Should data centers pay a carbon tax?
Two things voters hate, together at last!
An intelligent and sophisticated minority of people (the kind of people who read Slow Boring, no doubt) know that most of the anti-data center hype is massively overblown. There’s plenty of water for data centers, with reasonable policy data centers don’t drive up electricity prices, and the local tax revenue generated by data centers is fundamentally valuable.
This still leaves two issues that I think are legitimate complaints.
One is the tax abatements that data centers receive. I think the backlash to this is somewhat misguided, since the purpose of the abatement is to generate a revenue-positive outcome. But it is genuinely bad that standard practice is for cities and states to write bad tax codes that are counterproductive for growth and revenue and then patch them with lots of abatements. I believe this whole process is actually much less corrupt and harmful than it seems to most people at first glance. But it’s not good, and the fact that it creates the appearance of corruption ought to inspire public officials to actually change what they do.
A more serious issue, though, is that data centers raise carbon dioxide emissions.
Now what’s true is that as with their water or land use or anything else, the emissions associated with data centers are not particularly large compared to the emissions associated with agriculture or industry. It’s not like data centers are some unique environmental hazard the likes of which the world has never seen. A lot of the rhetoric on this subject is misleading. Still, just pointing this out doesn’t fully answer the objection.
The problem is that if you start with a temperature target of 1.5 or 2 degrees Celsius, and then derive from that a global emissions target, and then divide that into national and local emissions targets, the status quo is a huge problem. There’s a lot of decarbonizing that has to be done to hit these targets on the appropriate schedule, and a big new source of energy demand is very inconvenient for that.
Say you did a lot of work back in 2019 on a PDF about how to get to 100 percent clean electricity in 30 years. Well, if you managed to generate all the clean electricity your PDF promised but now electricity demand is higher than you predicted, you’re going to fall short of the 100 percent target.
The right way to think about this is that the entire exercise of reasoning backwards from targets is misguided, and that’s especially true when you’re dealing with targets set by individual U.S. states. But lots of organizations, institutions, individual politicians, and even entire states have in fact adopted these targets. And institutions don’t like to change things.
This creates a powerful desire to just make the data centers go away rather than think seriously about the climate issue.
The climate/A.I. bubble
If you think about it rationally, inventing a new economically useful way to consume electricity simply means that the globally optimal level of greenhouse-gas emissions and total warming has gone up.
Suppose someone invented a National Cancer Shield that could, as long as it was operational, prevent any American from developing any form of cancer.
The N.C.S. would be incredibly useful. If you discovered that running the N.C.S. would require 100 terawatt hours of new electricity generation, you’d say the United States had better build the 100 terawatt hours. Green-inclined people would push for as much of that as possible to come from renewables, which would be fine, but obviously the fastest and cheapest way to stand up the N.C.S. would be to use plenty of natural gas. And with state-of-the-art highly efficient combined-cycle turbines in short supply, it would still make sense to use plenty of less efficient turbines.
What’s more, even if a huge share of the N.C.S. power supply were supplied by renewables, that would still be diverting resources — sunny or windy land, transmission lines, physical construction capacity — that could otherwise be used to crowd out existing fossil-fuel usage.
The point is, the invention of the N.C.S. technology implies significantly higher greenhouse-gas emissions than in a no-N.C.S. universe. Obviously, though, the N.C.S. is still good on balance.
But if your whole life’s work is creating and enforcing climate targets, that would still be very inconvenient for you. You wouldn’t want to believe that something genuinely useful and beneficial had come along that just meant the world was — with good reason — going to move further from the targets.
L.L.M.s, of course, do not cure cancer (yet), but, on its face, artificial intelligence seems to clearly be an economically valuable technology. After all, if it weren’t economically valuable, then nobody would be building all these data centers. The whole pace of construction and rising electricity demand is premised on the fact that investors think they’re going to make money.
Someone invented a new economically useful way to use electricity, and that is rightly going to push the emissions trajectory upward.
What complicates this is that there are multiple interests at stake with climate change. A.I. is not helpful to polar bears. And polar bears have much less ability to adapt to a warmer planet than humans do. So a new economically useful (to humans) technology is really bad news for polar bears, for coral reefs, and other non-human animals that stand to lose out from a warming planet. If you have a deep ethical commitment to polar bears, this is a troubling development. But there’s also going to be resistance to admitting that you’re asking people to give up something useful for the sake of polar bears.
And I think this helps drive a lot of the “A.I. bubble” narrative. Given that rising A.I.-related electricity demand is bad news for climate targets, the most expedient way to defend the targets is to insist that the demand is just a big mistake, that actually this isn’t economically useful at all, that nothing is being given up by not building it.
I’m not saying this bubble argument is only ever offered in bad faith.
But I do think there’s motivated reasoning happening here. I don’t think I’ve ever seen someone say, “Unexpectedly rapid progress in A.I. means the cost of hitting previous climate targets has gone up, but I think everyone should accept that increased price for the sake of the polar bears.”
Yet, clearly, the environmental movement has also not said, “Unexpectedly rapid progress in A.I. means the cost of hitting previous climate targets has gone up, so everyone should accept more warming.” But unless it’s a bubble, it has to be one of those two things, neither of which is palatable to say. So the bubble solves an acute problem, and that gives people strong motivation to believe in it even though everyone should actually be confronting the issue squarely.
Local benefits, global harms
I am not personally all that upset about the polar bears.
I was, however, struck by Bernie Sanders’s proposal to half-nationalize frontier A.I. labs on the grounds that this is the way to “guarantee that the trillions created by A.I. are used to improve the lives of all of us.”
Step back from the specific policy design for a moment. The aspiration to ensure that the wealth generated by A.I. benefits all of us is clearly correct. But who is “us”? One of the virtues of Bernie’s proposal is that it’s very explicit about this: “Us” is the American people, the ones who would own the labs.
But the cost of the additional warming is not just borne by us. It’s borne by the population of the entire world. In fact, it is disproportionately inflicted on the people who live in the tropics, where marginal increases in average temperature greatly impact the risk of extreme heat events. Of course, many parts of the United States are very hot too, but nobody’s forcing people to move to Houston and Phoenix every year. America is full of cold cities that people could move to if they wanted to. Nigeria and Bangladesh are not.
This is something I am pretty worried about, not just in the specific context of data centers and climate change, but in terms of the broader conversation about who will benefit from A.I.
Clearly, “a small number of early investors reap all the gains” is not the right outcome. But while you can imagine democratic political action forcing the gains to be broadly shared domestically, it’s hard to imagine that approach will work for the global poor, even though they’re the ones who’ll pay the price for higher emissions. So while on one level this is another “green groups are out to lunch” take, at the end of the day they are correct that climate change is a serious problem and that the climate-A.I. nexus is worth taking seriously.
Two wrongs make a right?
Mostly, this underscores the need for a politics of clean energy abundance.
Not a politics of blindly intoning “clean is cheap,” but a politics of genuinely doing everything in our power to accelerate the deployment of utility-scale renewables and to facilitate potentially game-changing investments in geothermal and nuclear power. These are all promising technologies, but the impediments to actually implementing them quickly and at large scale remain daunting.
But here’s my other thought.
We know voters hate carbon pricing. And we also know that voters hate data centers. So what if we … made data centers pay a carbon tax?
Normally, people don’t like carbon pricing because they don’t want to pay more. But there would be no consumer-facing price here at all. The tax would be a burden on economic development, but specifically on the development of a sector that the public has a somewhat irrationally negative view of.
We could set the tax to account for the global social cost of carbon. And then, because the United States is such a dominant player in the tech industry, we could layer on mechanisms to penalize American companies that utilize foreign non-priced data centers, thereby encouraging the whole world to level up. There is a technical issue here in terms of assessing exactly what emissions are attributable to a given data center. But I think this is solvable, and also it doesn’t need to be calculated perfectly to be a good idea.
I know carbon pricing has been left deader than the dodo in the policy community, but I think data centers offer a genuine opportunity to revive a policy that could really do a lot of good.
For me, it's one thing I hate (Pigouvian taxes), and another I'm ambivalent about (data centers). Just throw the kitchen sink at building as much cleaner energy as possible.
But I think the ultimate problem that data center perception faces is that many, perhaps most, people don't agree with Matt that "on its face, artificial intelligence seems to clearly be an economically valuable technology". Their reasons seem to be all over the place beyond using electricity--water consumption, IP infringement, job loss, particularly to the creative class, output deemed to be bad, investors deemed to be bad, existential risk to the human race. Any of these may not be well founded, but the fact that they are present suggests that not enough people do, in fact, find that this technology is clearly economically valuable. As for me, I find the value to be plausible, but not clearly demonstrated, hence my ambivalence. Perhaps that means that the companies building the data centers should work harder on their PR to demonstrate to the public that they are, indeed, creating economic value.
The single most under-discussed fault lines in left of center spaces is between taxes+transfers vs regulation as the primary government tool for addressing problems.
I like the idea of taxing data centers if they are producing a negative externality. But those on the other side of the debate just want them banned or want their buildout slowed down via red tape. Both approaches largely do the same thing, but only one raises revenue.