The Key Input

The machine does not isolate man from the great problems of nature but plunges him more deeply into them.

A cheap input matters when people begin designing for its abundance. Cost decline alone is not enough. The input must be available with sufficient reliability, admit uses that could not be organized around its former price, and arrive alongside institutions capable of financing and absorbing the change.

Carlota Perez calls such an input a key factor within a techno-economic paradigm.¹ Her distinction between installation and deployment concerns the relation between financial capital and productive capital. During installation, finance races ahead of proven use and funds the new infrastructure. During deployment, applications diffuse, organizations adapt, and the new cost assumptions enter ordinary design. This is a historical pattern, not a timetable. Revolutions overlap, crucial inputs come in bundles, and many cheap technologies never reorganize an economy.

Mechanized cotton spinning shows why the concept remains useful. The water frame did more than substitute a machine for a hand movement. Cheap thread at scale altered the economics of weaving, bleaching, dyeing, factory location, labor discipline, and finance. None of those consequences was contained inside the machine. They followed because a new productive capacity entered a web of complementary techniques and commercial demand.

The microprocessor produced a different reorganization. Computation had existed in large machines, but cheaper and smaller processors allowed designers to place it inside products, factories, vehicles, offices, and eventually networks. Value did not arise from transistor switching in the abstract. It arose from the devices, software, organizations, and connections built around a cost curve that kept opening new feasible uses.

Foundation models may mark another such change, but the relevant input needs a careful name. It is not computation in general. Calculators and databases have performed computation for decades. Nor is it electricity by itself. The new capacity is learned inference supplied by trained representations: case-specific outputs generated from patterns fitted across data rather than from an explicit rule written for each case.

That capacity can be used as a service and consumed at once. It can also participate in code, designs, evaluations, workflows, and other structures that remain available for later use. The distinction matters more than the familiar contrast between software and labor. A runtime output may vanish after answering a question. A trained model may be copied and called again. A generated design may become plant. Similar physical throughput can therefore enter economic life in different states.

Learned inference does not originate authority. Institutions decide where it may act, what evidence counts, and who answers for error. Yet they can delegate part of judgment unfinished: which features matter in this case, which option to select, whether a threshold has been crossed. That is enough to alter production even when the computational system owns nothing and possesses no legal personality.

Its abundance remains conditional on a heavy physical system. Computation requires chips, memory, networks, cooling, skilled operations, and power delivered at the right place. In the United States, generation and storage projects have faced large interconnection queues and long development times.² Transformers and other grid equipment have likewise carried extended lead times.³ Money can bid for scarce capacity, but it cannot make a permitted substation or advanced package appear instantly.

This combination makes the present transition easy to misread. The marginal cost of one more output may fall while the capital required to supply dependable output at scale rises. A widely available model can coexist with concentrated fabrication, cloud, power, and distribution. The key input may diffuse at one layer while the complementary assets beneath it become more strategic.

Perez's framework earns its place here because it directs attention to redesign. The question is not whether a date such as 2022 began an inevitable age, or whether computation fits every feature of earlier revolutions. It is whether firms and institutions are reorganizing processes around the expectation that some cognitive work can be purchased from machines at declining cost. Evidence will differ by sector. A call center, a laboratory, a software firm, a hospital, and a power grid do not absorb the same capability in the same way.

The working hypothesis is therefore narrower than a paradigm verdict and more consequential than a product forecast. Physical throughput can now be routed through computation into cognitive work that is immediate, reusable, or embodied in another productive arrangement. If organizations continue redesigning around that possibility, the economic state of work changes. The next chapter asks what happens to the capital already built for the previous state.