9 ms·

Should this apply only to AI compute facilities or all industries? I have no real evidence for this but I think a general principle of good governance is that

Should this apply only to AI compute facilities or all industries? I have no real evidence for this but I think a general principle of good governance is that the policy is agnostic to the industries which it affects and robust under any set of conditions, so for instance, the price that electric utilities charge should be enough to pay for expansions, either prospectively or retroactively (via loans secured before the construction, taken out against projected future revenue), rather than subsidizing the expansion with taxes collected by industries that won't be using the electricity, or exercising discretionary control over which types of industrial expansion are accommodated through investments in expanding the energy infrastructure. Huge expansions of energy production have occurred in the past to meet growing demand for energy from then-emerging industries. And I'm not aware of any past campaigns to stop the various industries that needed energy on account of the allegation that their increase in energy usage would burden other sectors of society. I have concern that the treatment of AI compute facilities is based on public opinion rather than rule of law and consistent protection of private property rights.

joshspankit 6d ago
AI compute is a unique facility: can be placed anywhere, takes in mainly energy, does not produce the same volume or composition of waste, has noise pollution, and needs minimal access. I am not a regulator but that feels like something that needs new rules
- ETH_start 6d ago
  AI compute facilities can inspire new regulations, yes, but what I'm advocating is that those regulations should not concern whether the facility is generating AI inference or aluminum or chemicals. All it should care about is the metrics that are relevant to the community like positive externalities, and negative ones like noise pollution, industrial pollution, etc. And all uses of private property should be judged on the basis of those criteria, not what industrial category the use falls under and the generalizations that are made about that category. My own cursory take on AI compute facilities is that they're mostly much better for the community than previous industries. They emit zero industrial pollution, and only require water for cooling and energy from any source. The water used can also be recycled. And even in the early designs that don't use recycled water, the water consumed is orders of magnitude less than agriculture. On the downside, they produce fewer jobs, but they encourage and fund infrastructure upgrades in telecom links and power transmission, which has second-order benefits for the local economy. And they consume very little government services, so the taxes they pay produce a much larger fiscal surplus for local governments. They also allow for energy that would otherwise be uneconomical to harvest, to be economically harvested. Let's just take an example of a desert which has very little wildlife, and no major nearby populations. AI compute facilities can make it economical to harvest the sunlight in those regions and make economic use of it.
  - Hikikomori 5d ago
    In general they just use a shitload of energy so prices go up for everyone and produce very little jobs. The worst part is that they don't generate any taxes for locals or Finland in general from services produced by datacenters there, which is very different from other industries.
    - ETH_start 5d ago
      They pay for the energy they use, which leads to the energy infrastructure scaling up, which generally reduces per unit electricity costs. They also fund, either directly or indirectly through their taxes, upgrades to transmission lines, which also improves the energy infrastructure. They don't create many permanent jobs on site, it's true, but they also don't consume many government services. So of all the industries, they produce the highest fiscal surplus for local governments. It's a net benefit to the community.
      - Hikikomori 5d ago
        What surplus? They basically only pay property taxes and for services they buy like electricity, like any other company does. But taxes for services they produce go to Ireland. A local industry would have loads of employees and also pay much more taxes that benefits the community. The community benefit is nowhere close and I'm not sure why you're trying to pretend it is.
        ETH_start 4d ago
        I'll try to find the analysis. It compared industries and showed the amount of taxes AI compute facilities pay (which is primarily in the form of property tax) relative to the cost of government services they use is the highest ratio of all industries. This surplus is why the county with the highest concentration of data centers, which is Loudoun County, has been able to reduce its property taxes year over year. The property taxes the AI facilities pay is what enables that.
        4d ago
typ 6d ago
I think it should apply to all industries. But for many (or most?) countries where power generation is privatized in the first place, this shouldn't be a problem that we need to fix. The trickier question might be about the grid, which in most cases is publicly funded.
AnthonyMouse 6d ago
> I think a general principle of good governance is that the policy is agnostic to the industries which it affects and robust under any set of conditions The trouble with this is that people frequently don't like the consequences of that. For example, a lot of the costs of a grid connection are non-linear if not totally independent of the amount of power you use. When a tree falls on a power line and a crew has to show up and install a new pole, the cost of that has nothing to do with the average percent utilization of those wires. Likewise, a line that can deliver 100MW to a single customer costs more than one that can handle 100kW, but it doesn't cost anywhere near 1000 times as much or consume 1000 times more land etc. The result of this is that the "works under any set of conditions" pricing for a grid connection would be a flat monthly fee for the capacity of the line totally independent of how much you use, and a line with 1000 times the capacity would only cost e.g. 10 times as much. But then a typical household electric bill would have a fixed charge of ~$100/mo for a grid connection regardless of how little electricity you use. Likewise, if you have a high proportion of renewable generation in the grid then the price per kWh during daylight hours when there are clear skies would be essentially zero, but the price on a cold still night in winter would be oppressively high and anyone who failed to turn off their electric heat on that specific day would end up with a $5000 electric bill. So in most places what happens instead is that the power company charges a fee per kWh which is used to pay line crews and other expenses that don't actually depend on usage and then have a much lower fixed connection charge. And charge a certain amount per kWh during days when supply is sufficient and then use the money to subsidize consumption during shortfalls, even though that actually raises total costs, because otherwise people would scream about prices hitting $500/kWh for a few hours every year or two. But once you're doing those things, a customer that wants megawatts of power is reasonably going to object to that rate structure, because the few percent inefficiency those things induce is millions of dollars in their case. At which point people start complaining about subsidies, even though not doing those things would be the absence of subsidies.
- ETH_start 6d ago
  That's a fair point, but even this kind of "exception" to the electricity tariffs (which gives the appearance of a subsidy and elicits public opposition) can be done in a more agnostic way. E.g. a 100 megawatt AI compute facility can be treated the same as a 100 megawatt aluminum smelter, can be treated the same as a 100 megawatt manufacturing plant. Large concentrated consumers of electricity can be treated differently than smaller consumers, but should all be treated the same as each other.
mitxela 6d ago
The principle prevents governments from taking actual realities of any specific scenario into account, so it seams bad.
- ETH_start 6d ago
  The principle means reasoning about the harms that you want to reduce, and the benefits you want to maximize, and designing policy thoughtfully around optimizing on those two fronts, instead of basing policy on vague generalizations about industrial categories.
impossiblefork 6d ago
So, the problem in this case is that there's no gain for the country building the datacentre. If you're in Finland and there are two possible uses for electricity production, let's say, either a steel plant or a datacentre. The steel plant will employ a bunch of people locally. A datacentre will employ a bunch of people in California. So if you are to build a datacentre, the deal must necessarily be that the R&D for the models that are to run on it must happen locally. Otherwise there's no reason to give them the allocation over the steel plant.
- ETH_start 6d ago
  Okay, that is a fair argument, but the policy that that argument inspires should be implemented in a more agnostic way. There should be some objective measure of "positive economic externalities generated per megawatt of electricity consumed". And projects should be judged on that basis, with projects below a certain threshold either being disallowed or being forced to pay more per megawatt of power consumed on that basis. It shouldn't just be based on some vague intuition, that is largely a product of the virality of the arguments that circulate on social media platforms (like this one).
  - impossiblefork 6d ago
    It's always good to set up objective criteria for decision so as to prevent corruption, yes. But when ad-hoc deals are made anyway, then it is more important that principles like those have been applied than that everything is ideal.
    - ETH_start 6d ago
      I'd err on the side of allowing development until a proper regulatory process can be implemented or unless there's a concrete reason to prevent it. If allocating scarce electricity resources to an AI-compete facility prevents the building of a steel plant, that would be a reason worth considering it. But there would need to be a concrete trade-off, not just a default assumption that the effect is net negative on account of the industry that the project falls under. More generally, the bias is to block development until safety/fairness issues are addressed, but it shouldn't be in my opinion, because that overlooks the risk of inaction. When development is blocked, what that does is reduce visible risks. What it usually increases however is total risk. We are already under constant threat from deterioration: aging, depreciation and decay. Entropy is the default. Action is what pushes back against it. We need to weigh any risks restriction prevents against the risks it leaves us less equipped to mitigate.

joshspankit 6d ago

AI compute is a unique facility: can be placed anywhere, takes in mainly energy, does not produce the same volume or composition of waste, has noise pollution, and needs minimal access. I am not a regulator but that feels like something that needs new rules

ETH_start 6d ago
AI compute facilities can inspire new regulations, yes, but what I'm advocating is that those regulations should not concern whether the facility is generating AI inference or aluminum or chemicals. All it should care about is the metrics that are relevant to the community like positive externalities, and negative ones like noise pollution, industrial pollution, etc. And all uses of private property should be judged on the basis of those criteria, not what industrial category the use falls under and the generalizations that are made about that category. My own cursory take on AI compute facilities is that they're mostly much better for the community than previous industries. They emit zero industrial pollution, and only require water for cooling and energy from any source. The water used can also be recycled. And even in the early designs that don't use recycled water, the water consumed is orders of magnitude less than agriculture. On the downside, they produce fewer jobs, but they encourage and fund infrastructure upgrades in telecom links and power transmission, which has second-order benefits for the local economy. And they consume very little government services, so the taxes they pay produce a much larger fiscal surplus for local governments. They also allow for energy that would otherwise be uneconomical to harvest, to be economically harvested. Let's just take an example of a desert which has very little wildlife, and no major nearby populations. AI compute facilities can make it economical to harvest the sunlight in those regions and make economic use of it.
- Hikikomori 5d ago
  In general they just use a shitload of energy so prices go up for everyone and produce very little jobs. The worst part is that they don't generate any taxes for locals or Finland in general from services produced by datacenters there, which is very different from other industries.
  - ETH_start 5d ago
    They pay for the energy they use, which leads to the energy infrastructure scaling up, which generally reduces per unit electricity costs. They also fund, either directly or indirectly through their taxes, upgrades to transmission lines, which also improves the energy infrastructure. They don't create many permanent jobs on site, it's true, but they also don't consume many government services. So of all the industries, they produce the highest fiscal surplus for local governments. It's a net benefit to the community.
    - Hikikomori 5d ago
      What surplus? They basically only pay property taxes and for services they buy like electricity, like any other company does. But taxes for services they produce go to Ireland. A local industry would have loads of employees and also pay much more taxes that benefits the community. The community benefit is nowhere close and I'm not sure why you're trying to pretend it is.
      - ETH_start 4d ago
        I'll try to find the analysis. It compared industries and showed the amount of taxes AI compute facilities pay (which is primarily in the form of property tax) relative to the cost of government services they use is the highest ratio of all industries. This surplus is why the county with the highest concentration of data centers, which is Loudoun County, has been able to reduce its property taxes year over year. The property taxes the AI facilities pay is what enables that.
        4d ago

typ 6d ago

I think it should apply to all industries. But for many (or most?) countries where power generation is privatized in the first place, this shouldn't be a problem that we need to fix. The trickier question might be about the grid, which in most cases is publicly funded.

AnthonyMouse 6d ago

> I think a general principle of good governance is that the policy is agnostic to the industries which it affects and robust under any set of conditions The trouble with this is that people frequently don't like the consequences of that. For example, a lot of the costs of a grid connection are non-linear if not totally independent of the amount of power you use. When a tree falls on a power line and a crew has to show up and install a new pole, the cost of that has nothing to do with the average percent utilization of those wires. Likewise, a line that can deliver 100MW to a single customer costs more than one that can handle 100kW, but it doesn't cost anywhere near 1000 times as much or consume 1000 times more land etc. The result of this is that the "works under any set of conditions" pricing for a grid connection would be a flat monthly fee for the capacity of the line totally independent of how much you use, and a line with 1000 times the capacity would only cost e.g. 10 times as much. But then a typical household electric bill would have a fixed charge of ~$100/mo for a grid connection regardless of how little electricity you use. Likewise, if you have a high proportion of renewable generation in the grid then the price per kWh during daylight hours when there are clear skies would be essentially zero, but the price on a cold still night in winter would be oppressively high and anyone who failed to turn off their electric heat on that specific day would end up with a $5000 electric bill. So in most places what happens instead is that the power company charges a fee per kWh which is used to pay line crews and other expenses that don't actually depend on usage and then have a much lower fixed connection charge. And charge a certain amount per kWh during days when supply is sufficient and then use the money to subsidize consumption during shortfalls, even though that actually raises total costs, because otherwise people would scream about prices hitting $500/kWh for a few hours every year or two. But once you're doing those things, a customer that wants megawatts of power is reasonably going to object to that rate structure, because the few percent inefficiency those things induce is millions of dollars in their case. At which point people start complaining about subsidies, even though not doing those things would be the absence of subsidies.

ETH_start 6d ago
That's a fair point, but even this kind of "exception" to the electricity tariffs (which gives the appearance of a subsidy and elicits public opposition) can be done in a more agnostic way. E.g. a 100 megawatt AI compute facility can be treated the same as a 100 megawatt aluminum smelter, can be treated the same as a 100 megawatt manufacturing plant. Large concentrated consumers of electricity can be treated differently than smaller consumers, but should all be treated the same as each other.

mitxela 6d ago

The principle prevents governments from taking actual realities of any specific scenario into account, so it seams bad.

ETH_start 6d ago
The principle means reasoning about the harms that you want to reduce, and the benefits you want to maximize, and designing policy thoughtfully around optimizing on those two fronts, instead of basing policy on vague generalizations about industrial categories.

impossiblefork 6d ago

So, the problem in this case is that there's no gain for the country building the datacentre. If you're in Finland and there are two possible uses for electricity production, let's say, either a steel plant or a datacentre. The steel plant will employ a bunch of people locally. A datacentre will employ a bunch of people in California. So if you are to build a datacentre, the deal must necessarily be that the R&D for the models that are to run on it must happen locally. Otherwise there's no reason to give them the allocation over the steel plant.

ETH_start 6d ago
Okay, that is a fair argument, but the policy that that argument inspires should be implemented in a more agnostic way. There should be some objective measure of "positive economic externalities generated per megawatt of electricity consumed". And projects should be judged on that basis, with projects below a certain threshold either being disallowed or being forced to pay more per megawatt of power consumed on that basis. It shouldn't just be based on some vague intuition, that is largely a product of the virality of the arguments that circulate on social media platforms (like this one).
- impossiblefork 6d ago
  It's always good to set up objective criteria for decision so as to prevent corruption, yes. But when ad-hoc deals are made anyway, then it is more important that principles like those have been applied than that everything is ideal.
  - ETH_start 6d ago
    I'd err on the side of allowing development until a proper regulatory process can be implemented or unless there's a concrete reason to prevent it. If allocating scarce electricity resources to an AI-compete facility prevents the building of a steel plant, that would be a reason worth considering it. But there would need to be a concrete trade-off, not just a default assumption that the effect is net negative on account of the industry that the project falls under. More generally, the bias is to block development until safety/fairness issues are addressed, but it shouldn't be in my opinion, because that overlooks the risk of inaction. When development is blocked, what that does is reduce visible risks. What it usually increases however is total risk. We are already under constant threat from deterioration: aging, depreciation and decay. Entropy is the default. Action is what pushes back against it. We need to weigh any risks restriction prevents against the risks it leaves us less equipped to mitigate.