March 2026

Revisiting Predictions

Last year I wrote a set of predictionsabout how the tech ecosystem would evolve. Those predictions shaped how I thought about building my new company, which I started roughly ten months ago. A lot has changed since then, so it's worth revisiting what I got right, what I got wrong, and where I think things are heading.

What I still believe

Most of my original predictions hold. I continue to think the following areas represent genuine, durable moats:

• AI for scientific discovery. Being at the frontier of scientific knowledge is hard to replicate and compounds over time.
• Hardware. Barriers to entry are real: patents, raw material constraints, supply chain complexity, regulatory compliance, and demand that shows no sign of slowing.
• AI safety. AI will become ubiquitous. Making it safe is an existential priority, and humanity will direct enormous resources toward solving it for as long as that remains true.
• Crypto. Crises are a permanent feature of the world. An independent store of value will always find demand.

Where I was wrong: network effects

I no longer believe network effects constitute a meaningful moat, and I think the reason is worth spelling out carefully.

LLMs have become genuinely capable of writing production code and operating computer systems at a genius level. AI agents can now interact with software autonomously and, critically, they can communicate and coordinate directly with one another. This changes the logic of network effects entirely. The reason platforms like WhatsApp or Telegram have moats is that people are on them - switching requires convincing your entire network to move. But if agents can establish communication directly, regardless of which platform either party uses, that lock-in disappears. There is no longer a need to be on the same network. Agents are capable of maintaining their own trusted infrastructure and ranking the trustworthiness of counterparts autonomously, without relying on centralized third parties.

Adoption is also happening faster than most people expect. ChatGPT's growth was already historically rapid. Each successive wave of transformative technology tends to be adopted faster than the last - early internet adoption was slow because it required physical infrastructure (fiber, antennas, hardware) to catch up with demand. That constraint no longer applies to software. It is entirely plausible that within a few years, every person will have access to a personal AI agent (OpenClaw is early evidence of such a trend), much as everyone now has an email address or a smartphone - and that agent will be, in practical terms, a highly capable autonomous genius.

How incumbent platforms will adapt

This does not mean companies like Airbnb or Booking will disappear. I expect their transition to be gradual and relatively smooth. Initially, agents will simply index and interact with these platforms on users' behalf. Over time, the platforms will adapt their business models: competing on infrastructure efficiency to lower costs, or shifting toward services where their core value lies in risk assessment and insurance rather than in acting as an intermediary.

The wrapper problem, revisited

I noted previously that companies building thin wrappers around LLMs could generate significant revenue quickly, even without a durable moat, making them attractive to investors seeking short-term returns. I still think that's true in the near term, but the more interesting model going forward is royalty-based investment - taking a stake in revenue rather than equity - which better matches the risk profile of businesses that are profitable but structurally fragile.

Where real moats will form

The most defensible positions will belong to those who use AI to produce things that are hard by definition: formal proofs, algorithms with proven performance guarantees, genuine scientific discoveries. These outputs are difficult to replicate, and the cost of reproducing them independently is high enough that licensing or paying for access simply makes more sense. That difficulty is the moat.

The future of “simple” software

I think the value in software is shifting toward ideas. As AI makes it trivial to generate functional code, the scarce resource becomes clarity and creativity - software that does something genuinely useful, does it simply, and does nothing else.

My expectation is that people will increasingly pay voluntarily for software that is focused, well-designed, and free of unnecessary complexity. Open source becomes the default distribution model, and payment becomes optional - a signal of appreciation rather than a gate. The downstream effect is that simple, well-crafted software no longer requires outside investment to sustain itself. Founders can become self-sufficient. The capital-intensive model that currently dominates software development becomes less relevant as the cost of building falls and the value of a good idea rises.

Extra thoughts

Late-stage investors have traditionally managed risk by backing companies large enough to have predictable revenue, defensible market positions, and institutional staying power. That advantage is eroding. The cost of building software is collapsing, and small teams - in some cases a single person - can now build and ship products that would have required dozens of engineers just a few years ago. Scale is increasingly a liability and large organizations carry indefensible overhead. The behemoth model is getting cooked.

That said, some sectors will likely remain structurally resistant to this shift. In areas defined by high regulation, mission-critical reliability, or relationship driven trust with extremely high stakes - defence, finance, healthcare - size and institutional credibility still matter. These may be where the classic late-stage investment playbook remains most relevant - but I would still be concerned.