The Only 3 Moats Left in AI

Detailed Summary

Between mid-January and mid-February 2026, $2 trillion was wiped from the software market cap in just 30 days.
This crash represents an investor realization that traditional software valuations have peaked because software itself is no longer defensible.
AI tools like Claude Code and Opus 4.6 have "crossed the Rubicon," making it possible to build SAS products and database migrations instantly, destroying the moat of "hard-to-build" software.

Proprietary data only functions as a moat if it adds real ROI and produces a differentiated output that an LLM couldn't generate on its own.
The "data bubble" for training models is bursting; model providers now prioritize high-quality, curated reasoning data over raw volume due to the rise of synthetic data.
Expert consultants have a massive opportunity to scale by codifying their "playbooks" into markdown files and skills that an agentic system can execute.
Case study: A 30-year litigation attorney's expertise was codified into an agentic legal assistant, creating a moat that a person "off the street" with the same AI cannot replicate.

Aggregating the "crowd" creates a dataset where the consensus is more accurate than any individual expert (e.g., a jury vs. a single judge).
Eric Schmidt predicted in 2016 that the next trillion-dollar companies would use the crowd to learn and then sell that learned intelligence.
Examples include Polymarket for predictions, Reddit for human-verified information, and Waze for real-time traffic data.
Companies like X (formerly Twitter) and Reddit are putting up "API walls" because they realize their crowdsourced human data is their only remaining value.

Hardware is the "hardest" moat due to the opex required for manufacturing, supply chains, and inventory management.
The real value of hardware in the AI era is capturing proprietary data that doesn't exist on the internet (e.g., John Deere soil sensors, Apple Watch biometrics).
Future startups will use hardware (like robot dogs mapping rainforests) to feed "world models" that understand physical environments.
The hardware flywheel: More devices lead to more proprietary physical data, which makes the AI powering the device better, leading to more sales.

Features that are not moats include: better UI, more features, per-seat pricing, being first to market, and even brand loyalty.
In 10 years, there is a 99.9% chance that AI will be on a full exponential path, rendering the digital software space completely collapsed.
Strategy for the immediate future: Focus on acquiring or partnering with expert knowledge, as it is the easiest moat to plug into agentic systems right now.