Where Tokens Meet Atoms

Everything is very simple in war, but the simplest thing is difficult. These difficulties accumulate and produce a friction, which no man can imagine exactly who has not seen war.

An inference can finish before a truck reaches the next traffic light. At first this appears to be a difference in speed. It is more disruptive than that. The inference and the truck inhabit different regimes of consequence, and an economy that makes the first abundant does not thereby make the second obedient.

Call the passage between them actuation: the movement from a computational judgment into a change in someone's world. An order is placed. A payment settles. A robot moves. A permit is refused. A molecule is synthesized. A contract binds. Some of these acts can be reversed and some cannot, but all cross a boundary that an answer alone does not cross. They consume resources, alter positions, or expose someone to loss.

This is where computational abundance meets its resistance. A system may produce a thousand designs while a factory can test only three. It may draft a contract in seconds while authority to sign it remains contested. It may find a route around congestion while the truck still moves through weather, labor rules, loading docks, and streets. The cheapening of proposal does not cheapen consequence at the same rate.

That difference is not a residue awaiting a better model. It is part of the political economy of the model's use.

Different Difficulties

Moravec observed that tasks easy for human bodies can be difficult for machines even when abstract calculation is easy.¹ The contrast is not a permanent ranking of abilities. It directs attention to the density of feedback involved in moving through an unprepared world. A chessboard declares its state. A warehouse, clinic, road, or construction site does not.

Baumol's account of unequal productivity growth points to another consequence.² When the cost of one activity falls much faster than the cost of a complement, expenditure migrates toward the complement even if that activity has not become less important. Cheap cognition can therefore enlarge the share of cost attached to fabrication, inspection, care, authorization, and repair. Abundance in one layer reveals the price of the next.

Several difficulties recur at that boundary, but they should not be mistaken for a universal stack. Their order changes with the work.

Physical throughput is the clearest. Chips must be packaged, transformers built, facilities connected to a grid, goods carried, and machines maintained. Reports on advanced packaging, large power transformers, and interconnection queues describe delays measured in years rather than software-release cycles.³ Those reports are dated observations from particular markets, not irreducible constants. Capacity can be built. What matters is that intellectual acceleration does not build it by declaration.

Access creates a different friction. A recommendation acquires force only when a system can reach the account, machine, database, or legal power needed to execute it. Authentication proves control of a credential. Authorization specifies what that credential may do. Neither explains whether the act served a legitimate purpose. Large organizations have accumulated these distinctions in permissions, approvals, audit systems, and divided offices because the ability to act is itself a source of risk.

Then the world must answer back. A returned API status may establish that a command was received. It does not establish that a package arrived intact, a weld will hold, a diagnosis was sound, or a person understood the bargain. Sensors can make physical states more legible, but their placement, calibration, interpretation, and legal significance remain institutional achievements. More observation can narrow uncertainty without eliminating the dispute over what the observation means.

Finally, consequence needs somewhere to attach. A terminated runtime cannot compensate an injured party or appear in court. The relevant principal may be the firm that deployed it, the owner of the asset it controlled, the professional who adopted its judgment, or another institution that accepted the risk. Identifying that principal is not a concession to human ceremonialism. It is how assets, duties, evidence, insurance, and remedial power are brought within reach of the act.

Autonomous driving makes the conjunction visible. Waymo has expanded service through bounded operating domains rather than by releasing a general driving capacity everywhere at once.⁴ Capability matters, but so do roads, weather, maps, fleet operations, local permission, insurance, incident response, and the conditions under which the company will accept a ride. The boundary of deployment is made from all of them.

What Can Be Verified

Chapter 16 proposed V/C as one candidate ordering variable: value at stake relative to the cost of checking an output to acceptable confidence. Actuation shows why the ratio cannot order deployment alone. Checking a proposed route and discovering that a shipment was spoiled are different acts. So are detecting an error, preventing it, assigning the loss, and repairing the injury.

A compiler can reject malformed code before execution. A temperature sensor can preserve evidence after a cold-chain failure. A court may decide years later who bore the duty. Each reduces a different uncertainty, at a different time, for a different party. Compressing them into one denominator would conceal the institutional design the comparison was meant to reveal.

The useful question is narrower and harder: which uncertainty prevents this decision from becoming an act? In one setting the obstacle may be a missing sensor. In another it may be a permit, an insurer, a fabrication slot, a counterparty willing to rely on the output, or a principal with assets exposed to remedy. If the same obstacle persists across otherwise comparable deployments, it begins to look like an economic constraint rather than an anecdote.

This also explains why “human in the loop” is a poor measure of control. A person may be present but unable to inspect the evidence, refuse the recommendation, or alter the outcome. Another system may act without contemporaneous approval while a durable institution remains fully responsible, preserves the record, and can reverse or compensate for the consequence. What matters is not the visibility of a human gesture but the distribution of knowledge, authority, and remedy.

Electricity at the Boundary

The Joule Standard does not travel intact into this chapter. Bitcoin mining and model inference may compete for electrical capacity at a site where the operator can perform either use, but most acts do not share that conversion path. A pharmaceutical shipment, a steel mill, and a legal authorization cannot be priced by asking what their energy would have earned in a mining machine.

Even inside a data center, the comparison is conditional. Mining requires specialized hardware and access to a protocol and market. Inference requires suitable accelerators, software, customers, and service obligations. Power, cooling, buildings, and some operating knowledge may be shared, while the commercial systems are not. The mining return can matter to an operator with a feasible choice. It cannot become a universal tariff upon cognition simply because electrons enter both machines.

What the comparison does reveal is an allocation problem. Computational infrastructure can direct the same constrained inputs toward immediate service, reusable capability, protocol-recognized settlement work, or another experiment. As systems participate in making that choice, they enter the loop by which capital decides what will exist next. Actuation begins before a robot moves. It begins when a proposal gains the authority to consume a scarce resource.

More Than Two Clocks

It is tempting to divide the economy into a fast digital layer and a slow physical one. The distinction is useful until it becomes too clean. Digital execution can be followed by years of litigation. A physical sensor can report in microseconds. A protocol may settle according to its own rules while ownership remains disputed elsewhere. A hospital may change practice more quickly than a software firm can replace a deeply embedded system.

Perez's history of technological revolutions nevertheless gives the mismatch economic significance.⁵ New capabilities can attract capital before complementary infrastructure and institutions are ready to use them widely. Valuations then move on one clock, grid connections on another, organizational practice on a third, and public law on a fourth. A correction may follow, but no historical template fixes its date or form.

The strongest thesis is not that atoms are always slow and tokens always fast. It is that computation can expand the supply of possible action faster than the surrounding world expands its capacity to select, authorize, absorb, and answer for those possibilities. Pressure then accumulates at whichever boundary resists: energy, fabrication, access, observation, law, or trust.

That pressure can produce rent. It can also expose work that was never dispensable. A professional signature may conceal an obsolete monopoly, or it may identify the person who must examine uncertainty and remain answerable. A logistics network may exploit its position, or it may coordinate vehicles, workers, claims, and recovery in ways a route optimizer does not. The task is to distinguish the gate from the office it was built to perform.

When computational systems can initiate more acts, that distinction becomes harder to postpone. The next chapter follows the transaction across the boundary. Its subject is not whether a runtime can sign a contract, but which principal has already decided that its signature, assets, and authority may stand behind what the runtime does.