AI Labs Will Trap You or Replace You - Dillon Rolnick, Nous Research

An engineer can try a new model as soon as it looks promising. A financial firm that has spent years building an investment process around an agent faces a different decision. Changing the model might be easy. Moving the process means preserving how the firm brings together its data, its code and its analysts’ judgment. Dillon Rolnick, CEO of Nous Research, argues that the gap between those two switching costs explains why AI labs have such a strong incentive to own the agent layer.

Rolnick pointed to charts showing that the average price of a token was falling even as total spending on tokens rose. He described the infrastructure costs associated with serving those tokens as roughly flat. In his reading, customers were buying more intelligence while competition was squeezing the premium a lab could charge for a capable model. For a lab trying to recover an infrastructure and research investment that Rolnick puts in the hundreds of billions of dollars, growing token demand answers only part of the business question.

One response is to move closer to the user. Businesses make especially attractive customers because they put tokens to economically productive use. As employees learn a platform and build processes around it, the price of a competing model becomes a poor measure of what a switch would cost. The incumbent gains room to raise prices precisely because the customer has done so much work of its own.

Rolnick asks executives to imagine that, two or three years from now, agents account for 70 percent of their company’s output. The figure is his forecast, and the unit matters. He is talking about the share of work produced. Even a firm that retains its people could depend on an outside platform to organize much of what those people deliver. A change in the platform’s terms would then reach well beyond the software budget.

Finance illustrates a second incentive. A lab that believes its systems can solve a lucrative financial problem can sell a firm the intelligence to tackle it, or it can pursue the returns itself. Rolnick thinks labs will eventually favor the latter. He also identifies what makes that difficult today. Labs lack some of the context, know-how and experience that financial firms take for granted. He points to labs hiring investment banking analysts and partnering with established firms as ways to acquire that expertise.

Whether those partnerships develop into direct competition is a prediction. The engineering decision comes sooner. Every time a team encodes another part of its investment process, it chooses where that knowledge will live and what it would take to carry it elsewhere.

Rolnick offers an example from his own team to explain why a proprietary feature may buy little time. When xAI released Grok bots, the Nous team liked the interface. According to Rolnick, they used Hermes Agent to reproduce it over a weekend and released the result as an open-source Hermes feature. The scope of the story matters. What Nous wanted and says it reproduced was a particular user experience. Engineers comparing the two products would still have to assess how reliably each one delivers it.

Rolnick sees a related pressure on model developers. In distillation, one model supplies outputs that can become training data for another. He argues that much of a model’s advantage comes from the quality of its data and the scale of training, and that teaching from a capable model can help another approach its performance. He describes labs objecting to unauthorized distillation and training models to decline assistance with building other models. His broader bet is that, as features and capabilities become easier to reproduce, reliability, reputation and network effects will account for more of a supplier’s staying power.

Rolnick compares an agent to an operating system or an application framework. Consider again what goes into an investment decision. A firm has data sources, routine code that operates on those inputs, people with opinions, and a point at which someone has to exercise judgment. An agent gives the application a place to combine repeatable computation with reasoning whose result can vary. A chat interface exposes only a small part of that arrangement.

That distinction suggests a practical test for the architecture. Trace an investment decision from its inputs through the routine code to the point where the application calls on a model. If the firm chooses a different model, can it still bring the same sources and analysts’ views to bear? The new model may reason differently, which may be the point of switching. The firm should still control which evidence reaches it and how its reasoning fits into the larger decision.

Maintaining an entire agent stack has a cost of its own. A financial firm that builds every agent capability internally takes on the job of updating that machinery as the field advances. Rolnick draws the boundary at the work the firm is in business to do. Keep control of the investment process and the choice of model. Build on a framework through which the team can inherit improvements to the common machinery.

Hermes Agent is Nous Research’s implementation of that argument. Rolnick says teams can choose where to run it, modify it and continue using it if their relationship with Nous ends. His commercial ambition is for Hermes to become a standard and for customers to seek out Nous for support, safety and help operating at scale. That makes the question concrete for a buyer evaluating any agent platform. If the supplier changed its prices tomorrow, which parts of the application could the team run for itself, and which services would it still willingly pay the supplier to provide?