Newshot 
I wrote a post in March, where I argued that intelligence was becoming a commodity, and that the context we fed the models was the real moat. Things have evolved a lot since then (including my thinking), but what we can all agree on is that intelligence is wide-spread, and have indeed become a commodity. You can feel this is the case when AI labs are praising the combination of harness+model (i.e. their vertical integration), as the way one can get the most out of the intelligence their models provide. Their sell: not all tokens are equal if you use the right harness.
In that post, I argued that context (and its underlying know-how) was what would become the moat in a world of wide-spread AI. I think there’s still some truth in this. Even more, this is (to me) a big reason why subscriptions are currently being subsidised. This post perfectly describes a thesis that I also share: subscription tokens have a discount because, through them, we are sharing a lot of the know-how of our jobs and our personal context (we are part of the training feedback loop that is allowing labs to become better and better at certain tasks). Each output we correct and each decision we make inside the lab’s harness is the kind of work labs buy from vendors like Mercor and Surge. So the discount is how the labs pay for our context, and once it is in their models, it stops being a moat for us.
With all the recent model developments and releases, I myself came back to this question about moats. If context and know-how are not a moat anymore (or at least not ours), what are the moats that businesses can build to be durable in the age of AI, where the cost of building software and technology is approaching zero (or the cost of a few tokens)? Of course, there have been heated discussions around this in the Twitter-sphere lately, with Lenny Rachitsky asking “Moats! Does anyone have one anymore?”, so I decided to write down my own view, as a way to trigger a discussion, and to remind “future me” what my thinking was at this point in time.
Are these really moats?
If you read all the lists out there about moats in the age of AI, you’ll see the same things over and over: the physical world (anything made of atoms that a model can’t touch), distribution channels, an established brand, switching costs, network effects, data effects, access to capital, an audience, etc. Dan Hockenmaier’s post keeps four of them (switching costs, network effects, scale economies, and data effects), and a checklist called “Is your business AI proof?” as a way to grade your projects and ideas: you need atoms, you need permission, people come for you, people come for each other, you own data nobody else has, and you sell the result.
But are these really moats? To me, these lists mix the things that protect you with the things that help you go faster or have a headstart. A licensed electrician scores two points in the checklist, because someone has to show up and the job needs a licence. That keeps the AI out, but it does nothing against the electrician down the road. So before deciding what goes on my list, I had to define what I understand by a moat.
What is a moat?
For me, a moat is something that temporarily protects your business from the competition coming for your market and your customers, even if you don’t do anything new. You don’t add more capital, you don’t ship new features, you don’t run more campaigns, and the moat still defends your business.
A good example of this is a strong brand. A brand like Coca-Cola is hard to replace even if no more investment is added to it, and a competitor would need decades to build one like it from scratch. Licences are another one. Once you get one, it allows you to operate without additional investment, and a competitor needs time and capital to get their own (if they ever get it). I see this at Shellment, anyone could just copy the platform, but nobody can replicate the authorisation, the licenses, and the legal framework that makes this idea possible.
Moats are defined by their height, i.e. how much time and money a competitor needs to replicate them and attack you. Mayo Clinic’s move to Epic cost them over a billion dollars across five years, which gives you an idea of how high Epic’s moat is. The electrician’s licence is also a moat, but a low one that many people go over every year.
So the test I’ll use for the rest of the post is simple: if you stopped adding resources to your business tomorrow, would this still protect you?
Fuel
Then there is fuel, i.e. the gas and energy you need to build the business and, eventually, a moat around it. Capital allocation, technology that doesn’t challenges the laws of physics, execution speed, distribution channels, and an audience are all fuel. They help you win a market and start building a moat around it, but they don’t protect you in themselves. Hockenmaier puts speed, distribution, brand, the team, and funding in the same bucket, and this is the idea from his post that I agree with the most (with brand I only half agree: building a brand is fuel, but once it is established it is a moat).
The difference between fuel and a moat shows the moment you stop adding resources. More gas lets you make progress towards your moat, but if you stop investing there is no intrinsic protection, and a competitor that keeps adding fuel can get ahead of you and eat your business. If I stop writing this newsletter for a few months, my readers will move on to whoever is still writing and adding value to them. I still have a brand and a distribution channel that I can probably leverage, but it is competing with the new shiny content they may have found.
Economies of scale are also fuel, even if many count them as a moat. Tesla is a good example of this. As it scaled, it had to build more factories and open more points of sale, and that made each car cheaper to build and easier to buy. But factories and stores are not a moat, because with enough money a competitor can build them too, and that’s what happened: BYD sold more electric cars than Tesla in 2025. Even some of the inventions Tesla used to improve its processes, like casting big parts of the car in a single piece with giant presses, were copied by Chinese carmakers that bought the same presses. The scale was the fuel that helped Tesla build its real moat around the product and the brand, and whether that moat holds now that the competition has the same factories is still an open question (to me, the interesting part of Tesla today is whether the data from its fleet becomes a loop, but that’s a story for another post, and one that we’ve already discussed briefly in this newsletter).
Not long ago (many may remember those times) software and technology could be a moat in itself. But I am seeing that change a lot in my day-to-day. These days, when I see a piece of paid software I like, my first instinct is to build my own version with my agents and adapt it to what I need, instead of paying for someone else’s that I don’t have any control over. When a customer can rebuild your product over a weekend, your technology stops protecting you. You can already see the flood in games, where around half of the new releases on Steam now declare some AI-generated content. We are living the time in history when it is easiest to build software, but the hardest to make money out of it.
This is also why I now think that context and know-how are fuel, and not a moat. They give you a head start, but if you stop feeding them into your business, a competitor using the same models (trained, among others, on your corrections) will eventually catch up.
When I was thinking about this a few weeks ago, my own list of potential moats had audience and capital on it, as two scarce resources that protect you against the competition. However, the more I think about it, the more I think they are also fuel. That being said, for someone building alone, I still think audience and capital matter more than anything else on that list. When you run experiments on the side, savings and people who listen to you are what let you try again after something fails. They aren’t a moat (they stop the moment you stop), but they keep you in the game long enough to find one, and is an invaluable source of feedback (and this is why I am so grateful to every one of you that spend the time to read whatever I have to write about every week, thank you).
Loops
Reinforcing loops are a good way to remove the need for constant fuel and let the moat grow higher. A loop is when each new customer makes the next one easier to win, without you having to add more resources. Hockenmaier calls them compounding advantages, and his four (switching costs, network effects, scale economies, and data effects) are all loops when they work.
A good example of this is Stripe. Most of the payments that go through Stripe have already been through Stripe before, so every transaction improves its fraud detection, which brings more merchants, which brings more transactions. No newcomer can buy that, however much fuel they have. This is also how economies of scale can stop being fuel and become a loop. Zipline’s cost per delivery has fallen from around $300 to $12, that saving goes back into cheaper deliveries, and cheaper deliveries bring more deliveries. A rival that starts today inherits none of that and needs to build that scale effect (and the underlying loop) from scratch.
Loops can be a moat in themselves, because they reduce the need for more fuel. Once the loop is turning, the business gets harder to attack while you focus on growing and improving other fronts, and a competitor not only has to copy where you are, but also catch something that keeps moving away from them. The moment where acquisition costs, capital allocation, etc. don’t need to grow proportionally to the new customers you get, and the business keeps growing more autonomously, is when you know you have product-market fit.
Of course, AI doesn’t affect all loops the same way. Network effects are the ones that hold up best, because an agent can rebuild the product of any marketplace you can think of, but all you get is an empty room. Data loops are also extremely valuable, but narrower than many founders think. As Casado and Lauten put it, “the value of incremental data goes down”, while collecting the data that really makes a difference gets more expensive.
Switching costs are where AI is doing the most damage, because it removes the effort of moving. In October I wrote that leaving an API takes one line of config, and Lawrence Lundy-Bryan has a great thread on how the same is happening with chips. CUDA, the code that kept everyone on Nvidia for years because nobody wanted to rewrite it, is being eroded by AI coding tools, and Google’s new tools plug into Claude Code and Cursor to get people onto TPUs in minutes. What AI can’t remove is the consent. An agent can migrate a hospital’s patient records, but it can’t collect the sign-off from everyone who has to approve the move, or take the blame if a record goes missing. As far as I can tell, nobody has measured yet how much lock-in agents actually dissolve, so if companies start moving off their core systems in the next couple of years, I’m wrong about this.
And this takes me back to the labs and their subsidised subscriptions. To me, the discount is fuel spent to start a loop. Hockenmaier puts the frontier labs under scale economies, but I’m not convinced, because when a lab gets cheaper at serving tokens, the price war hands the saving to users. The loop they are after is the one Hockenmaier describes for Cursor, where “with each accept or reject from users, the model gets better at understanding what good code looks like”. Even more, labs are better positioned than Cursor, because they sit downstream of that data loop (they own the models). Our corrections make the model better, a better model brings more users, and more users bring more corrections. If that loop turns, the discount will have been the cheapest moat ever bought. I don’t know yet if it will, because open models keep closing the gap.
Fuel or moat?
So, with all of this, this is how I would classify everything we’ve discussed:
Most of what we call moats in the age of AI is fuel, and AI is making fuel cheaper for everyone. That doesn’t mean it’s useless. My weekly experiments don’t have a moat, and they don’t need one, because their job is to teach me something. But the moment you decide to scale something, what matters is whether the fuel you are burning is starting a loop.
So, if you stopped adding fuel to what you are building tomorrow, what would still protect it? I would love to hear everyone thoughts about moats and fuel. Let’s kick off the discussion.
Until next week!