Local models acting as autonomous dispatchers will drive the next leap in AI token volumes

Thomas Sohmers places the next order-of-magnitude growth in AI token volumes not in more humans typing into chat interfaces, but in local models acting as autonomous dispatchers for more capable cloud systems. The mechanism he describes is trust: once a user trusts the model on their laptop or phone to decide when and how to call a frontier model, the number of tasks routed through AI grows well beyond what any individual would initiate manually.

The call is specific enough to be checked. It requires that local model capability cross a threshold where users delegate routing decisions rather than making them. It requires that the cost and latency of local inference stay low enough to make that delegation feel safe. And it requires that the resulting token volumes, driven by machine-to-machine prompting rather than human-to-machine prompting, dwarf what current usage patterns produce.

The volume implication is the part that matters most for infrastructure planning. When a human prompts a model, there is one call. When a local model decides to break a task into subtasks and farm them to a cloud system, the call count multiplies before the user sees a result. Sohmers is describing exactly the mechanism that makes that multiplication happen at the individual device level, compounding across every device running such a local model simultaneously.

The trust problem Sohmers identifies is not a marketing problem. It is a reliability and transparency problem. Users will delegate routing only when they believe the local model will not silently make expensive, incorrect, or privacy-compromising calls on their behalf. That means the systems that win this transition will need to surface what the local model is dispatching and why, at least until the delegation feels routine. The analogy is email filters: few people inspect spam folders daily once they trust the filter, but they needed evidence of accuracy first.

The earliest version of the dynamic he describes is already visible in agentic coding tools, where a lightweight local process decides which API calls to make and in what order. What Sohmers does not specify is a timeline, which is the honest position given how much depends on hardware cost curves, model efficiency improvements, and user behavior that is genuinely hard to predict. Whether that pattern generalizes to a wide enough set of tasks to produce the volume jump he is anticipating is the open question.

The infrastructure bet embedded in his call is significant regardless. Anyone sizing compute capacity, API pricing, or network architecture on the assumption that token volume grows proportionally to human user counts is planning for a curve that Sohmers says will break. If local models become reliable dispatchers, the marginal token is no longer generated by a new user signing up. It is generated by an existing user’s device running autonomously while they do something else. That changes the economics of both the cloud providers receiving the calls and the edge hardware makers whose devices would be doing the routing.