ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial4 min readDisconfirmAI Editorial

The Harness-Model Flywheel: Why AI Application Companies Are Not All Depreciating

An investment newsletter argues that harness logic adding intelligence is a depreciating asset, but harnesses selling infrastructure, trust and efficiency appreciate as models improve, and proposes a five-trait test for distinguishing the two.

2026-11-23 03:55 GMT+8

Every model release eats some of the intelligence that a harness used to patch in. That much is true and has been true for two years. The question Rob May poses in his November 22 newsletter is whether this means the harness business itself is dying. His answer is no, and the distinction he draws between two claims investors tend to merge is the most useful part of the argument.

The first claim is that harness logic which adds coding intelligence is a depreciating asset. May agrees. The second is that harnesses are a depreciating business. He disagrees, and his reasoning runs through what he calls the Model-Harness Flywheel. Harness engineers find where models break and patch it with code. Those patches generate usage data. Labs train on that data, and the next model no longer needs the patch. The harness discovers a skill, the model absorbs it, and the harness must move to the next frontier.

May grounds this in a hardware precedent. In 1989 Intel's 486 pulled the floating-point co-processor onto the main chip. The standalone math chip business faded, and software that leaned on fast floating point became the norm. Absorption killed one layer and fed the next. The analogy is illustrative rather than quantitatively verified, but it captures the structural point: when a capability moves from an external component into the core system, the component vendor loses revenue while the ecosystem built on top of the now-cheaper capability grows.

What makes the harness business different from the floating-point chip business, in May's framing, is that some of what harnesses provide are not skills at all. They are properties of the environment the model runs in. A model cannot train its way into being your audit log, your permissions system, or your cloud bill. Infrastructure, trust and security, and efficiency are not things that get absorbed into model weights. They are things that get more valuable as the model becomes more capable and more widely deployed, because there is more to secure, more to route, and more to optimize.

Sign in / Sign up
ReadingResearchRadarInvestment framework

May is explicit about his bias. He runs TaskRouter, a company focused on inference efficiency, and he links to it in the piece. His efficiency argument rests on a hypothetical: a company spending $10 million a year on inference could cut its bill to $3.7 million if its harness routes 70 percent of calls to a task-specific model costing a tenth as much. That is $6.3 million of margin, and the savings grow as small models improve. He also notes that labs have little incentive to route traffic to a competitor's cheaper model, which leaves room for neutral efficiency layers that a lab cannot easily own.

The practical contribution is a five-trait checklist. May proposes that if a harness or agent company has at least three of these, it is positioned to benefit from model advances rather than be destroyed by them. Proprietary failure data: its usage shows where models break before the labs see it. Revenue tied to infrastructure, trust, or efficiency rather than to patching model weaknesses. Model-agnostic: better and cheaper models lower its costs instead of replacing its features. Frontier speed: it moves into new domains as fast as models absorb the old ones. Trust lock-in: security reviews and compliance work that take months to redo.

The strongest objection is that this framework has no testable track record. The five traits are plausible but unvalidated, and the author's commercial interest in the efficiency layer means the checklist conveniently highlights the category he operates in. The Intel analogy is suggestive but not conclusive: floating-point was a narrow computational primitive with clear performance metrics, while infrastructure, trust and efficiency are heterogeneous and harder to isolate as a single value proposition. Whether a company with three of these traits actually outperforms one with two remains an empirical question that May does not attempt to answer.

Still, the core distinction is sharp and useful. When evaluating an AI application-layer company, the single most important question is what happens to its product when the next model ships. If the answer is that it gets better, the company may be a flywheel beneficiary. If the answer is that the team must rebuild, the company is renting time against the model roadmap. That is a concrete test a reader can apply without needing to verify any of May's specific numbers.

Sources

  • investinginai.substack.com↗
Continue researching →