Two years ago, building a working AI product was the hard part. Now it's often the easiest part of the business.
Foundation models keep getting better and cheaper to call. What took a specialist engineering team six months to build in 2023 now ships in an afternoon with the right API keys. A founder can get from idea to working demo faster than at any point in software history.
You'd think that's good news for competition. It isn't, particularly. If any team with a laptop can build a plausible AI product in a weekend, "we built an AI feature" stops being a pitch. Investors, customers and competitors all know it. The technical achievement that used to signal a defensible business now signals very little on its own.
So what's left to differentiate on? Not the model, and not much of what's built around it. What still matters: proprietary data that gets better the more it's used, distribution into a market a competitor can't easily reach, and switching costs built from real workflow integration rather than a slicker interface. Mach42, one of our portfolio companies, is a case in point: its AI sits on top of years of engineering simulation know-how, not a thin layer over someone else's model. Alloyed is another: translating models into novel metal alloys that have to perform in critical, real-world applications, including defence, is a very different proposition from generating plausible outputs on a screen.
There's a second pattern underneath this, and it's less comfortable. Most of the economic value created by this AI cycle has gone to a small number of companies building the infrastructure everyone else depends on: the compute, the foundation models, the chips. What gets built on top of that infrastructure is still an open question, competing in a market that resets every time the underlying models improve.
It's why, at Future Planet Capital, we ask what a company solves before we ask what it's built. Technology is a moat when it's genuinely hard to replicate: deep engineering, real IP, years of accumulated know-how. It's rarely a moat when it's a thin layer over someone else's model. Increasingly, most AI products are the latter.
None of this makes AI a bad place to invest. It makes it a much harder place to tell a real business from a very good demo. | 
