The Floor of Intelligence

A bit costs a watt

A bit looks weightless. But flipping one costs energy — at minimum kT ln2, a hard thermodynamic floor (Landauer's principle). Thinking, in the end, is the moving of bits, and moving bits sheds heat. One operation is tiny. The problem is scale.

For a long time, the price kept falling. The number of computations you get per unit of energy has historically doubled roughly every 1.6 years (Koomey's law). The exchange rate between watts and bits kept improving. So most people assume intelligence will keep getting lighter. For a while, I assumed it too.

But the total explodes

When efficiency rises, demand is unleashed (Jevons' paradox). Each gain in cheapness gets spent many times over. And lately the engine of cheapness — Koomey's law — has been slowing. Unit costs improve more slowly; usage grows faster. The arithmetic is brutal: total watts swell.

The swelling has a name now. Stargate — a plan to build up to $500 billion of AI infrastructure in the United States, targeting 10 gigawatts. The flagship campus in Abilene, Texas, alone will draw 1.2 GW — the power of about 750,000 homes — to run some 400,000 GPUs. To produce thought, we now build things on the scale of cities.

Watts can flee to orbit — chips can't

There's an escape hatch. Space data centers. In orbit the sunlight is endless — no night — and cooling is nearly free, dumped into the cold of space. Power and heat, the two constraints that bind AI on the ground, you can shed both at once. Putting compute in orbit is no longer pure science fiction.

But you can move the watts to space and still not move the machine that makes the bits. You need the chips. And the chips are the hardest knot on Earth. The most advanced logic is made overwhelmingly by one company — TSMC, around 90% of the leading edge. And even TSMC can only buy the machine that prints those chips, the EUV lithography system, from ASML — the only company on the planet that builds it, at roughly $400 million each. No EUV, no advanced chip. You can't lift a fab into orbit. Beneath watts, there is silicon — and it does not move.

Beneath silicon, stone and gas

And what is silicon made of? Keep descending. Ultra-pure silicon begins as quartz — and the crucibles that grow the ingots need quartz of almost unimaginable purity. Between 70 and 90% of the world's high-purity quartz comes from two mines in one small town: Spruce Pine, North Carolina, where 380 million years ago a continental collision happened to leave a stone of unmatched purity. You cannot make more of it. There is no second hill. (When Hurricane Helene hit the town in 2024, the whole semiconductor world held its breath.)

The neon for the lithography lasers was about half sourced from Ukraine — a by-product of old steel mills — until a war made the thinness visible. The gallium and germanium nearest the metal of the chip are refined mostly in China. The deeper you go, the narrower, the harder, the more fixed to one specific piece of ground.

So nations fight over stone

Lay the map out and something strange appears: no single place holds the whole stack. Taiwan has the fabs; the Netherlands has EUV; Japan has the materials and chemicals and tools (it makes about 96% of the resist-coating equipment, 53% of advanced photomasks); South Korea has the memory; the United States has the design and the capital — and the quartz. China has the minerals. Ukraine had the gas. Each one's hand rests on a different valve.

And lately the valves are being turned into weapons. China throttled gallium exports to the U.S. — official figures show shipments collapsing from 6,876 kg to 227 kg in months. Japan has done the same with materials. The U.S. blocks China through design and tools. The great game of the AI age is, at bottom, a scramble over stone and gas buried in particular ground.

The strange grace of it: because no one holds the whole stack, closing your own chokepoint completely invites retaliation on the one you depend on. Interdependence is a thin, mutual deterrent — a chain of peace. The first to weaponize fully probably loses the most.

The line between company and state

The funnel has a second reading. The upper layers are held by companies; the lower layers, by states.

At the top — models, software, capital — the holders are firms, and those firms have outgrown nations. Stargate's $500 billion exceeds the annual budget of most countries on Earth. Seen from the surface, intelligence looks like a corporate empire. But descend, and the holders change. The deepest layers — the minerals, the land, the power permits, the physical security of the fabs themselves — are not corporate. They are territorial. China doesn't ask a company's leave to ban gallium; Washington doesn't consult Nvidia before it blocks a sale; no company defends Taiwan. The more physical the layer, the more it belongs to the state.

And the company has every reason to push down. It pays to own not just the model but the chip beneath it, not just the chip but the material beneath that — because the deeper you reach, the more of the un-movable you hold, and the un-movable is where lasting leverage lives. So firms integrate downward, layer by layer, until they meet a floor they cannot buy. That floor is where the boundary gets drawn.

So the boundary between company and state runs as a horizontal line through the stack — drawn exactly where the abstract turns territorial. Intelligence began as a company's game of algorithms and capital. But as its binding constraint sank into atoms and ground, it crossed into the state's domain. The state is returning to technology — not by choice, but because the floor of intelligence is made of the one thing states have always governed: ground.

Where the age is heading

As chokepoints are weaponized, the world fractures into blocs — allies assembling a full stack between them, an echo of the 1930s. But it can't fracture cleanly. There is one ASML, one TSMC, one quartz town. No bloc is self-sufficient. So the world stays deeply entangled even as it splits — de-risking, not decoupling.

And here the intuition flips. The state does not shrink. It grows — in two opposite directions at once. Its reach expands: it reclaims what it had let drift — supply chains, compute, energy, minerals. But its autonomy shrinks: no nation holds the whole floor. Reach up, autonomy down — and that contradiction lifts the effective unit from the nation to the bloc. What grows is not the great power but the chokepoint: Taiwan, the Netherlands, a handful of firms — small, yet impossibly heavy.

Those chokepoints are also single points of failure. One island, one hill. The smartest future is being held hostage to a few square miles.

But to read the ground only as power is one reading. What waits at the bottom is not just something to be seized; it is also common capital — the stock a society rests on. Nature, like air and water. Infrastructure, like roads and power. And institutions, like law and trust. The floor of intelligence holds all three.

Which means the deepest layer may not be physical at all. Beneath the stone and the gas lies another floor — institutions. Whether the same chokepoint becomes a flashpoint or a foundation is decided less by who holds it than by the institutions that govern it. Who owns the ground is a question of force; whether it can be tended as common capital is a question of institutions.

What should be done at that boundary? A little can be said. Treat the floor of intelligence not as territory to be seized but as common capital. Leave it to the market and you miss the strategic floor; let the state swallow it whole and you smother the life above. Common capital is held neither by the market nor by state command, but in trust, by professional institutions. So what is really needed is less to win the scramble for ground than to build — before the shock — the institutions that govern the chokepoints as a commons. Redundancy is cheap only in advance; institutions come too late if built after the crisis.

Which way it breaks — toward redundancy bought at the speed of new fabs and new mines (slow, measured in years), or toward concentration held in check by mutual dependence — I won't pretend to know. It's still moving.

What I found in the basement

One thing is clear. The most futuristic thing we have built — artificial intelligence — is dragging the world back to its oldest logic: territory, resources, blocs, and the contest over them. The twenty-first century's race for intelligence rhymes with the nineteenth century's race for resources, and the bloc economies of the 1930s.

By descending into the ground, intelligence pulled itself back under the state — the oldest organizer of land and force. The newest thing returned to the oldest order — an order that still allows two answers: to seize the ground, or to tend it as a commons. Which one we choose is not yet written.

We spent a long time believing thought was weightless. But go down, and bits are made of watts; and beneath the watts, silicon; and beneath the silicon, stone; and beneath the stone — a particular patch of ground, and the old, long quarrel over who holds it. Each time we grow smarter, we dig one layer deeper. I wanted, just once, to see what is buried down there. That's the whole of this note.