Brian Albrecht’s A compute tax is a REALLY dumb idea disputes a claim that I made in Safety Net When AIs Take Our Jobs about a possible need for a tax on GPUs and similar hardware.
Albrecht and I are answering different questions. He asks what’s the optimal tax for normal times. Whereas I ask what to do if optimal taxes are already collecting nearly the maximum feasible revenue, and there’s still a short-term risk of a debt spiral.
He says:
If you care about things like increasing output, total surplus, minimizing deadweight loss, those types of things, stay away from this.
Albrecht is correct that the compute tax fails to optimize the standard goals of economists. There are significant deadweight costs. My safety net post endorsed some unpleasant methods to deal with enforcement problems.
One point that I neglected was the risk that a compute tax would cause compute to relocate to other countries. This is a significant cost over the long term, but datacenters take years to move, and are partly constrained by latency issues.
For most of my life, economists were correct to focus on the goal of increasing wealth, given decent institutions. I expect that within a few years, it will become rather inappropriate to judge policies mainly on that goal. Avoiding ruin is a closer approximation to how I’d describe what I want from policy — not stability, which would preserve bad arrangements along with good ones, but avoidance of the outcomes we won’t recover from.
My safety net post focused on a risk that a temporary drop in tax revenues due to a labor shock could cause a government debt crisis. Such a crisis would trigger a variety of political risks.
This is part of a more general concern about political instability. Economists often take property rights for granted. Yet an era in which many people see serious risks to their main source of income is a time when voters may get desperate for other ways to grab wealth. If we get a sudden mass-replacement of human labor by robots, it will be hard to assure voters that a safety net can adequately replace their jobs. A compute tax will help stabilize the government’s ability to provide a saftey net, while slightly slowing the adoption of robots. Slowing their adoption is a minor bonus that mildly reduces the political risks.
If the talk of a compute tax is due to misplaced panic over AI, then Albrecht is correct that the tax would be bad. So we should think carefully about how big the effects of AI will be, and to the extent they’re big, it becomes important to have draft legislation that is prepared to deal with it.
I want the compute tax to be quite temporary. It’s only designed to handle a brief COVID-like shock that causes a sudden drop in tax revenues and an unusual demand for welfare spending. But regardless of whether the tax ends up being temporary, I expect that a few years after the labor shock the world will be wealthy enough that the costs of an inefficient tax will be minor compared to the risks we face before we become wealthy.
My safety net post only proposed the compute tax after proposing a 30% VAT (capped by black-market risk), a 2% LVT (capped by bank solvency), a 35% corporate rate (capped by historical evasion), and still seeing a trillion-dollar gap. Maybe there are better additional taxes that will provide comparable safety from ruin.
But most critics of a compute tax seem to assume away the risks that concern me, rather than proposing a better plan to deal with them.