it, or it stays pure and hands that value, and some leverage, to whoever does.

Because no human is in the truck, these aren’t truck stops. They’re where the truck gets refueled or recharged, where cargo transfers from the long-haul leg to the local one, and where inspections and maintenance happen to keep the thing running. That’s a real operation. There are three ways it can shake out:

1. Aurora owns it. Most control and the fattest margin per site, but it torches the asset-light story and ties up enormous capital. Given everything Aurora has said, this is the least likely path.

2. Aurora franchises or licenses it. Aurora sets the standard, certifies the locations, takes a fee, and lets partners put up the capital. This is the real Supercharger parallel, and it keeps the model asset-light while leaving Aurora strategically in the middle.

3. Somebody else owns it entirely. A Pilot, a Love’s, or a logistics landlord builds the network and Aurora just plugs in. Lowest capital for Aurora, but it captures none of that layer and risks the terminal owner having leverage over it.

Option three does nothing for Aurora. One and two both do, and option two is the interesting one, because it’s the path that fits the business they say they want to be.

So what could licensing actually be worth? Let’s noodle it. Picture a mature network of, say, 150 terminals at the major over-the-road handoff points across the country. If Aurora licenses its standard and takes something like $1 million per terminal per year, in fees, certification, software, and a cut of the services, that’s $150 million a year. Push it to 400 terminals as the network fills in and it’s $400 million. That is real money, and notice what kind of money it is: high-margin, recurring, and asset-light, because the partners funded the buildings. It would sit on top of the per-mile Driver fee, not replace it. None of these numbers are Aurora’s. I made them up to size the prize. But they show the shape of it, a second recurring revenue line that compounds as the truck network grows, without Aurora owning the dirt.

And here’s the part that gives the whole idea teeth: it gets far more powerful if these trucks go electric. A diesel terminal is a service center, useful but not a moat, because a truck can fuel anywhere. An electric terminal is a charging depot, and now you have the actual Supercharger dynamic: proprietary charging, a captive fleet that has to stop where the chargers are, and a location advantage that compounds. So the terminal-as-Supercharger idea is really a bet on electrification as much as autonomy. Today it’s a service network. In an electric future, it’s infrastructure.

I want to be clear that this is speculation, not a forecast. Aurora has not announced any of this, and the asset-light purists would argue they should never touch it. But the terminal layer is coming whether Aurora builds it or not, and a company that already understands these trucks better than anyone, and is already standing up terminals in Texas and Phoenix, is about as well positioned as it gets to be the one that sets the standard.

Read more

The company is Applied Optoelectronics, or AOI. That company is largely based a little bit down the road from NASA’s Johnson Space Center today, in Sugar Land, Texas.

This is a company that has run from $15 a share to $200 plus per share in the last year. These things take me a few days to put together and publish, so there’s a good chance that by the time you’re reading it, it’s nowhere near $200 a share. Could be back down at $15, could be up at 600. We just don’t know.

But part of this publication is not always finding the perfect companies to buy. It’s digging in and understanding what these companies are, what they do, and then ultimately making a decision as to whether now is the right time, or whether maybe there’s just not enough juice left in the squeeze. That’s a perfectly fine outcome too.

So ultimately, we’re looking to answer whether there is a valid business under all this, or whether it’s all smoke and mirrors that will eventually lead to ruin for those that are getting in at this $200-per-share mark.

What is AOI?

AOI is a vertically integrated optical maker. If that means nothing to you, don’t worry, because hopefully by the end of this section it will.

Now all around us we have cables, and those cables are sending data back and forth. As I write this I am watching the Stanley Cup Finals on a TV that has an HDMI cable. That HDMI cable sends a signal back and forth. It has Ethernet plugged into the back of it, and that Ethernet is sending a data signal back and forth.

Now most cables that we interact with, most cables that we’ve dealt with over our lives, have been made of copper. Copper is great. Copper can send these signals for decently long distances and it has a lot of strengths. Ultimately, when it comes down to it, copper just can’t handle fast signals, and the world that we are rapidly heading towards requires fast signals.

If you look at the data center buildouts that are happening, these data centers all have racks and racks of GPUs. To be able to build the types of models that we are seeing from these AI powerhouses, you need to be able to connect GPUs, and you need those connections to be fast. That’s where optical makers come in. They are not building these connections out of copper wires any more. They are building these connections out of lasers and light.

And that’s where AOI sits. It helps make these transceivers and the technologies that are required to send a signal back and forth between different racks in these data centers. Sounds simple, but there’s actually a lot of tech that goes into it.

Right now the product that is really driving the price in AOI up is called an EML transceiver. This particular piece of technology helps send a signal back and forth between different racks, or between wherever you need that signal to go. Right now 800G is the big buildout. All the data centers want 800G. This is a 100Gb connection across eight different lanes, which means you can send a ton of information back and forth. A lane is just one of those parallel streams of light. Stack eight of them at 100 gigabits each and you land at your 800 gigabits.

There’s a lot of fun physics that I got into as part of researching for this article, which I won’t bore you with. I was actually pleasantly surprised by a primer on the physics that Claude provided me. So, yes, I definitely used some transceivers in the making of this article.

Alright, where were we. AOI. Well, the thing I want to call out here is that this is not AOI’s first rodeo when it comes to catching onto a massive buildout trend, and it’s actually one of the most alarming things I found about this company. AOI is a pretty old company. It’s been around for a good long while. It’s been in the cable business. It’s been in all kinds of businesses. It’s been around that long.

In 2017 the stock had explosive growth on the back of Amazon. So Amazon shows up, they’re building AWS, and AOI was the company that was helping them build the transceivers for it. Instead of 800G this was 40G and 100G at the time. This sent AOI stock up to $100 a share into 2017, and it ultimately ended up busting down to $1.48 when it turned out that Amazon didn’t need anywhere near the volume everyone had built up for.

I mention this not as a guarantee of the path that we’re on, but just to highlight that we have been on this exact path before. The parallels are eerily similar. In 2017 Amazon was by far the largest customer, and the ten largest customers AOI had made up greater than 95% of the company’s revenue. In 2026 the concentration is just as extreme. The top ten customers again make up more than 95% of revenue, and just two of them, a cable distributor named Digicomm and Microsoft, are better than 80% of the total. Different names, same shape. Not a guarantee of things to come, but a path to be wary of.

And this kind of thing happens because of cyclicality. We will touch on cyclicality later in the article.

The Technology Transition

800G, as we have mentioned, is EML. There are a few different types of products that exist in this space, but 800G is dominated by EML, and EML happens to be exactly what AOI is built around.

What is going to happen as we move forward is a significant shift in how the market works. This shift happens at 1.6T, and especially at 3.2T, where silicon photonics takes share and the value migrates from this high-margin EML to cheaper CW lasers.

The next question you probably have is whether AOI is participating in 1.6T. Yes, AOI is a real 1.6T participant. But what we need to be aware of is that it will not be a dominant participant in CW lasers, where there are already significant competitors. It’s also a world with lower margins, and a world that is more contested.

The good news, though, is that one of AOI’s significant advantages still plays out in the CW laser space. Their vertical integration. In late 2025 they announced their own high-power CW pump laser aimed right at silicon photonics and co-packaged optics, which is their way of trying to keep a foot in the next generation rather than just defending the last one.

We touched on vertical integration above but did not go into it in depth. A thing worth knowing about AOI is that it is incredibly vertically integrated. It goes all the way from growing the crystals that are required in the end product, all the way to delivering the end product. Competitors have to rely on others for different parts of this. AOI simply does it all itself.

And AOI is building the newest of this capacity, the 800G and 1.6T lines, inside the United States, down in Texas. Right there are two things that work in AOI’s favor versus its competitors.

AOI is not the market leader in high-end EML lasers. That title belongs to Lumentum. It is not the leader in 800G module volume. That title belongs to InnoLight. But AOI’s real edge is that it makes its own lasers and its own modules, in the United States, from the ground up. It’s not waiting in Nvidia’s allocation line, and it’s geopolitically very clean.

Competitors

There are a number of competitors in the space. You have the aforementioned Lumentum. Lumentum is the EML laser leader. This is the 800-pound gorilla in the room when it comes to EML lasers. They own about 60% of the market and they’ve been rewarded heavily for it. This is roughly a $70 billion company at the time of writing, and it has run more than tenfold over the past year. It was one of the best-performing stocks in the entire S&P 500 this year, and it just joined the Nasdaq-100.

Coherent is another competitor. Coherent broadly builds EML plus SiPh. SiPh is cool, or silicon photonics. This is the technology of building the optical circuits directly onto standard silicon microchips.

InnoLight and Eoptolink are the volume module leaders out of China, and they compete on a pure cost basis. These are the companies that pump stuff out cheap. Sometimes cheap is good, but this is also a supply chain risk, which is part of why AOI stands to benefit from being stateside.

I called Lumentum the gorilla in the room, but the true gorillas are companies like Broadcom. Broadcom makes the DSP chips and the switch silicon that sit at the very center of these networks, which is a different layer from the lasers and modules AOI builds. DSP stands for Digital Signal Processing, and it is the chip inside a transceiver that cleans up the signal so it survives the trip. Broadcom builds those for 800G and 1.6T, and it is also out front on CPO.

And CPO is another term you’ll want to be familiar with if you are interested in investing in semiconductors. CPO stands for co-packaged optics. Traditionally a switch chip uses electrical traces to send data to pluggable lasers on the front edge of the server rack. CPO integrates the optical silicon directly next to the switch ASIC on the same package. So it eliminates the copper travel distance, slashes power consumption, and reduces latency for these massive hyperscale networks.

Amongst all these names, AOI is the small, vertically integrated Western player. Of all the Western InP laser makers it is the smallest, which generally means it’s got more room to grow.

The Financials

Let me walk you through the numbers, because this is where the story gets real on both sides.

Start with the top line, because the top line is the whole reason anybody is paying attention. In 2024 AOI did about $249 million in revenue. In 2025 it did about $456 million. That’s an 83% jump. For 2026 management is guiding to more than $1.1 billion, which would be better than doubling again. So you’ve got a company going from a quarter of a billion to over a billion in revenue in two years. That is the kind of slope that takes a stock from $15 to $200.

The most recent quarter, Q1 2026, reported on May 7, was the fourth record quarter in a row. Revenue was $151 million, up 51% from a year earlier. Data center was $81 million of that, more than half the company now and up over 150% year over year. Cable chipped in $67 million. So the AI data center business has officially become the main event.

AOI still loses money, and I am not going to dress that up. Q1 was a GAAP net loss of $14 million. Gross margin was about 29%, which actually slipped a little from the quarter before as the data center mix grew. One detail the headlines skipped right over: of that $81 million in data center revenue, only $4.6 million was actually 800G. The big 800G ramp hasn’t really happened yet. What grew in Q1 was the older 100G and 400G product. The 800G wave, the one this whole thesis rides on, is a second-half-of-2026 event.

This matters, because AOI has never put together a sustainably profitable year in its entire life as a public company, going all the way back to its 2013 IPO. Full year 2025 was a loss of about 26 cents a share.

So why does anyone care? Because management says that changes this year. They are guiding 2026 to more than $140 million in non-GAAP operating profit, and they expect to flip to non-GAAP profitability starting in the second quarter. They’ve laid out 60% to 80% sequential revenue growth in the back half as new Texas capacity comes online, with gross margins climbing to 35% by year-end and over 40% in 2027. A company that crosses from years of losses into real profit while doubling revenue can re-rate hard. That’s the bet.

And the number management keeps pointing at sits even further out. On the call the CFO put a hard figure on mid-2027: roughly $471 million a month in data center transceiver revenue. Annualize that and you get about $5.65 billion in transceivers alone, before a dollar of cable. The current full-year guide is $1.1 billion. So management is telling you the exit speed eighteen months out is something like five times this year’s number. The CEO literally called this year’s $200 million-plus orders small compared to what is coming.

Last, the balance sheet, because building laser factories is not cheap. AOI exited Q1 with about $449 million in cash, but a big chunk of that came from a $382 million stock raise during the quarter. There’s also $125 million of convertible debt due in 2030. The company is spending heavily on capex, burning cash, and funding the buildout by issuing stock. That keeps the lights on and the factories going up, but it means the share count keeps creeping higher, and every new share is a slice of the pie taken out of the existing owners’ hands. Hold that thought for the bear case.

Cyclical or Structural

I promised we’d come back to cyclicality, so here we are. Everything in the bull and bear cases below hangs on it.

The optical business has always been cyclical. Wildly so. If you’ve ever looked at memory chips, at companies like Micron or SK Hynix, you know the pattern. Demand grows for a couple of years, everybody races to add capacity, capacity overshoots, prices crash, and the whole industry has a miserable year before it does it all again. Optical transceivers have run on that same rhythm for two decades. The firm LightCounting, which tracks this market more closely than anyone, has pointed out that every two or three good years gets interrupted by a flat or negative one. And the down ones are not gentle. The last big cycle, the 2017 to 2018 one that AOI rode and then got crushed by, saw these stocks fall 50% to 70%.

So how does that happen, mechanically? It usually isn’t pure overbuilding by itself. It’s the inventory whipsaw on top of it. When parts are scarce, customers panic and over-order from every supplier they can find, building a stockpile just to guarantee they get their allocation. The moment the shortage eases, they cancel the duplicate orders and start working down that stockpile. Reported demand falls off a cliff even though the real end demand is fine. Then, because everyone added capacity during the boom, there’s now too much of it, and suppliers start competing on price. That’s the cycle. And the dangerous part is that it tips right at the moment the shortage ends.

LightCounting expects the current laser shortage to ease around the middle of 2026. They are explicitly calling for a flat quarter or two in late 2026 as the supply chain finds its footing. That is the next test.

Now the other side. This time might genuinely be different, at least in degree. In every past cycle, one speed generation ramped, peaked, and got replaced by the next, in sequence. This time the generations are stacking on top of each other. 800G is ramping, 1.6T is ramping right behind it, 3.2T is coming, and co-packaged optics adds a fourth layer behind that. Because demand is spread across so much of the stack at once, no single generation’s inventory hangover can sink the whole thing the way the 100G glut did in 2018. That is the single best argument that the next downturn could be shallower than the last.

The demand itself has a self-feeding quality this cycle. Every new AI model generation increases the amount of inference running, which means more hardware, which means more optics, on a loop. And the network is eating a bigger and bigger share of every data center dollar as clusters get larger. So the trend line is real, and it is steeper and more durable than memory’s ever was.

But, and this is the honest synthesis, structural growth and cyclicality are not opposites. Memory demand has compounded for forty years and memory is still the most violently cyclical business in tech. Both things are true at once. The market growing to $35 billion or more by 2030 and the industry having brutal down years on the way there are not in conflict.

The cleanest tell that even the biggest players expect the cycle to persist is this. Nvidia reportedly asked the laser supply chain to build 20 times more capacity by 2030, and the suppliers pushed back and agreed to only 12 times (per Rosenblatt Securities’ May 2026 indium phosphide supply model). People who believed in permanent, non-cyclical demand do not deliberately under-build. They are hedging against the down year they know is coming.

So where do I land? This is structural growth wrapped around a cycle. The decade trend is up and to the right, steeper than memory. But there will be air pockets, the demand sits in the hands of a few spending-cyclical buyers, and the first real test arrives in the back half of 2026. For a company like AOI, which we’re about to see is concentrated, richly valued, and not yet profitable, the cyclicality is not a footnote. It’s the main risk.

Bull Case

Alright, let me lay out the bull case as strongly and as honestly as I can, because there is a real one here.

It starts with a moat that’s easy to miss. Everybody in this industry can design an 800G transceiver. The incumbents can all do it. The hard part is not designing one. The hard part is building enough of them, precisely, at a cost that scales. Inside every transceiver is a microscopic assembly of lasers and waveguides where a misalignment of a fraction of a hair kills the signal. And remember what we covered in the technology section: silicon cannot make light, so every one of these things needs an indium phosphide laser, and those lasers are the single most supply-constrained part of the whole industry. There aren’t enough of them, and the furnaces that grow them are booked solid.

AOI is different because it makes its own lasers. In-house. The exact part everyone else is fighting to source, AOI grows itself, from the crystal up. And it goes further than that. It builds its own automated production lines, and, more importantly, much of its own manufacturing equipment. So when the CFO got the obvious bear question on the call, which is, fine, demand is huge, but can you actually build it, his answer was that equipment availability hasn’t been a problem because they make most of the equipment themselves. While competitors line up for the same furnaces, the same merchant lasers, and the same third-party tools, AOI is not standing in any of those queues. The moat here is not a patent. It’s throughput nobody can copy on a short timeline.

Then there’s the order book, and these are signed orders, not a slide deck. In March, AOI announced its first volume 1.6T order from a major hyperscaler, worth more than $200 million, big enough on its own to put that one customer back above 10% of revenue. It booked a $53 million 800G order, then another $71 million one. That’s over $324 million from a single hyperscaler in about four weeks. And when an analyst asked why that $200 million 1.6T order stretches into 2027, the CEO’s answer was that it’s small compared to what’s coming in 2027.

Stack the capacity ramp behind it. AOI exited Q1 making about 100,000 of these units a month. The plan is 150,000 this quarter, over 650,000 a month by the end of 2026, and over 930,000 a month by the end of 2027, with more than half of that built on US soil. Management calls it the largest US production capacity for 800G and 1.6T.

Put it together and the bull case is this: a company that owns the scarce input, can build at a speed nobody else can match, is sitting on signed hyperscaler orders, and is about to cross from years of losses into real profit while its revenue multiplies. Add a genuine strategic edge in US manufacturing, which the hyperscalers increasingly want for supply security and tariff reasons, plus an Amazon that has tied its own stock warrant to how much product it buys.

But I want to state the bull case fully, including its shape, because the shape is the risk. The bull case, completely stated, is a pre-paid, high-beta execution bet on a 9x capacity ramp landing on schedule, into a demand environment that has to stay vertical. Every word in that sentence is doing work. Pre-paid, because you are paying for it now, today, before it happens. Execution bet, because the capacity actually has to get built. And demand that has to stay vertical, because if the AI spending cools, all of it cools with it.

Bear Case

Now the other side. A thesis you can’t argue against is not conviction, it’s hope. So here’s the honest bear case.

Start with the price, because the price is the problem. At around $200 the stock trades at roughly 15 times this year’s sales, for a company that has never strung together a profitable year. A lot of very good news is already baked in. As we’ll see in the valuation, even a genuinely good outcome is largely priced in today. That leaves very little room for error, and this is a stock that does not handle error gracefully.

Then the concentration, which is the biggest risk by a mile, and the one that should make the hair on your neck stand up given the history. The top ten customers are about 97% of revenue. Two of them are more than 80%. And AOI sells on purchase orders, not binding long-term contracts, so there’s no floor under any of it. The famous $4 billion Amazon figure is a discretionary ceiling, not a commitment. We’ve watched this exact movie before. In 2017 one giant customer was most of the revenue, the orders dried up, and the stock went from $100 to $1.48. Nobody is saying that happens again. But the structure that made it possible is sitting right there on the page, unchanged.

Next, the technology, which is the slow risk rather than the sudden one. AOI’s strength is the EML, and the EML is exactly right for 800G. But the industry is moving toward silicon photonics and cheaper CW lasers at 1.6T and beyond, and in that world the value drains out of the expensive part AOI is best at. AOI is in that race, but by Rosenblatt’s own numbers it is the smallest of the Western laser makers, well behind Lumentum and Coherent. So the bull case is partly leaning on a technology whose long-term share is genuinely in question.

Add that the profitability is still a promise. The 35% margins and the $140 million of operating profit are forecasts, not results. And the most recent quarter actually missed on revenue versus what Wall Street wanted, while the next quarter’s guidance came in a touch light. This is an execution bet, and the execution hasn’t been flawless.

Then the dilution we flagged earlier. The buildout is being funded by selling stock, and the convertible notes and the Amazon warrant are more shares waiting in the wings. Every per-share number in the bull case quietly assumes those new shares don’t pile up too fast.

A couple more. The insiders have been selling into this run, including the CEO offloading around $10 million worth and a director selling a block. Some of that is scheduled and perfectly normal, but into a vertical move it’s worth clocking. And the laser advantage cuts both ways. That 350% expansion of the laser fabs, plus a shift from four-inch to six-inch wafers, is exactly the kind of manufacturing transition where yield problems hide, and a laser yield problem is a revenue problem.

Finally, two longer-term clouds. There’s a quiet competitive threat from copper and from co-packaged optics. Active electrical cables can now handle the very shortest links (a point Credo hammered on its recent earnings call), and as those and CPO take hold, they can chip away at how many of these pluggable optical modules a cluster even needs. And the whole thing rides on the cyclicality question from the last section. If hyperscaler spending takes a breather, or the inventory whipsaw hits in late 2026 the way LightCounting expects, a richly valued, high-beta, concentrated name is precisely the kind that takes the 50% to 70% drawdown.

Read more

Like seemingly everybody in 2020, I started a YouTube channel. On that channel, I posted videos breaking down different kinds of companies, talking about investing, and occasionally teaching you how to build financial models. Some of those videos were on a company called Rocket Lab.

I have always been a fan of space, and I have always been a fan of investing. Rocket Lab caught my fancy. I pondered what could happen to this SPAC that had recently gone public and was being beat down along with every single other SPAC at the time. I saw a great future ahead for the company.

My problem was, especially if you look at some of the comments on the original video, I was far too conservative. I even had to defend a couple of my positions on Reddit, saying, “Hey, if I am too conservative, predicting here a more than 100% return in stock, then that is incredible.”

Rocket Lab has returned more than 3,000% since I made those videos saying that I was buying. My only problem right now is I wish I had bought more.

Here we are. I am writing this in May 2026. We are going to start out, right here on SpaceBreakdowns, with a deep dive into Rocket Lab. What has changed about Rocket Lab over the past few years, what the future looks like, and whether at $120 plus per share there is still room to run in this name, or whether it is finally time to sell.

Where Past Analysis Was Incorrect

In April 2024, I told you that excellent execution would yield a $20 stock. I was wrong. I modeled a traditional industrial business, but Peter Beck built an unassailable vertical monopoly. My models caught the trend, but missed the scale of institutional premium that we see today.

The Multiples Trap

In the April 2024 video linked here, I arrived at a long-term target price of $20 per share, assuming Rocket Lab had excellent execution. If Rocket Lab ended up at $20 a share now, I would be mind blown and extremely disappointed.

My thinking back then was that Rocket Lab should be modeled like a traditional high-growth industrial hardware company that would face multiple compression as its capital expenditure scaled. Instead, the market gave Rocket Lab a software-style scarcity premium, which is completely fair.

It is fair because Rocket Lab emerged as the only functional Western alternative to SpaceX. Wall Street threw standard EBITDA multiples out the window and re-rated them as a tier 1 national security asset. At $120 plus per share, I was wrong because I assumed the market would remain rational about valuation multiples for a hardware monopoly.

The Neutron Pricing Floor

In the first video I made about Rocket Lab in January 2024, I caught a lot of flak in the comments for modeling Neutron at a hyper-conservative $30 million per launch. The same feedback was on Reddit posts, and I made a second video that reran the numbers at the full $50 million retail price just to see the upside.

This was wrong, because even at $50 million, the model is treating launch as a static, commoditized line item. What actually happened by 2026 is that Rocket Lab began commanding premium block deals and National Security Space Launch phase 3 task orders. These defense missions do not just pay for the freight, they pay massive high margin premiums for custom orbit insertion, mission integration, and secure payload handling.

Rocket Lab being able to integrate all these parts was called out in those previous videos, but I never did acknowledge the value that would have on the pricing of Neutron.

Velocity and Margins of Space Systems

I correctly identified that Space Systems was the secret weapon, but I underestimated the sheer velocity of Rocket Lab’s vertical integration strategy.

By aggressively swallowing up the supply chain, capping it off with major moves like the Motiv and Mynaric integrations, Rocket Lab has turned space systems into an end-to-end captive powerhouse.

To be fair to my own modeling, I don’t believe anybody in early 2024 was modeling a 200 million dollar single quarter revenue run rate with 38.2% GAAP gross margins this early in the cycle.

Rocket Lab has very quickly ceased just being a component vendor and has become a prime contractor much sooner than expected.

The Q1 2026 Setup

I do not think there is any dispute that Rocket Lab had a fantastic Q1. The company had record revenue of $200.3 million. They now have a record backlog of $2.2 billion. They were able to sign 5 Neutron launches, and that manifest is now filling up through the end of the decade.

In addition, there were 31 Electron and HASTE missions booked, which is the highest booking Rocket Lab has ever had in a single quarter.

So far in 2026, there have been 8 launches year to date, and there are 70 more in the backlog. The 100th Electron launch for Rocket Lab should occur this year. There have been big partnerships with the likes of the Department of War, which signed a new launch contract for HASTE, and partnerships with Anduril to deliver new hypersonic capabilities for their national security program.

The key thing to pull out is the acquisition of the company Motiv and the completion of the Mynaric acquisition. Both of these acquisitions help expand how far Rocket Lab’s vertical integration can go.

Motiv brings space-proven robotics to the company’s offerings. As the company behind the Mars Perseverance Rover’s robotic arm and CADRE lunar rovers, they have put robotics systems on other planets that have been usable. This capability will now be available to Rocket Lab customers.

Along with robotics, Rocket Lab also added 7 new high-value products to its space systems lineup:

1. Electric propulsion, known as Gauss, which is a new in-house developed system for high-volume production

2. Optical payloads

3. Laser comms

4. Silicon solar

5. Star trackers

6. Robotics

7. Precision mechanisms

A popular line from Peter Beck is, “He who owns the payload owns the mission.” We can see from the steps Rocket Lab is taking that this is becoming more true over time. It is not new that Rocket Lab is trying to become the vertical integration play of space, but it is expanding how large that vertical offering actually is.

Over the years since I started covering Rocket Lab, it has gone from something that could put a satellite into space to a company where you can go and say, Hey, I want to build a robot that will go digging on Mars. Can you help me with that? And the answer is yes, 100%.

Rocket Lab can now design that robot, launch that robot, deploy that robot, and monitor and maintain that robot. All you need is the idea and the money to make it happen, and Rocket Lab will do that for you.

Overall, it was a fantastic quarter for Rocket Lab. They are the dominant Western alternative to SpaceX. SpaceX is the 1,000-pound gorilla in the room, but Rocket Lab is showing that they have the ability to be a major space company right alongside SpaceX. They may not dominate launch like SpaceX, but they are tackling a piece of the puzzle that nobody else is. That in itself is valuable.

The Bull Case

At the time of writing, Rocket Lab is a $72B company. You do not have to look back that far in the past to a time when it was a $10B company. To have gone from those levels up to $70B in such a short time is incredible.

At those levels, Rocket Lab is trading at approximately 25x forward 2027 revenue. The question is not whether that is expensive in absolute terms, as it obviously is. The question is whether Rocket Lab can continue to grow from these levels and prove that it is a company that is worth that much.

While we are not making buy or sell recommendations as part of this publication, it is worth talking about the price of Rocket Lab for just a minute. There are, I believe, 3 pillars that justify the valuation Rocket Lab is trading at today.

Pillar 1: Neutron

The first pillar is, of course, Neutron. Neutron is the shiny new thing and rightfully so. It is Rocket Lab’s medium-lift, partially reusable rocket that’s designed to put 13,000 kg into low Earth orbit. Makes it a direct competitor to SpaceX’s Falcon 9. A first launch is currently expected no earlier than the last quarter of 2026. The first stage is designed to be recovered on an ocean barge that they’re calling Return On Investment, for reuse.

Now Neutron matters but it’s not just for the launch revenue. The biggest story around Neutron for me is the flywheel. If you are a customer coming to Rocket Lab for a medium-lift launch, you are already more likely to use their components on your satellite. You’re more likely to use their software and you’re more likely to integrate with their broader space systems offering.

The inverse of that is also true. If you are already a Rocket Lab Space Systems customer, the launch decision becomes a natural fit. Neutron is not just a rocket; it is the gateway product that pulls everything else through.

This is a meaningful difference between Rocket Lab and a traditional launch provider. Rocket Lab is not selling Neutron the way ULA sells Vulcan or Arianespace sells the Ariane 6. They are selling Neutron as part of a bundle that competes on integration rather than just pure price per kilogram.

Pillar 2: Space Systems, The Prime

The second pillar is the one I underestimated the most in my early modeling. Rocket Lab is fast becoming the prime contractor for getting to space. If you have an idea for a mission, you go to Rocket Lab. It will design the satellite, build the components, integrate the payload, launch the rocket, and operate the spacecraft in orbit.

The Motiv acquisition is the move that crystallizes this thesis. Motiv built the robotic arm on NASA’s Perseverance rover and is delivering hardware for the CADRE lunar rovers. Rebranded as Rocket Lab Robotics, this isn’t just a component-supplier acquisition; it’s a full-on credentials acquisition. Rocket Lab can now credibly bid as a prime on planetary missions that previously belonged to JPL contractors and legacy primes. You combine that with the Mynaric acquisition for laser communications and the introduction of seven new high-value products in their space systems line up in Q1. Rocket Lab has closed most of the remaining gaps in their vertical integration play.

Peter Beck has always been explicit about the playbook: “Our acquisition strategy is simple but proven and effective. We identify the best space technologies that have struggled to scale and we bring them into the Rocket Lab ecosystem. By applying our resources, expertise, and manufacturing scale, we make these technologies more accessible and affordable for the global space industry.”

The Motiv deal is Rocket Lab’s seventh acquisition since 2020. Seven acquisitions in five years is not a side project. It’s a deliberate roll-up strategy that is aimed at making Rocket Lab the only place you need to go if you want to get to space.

If Rocket Lab succeeds in becoming the default prime for space access (in the same way that Heico became the default prime for aerospace replacement parts), the addressable market is substantially larger than launch alone. Space Systems revenue is growing faster than launch revenue. It carries a higher margin and benefits from the kind of switching costs that pure launch businesses alone do not have.

Pillar 3: Flatellite and the Constellation Play

The pillar 3 is the fun one and it’s the one that most people do not seem to be pricing in or talking about with Rocket Lab. In 2025, Rocket Lab announced Flatellite. It’s a flat stackable mass-producible satellite that is designed for large constellations. Beck himself described it as “a bold strategic move toward Rocket Lab operating its own constellation and delivering services from space.”

And Rocket Lab has rarely been shy about indicating that they will build their own constellation one day. Reporting from SatNews suggests that Flatellite is being aimed at 5G non-terrestrial networks, with industry chatter putting the minimum viable constellation in the 150 to 200 satellite range rather than the thousands required for Starlink-class build out. Beck’s own framing is: “By having our own rides to space with Neutron and Electron and being able to build our own spacecraft in high volumes, we have a distinct advantage when it comes to deploying constellations with speed and cost efficiency.”

To return to a quote we have earlier in the article, “He who owns the payload owns the mission.” The natural endpoint then of Rocket Lab’s vertical integration strategy is not just providing the rocket, the components, the spacecraft, and the operations to other people. It is operating its own services from space and capturing the revenue that flows to constellation operators rather than to their suppliers.

And this is obviously the part of the story that is hardest to value because it is the furthest from today’s revenue but it’s also the part that justifies multiples that look insane today. Starlink is on track to do somewhere in the neighborhood of $12 billion in revenue this year. If Rocket Lab can build even a meaningful niche position in 5G non-terrestrial networks over the next decade, that is not a $70 billion company. It is something much larger.

Of course the competition is real. Starlink dominates, Kuiper is coming, AST Space Mobile is in the market but this planet is large. The demand for connectivity is growing and the regulatory and capital barriers favor companies that already have the launch and manufacturing infrastructure in place. Rocket Lab is one of perhaps four companies in the world that can credibly stand up a constellation from scratch.

The Bear Case

There is a lot of expectation riding on Rocket Lab at these levels. At roughly 25x forward 2027 revenue, the market is pricing in flawless execution on three separate fronts that all still carry meaningful risk. Anyone that is buying or holding Rocket Lab at $120 plus per share owes themselves an honest accounting of what could go wrong.

Neutron Slippage

The most obvious risk is that Neutron slips again. The rocket was originally targeted for a 2024 launch that moved to 2025. It then moved to Q1 2026 after a stage one propellant tank ruptured during the hydrostatic test in January 2026. The first launch is now expected no earlier than the last quarter of 2026, with some analysts arguing that even that timeline looks optimistic.

Neutron development is not cheap. Spice, the Rocket Lab CFO, disclosed on the Q3 2025 earnings call that the company will have spent approximately $360 million on Neutron development through the end of 2025, with workforce costs running about $15 million per quarter. The original development estimate was $250 million to $300 million. That overrun is meaningful for a company that is not yet profitable.

Every quarter that Neutron slips is a quarter where Rocket Lab burns development cash without booking the launch revenue that justifies the investment. It is also a quarter where the NSSL lane 1 onramp gets pushed, where commercial Neutron manifests back up, and where the multiple compression risk increases. If Neutron does not fly until 2027, the bull case has to absorb a year of additional cash burn against a flat revenue ramp on the launch side.

Reusability is hard

SpaceX makes it look easy but the reality is that reusability is hard. Only two companies have ever landed and re-flown an orbital-class booster: SpaceX and Blue Origin. As SpaceX took years and many failures to make Falcon 9 reuse reliable, Blue Origin’s New Glenn has flown but is still early in its reuse program. Reusability is not a checkbox. It’s an engineering capability that takes dozens of flights to mature and it is part of the unit economics that justify Neutron’s $50 million price point.

If Neutron’s first stage proves harder to recover and refurbish than expected, if methalox cleanliness advantage is smaller than expected, if the canard and landing leg system has stability issues, if turnaround times stretch to months rather than weeks, then the economics of Neutron quickly deteriorate. A Neutron business case is heavily leveraged to reuse. Expendable Neutron is a much harder sell against the mature and reusable Falcon 9.

Scarcity Premium Can Evaporate

A lot of what is driving Rocket Lab’s premium valuation and SpaceX is scarcity. Rocket Lab is considered the only functional Western alternative to SpaceX and that scarcity justifies multiples that would not apply to a normal industrial hardware business.

And scarcity premiums are real but they are also very fragile. They depend on a competitive landscape being static and space is not static.

There’s Blue Origin’s New Glenn that flew successfully in 2025. It is ramping production. There’s Stoke’s Nova vehicle, targeting full reusability, with an architecture that, if it works, has better unit economics than Neutron. Relativity’s Terran R is in development. Firefly has an MLV in development with Northrop Grumman. ULA has the Vulcan, which is now operational. None of these are guaranteed to succeed but the field of medium-lift competitors is meaningfully larger today than it was when the market re-rated Rocket Lab.

If even two of these programs reach operational cadence over the next three years, Rocket Lab is no longer the only functional Western alternative to SpaceX. It ends up being one of several. At that point the scarcity premium would compress and Rocket Lab gets valued like a normal aerospace hardware business with a strong space systems franchise, which is real and a good business. It likely does not stay valued at 25x forward revenue.

The Verdict

Rocket Lab today is a different company than it was when I first covered it in 2024. A 3,000% return for investors reflects real operational achievement, record revenue, record backlog, the only functional Western alternative to SpaceX in development on Neutron, and a vertical integration strategy that has executed on seven acquisitions in five years and is now closing the final gaps in the stack.

At the same time, the valuation has run ahead of fundamentals in ways that require flawless execution to justify:

• Neutron has to fly

• re-use has to work

• scarcity premium has to hold against a growing field of competitors

• and the SpaceX IPO has to land well for the sector

Any one of those breaking lowers the multiple. Two of them breaking would compress it significantly.

The honest framing is that Rocket Lab today is no longer the asymmetric setup it was at $4, $10, or even $30 a share. The risk/reward has changed. The bull case is intact and arguably stronger than it has ever been, but it’s mostly priced in. The bear case is real but requires specific catalysts to manifest. The base case that Rocket Lab is a strong company executing well and a stretched valuation supports the current price. It does not necessarily support significant further multiple expansion from here.

This is the kind of company where the next 12 months matter much more than the previous 12. Neutron’s first flight, the Q2 and Q3 earnings prints, and the Golden Dome program flow will tell us whether the $120+ per share is a floor or a ceiling. Until then, every Rocket Lab holder is being asked the same question: Do you believe the execution risk is worth the price you are paying for the optionality?