| url | https://blog.eladgil.com/p/ai-startup-vs-incumbent-value |
|---|---|
| raw | raw/highlights-ai-startup-vs-incumbent.json |
Elad Gil’s framework for thinking about which technology waves favor startups and which favor incumbents — written just as the post-ChatGPT scramble was getting underway. The most useful single piece for figuring out where to actually start an AI company.
The historical pattern
Gil maps the splits:
- Internet (1995–2010): ~70:30 startup-to-incumbent. Almost everyone with a serious internet business was new.
- Mobile (2007–2015): ~20:80 incumbent-favored. Incumbents just shipped mobile apps. The breakout pure-mobile startups were exceptions.
- First wave of AI (pre-LLM, 2015–2020): went heavily to incumbents. An incumbent that was 50% as good still won when bundled into an existing product with an existing user base.
His read on the LLM wave: this one swings back toward startups, but only in specific market structures. Google was in the obvious default position to win generative AI and fumbled it badly enough that Gil called it an “Xerox PARC moment” for OpenAI — a generational research lead that the holder failed to convert into a product position.
Where AI startups actually win
Gil’s beachhead criteria:
- Highly repetitive, highly paid tasks. The economics work out when the human is expensive and the task is boring.
- Imperfect fidelity is acceptable. Human-in-the-loop is natural; the AI doesn’t have to be perfect to be useful.
- Weak or nonexistent existing workflow tools. An incumbent with deep software won’t lose easily. An incumbent with a clipboard absolutely will.
- Summarization or generation is core to the work. This is the part LLMs are unambiguously good at.
These are the beachheads. Pick a job that fits all four and you’re inside a defensible niche before the incumbents wake up.
Why incumbents usually win
Bundling. An incumbent can add AI features to an existing product with millions of users for nearly zero distribution cost. A startup has to build the product and acquire users. This is why so many first-wave AI startups failed — they competed head-on with incumbents who could “just add AI” (counter-positioning). The exception is when the incumbent’s existing product is itself the obstacle, which is when counter-positioning becomes available.
The hammer-looking-for-nail trap. This is the warning Gil keeps hitting: exciting tech ≠ real product. “Identify actual end-user needs and unserved product/markets.” The idea-maze-concept matters more, not less, when the underlying technology is this general-purpose. If you can build anything, the only question that matters is what.
The ChatGPT paradox
Gil’s framework would have predicted that an incumbent with a research lab (Google) should have won generative AI. The framework was right on the framework, wrong on the prediction. ChatGPT’s success is the case study in what happens when an incumbent fumbles a generational handoff — and what happens when a startup ships a research preview into a market it didn’t realize existed. The framework now needs an addendum: incumbent advantage is conditional on the incumbent actually shipping. They often don’t.
Connections
chatgpt-pmf · counter-positioning · crossing-the-chasm-concept · moats · google-no-moat · cornered-resource · data-moat