Everyone can make fried chicken now: five moats for the AI arms race

We build with AI every day now, and I still get a kick out of how quickly an idea turns into something you can click on. A working prototype that once needed a team, a budget and most of a quarter can be in front of people within a couple of weeks, which is brilliant news for anyone with a list of things they've always wanted built, and slightly less brilliant news for anyone whose business depends on nobody else being able to build them.

The people making the models saw this coming first. In 2023 a Google engineer wrote an internal memo, later leaked and published by SemiAnalysis as 'We Have No Moat, And Neither Does OpenAI', arguing that open-source models were closing the gap faster than either company could widen it. 'We aren't positioned to win this arms race and neither is OpenAI', it admitted, and three years on that race has spilled out of the labs into every market that runs on software.

Roughly 560,000 new apps reached Apple's App Store in the first half of 2026, nearly as many as in the whole of 2025, while downloads grew by about 2% (Sensor Tower's figures, via the New York Times). When OpenAI launched its Deep Research agent in February 2025, Hugging Face published an open-source rebuild two days later - not quite as good, but close enough to make the point. Building has stopped being the hard part, which leaves every business with an uncomfortably simple question: if somebody copied your product tomorrow, what would they still not have?

Difficulty used to be the moat

For most of the history of software, difficulty was its own protection. Building something properly took specialist people, real money and a lot of time, so a good product with a head start could hold its lead for years. That protection is evaporating (in early 2025 a quarter of Y Combinator's newest startups had codebases about 95% written by AI), and Jasper shows what happens to a business that was relying on it.

Jasper was an AI copywriting company valued at $1.5bn in October 2022, built on OpenAI's GPT-3, and a few weeks later its own supplier launched ChatGPT. Within a year it had laid off staff, reportedly cut its internal valuation and replaced its chief executive. It's still going, repositioned around enterprise marketing, but it learned the hard way that a product anyone can approximate isn't a moat, however good it is.

None of this means the product doesn't matter, because a bad one still loses. It means the product gets you into the race and something else keeps you in it, and looking at the businesses holding their ground (and a few that aren't), I think five things do that work - for the established business worried about a nimble challenger as much as for the founder worried about a giant noticing their idea.

1. Distribution: when the product is a commodity, reach decides

When two products are roughly as good as each other, the one already in front of the customer usually wins. Microsoft launched Teams inside Office 365 in 2017, and by April 2020 it had more than 75 million daily users against the 12 million Slack had last reported. Slack, which had the head start and arguably the better product, complained to the European Commission, and when the Commission closed the case in 2025 its view was that tying Teams to Office had given it an unfair advantage in exactly one thing: distribution.

AI is following the same script. Meta AI reached a billion monthly users in 2025 by being built into WhatsApp, Instagram and Facebook, and Google's AI Overviews reached more than 2 billion people a month by sitting at the top of its search results. Reach isn't a guarantee, mind you: of Microsoft's 450 million-plus paid commercial Office seats in early 2026, only about 15 million were paying for Copilot, so being in front of somebody gets you seen rather than chosen.

Closer to home, Checkers Sixty60 turned the stores Shoprite already had into its delivery network, and it now delivers from 976 of them, with sales up 34.5% to R25.5bn in the year to June 2026. Pick n Pay's asap! runs from around 620 stores with about a quarter of the drivers, and in on-demand groceries the shop around the corner is most of the product.

For an established business this is usually the moat you already have and underuse, so put whatever you build next where your customers already are. If you're starting out, treat distribution as part of the product from day one, because a clever product with no route to customers is exactly what a bigger competitor will copy and hand to theirs.

2. Execution: anyone can build it, few deliver it every time

Distribution gets you to the start line, and execution decides who keeps the customer. Cursor started life as a version of Microsoft's own open-source code editor with AI models anybody could rent built in, up against GitHub Copilot and all of Microsoft's distribution, and it simply made the experience better. It passed $1bn in annual revenue by late 2025 and SpaceX bought it at a $60bn valuation in 2026, while Windsurf, a rival built on the same foundations, was split up and sold the year before, so the foundations clearly weren't what made the difference.

Execution is also keeping the promise every single time, which is much harder to copy than a feature. Costco has sold its hot dog and soda for $1.50 since 1985, and more than 90% of its US and Canadian members renew because they trust the promise to hold. Capitec did the same in South African banking with a promise of simplicity, and now has 25.8 million clients, almost double the next biggest bank.

AI makes this moat more valuable, because when customers can't tell products apart by what they do, they tell them apart by how they feel to use, how reliably they work and how they're treated when something goes wrong. Established businesses should use AI to make their promise more consistent, not just cheaper, and founders should take heart that a small team that sweats the details can still embarrass a giant.

3. Integration: the deeper you're wired in, the harder you are to unplug

Every system a product connects to is another reason not to replace it - what Hamilton Helmer calls switching costs in 7 Powers - and my favourite example is Epic, which runs the records of more than 40% of America's acute-care hospitals. When a large hospital system chooses a new records system it tends to choose Epic (both that did so in 2025 picked it), often so it can share data with the hospitals around it. When a wave of AI scribe startups arrived, Epic built its own scribe straight into the record, where it can draw on a patient's whole history without a separate contract, security review or fee.

AI deepens this moat, because a model is only as useful as the context it can reach. Microsoft prices Copilot on exactly that basis - only the paid version grounds its answers in your email, files, meetings and calendar - and whoever already sits in the middle of a customer's data starts the AI era with a head start nobody can prompt their way around.

Discovery has turned this into an art form. Vitality ties health, life and car insurance, banking and wearables together so that each extra product makes the others more rewarding (a Discovery Insure client who also banks with Discovery earns up to 20% more Discovery Miles for good driving), and the group is pulling all of it into one banking app. Leaving one part means giving up value in all the others, which is far stickier than a slightly better price.

If you're established, count the systems your customers would have to unplug to leave you, because each connection is a moat you've probably never put a value on. If you're a founder, plug into the systems your customers already live in from the start, since a brilliant standalone tool is precisely what gets copied.

4. Customisation: specific beats generic in a sea of copies

When anything generic can be produced with a prompt, generic is what everybody ends up with. Two months after OpenAI let anyone build their own version of ChatGPT, people had made more than three million of them, and a couple of months after that the store was filling up with knock-offs. The more of that there is, the more it's worth being specific, shaped around one business's customers, data and quirks in ways a competitor can't reproduce because they don't know what you know.

MIT's NANDA initiative looked for where AI was paying off in its 2025 report, The GenAI Divide. It has its critics (the sample is small and its headline figure has been quoted far too confidently), but its central finding rings true: only about one organisation in twenty had a task-specific AI tool in production with a measurable effect on the bottom line, and the barrier wasn't the models, it was that generic tools don't adapt to how a business actually works. One lawyer in the study kept reaching for a $20-a-month ChatGPT subscription over a $50,000 specialist contract tool that was too rigid to fit the work, which is a good reminder that specialised and customised aren't the same thing.

The irony is that the companies selling the most generic AI are now paying a fortune to customise it on site. Palantir built its business on engineers who embed with a client for months to fit its systems around how it really works, and in May 2026 OpenAI launched a company with more than $4bn behind it to do the same, a week after Anthropic announced a similar venture. If the people who make the models think the value is in the fitting, I'd take that seriously.

Established businesses are sitting on the raw material for this - their data, their processes, the way their customers actually talk - and it's worth more now than ever. Founders should pick a narrower customer and fit them perfectly, because 'good enough for everyone' is the one market the prompt has already won.

5. Complexity: everyone can make fried chicken

Everyone can make fried chicken, and very few people can make it taste like KFC's. KFC says Colonel Sanders' 1940 recipe of eleven herbs and spices sits in a safe at its Louisville headquarters, and that the blend is split between suppliers so no one holds the whole formula. When a relative of the Colonel showed the Chicago Tribune what might have been that recipe in 2016, KFC said it wasn't. Whether every detail of the legend is true matters less than what it shows: anyone can fry chicken, but nobody has copied the system that fries it the same way in nearly 34,000 restaurants, more than 1,000 of them here.

Toyota proved the same thing at industrial scale. In 1984 it opened a joint venture with General Motors in a shuttered GM plant in California, ran it on the Toyota Production System and had quality matching its Japanese factories within months. GM had a front-row seat for years and still couldn't make the system work in its own plants, because what made it work (the trust between managers and workers, the habits, the thousands of small routines) couldn't be photographed and taken home, which is exactly the picture for an era in which AI can copy anything you can see.

Michael Porter put numbers to this in 1996: if a rival has a 90% chance of matching any one thing you do, the odds of matching two drop to 81% and four to 66%, and they keep falling with every one you add. Sixty60 turns up again here for a reason - the app is the part anyone could copy, while the 29 distribution centres, the delivery business Shoprite bought out and the 10,000 or so riders and 8,000 in-store pickers behind it are not.

Established businesses often treat their complexity as baggage, and some of it is, but the parts that make you hard to copy are worth protecting rather than simplifying away. If you're starting out, pick the one thing you'll do better than anybody and build the hard, unglamorous machinery behind it that no prompt can produce.

The strongest businesses stack them

The five moats aren't alternatives, and the businesses I'd bet on have several at once, each one making the others harder to cross. Discovery is integration and complexity in the same breath, Sixty60 is distribution and complexity, and Microsoft has spent decades showing what distribution and integration do together. A competitor might copy any one of those, but copying them all at once is a different job entirely.

If you want to test your own position, these are the questions I'd ask, as if a well-funded competitor had copied your product last night:

• Distribution: could they reach your customers as quickly and cheaply as you can?
• Execution: would your customers notice the difference in how it's delivered, every single time?
• Integration: how many of your customers' systems would they have to unplug to leave you?
• Customisation: what do you know about your customers that no prompt does?
• Complexity: how much of what you do could somebody work out by watching from the outside?

AI has made the fried chicken easy, which is genuinely exciting, so the work from here is the recipe, the restaurant on every corner and the reason people queue anyway.