← Writing · Thesis

Commoditized magic: what still has a right to exist

The framework behind every ellipsis investment

In 2024 we coined the term commoditized magic to describe the future we saw AI painting. Technology and products are becoming truly magical, unlocking capabilities that were previously impossible, yet they are almost completely commoditized by frontier models. We remain optimistic about the "magic" part: it introduces a massive economic opportunity by unlocking value that was previously inaccessible. But the commoditization risk is real and disruptive, and it makes entire areas uninvestable.

What changed

AI's capability to generate high-quality code brings the marginal cost of creating software to near zero, and it reduces the time and expertise needed to build it. Companies used to buy software and spend years customizing it; that complexity created lock-in. AI flips this entirely. Commoditization now rushes in from all sides at once: in a world where a twenty-dollar model can write code and copying a great idea takes days, writing code is no longer a lasting moat, nor is fast execution or a first-mover advantage. The way software is bought is shifting toward agentic automation, where brand and trust matter less over time in favour of automated, measurable decisions.

Intelligence itself is becoming a utility, like electricity or water. The price of a fixed level of intelligence has collapsed by two orders of magnitude in two years, open-weight models run always-on agents on a single consumer GPU, and enterprises route every request to the cheapest model that can do the job. Unless a company has a clear data or expertise moat, it will eventually compete on margins.

The most basic rule of venture has not changed

A company still needs differentiation and defensible moats to sustain high-margin success at scale. What counts as a defensible moat has shifted dramatically, with the bar rising to a much higher level. If a business lacks a genuine moat, whether proprietary data or unique expertise that can withstand an army of highly skilled AI agents, it will face disruption inside the commoditization kill zone. Many startups today are distribution layers for the model providers: capital flows in and leaves just as fast in API costs and infrastructure fees. That is not a moat. Brand helps a little, but in generative AI, where new winners appear every week, only a few names stick.

What we look for

There is clearly a massive opportunity for new unicorns, just with a higher bar for defensibility. In previous technology revolutions the biggest winners were not doing old things slightly more efficiently; they created entirely new businesses. We pursue two problem spaces:

  1. AI unlockers. High-value problems that were previously considered impossible and are now unlocked by recent AI capabilities, with a strong "why now".
  2. Disruptive AI-first technology. Small teams of world-class experts building critical parts of the emerging AI stack that Big Tech will not build but that are needed to unlock AI at scale. If AI had been invented before databases and APIs they would look completely different, more native to vector space; the startups thinking this way are building something incumbents are not even running toward.

For both we require clear defensibility against the commoditization forces. We believe the durable moats are data (problems that require unique proprietary data models do not have and cannot easily scrape or synthesize, with a flywheel that makes the model measurably better within months and creates switching costs), expertise (rare, hard to define or simulate, where small teams can still lead: security, optimization, edge computing, identity, payments, science), and, increasingly, compute (problems complex and valuable enough to justify expensive intelligence). We love companies that combine AI deep tech with deep industry or science expertise.

Two examples from our portfolio: a company automating and optimizing manufacturing plants, where the training data only exists if you install sensors in factories and every deployment generates more of it; and a company designing complex proteins for new immune therapies, where existing datasets are tiny and hard to expand and only a small subset of possible proteins has ever been explored.

The question we ask every founder

What right will you have to exist with healthy margins three years from now, once the models have improved and the incumbents have noticed? It has never been easier to start a company, build a prototype and run a proof of concept with a large enterprise; a lot of that POC money looks like revenue but is not tied to real scale. Speed matters only if you are moving in the right direction. Some of our companies take longer to reach first revenue, and when they do, it is far more defensible.

And when the moat simply is not deep enough for a venture-scale outcome, we say so. The same forces that make unicorns harder make another path possible: a small team using agents to run a herd of profitable niche businesses without outside capital. We wrote about it in Donkeys, Not Unicorns.

Further reading