The hardest thing for AI to touch is the physical world.
A lot of smart people are arguing about what AI does to SaaS. I’m not going to litigate that here. The more interesting question is the one nobody’s finishing: if software is what AI eats first, what does it eat last, and how long do builders there have before AI catches up?
The version of that question that actually matters depends on who’s asking it.
Three different questions, three different answers
When people argue about “how long before AI eats hardware,” they’re usually mixing three different questions that have three different answers.
The frontier capability question: when does AI become technically capable of compressing the moats around hardware? This is what most accelerationist commentary is about. The honest answer is probably 3–5 years on several moats. Frontier capability is moving fast.
The median deployment question: when does AI compression reach the average factory, the average regulatory pathway, the average supply chain? This is where most of the diffusion-gap arguments live. The honest answer is closer to 10–15 years and counting, depending on category and geography.
The founder window question: if I start a defensible consumer hardware company today, how long do I have before AI compression becomes material enough to threaten the moats I’m building?
That last one is the question that matters if you’re allocating capital or starting a company. It’s also the one this thesis is actually about.
Where the founder window sits
The founder window isn’t the frontier number. It’s not the median deployment number either. It sits between them, and it’s weighted toward the realities of the businesses your readers actually build — typically low-to-mid-MOQ specialty consumer hardware, manufactured in geographies where deployment lag is significant.
My central case: 5–10 years, with seven as the midpoint. Roughly one fund cycle.
That’s not a capability claim. It’s a claim about how long the moats hold up at meaningful strength under realistic assumptions about how fast capability translates into deployed reality at the factory floor your hardware is actually being made on.
The four moats, briefly
Four structural moats determine where any given category lands in the 5–10 year range:
Iteration cycles — software iterates at the speed of a deploy; hardware at the speed of tooling, certifications, and container ships. AI compresses this fast, but it has a real floor. Compression timeline: 4–6 years.
Supplier reliability under stress — AI agents can source vendors in a week. They can’t earn trust at 2am when a batch fails QA. The frontier-vs-deployment gap is widest here. Compression timeline: 8–10 years for the categories most readers operate in.
Regulation — a rate-limiter, partially compressible. AI compresses prep; it doesn’t compress regulator review or liability frameworks. For health hardware, the moat is strengthening in real terms, not weakening. Compression timeline: 7–10 years.
Brand and taste — strengthens in premium and identity-driven categories, commoditizes in algorithmically distributed categories. Different timelines depending on which side of that line your category sits on.
Each of these gets its own post over the next four weeks. The math behind seven specifically — including the three numbers I have the lowest confidence in — is the post after that.
Why this is the right framing
Stating the thesis as a founder-window claim rather than a capability claim does two things.
First, it pre-empts the strongest version of the bear case. When someone says “but AI capability is moving faster than that” — they’re right, on the frontier. The founder window doesn’t depend on the frontier; it depends on deployment. Capability and deployment are running on different clocks, and the gap between them is wider than capability arguments tend to assume.
Second, it makes the thesis actionable. A capability claim is interesting commentary. A founder-window claim is a basis for allocation. If the window is seven years, you can build something defensible inside it. If the window is two, you probably can’t. The math determines the bet.
A bet placed in 2026 harvests in 2033 — central case. That’s roughly one fund cycle. Probably not coincidence.
Next: why hardware iteration has a floor, and what Nvidia just disclosed that almost made me revise the number.