What will actually matter when intelligence costs nothing?
The MOAT after AGI
In 1995, companies put “visit our website!” in press releases. Analysts raised price targets. Consultants sold “internet strategies” for six figures.
By 2005, a website was like a phone number. The companies that built their moat on having the internet had vanished. The ones that built on top of it — Amazon with logistics, Google with the intent graph — owned the next decade.
“We use AI” is today’s “we have a website.” And just like in ‘95, most companies are confusing access to the technology with a competitive advantage.
AGI is an infrastructure layer. It will commoditize. In ten years, having an AI system will be table stakes. So what will be a reason to win?
The intuitive answer: things AGI “can’t do.” Trust. Taste. Brand.
I don’t buy it. Betting against AGI capability has been a losing trade. If it’s truly general, assume it can reconstruct anything given enough time and data.
The better question: what can’t you build from a standing start when your competitor has the exact same AGI you do?
When everyone has unlimited intelligence, intelligence cancels out. The only differentiator is where you already are.
The real moat is a self-evolving system coupled to real-world feedback that has reached escape velocity before anyone else.
Self-evolving, not static. Most companies will use AGI as a calculator. Query in, answer out. The winners will build systems that run experiments, self-evaluate, accumulate context, and compound autonomously.
Coupled to reality. A loop running on synthetic data can be cloned and fast-forwarded. A loop feeding on real-world signals — customer behavior, market outcomes, regulatory responses — cannot. You can’t fast-forward reality.
Escape velocity. The system with the best architecture, the most compute, and the best seed data compounds faster. Better experiments → richer data → faster cycles. It’s not linear — it’s exponential divergence. Two systems launched six months apart aren’t six months different. They’re in fundamentally different states.
Amazon didn’t win by launching first. It reached escape velocity on the flywheel: selection → traffic → sellers → selection. Once past the tipping point, no capital could close the gap. Google, same story. Queries → data → relevance → queries. Same exponential divergence. Same unbridgeable moat.
Tesla has done the same with navigation data from Tesla’s cam & sensors.
The AGI era will produce the same pattern. Faster, and more violently.
Two layers of defense:
Accumulated state. The system’s evolved heuristics from thousands of real-world cycles. Architecture is copyable — it’s code. State is not — it’s the residue of time spent in contact with reality.
The fitness function. What you optimize for. Two identical systems with different fitness functions diverge radically. Choosing the right one is taste applied to systems engineering — possibly the last irreducibly human decision in the stack.
The test is simple. Ask: Could a competitor with unlimited AGI and unlimited capital, starting from zero today, catch me?
If yes — you haven’t reached escape velocity. Move faster.
If no — that’s the moat. Not the AI you use. The compounding machine you built on top, and how long it’s been running.
Start your loops now. Feed them reality. Choose your fitness function carefully. And reach escape velocity before someone else does.
If this made you think — or disagree — share it.
There are perhaps different moats and I'll now elaborate based on personal experience…
Step 1/
Develop the skill to focus on a single thing for an extended period of time while working on Your metacognition. Metacognition as being able to reflect on Your own thoughts and iterate on new ideas.
Step 2/
Understand how to parallelize Your work. Parallelize as in understanding how to make multiple agents work on a given project/idea at the same time.
Final Step/
Be disciplined and create a log of what You learn and understand that the natural result of Your iterations should be a SYStem that You can follow to get the desired result across multiple domains.