or as claude explains when i walk it thru basic econ 101 analysis and ask it to state it back to me with verification:
the "maximize the pie and redistribute, but the record isn't great" line is the core mistake. the record is bad because we either didn't redistribute at all or did it with distortionary tools. that's no evidence against separating efficiency from distribution. the second welfare theorem says you can land on any distribution you want at an efficient allocation with lump-sum transfers. the practical version is taxing things in fixed supply, mainly land rent, where there's no behavior to distort.
a compute tax fails for a concrete reason, not a timing one. it's a tax on an intermediate input. production efficiency theory says don't tax intermediate goods, because firms substitute away from the taxed factor and you eat deadweight loss with nothing to show for it. waiting for the industry to mature doesn't fix that. it's the wrong base permanently.
equalizing rates on labor and capital doesn't help either. both are distortionary bases. setting them equal just spreads the deadweight loss around instead of shrinking it.
the distribution problem with ai is a rent problem. returns pool in owners of scarce factors, network effects, market power, and eventually land. tax those rents, which you can do at zero or negative deadweight loss, and pay it out as cash. that's the efficient fix and about the only thing on your list that survives contact with the theory.
"giving people equity in ai" is the right instinct. it's a citizen's dividend funded by rent capture in different clothes. cash also beats wage insurance, retraining, and mandated shorter weeks for the same reason every time: people spend it on their own highest-value use with no distortion, while the others bolt on moral hazard or labor-market distortions.
you don't have to choose between the pie and the slices. you only think you do if you assume the only way to fund redistribution is to tax production.
If the only danger from advanced AI were job loss then I would agree. But we are not talking about a calculator here. Or a tractor. We are talking about a machine that has just solved a bunch of long-standing mathematical problems.
Once AI has advanced robotics far enough to be able to displace all human workers, human extinction is pretty much inevitable.
Shift from taxing "labor" (income) to consumption. If some kinds of consumption are AI intensive, that would in asense be a "tax on AI." But why beat around the bush? Just tax consumption, a VAT to finance as big a welefare state as we want and a progressive consumption tax for other current expendituress and for redistribution?
Thanks for writing this! I hadn't considered the fact that a B2B exemption might make a token tax close to ordinary VAT before; this was an interesting point.
> In principle, you could also tax the output. [...] I don’t think we should call that a compute tax.
This seems like a strawman, particularly given that (as you note!) Korinek and Lockwood favour a tax on tokens to one on the stock of compute, while the main theoretical results only apply to the stock-based version of the tax.
And when talking about AI, people very often use "compute" to refer to the flow (e.g. "inference-time compute" is clearly about the use of tokens, not the physical GPUs sitting in a datacentre), so it seems perfectly reasonable to call such a scheme a compute tax.
It's not obvious to me that you would need to implement a B2B exemption (although I appreciate that this is what K&L advocate for) -- we already have payroll taxes on human labour, and insofar as AI is just displacing human labour, presumably we want to tax it similarly, rather than subsidising substitution of human labour with AI labour because of more favourable tax treatment on the latter. (There are also negative externalities from automating away workers that the firm doesn't internalise, e.g. search costs, etc.) Even if you did implement a B2B exemption, it would be a pretty significant change to the tax code given that the US doesn't have any federal VAT. I think it could've been more fruitful if you did engage with proposals of this sort rather than dismissing them immediately.
> U.S. compute spending is large in absolute terms but small relative to GDP.
Isn't a large part of the motivation for thinking about compute taxes that spending on tokens might soon become a potentially very large fraction of GDP, displacing large amounts of earned income (which is taxed as a flow)? E.g., see https://www.anthropic.com/research/economic-policy-responses.
Good article that explains the fundamental economic issues with AI taxes. However, even with cheaper AI services there will still be plenty of anxiety over job automation and the impact of data centers on local environments. This public perception towards AI needs to be addressed, otherwise we will keep seeing politicians getting elected on anti-AI platforms.
I do not understand the comparison between a company using 1,000 workers and 1,000 machines (and the article is paywalled and I am not very familiar with the US tax code). Presumably in practice this means the second company pays the same as the first company's wage bill in each period to rent / maintain it's stock of machines? In both cases these costs are deducted from revenue before profits are taxed. Presumably you aren't arguing that the second company should pay the equivalent of payroll taxes on it's depreciation costs, or that payroll taxes should be eliminated? What am I missing / misunderstanding?
Well, a compute tax would certainly be one way for the governments to get back some of the excesses they’ve allowed in the AI space. So first you let the hyperscalers build with government support at the expense of the tax-paying public then you tax the them. Maybe I’m missing something, but it sure sounds like making money on top of money and still coming out with the deficit.. I mean, the wars have to be financed somehow.
I agree in the current environment, but what if AI becomes slightly more efficient than at least some kinds of human labor and starts to cause structural unemployment?
Imagine in the future that an AI costs a company 90% as much to "employ" as a human for the same job. If the company replaces the human with the AI (and this happens en masse, where many humans are unable to find replacement jobs), they save money. But from a societal perspective, we're paying the compute cost of the AI, but we also need to provide for the human, which in total takes more resources than simply employing the human at this job.
or as claude explains when i walk it thru basic econ 101 analysis and ask it to state it back to me with verification:
the "maximize the pie and redistribute, but the record isn't great" line is the core mistake. the record is bad because we either didn't redistribute at all or did it with distortionary tools. that's no evidence against separating efficiency from distribution. the second welfare theorem says you can land on any distribution you want at an efficient allocation with lump-sum transfers. the practical version is taxing things in fixed supply, mainly land rent, where there's no behavior to distort.
a compute tax fails for a concrete reason, not a timing one. it's a tax on an intermediate input. production efficiency theory says don't tax intermediate goods, because firms substitute away from the taxed factor and you eat deadweight loss with nothing to show for it. waiting for the industry to mature doesn't fix that. it's the wrong base permanently.
equalizing rates on labor and capital doesn't help either. both are distortionary bases. setting them equal just spreads the deadweight loss around instead of shrinking it.
the distribution problem with ai is a rent problem. returns pool in owners of scarce factors, network effects, market power, and eventually land. tax those rents, which you can do at zero or negative deadweight loss, and pay it out as cash. that's the efficient fix and about the only thing on your list that survives contact with the theory.
"giving people equity in ai" is the right instinct. it's a citizen's dividend funded by rent capture in different clothes. cash also beats wage insurance, retraining, and mandated shorter weeks for the same reason every time: people spend it on their own highest-value use with no distortion, while the others bolt on moral hazard or labor-market distortions.
you don't have to choose between the pie and the slices. you only think you do if you assume the only way to fund redistribution is to tax production.
If the only danger from advanced AI were job loss then I would agree. But we are not talking about a calculator here. Or a tractor. We are talking about a machine that has just solved a bunch of long-standing mathematical problems.
Once AI has advanced robotics far enough to be able to displace all human workers, human extinction is pretty much inevitable.
Shift from taxing "labor" (income) to consumption. If some kinds of consumption are AI intensive, that would in asense be a "tax on AI." But why beat around the bush? Just tax consumption, a VAT to finance as big a welefare state as we want and a progressive consumption tax for other current expendituress and for redistribution?
Thanks for writing this! I hadn't considered the fact that a B2B exemption might make a token tax close to ordinary VAT before; this was an interesting point.
> In principle, you could also tax the output. [...] I don’t think we should call that a compute tax.
This seems like a strawman, particularly given that (as you note!) Korinek and Lockwood favour a tax on tokens to one on the stock of compute, while the main theoretical results only apply to the stock-based version of the tax.
And when talking about AI, people very often use "compute" to refer to the flow (e.g. "inference-time compute" is clearly about the use of tokens, not the physical GPUs sitting in a datacentre), so it seems perfectly reasonable to call such a scheme a compute tax.
It's not obvious to me that you would need to implement a B2B exemption (although I appreciate that this is what K&L advocate for) -- we already have payroll taxes on human labour, and insofar as AI is just displacing human labour, presumably we want to tax it similarly, rather than subsidising substitution of human labour with AI labour because of more favourable tax treatment on the latter. (There are also negative externalities from automating away workers that the firm doesn't internalise, e.g. search costs, etc.) Even if you did implement a B2B exemption, it would be a pretty significant change to the tax code given that the US doesn't have any federal VAT. I think it could've been more fruitful if you did engage with proposals of this sort rather than dismissing them immediately.
> U.S. compute spending is large in absolute terms but small relative to GDP.
Isn't a large part of the motivation for thinking about compute taxes that spending on tokens might soon become a potentially very large fraction of GDP, displacing large amounts of earned income (which is taxed as a flow)? E.g., see https://www.anthropic.com/research/economic-policy-responses.
Good article that explains the fundamental economic issues with AI taxes. However, even with cheaper AI services there will still be plenty of anxiety over job automation and the impact of data centers on local environments. This public perception towards AI needs to be addressed, otherwise we will keep seeing politicians getting elected on anti-AI platforms.
This is a scary tax and needs to be ridiculed at every opportunity 😎
Brian, I'd be interested your comparison between the compute tax and Paul Romer's idea to tax digital advertising: https://adtax.paulromer.net/
I do not understand the comparison between a company using 1,000 workers and 1,000 machines (and the article is paywalled and I am not very familiar with the US tax code). Presumably in practice this means the second company pays the same as the first company's wage bill in each period to rent / maintain it's stock of machines? In both cases these costs are deducted from revenue before profits are taxed. Presumably you aren't arguing that the second company should pay the equivalent of payroll taxes on it's depreciation costs, or that payroll taxes should be eliminated? What am I missing / misunderstanding?
Well, a compute tax would certainly be one way for the governments to get back some of the excesses they’ve allowed in the AI space. So first you let the hyperscalers build with government support at the expense of the tax-paying public then you tax the them. Maybe I’m missing something, but it sure sounds like making money on top of money and still coming out with the deficit.. I mean, the wars have to be financed somehow.
https://youtu.be/GtZtEE7ZPoc?si=_G9KrYz7j4LB5V1b
https://youtube.com/shorts/5h38su8LRek?si=HjMXqTEu0dAE0PDQ
https://youtube.com/shorts/L6iLP_QLJnk?si=6hYK4p64lqpmdzD0
I agree in the current environment, but what if AI becomes slightly more efficient than at least some kinds of human labor and starts to cause structural unemployment?
Imagine in the future that an AI costs a company 90% as much to "employ" as a human for the same job. If the company replaces the human with the AI (and this happens en masse, where many humans are unable to find replacement jobs), they save money. But from a societal perspective, we're paying the compute cost of the AI, but we also need to provide for the human, which in total takes more resources than simply employing the human at this job.