The Silicon Metabolism
Artificial intelligence is scaling like a civilisation, not software
AI seems like software. It replicates across devices. Models update with a tap. It scales with the frictionless ease of other popular apps.
But frontier systems depend on chips manufactured at the edge of physics, on power stations and transmission lines, on cooling loops, water rights, concrete, planning permissions, and grid connection queues. They sit in copper windings, data halls and substations. The builders of AI are not scaling code. They are constructing a new layer of civilisation.
The build-out behaves like a metabolism, not a consumer product. The builders sell the promise of abundance to fund today's infrastructure. The infrastructure makes the models stronger, and the stronger models make the promise easier to sell. The appetite is always for the next world, never this one. And as long as intelligence keeps scaling with the build-out, AI will keep rebuilding the world in its own image.
Historical transitions worked the same way. The agricultural revolution was not about better seeds. It reordered land and labour. The industrial revolution was not about efficient machines. It reconfigured geography and time. The AI transition is not about models getting better at language. It is about civilisation reorganising itself around machine cognition and paying the physical cost of scaling it.
For most of history, demand was tethered to human rhythms. A loom waits for the weaver. A factory floor stops when the shift ends. Even early computing featured idle servers.
AI loosens that tether. Once deployed, frontier systems do not pause. They monitor, generate, critique, retrain, and spawn tasks in self-reinforcing loops. The main customer of advanced AI is other AI: models training models, agents optimising agents, systems auditing systems.
Demand has become endogenous. It does not lessen as human needs are met. It keeps increasing because the loop is internal. AI is an engine that does not sleep, does not tire, and has no biological limit to its appetite.
Think of an insect colony building a superstructure without any individual insect holding the blueprint. Or an ecosystem settling into a shape through constraint and reinforcement rather than central planning. We may be doing something similar. Not in the banal sense that humans are building data centres, but in the deeper sense that a civilisation might be compelled to build an infrastructure of intelligence because that is what civilisations do when they reach a certain level of complexity.
Human civilisation runs on multiple clocks. The silicon clock runs in milliseconds and training runs, impatient and quasi-exponential. The corporate clock runs in quarters and investor horizons. The political clock moves through consultation and election cycles. Beneath them sits the infrastructure clock: the years, often decades, needed for grids, dams, reactors, and transmission corridors.
These clocks do not harmonise. When the fast technological clock leans on the slow infrastructure clock, the result is not smooth adjustment. It creates bottlenecks. Local backlash. Regulators tightening connection rules. Firms revising plans because power, land, and permitting have become binding constraints.
The signals are already visible. In the United States, data centres are driving a sharp rise in grid needs. The International Energy Agency projects global data centre electricity demand more than doubling by 2030. Ireland has restricted new connections. Canada’s $2.4 billion sovereign compute push looks like a race to secure domestic capacity before bottlenecks harden.
Those stresses feed back into strategy. The compute arms race resembles a security dilemma. In an environment where larger clusters translate into capability and advantage, one actor’s attempt to increase their own power reduces everyone else’s relative security. Escalation becomes rational, even if it risks damaging the infrastructure everyone depends on.
As this dynamic escalates, centralisation will follow. Only the players who can finance at immense scale, secure long-term power, and absorb volatility will gain an edge. Smaller actors will rent access, queue for capacity, or accept second-order status. Governments will treat compute as a strategic asset rather than a commercial input, because capacity will become a form of geopolitical leverage in the way that nuclear capability once was.
If you make a mistake in software you can roll it back. In infrastructure, you live with what you build. Once the concrete is poured, transmission lines routed, and capital sunk into a new geography of compute, the structures become difficult to rewrite. We will all inherit the layout of this arms race.
Coordination is possible. But it is hard to coordinate once a system is locked into a build-out cycle. The most decisive moments in the AI story may not be breakthroughs in machine intelligence. They may be the outcomes of zoning battles, grid constraints, community opposition, and the struggle to build at the pace the silicon clock demands.
Meanwhile, the mind-like surface of AI keeps most of our attention. We frame AI as a drama of thought exceeding thought. We debate creativity, agency, the oracle’s soul. This keeps the human intellect at the centre of the story, as if the main event is a machine mind becoming better than a human mind.
Like the steam engine misunderstood as an iron horse, we fixate on the mimicry while the mines deepen, grids strain, and land and water are reallocated. The decisive question is not whether speaking to AI feels like speaking to a person. It is what the world becomes when intelligence demands its own external existence.
The minds the AI builders are creating are not forming a new digital layer. They have an outer structure. The silicon metabolism is reshaping the planet into a substrate for intelligence, one transmission corridor and cooling system at a time.
And the static infrastructure is only the first phase. These systems, once sufficiently complex, will be deployed with increasing agency. This will involve bodies. Many bodies. By the time we understand the implications behind the location of intelligence, we may be caught unprepared for its leap into motion.
Embodied AI is already entering our shared physical spaces. Autonomous vehicles. Logistics. Robotics. The silicon clock is entering our streets and workplaces. By the time the investor and political cycles have negotiated the price of the electricity that feeds these machines, the terms of our coexistence with them may already have been set.