The Invisible Trillion · How Much Do We Actually Need?

The Physical Version of the Same Question

Everything to this point has been about compute — how much inference and training the buildout is sized for, and how much of that demand is real. There is a second axis the compute framing leaves out, and the chapter’s title question — how much do we actually need? — applies to it just as hard. The buildout is not abstract. It is physical: chips, racks, buildings, transformers, cooling, and, in the most expansive versions of the vision, the machines the compute is meant to enable.

Those expansive versions are worth taking seriously, because they are doing a lot of the work of justifying construction far ahead of present demand. The justification is a projected future of physical abundance — humanoid robots, re-industrialization, machines building machines. The genre has a characteristic artifact: the confident viral figure for how quickly a swarm of humanoid robots could rebuild a city, passed around and amplified, approvingly, by the people building the robots. Set the number aside. Ask what it assumes. Every robot, every chip, every data center, every megawatt of new generation begins as mined material — copper, lithium, cobalt, nickel, rare earths, silica, steel, concrete — and ends, eventually, as physical waste. The chapter on mining and e-waste followed both ends of that line.

This is not hypothetical, and there is one case where the dynamic has already run all the way to completion. Beginning around 2016, China’s bike-sharing companies — Ofo, Mobike, Bluegogo, and at the peak nearly seventy others — flooded the cities with shared bikes. The idea was sound and the early numbers were real: ridership rose, car trips fell, registered users ran into the hundreds of millions. But the build was paced to a race for market share, funded by venture capital — Ofo alone raised more than two billion dollars — and by early 2018 the firms had put some twenty-three million bikes on the street, faster than any plausible accounting of how many were needed, with rides priced below cost to grab users before the money ran out.

Then the funding turned. Bluegogo went first, then a couple of dozen others, then Ofo itself, with something like sixteen million customers queued for deposits that mostly never came back. And the bikes — the steel, the aluminum, the physical objects — had to go somewhere. They went into what the photographs of the period made famous as “bicycle graveyards”: fields of hundreds of thousands of bikes, in more than twenty cities, stacked in heaps the size of stadiums and left to rust. A useful idea, built on a genuine thesis, overshot real demand by close to an order of magnitude, and the overshoot became one of the defining images of industrial waste of the decade.

A shared bike weighs about twenty kilograms. A data center, a chip fab, a fleet of humanoid robots, and the power plants to run them are the same dynamic with the mass turned up by orders of magnitude — and the most expansive versions don’t only add, they subtract first. To rebuild a city you have to take down the one already standing, and that city is also material: millions of tons of concrete and steel and glass that become demolition debris the instant the rebuild begins. The vision counts the construction. It almost never counts where the old city goes.

The fair rebuttal is the one the buildout’s defenders always make, and it has real force: efficiency improves, recycling improves, designs dematerialize, some material may eventually come from off the planet, and the things being built have enormous value. All true. But the material accounting is precisely the part the projections leave out, and leaving it out does not make it stop applying — and the bike graveyards are a reminder that the recycling is supposed to happen and frequently doesn’t, or doesn’t keep pace with the building. “How much do we actually need” is a question about atoms as much as about compute, and almost no one making the expansive case has done that arithmetic.

All 19 chapters of The Invisible Trillion · Library · All courses