Tokens Need Work, Not Users

Consumer AI demand is easy to see. People ask questions, draft messages, make images, and try new interfaces. Workplace demand looks convincing too: add an assistant to every seat, put it beside the existing tools, and measure how often employees open it. Both can consume a great deal of compute. Neither guarantees a durable production system.

Optional tools stay optional. A person can skip the assistant, rewrite its answer, or return to the old process when the result is awkward. The company may pay for access while the real work still moves through the same meetings, handoffs, and manual checks. More usage can mean the tool is useful. It does not mean the business has made model capacity necessary to an outcome it cannot avoid producing.

Our position is that durable token demand will not be built one curious user at a time. It will be built by factories that consume models as part of completing work. When the model sits inside the production route, demand follows the work instead of somebody's willingness to open another chat window.

A business does not need to be persuaded to keep fixing defects, meeting obligations, changing products, or responding to production. Those jobs already have consequences. Attach agents to them, give the agents bounded authority, and require evidence before the result moves forward. Model consumption now has a return path: it purchases an accepted change, a verified repair, or a decision made against fresh evidence.

That distinction matters. An assistant sells a possibility to a person. A factory buys capacity against a queue the business already owns. The first depends on habit and enthusiasm. The second depends on whether the complete route costs less, moves faster, or produces stronger proof than the human coordination it replaces.

This is also why a clever demo says so little about demand. A model can produce an impressive artifact without being trusted to finish the job. The spend becomes durable only when specification, permissions, verification, release, and feedback are strong enough to let the machine carry the work all the way through.

The useful customer is not the employee with a prompt allowance. It is the operating system that can route thousands of different decisions to the right amount of machine capability. Routine, reversible work takes the lean path. Ambiguous or consequential work earns more context, stronger models, independent checks, or a human decision. Failed attempts return evidence instead of merely restarting the meter.

That factory can change suppliers without changing its purpose. Models compete for routes based on cost, capability, latency, and the evidence their outputs survive. A price cut can open more work. A stronger model can earn a harder class of task. A weak result can lose authority immediately. Demand can grow while model names come and go because the durable asset is the controlled route, not loyalty to one worker inside it.

This is the part the model market often skips. Cheap tokens do not create their own buyer. Better models do not create their own operating authority. The factory does that by connecting capacity to real work, enforcing the boundaries, and preserving proof outside the model's account of itself.

Companies already spend heavily on repeatable knowledge work. The spend is hidden inside salaries, coordination, waiting, rework, and the managers needed to keep handoffs moving. Agent demand becomes economically serious when factories take that work out of the org chart. The token budget does not need to appear from nowhere. It moves from the payroll and software seats attached to execution the machinery now performs.

That means roles will disappear. A company will not keep paying people to route tickets, restate context, perform routine transformations, and narrate status after a governed system can execute those steps continuously and retain the process. The people who remain will choose destinations, own standards, decide acceptable consequences, and improve the machinery when reality finds a gap. Their judgment moves higher, but the old number of seats does not survive as a courtesy.

This is why per-user AI economics point backward. If the technology works, some of the users being licensed today are exactly the positions the factory removes tomorrow. Demand tied to those seats shrinks with success. Demand tied to completed work grows with every process the factory can responsibly absorb.

Do not begin by asking every employee to find a use for AI. Pick one consequential flow of work. Name the outcome, the authority the factory needs, the limits it cannot cross, and the evidence that earns acceptance. Meter the entire route, including retries, verification, intervention, and recovery. Then compare the result with the human system it replaces.

If the factory cannot finish without constant rescue, the demand is not durable yet. If it produces output nobody can independently trust, more tokens will only manufacture more doubt. Strengthen the specification, routing, controls, and proof until model capacity becomes a dependable input to production rather than a discretionary experiment beside it.

We predict the largest buyers of model capacity will be factories, not audiences. They will consume tokens because the business has work to complete, and they will stop buying any route that fails to justify itself. The winning model may change every quarter. The work keeps arriving. Build the machinery that can buy exactly enough intelligence to finish it.