Part IV · The Derivation

IV.A — Silicon Metabolism

10 min read · 1,817 words

Intelligence is physical; it has a thermodynamic cost; and the cost is what makes its advantages durable. But this leaves a question unanswered: why does computation deserve a special place in the production function, distinct from other energy-intensive industrial processes?

The answer is that computation is a general-purpose constraint-relaxer that can be turned back onto itself.

Steel production consumes enormous quantities of energy; so does aluminum smelting, cement manufacturing, and ammonia synthesis. What makes computation different? The answer lies in what computation transforms—and in what that transformation can be directed toward.

A steel mill takes iron ore and coke and transforms them into steel. The output is a specific material with specific properties: tensile strength, ductility, corrosion resistance. The transformation is physical in the most direct sense. Atoms are rearranged; chemical bonds are broken and formed; the product weighs as much as the inputs minus whatever was lost to slag and exhaust. Steel can be shaped into beams or rails or turbine blades, but it remains steel. The transformation is powerful and economically essential, but it is also bounded. Steel cannot become copper or plastic or semiconductor-grade silicon. The output is constrained by the chemistry of the input.

Computation transforms something different. It transforms constraints. A computer takes a pattern of inputs and produces a pattern of outputs according to rules encoded in its program. The inputs and outputs are physical in the sense that they are represented by voltages or magnetic orientations or optical states, but the transformation itself operates on the pattern, not on the substrate. The same computation can be performed on silicon or gallium arsenide or superconducting circuits or, in principle, on any physical system capable of representing and manipulating discrete states. The transformation is substrate-independent. This is what makes it general-purpose.

Claude Shannon, working at Bell Labs in the 1940s, gave this observation its mathematical foundation. His 1948 paper, “A Mathematical Theory of Communication,” defined information as a reduction in uncertainty and showed that information could be quantified in bits, independent of the physical medium carrying it.1 A bit is a bit whether it is stored in a transistor, punched into a card, or encoded in the timing of neural spikes. Shannon’s framework separated the logic of communication from its physical implementation, creating a science of information that applied equally to telegraphy, telephony, radio, and any future medium yet to be invented.

The implications went beyond communication. If information is substrate-independent, then so is computation, which is the systematic manipulation of information according to rules. A calculation performed on paper, on an abacus, on a mechanical calculator, or on an electronic computer produces the same result. The physical implementation affects speed, reliability, and energy cost, but not the logic of the operation. This is the sense in which computation is abstract: it is a pattern of relationships that can be instantiated in many different physical systems.

David Deutsch, a physicist at Oxford who helped lay the foundations of quantum computation, made the point explicit in a 1985 paper that reframed the Church-Turing thesis in physical terms.2 Deutsch argued that the question of what is computable is not merely a mathematical question but a physical one. The laws of physics determine which transformations of information are possible and which are not. A universal computer, in Deutsch’s sense, is a physical system capable of emulating the relevant dynamics of many other physical systems—not perfectly, but to sufficient precision that the emulation yields useful predictions. The computer is universal not because it transcends physics but because physics permits this kind of generality.

This generality is what distinguishes computation from other industrial processes. A steel mill can only make steel. A chemical plant can only make the chemicals its reactors and catalysts permit. But a computer, given enough time and memory, can substitute computation for some fraction of experimentation: it can model a steel mill, simulate airflow through a turbine blade, explore the folding landscape of a protein, or evaluate millions of candidate chip layouts before committing to fabrication. The substitution is not unlimited—many physical processes resist efficient simulation—but where it works, it compresses the cost of search and design.

The economic implications follow directly. A general-purpose technology can be applied across many sectors, and its improvements compound as they diffuse through the economy. Steam power was transformative because it could drive pumps, looms, locomotives, and factories. Electricity was transformative because it could power lights, motors, communication systems, and computation. Computation is transformative for the same reason, but with an additional property that sets it apart from prior general-purpose technologies.

Steam engines could not design better steam engines. Electrical generators could not optimize their own windings. But computation can be directed toward its own improvement in a closed loop. A compiler can be used to write better compilers. A chip placement algorithm can be used to design better chip placement algorithms. A neural network can be used to generate synthetic training data for other neural networks. Automated evaluation harnesses can assess model outputs at scales no human team could match. The feedback loop is intrinsic to the technology in a way that it was not for steam or electricity.

This is not a claim of permanent increasing returns. Bottlenecks eventually bind: the cost of training data, the physics of heat dissipation, the lead time for new fabs, the scarcity of engineering talent that can turn capability into deployment. The relevant claim is narrower: that over a meaningful range—the range we appear to be in now—the effective production function for computational capability is convex. The output of one generation of models becomes an input to the next. The factor that produces the factor is qualitatively different from a factor that produces only output.

Standard production functions treat computation as an intermediate good, produced by the capital and labor of the tech sector and consumed by other sectors as an input. This accounting is not wrong, but it obscures the recursive structure of the process. Computation is not merely consumed; it is used to produce more computation, and the computation it produces is used to produce yet more computation. An input that expands its own production frontier is qualitatively different from an input that produces only output.

The recursion is visible at every level of the stack. At the hardware level, electronic design automation tools—themselves running on chips—explore billions of potential layouts to minimize power consumption and maximize performance of the next generation of chips. At the software level, machine learning frameworks train models that improve training efficiency, optimize inference, and generate code for new applications. At the training level, synthetic data pipelines use existing models to produce inputs for training larger models, while automated evaluation systems use inference to assess quality at scales no human review could match.

None of these loops is fully closed. Human judgment remains essential at decision points: which architectures to explore, which benchmarks to trust, which trade-offs to accept. But the direction of travel is clear. The computational portions of the improvement cycle are becoming larger, and the human portions are increasingly focused on direction-setting rather than execution.

Consider the contrast with other factors. Labor can be trained and augmented, but a worker cannot directly produce another worker in the way that a machine can produce another machine. Capital can be accumulated, but a factory cannot design and build the next factory without human intervention. Land is fixed in supply. Energy is abundant in principle but requires infrastructure to convert and deliver. Computation, by contrast, can be directed toward its own improvement in a closed loop. The recursion is not perfect, and human judgment remains essential at many points, but the effect is to make the production of computational capability increasingly a function of computational capability itself.

This recursive property has consequences for how we model economic growth. Standard growth models treat technological progress as an exogenous residual or as the output of a separate R&D sector with diminishing returns. But if the technology that produces technology can improve itself, then the returns to investment in that technology may be increasing rather than diminishing—at least over some range, until the bottlenecks bend it back. The scaling laws observed in large language models, where capability increases predictably with compute, suggest that we are in such a range now. The question is how far the range extends and what happens when it ends.

The physics of computation sets the ultimate boundaries. Landauer’s principle establishes a floor on energy consumption per bit erased. Lloyd’s limits establish a ceiling on operations per unit mass and energy. The brain demonstrates that biological systems can approach efficiencies far beyond current silicon. But within these boundaries, there is room for enormous improvement, and the improvement is driven by computation acting on itself. The factor is producing more of the factor, and the factor it produces is producing still more.

This is what it means to say that energy is being structured into computation and computation into intelligence. The phrase is not poetic license; it is a description of a physical process. Energy flows into data centers and is converted into switching states in transistors. The switching states implement algorithms that manipulate patterns of information. The patterns encode learned models that reduce uncertainty about the world. Each stage of the transformation dissipates entropy and produces structure, and the structure has economic value because it does useful work.

We can now give Factor Prime a semi-formal definition: it is energy, structured through computation and disciplined by selection, that produces economically useful uncertainty reduction. The proxy metric is cost per successfully completed task at a defined quality threshold—where “success” means accepted by a downstream verifier, whether human or automated. This metric is imperfect, but it captures what the production function needs to track: the conversion of joules into economically relevant decisions.

The work is not physical in the way that lifting a weight or moving a railcar is physical. It is informational: predicting the next word, classifying an image, recommending an action, generating a design. But informational work has physical consequences. A better prediction reduces waste. A better classification improves allocation. A better recommendation saves time. A better design requires fewer materials or less energy to achieve the same function. The transformation of constraints, which is what computation does, translates into the transformation of matter and energy through the decisions it informs.

The next section introduces the mechanism that connects thermodynamic depth to economic value: the selection gradient that determines which structures survive deployment and which do not. Thermodynamic cost is necessary for value but not sufficient. What matters is whether the structure produced by that cost passes the test of use.