AI Labs Will Trap You or Replace You - Dillon Rolnick, Nous Research
Nous Research CEO Dillon Rolnick argues that as AI models commoditize, frontier labs will try to recoup their enormous investments by locking businesses into their platforms or competing directly in lucrative sectors like finance. He contends that proprietary software moats are disappearing, that agents behave more like operating systems than chatbots, and that companies should build on open foundations they own, such as Hermes Agent, so they can swap models and inherit new capabilities.
Dillon Rolnick, CEO of Nous Research, starts from a simple economic problem: the labs have spent hundreds of billions, perhaps trillions, building frontier models, and those models are commoditizing. Users switch between models freely, the cost per token keeps falling while total token spend surges, and the infrastructure cost of serving tokens has stayed flat, so the margin that once came from a top-tier model is narrowing.
Two Ways to Earn the Investment Back
Rolnick sees two obvious paths. The first is to find a different layer of the stack to compete in, either down at infrastructure or up at the user. The second is to build something far more capable than anyone else and monetize it. User-facing products offer stickiness, and stickiness means margin, which is why the competitive fight has moved to the agent layer. Consumers, however, are fickle customers. Businesses spend heavily on tokens and have economically productive use cases, so the real target is the enterprise. Once a product is embedded in a business, it is hard to remove. If agents come to perform most of a company’s output, a vendor controlling that workforce gains enormous leverage, and could effectively hold it hostage.
Rolnick argues that labs with the most capable systems will point them at the most lucrative problems, starting with finance. Labs already see markets as a solvable and tractable problem. They lack domain know-how today, so they are partnering with investment banks and hiring analysts, but he believes they will eventually go directly to the source. He frames the stakes as larger than ordinary lock-in and price increases: the target is meaningful double-digit percentages of the global economy.
Don’t Build on Quicksand
His warning to enterprise leaders is that whatever foundation they choose will harden. Organizations accumulate know-how, and processes and systems grow around the chosen tools, so the choice shapes the future of the business. His advice is to build on open foundations that you own and control, such as Hermes Agent.
Why Nous Research Exists
Rolnick says Nous was founded primarily for moral and philosophical reasons, to counter a seemingly inexorable trend toward centralized control of one of the most important technologies ever invented. He points to science fiction as a warning, noting that Skynet was, after all, a closed-source company.
Software Moats Are Gone
The marginal cost of reproducing software has collapsed. When xAI released Grok Bots, the team liked its interface, pointed Hermes Agent at it, and reproduced the feature in a weekend, releasing it as open source within Hermes Agent. Software used to require coordinated engineering teams and long hours to assemble open-source libraries into proprietary products sold as SaaS. Now anything can be reverse engineered, and motivated open-source builders can point their agents at competitors’ products. In his view, the moats of yesterday lived in proprietary IP, and the moats of tomorrow live in network effects, reputation, and reliability.
Distillation and the Fight Over Open Models
Rolnick argues that the labs have underestimated both their own tools and their users. Models can teach other models to become nearly as capable as the originals, and the advantage in a model rarely lies in a proprietary architecture. It usually lies in data quality and scale. Labs have labeled this teaching process “distillation” and called illicit distillation theft, while training their models to refuse to help train other models and calling that safety. He rejects the comparison to copyright infringement, arguing that it would imply every human’s creative work had been taken for training.
Agents Are Operating Systems
Agents, he argues, are much closer to operating systems and frameworks than to chatbots. Their core is a combination of determinism and non-determinism: data sources, routine code, human opinion, and judgment come together to form a new kind of application that can inject reasoning and intelligence into what used to be purely deterministic software. Companies should build these systems on foundations they own, be able to swap models in and out, and build ever more sophisticated layers of functionality on top of these primitives. Most companies should not try to become intelligence companies. They should focus on what they do well, and when the state of the art advances, they should inherit those capabilities rather than rebuild and maintain them. He stresses that these tools should make people better and should not be drop-in replacements for them.
Hermes Agent and the New Moat
In Rolnick’s framing, the moat has changed from entrapment to reputation. He says Hermes Agent is the most used and downloaded open-source agent in the world, with far more downloads than its closest competitors. He wants to set a standard that everyone can build on, arguing that “water will continue to flow downhill” and that as long as someone keeps open-sourcing these agentic systems, he wants Nous to be the downhill destination that offers support, safety, and scalability alongside the best open agent. His final point is that organizations can have state-of-the-art features and reliability without giving up autonomy, and that they should build on something they own.