AI Splits the Market

Welcome!

This website uses cookies to ensure you get the best experience.<br>We use them to remember your preferences and analyze site traffic.

Accept

Learn More

AI Splits the Market

June 09, 2026

Tagged in Artificial Intelligence, Musings

Comments [0]

Most predictions about the AI era assume one outcome across the whole software industry. Startups will eat incumbents. Greenfield will replace brownfield. Faster building means faster disruption. I don't think any of that is true uniformly. The same technology cuts in opposite directions depending on which market it lands in.

A recent NBER working paper by Mert Demirer, Leon Musolff, and Liyuan Yang, Writing Code vs. Shipping Code, tracked over 100,000 GitHub developers from 2022 through 2026 and found that sync agents increased lines of code by 741% and pull requests by 65%, while shipped releases moved only 20% and end-user downloads barely moved at all. The gain shrinks at every step from keystroke to user.

This speaks to a question I raised when DeepSeek's cheaper training run sent NVIDIA stock sideways and Satya pointed to Jevons paradox as the reason not to panic. I think the answer is split. Jevons holds for compute and for lines of code written. It fails for finished applications competing for human attention. The app stores show more supply, flat downloads, and a growing share of apps with effectively zero audience.

Which raises the question, if supply is up and consumer demand is flat, where is this 741% increase in code actually going? I think eventually it will end up in enterprise scenarios, starting with backlogs.

Every organization I have ever worked with curates a queue of deferred work that has been growing for years, filled with low and medium-priority bugs, unfunded accessibility audits, and internal tools that everyone needs but no team owns. None of those items ever shipped because the math did not work, not because the demand was missing. Now the cost side of that math has moved. Work that could not justify a developer-week three years ago can justify an agent-hour today, which is also why the ruthless prioritization I argued for matters less than it did. When the cost of saying yes collapses, the discipline of saying no loses its edge.

Consumer markets favor the startup. Switching costs are low, distribution runs on attention rather than procurement, and novelty can win. AI shortens the build cycle without changing any of that, so the startup advantage only gets sharper.

Enterprise is the opposite. The constraint there has never been the speed of building, it has been integration, compliance, sales cycles, data residency, audit trails, and the long tail of trust that might take a decade or more to accumulate. AI does not compress any of those. What it certainly does compress is the cost of working down an existing backlog, and the organization with the backlog is, by definition, the incumbent. The same technology that lets a startup ship a polished consumer app in a weekend also lets a fifteen-year-old enterprise vendor finally close the tickets they have been deferring since 2014.

If this split holds, AI will be a productivity shift, not a platform shift. Platform shifts (cloud, mobile, SaaS) created openings for new enterprise winners. Productivity shifts accrue to whoever already owns the category. The incumbents do not have to win. They just have to not lose.

The consumer race will be loud and visible. The enterprise race will be slower, less photogenic, and probably more durable. Enterprise is where the value will accrue, even as the consumer space gets all the attention.

Share on BlueSky Twitter Facebook LinkedIn

Success!

Comment Section

;:[\]]{2,}$" data-val-required="The Name field is required." id="Name" maxlength="60" name="Name" type="text" value="" /><br>Name

Email (will not be displayed or shared)

Home page (optional)

4+4=?

Post a comment

Submit

© 2026 - Mark Downie, All Rights Reserved.

Powered by Dasblog-Core commit 88a70f