Silicon Metabolism

There is a compelling analogy between biological organisms and industrial activities.

Why should computation occupy a special place among energy-intensive industries? Steel, aluminum, cement, and ammonia also depend on vast physical plants. Their products transform civilization. Material intensity alone establishes no new factor of production.

Computation differs in the range of patterns it can operate upon. Shannon separated the information in a message from the medium carrying it, while never denying that every message requires a medium.¹ A calculation may be instantiated in silicon, optics, mechanics, or another physical system, with radically different costs. Its logical relations can survive the change of substrate.

That portability gives computation an unusual industrial reach. A computer can help model a steel mill, search a turbine design, compare molecules, schedule a grid, or draft the instructions by which another computer will act. It cannot replace every experiment, and formal universality does not make every simulation tractable. Where the mapping works, however, one physical plant can supply cognitive work across domains whose material processes otherwise have little in common.

Deutsch sharpened the point by treating computability as a question constrained by physical law.² Abstraction and embodiment are not rivals here. The same logical operation may travel, while the speed, reliability, energy, hardware, and institutional consequence of each instantiation remain stubbornly local.

This reach includes parts of computation's own production. Electronic-design automation searches chip layouts. Models assist code generation, evaluation, data synthesis, and experiment design. Existing capability can enter the work of producing further capability, supplying proposals that change what the people and equipment around it have to do next.

A saved research operation changes the work left to do. It may remove an expense, permit more proposals to be examined, or make another constraint decisive. Generated code still requires integration; proposed designs must be fabricated; benchmark gains may not survive deployment. Objectives, equipment, power, capital, empirical tests, and authority remain outside any single runtime. Participation in this production loop does not make the system self-producing or establish a law of increasing returns. Its contribution has to be measured against a defined task and quality threshold, not inferred from the electricity consumed.

The result can be a service used immediately or a capability retained for further work. In either case, computational reach brings more than one kind of judgment into play. A prediction may reduce waste, a design may save material, and a recommendation may redirect capital; each has to enter an activity in which someone can discover whether it does so. The machinery can supply a candidate for the next decision. Selection, deployment and the authority to act govern what becomes of it.