← sources

Most knowledge work doesn’t need a frontier model, which could be a problem for the AI trade

source: reddit (reddit:ValueInvesting)published 2026-10-13 16:05ingested 2026-10-14 02:44

retained text

I wrote [a post](https://eastwind.substack.com/p/below-the-frontier-the-battle-for) about a part of the AI trade that gets less attention than it should: whether everyday knowledge work generates enough revenue to support AI capex, and who captures it.

Right now, frontier labs capture most of the enterprise AI economics, mostly through coding. But a lot of knowledge work are bounded: tasks like drafting an email reply or pulling data from a financial filing only need a fixed level of intelligence. Past that point, a smarter model doesn't do the job any better.

As open-weight models get good enough, most of this work (by volume) can run on cheaper models in the cloud, or on the user's own device.

Where this matters for us is the revenue math behind the AI trade:

* We're already [spending $1T globally](https://www.goldmansachs.com/insights/articles/global-investment-is-forecast-to-exceed-1-trillion-in-2026) on AI. If that reaches multiple trillions by 2030, inference revenues need to be several trillions a year
* Coding has carried enterprise AI revenue so far. If its growth slows, mainstream knowledge work has to fill the gap, otherwise most names in the bottleneck trade will be affected
* Work that moves to cheaper cloud models earns a fraction of frontier prices, and inference on a user's device never reaches the data center at all
* Labs may get pushed toward being neutral platforms as customers get more sophisticated (OpenAI is already testing a hybrid approach with Baseten), which means slower revenue growth & thinner margins

I still expect the labs to keep the bulk of revenue from unbounded work like coding and science. But mainstream knowledge work is still up for grabs.

I wrote a longer version, with a 2x2 matrix on where models run vs. where data lives, how labs / Big Tech / startups are each approaching "mainstream" knowledge work, and why the labs may eventually want to own the data layer (I wouldn't be surprised if their corp dev teams have looked at Dropbox / Box) 

[https://eastwind.substack.com/p/below-the-frontier-the-battle-for](https://eastwind.substack.com/p/below-the-frontier-the-battle-for)

document fetched 2026-10-14 00:44 · canonical https://reddit.com/r/ValueInvesting/comments/1x50pmy/most_knowledge_work_doesnt_need_a_frontier_model