Just when you thought concerns over datacenter power generation in the US couldn’t get any worse, the Trump administration gets rid of emissions restrictions on power plants.
You can listen to the latest episode of The Kettle on this page, Spotify, Apple Podcasts, or YouTube. You can also subscribe on those platforms to be notified when a new episode goes live.
The EPA is also considering getting rid of reporting requirements on small-scale polluters (e.g., datacenters with on-site power generation), meaning all those unrestricted emissions at new bit barns might not even be mentioned to the public.
But hey, at least Nvidia is riding to the rescue with some datacenter energy efficiency improvements, as self-serving as they may be for the company’s bottom line.
Join host Brandon Vigliarolo, systems editor Tobias Mann, and IT infrastructure reporter Dan Robinson as they discuss these developments and why coal may not come roaring back as a datacenter power source anytime soon.
A lightly edited transcript is included below:
Brandon (00:01):
Hi everyone and welcome to the latest episode of The Register‘s Kettle Podcast. I’m Reg Reporter Brandon Vigliarolo, and this week we have a couple of stories to review on the topic of datacenter power. AI is driving the construction of massive datacenters around the US, and people are beginning to push back. But recent moves by the Environmental Protection Agency are threatening to keep more people in the dark while making it easier to pollute in the name of progress. Some relief might surprisingly enough be coming from AI itself.
And with me to discuss this is Systems Editor Tobias Mann and reporter Dan Robinson. Welcome to you both.
Tobias Mann (00:35)
Good to be here.
Dan (00:35)
Hi.
Brandon (00:36)
So Tobias, let’s start with the story out of the AI Infra Summit this week from Nvidia. Apparently they’ve got some new iterations of some old technology that could make datacenters a bit more energy efficient, is that right?
Tobias Mann (00:48)
Yeah, it’s kind of interesting because you start looking at this and going, Well, why does Nvidia need to care about datacenters? But as it turns out, it’s a big bottleneck for their revenues. So this is, you know, all of a sudden something that they’re really interested in because their ability to sell GPUs is entirely dependent on there being enough power or people being able to get more out of the power that they already have.
Brandon (01:18)
Right.
Tobias Mann (01:18)
And so this has kind of been the focus of this DSX platform that’s been slowly rolling out since I think about around this time last year, they introduced this. And this is part of this AI factory push that they’ve been kind of dogging on for the last couple of years. and there’s and there’s some good news here because there’s a lot of technologies that have been introduced over the last couple decades.
You know, kind of based on the idea of as chips get hotter, we could be running datacenters more like smartphones. You know, you leave, you leave for the day.
Brandon (01:56)
Scaling up and scaling down, right? Yeah.
Tobias Mann (01:59)
And you don’t need to be running your air conditioning on full blast. You’re just wasting electricity. It’s basically the same idea of getting all of the pieces within the datacenter talking to each other so that you can have those kinds of automations and then the deeper that automation goes, you could have it based on a kind of demand response from the workloads themselves telling the rest of the datacenter how to adjust its settings to run efficiently, but also keep everything within their proper operating margins. That’s the entire idea behind what Nvidia has built with DSX. And this week, we got, prior to this, some hand-wavy kind of explanations. It all sounds good on paper, but this week we finally got some kind of proof of concept examples of how this is actually playing out in live deployments.
Brandon (02:54)
Okay, so that was not announced this week, but more like a ‘Hey, yeah, you guys knew this exists, but here’s some actual case studies or examples of this actually functioning in the real world.’
Tobias Mann (03:06)
Right. I mean, that’s really kind of the thing that everybody who might adopt this needs to see. Because we’ve seen versions of this where everything’s gonna talk to each other. Well, if you buy this thing from Schneider, we can’t work with that, or you gotta put it through our software platform and certain capabilities get locked out. This is if you’re in the Nvidia ecosystem, which everybody is at this point.
Whether it’s Vertiv or SN, because you have to buy their equipment and in order to deploy these GPU systems, you have to have power cabinets, you have to have cooling systems. Everybody has gotten on board, and so now you actually have a system where everything genuinely does talk to each other, and you can have these kinds of automation.
Brandon (03:54)
With, I’m assuming, AI.
Is AI kind of the orchestration layer here, right? As opposed to trying to build APIs to make these things talk to each other in a more old-fashioned way, there’s a little bit of an easier problem to solve here now that you have some machine learning to layer on top?
Tobias Mann (04:10)
I’m sure machine learning comes into play here. I don’t know the degree to which AI is actually influencing how the platform works. I’m sure it does.
Brandon (04:20)
Okay.
Tobias Mann (04:20)
Because Nvidia loves to shoehorn AI into everything just as another claim to do it.
Brandon (04:25)
Of course.
Tobias Mann (04:27)
But the two examples that they gave are pretty interesting because they demonstrate how you can get more out of the available capacity that you have. And on the flip side of that.
Not all datacenters are going to be running at full load all the time. And so they’ve also demonstrated how they can give back to the local grid during periods of you know high demand. You know, prior to us starting recording, we were talking about heat waves in London over the summer. This is a prime example of one of the things that they’re doing with this DSX Flex program, where the utility can send a demand signal to datacenter operators to say, ‘Hey, we’re gonna need a whole lot of power because it’s hot and everybody wants to run their air conditioning, and people are gonna be heading home early or starting them up before they normally would. Please cut your power.’ That works great for things like forges and other, you know, systems where it’s a switch and you can just say, ‘Okay, we gotta switch off half of the equipment in the facility.’
You go click, click, click, click. With datacenters, that’s not so simple because you can’t just shut off the machines. You gotta shut down the workloads. You have to let them get to this to a checkpoint where they can save. And so the the platform is built down to a granular level where you can have this demand signal come in and it automatically figures out what workloads can be shut down right away, which ones have to be kept running, which ones can be migrated to a different datacenter, kind of chasing the sun, so to speak in order to keep the grid stable.
And one of the big benefits that Nvidia sees in all this is again it comes back to money. I mean it it sounds really great for the you know you and I who are worried about brownouts during a hot summer day, but from Nvidia’s perspective, if the utility knows it can claw back, say 30 percent of the power that’s been allocated to the datacenter, they might green-light, say, a 150 megawatt datacenter buildout, knowing, if they’re not quite sure about, say, fifty megawatts of that capacity.
Whereas they might not have said, you know, we can give you a hundred, but we can’t give you a hundred and fifty. You know, so there’s you know, there’s that and every kilowatt that gets opened up is an opportunity to drive revenues for Nvidia.
Brandon (07:06)
Right, right. And of course, you know, I think as you mentioned toward the end of your story, this is not also an entirely benevolent move on Nvidia’s part either, right? But this is creating a lot more lock-in for them, right? Like you said, companies using other chips, other designs like AMD kind of setups and everything, they’re gonna have to probably invent their own version of this, you know, because Nvidia is creating an opportunity for people to to get more hardware more locked into Nvidia’s ecosystem, basically.
Tobias Mann (07:34)
Right. You know, the other example that they gave, this was working in collaboration with Lambda Labs, was this what is it called? DSX MaxLPS. And it was, you know, this example of where by knowing where all of the power is going in the datacenter, you can run a whole lot closer to the actual capacity available to the facility. Cause usually traditionally you might run at 80 percent of the available capacity because if you had too many things click on, or you had failures for redundancy, you didn’t want to end up in a situation where you were pulling more than you actually had access to. And so this visibility lets them run closer to that edge. Now, from an economic standpoint, somebody buying into this sees that they can get 19 GPU nodes running in a power envelope that previously supported only 16. That’s a huge economic driver and a very attractive one to go with Nvidia now.
And so they’re building this walled garden around this AI factory push where it’s like, well, it’s an Nvidia AI factory. So you can cram more GPUs into the same number of megawatts that you would have, and you can probably convince your utility to give you more megawatts than they otherwise would, because we have it so tightly integrated that you know it won’t be a problem.
Whereas if you go with AMD or any of these other AI chip startups, you might get some of those features from Schneider and others, but it might only work with their equipment. It might not be as feature rich. And so it then becomes just a you know a gamble to go with somebody else other than Nvidia. This is the sales pitch here.
So then you know, once you build this in a DSX spec Nvidia AI factory, are you going to risk putting in AMD GPUs? Now Nvidia says that they’re going…
Brandon (09:45)
Yeah, that won’t talk or to you know, yeah… I was gonna say that these GPUs might not talk to the rest of the stack now, right?
Tobias Mann (09:55)
And so there is definitely some walled garden building going on here and some lock-in. I don’t know how long it will last though, because I think Nvidia recognizes that that kind of lock-in is potentially problematic from a regulatory standpoint, but also they want everybody building to their spec. And so, in some cases, they have opened up the technology so that everybody can have like 80 percent of it, and then they get like the 20 percent on top of extra feature stack with their stuff is enough for them to be okay. They did this with their rack designs with MGX. So we might see something like that happen from them just to get more datacenters buying into the design. But it’s hard to say at this point. They may just use it to block out competition.
Brandon (10:49)
Yeah, we’ll see, right? I mean, but it seems like in general this might have a net positive effect on datacenter energy footprints, right? We’re talking running more efficiently, you know, getting more out of each kilowatt, out of each megawatt. This is a net positive for energy concerns when it comes to datacenters, it seems.
Tobias Mann (11:09)
I think I think this is absolutely the right direction that we should be going. Historically, and I know Dan has written about this quite a bit over the last couple of years. Historically, these power saving features don’t get used. They’re there and they look really promising, but when you have a datacenter that’s serving customers, there is a great deal of concern about meeting your SLAs.
And so the risk inherent to enabling these kinds of power savings features has always been that you might not, you know, you might lose performance, you might introduce latency, you might not be able to reach your your SLAs, and so you might end up not being able to charge your customers for the the work that has been done. And the power but the power consumption prior to what, 2022, 2023, has been pretty benign relative to the kinds of datacenters we’re talking about now.
Brandon (12:11)
Sure, yeah. Yeah, this is a huge increase. It is interesting that we have kind of like this opportunity for AI to solve a bit of the problem that didn’t exist before AI came along and and and caused this, right? So I guess we’ll see. There’s some good news here, but on the flip side of that, we have some recent announcements out of the EPA that make the datacenter energy problem, at least in the United States, seem like it’s gonna get a heck of a lot worse. Dan, you covered a story in late August about some proposed changes with public reporting requirements for air pollution. Is that correct?
Dan (12:45)
That’s right, yes. The story was that the EPA was proposing to eliminate basically federal rules that would require participation in minor source air permits is what it’s called but what it basically boils down to is to whether there’s any kind of public notification about things like datacenters that are going to be causing pollution and whether there’s any feedback on that, whether there’s any say, the public gets any say on it.
Dan (13:34)
So, at the moment, this thing exists as part of the Clean Air Act, I believe, but they want to undo that and basically it will go back to state level, so it will be up to the local authorities as to whether they actually bother to do that when there’s a new datacenter or something being built.
Brandon (13:58)
Right.
So when we’re talking about sort of minor or small emissions issues. This is we’re talking about like datacenters that are building their own on site energy production facilities, correct? I mean, this isn’t talking about grid connection or new construction being done for existing grids that datacenters are gonna connect to, or is that covering both here?
Dan (14:21)
Yeah, I mean it’s basically causing anything which might add extra pollution into the area.
It possibly doesn’t include things like power stations and things which is where the minor comes in. But a lot of datacenters these days are generating their own power and effectively having small on-site power stations in addition to the backup power generators that they all have. And so it’s just a kind of… They pitched it as a rolling back of power, deregulation, removing unnecessary regulation.
All these woke rules that people have to follow.
Tobias Mann (15:16)
I was gonna say that I feel like the residents of Boxtown, Tennessee probably would disagree with all of this already.
Brandon (15:22)
Mm-hmm.
Tobias Mann (15:23)
You know, the emissions coming off of these quote unquote portable gas generator plants, that they’re portable in the sense that they’re they’re an extra wide load that can be hauled behind a a large semi-tractor. and what if you looked at the Colossus datacenter that SpaceX AI built what a year and a half ago. They had 30 some of these running because they were waiting on a 150 megawatt interconnect to the utility.
Dan (15:59)
I think it was more than that at one point, wasn’t it?
Brandon (16:01)
Yeah, I think it ramped up. I think there’s maybe fifty or more there now. Yeah.
Tobias Mann (16:05)
Yeah, yeah, it’s increased now. But the challenge that happened was for a certain period of time, I think it might have been about a year that they could run these gas generator plants without all of the air handling equipment that would make them more you know, less polluting, and cut down on the smog-generating materials. And so they just didn’t.
They’re just like, well, you know, we’re not required to.
Brandon (16:33)
Right. I mean that’s the thing. There’s no requirement.
Yeah, exactly. That’s the thing is there’s no requirement that these companies or that datacenter operators who are bringing on their own site power have to choose the cleaner version, right? There are multiple kinds of gas turbines available. Some, like you said, include pollution mitigation technology, but there’s nothing that says you have to use that and they’re more expensive. So if you’re a datacenter operator that’s trying to keep their bottom line up, you’re probably not gonna opt for these if you don’t have to.
And now if you don’t have to report to the local community the potential emissions from this thing, there’s even less of an incentive to do that. Dan, do you know if this rule has been finalized or is it still the sort of thing that’s in discussion?
Dan (17:15)
This was only a month ago, I don’t think I have heard about it being finalized.
Brandon (17:21)
Okay.
Dan (17:23)
We heard of it because of a pressure group called the Environmental Protection Network, which is actually made up of former EPA employees who were raising the alarm about this because they were saying that apart from the greenhouse gases, extra greenhouse gases that might be emitted by these things. There’s actual pollution, know, the nitrogen oxides that generate smog and also, there can be heavy metals and other pollutants in there which are not very nice to breathe in.
Brandon (18:03)
No, definitely not. I mean, and so that’s still kind of in the maybe this will make things worse category. But in the actually making things worse category, I don’t know if you guys had heard this week that there was a G20 energy meeting in Houston, that was going on and the Environmental Protection Agency, which I mean, is barely even worthy of that name anymore, they decided this week to announce that they were eliminating these Biden-era greenhouse gas requirements for power plants, basically getting rid of restrictions on on power generation facilities that capped their greenhouse gas emissions, right?
And they basically were saying in the EPA’s press release, they said, you know, this was gonna be a cost-saving measure, it was gonna save $310 billion. and there were some additional proposals they rolled out to eliminate, quote, every remaining greenhouse gas standard for the power sector, unquote. And that they said would save an additional $370 billion in direct compliance costs, which, you know, fantastic, right? But now they’re saying that this is gonna probably cause an increase in coal production for power sector use by a factor of 10, right? So basically we’re being reintroduced to the era of soot covered cities and smog and, like you said, heavy metals and all this in the name of essentially.
I think a lot of this is probably gonna be rolled out to support the AI industry, right? I mean that’s what the power push is for right now. So why is the American government deregulating pollution requirements if for no other reason than to ensure that we win this supposed AI race that we’re in with China right now?
Dan (19:42)
Yeah.
Tobias Mann (19:42)
Yeah, this is one of those that the US government is sometimes disconnected from the realities of the power sector.
There is the potential that we’ll see increased coal use. I’m not sure that that will necessarily be true. One of the factors here is that getting coal out of the ground and maintaining those facilities is just getting really expensive in general relative to other sources of electricity. But the other part of this is that power utilities in the United States have been burned so many times on speculative markets that they’re really conservative to the point of being infuriating to the AI sector. They don’t want to add capacity based on demand that hasn’t materialized yet. Because if they do the ratepayers end up paying some of it, but they end up eating a lot of the cost as well, ’cause they can’t put you know, they can’t push the entire cost under the ratepayer. It’s just not possible to do that. And so while the vision of Blade Runner and and the blackened sky and the perpetually orange smog is a visceral one.
At this point the fuel costs have gotten to such that I don’t know that this is going to necessarily have the negative impact that we fear, but also I don’t necessarily think it will have the economic impact the US government is pushing with this.
Brandon (21:18)
Yeah, I mean Trump has been touting bringing back coal for a long time, right? And he’s gonna cause a resurgence in the coal industry. And it hasn’t happened, right? Because like you said, it’s prohibitively expensive. A lot of the mines have already shut down, or getting the remaining coal out is more and more work. I hope you’re right.
Dan (21:37)
Sure, sure. But some of those coal-burning power plants were due to have closed down and basically the demand for energy has meant that they’re being kept online instead of being shut down.
Brandon (21:54)
Right. Right. And if they are still online and they’re now less beholden to environmental regulations, I mean that’s probably a net negative, especially now if there’s at least enough demand to keep them fired and keep them fired with less for new requirements. And also on again, right, I’m sure this touches on-site power generation for datacenters too. Right. If you’re building your own datacenter and installing gas turbines – I mean, no one’s going to be putting coal-burning generators beside a datacenter anytime soon.
Tobias Mann (22:24)
One would help.
Brandon (22:25)
Yeah. But, you know, I mean they do mention natural gas in this too, right? So they’re basically saying that they want to eliminate clean energy requirements for natural gas turbines too. So, you know, even if you theoretically wanted to get these cleaner ones, that might get harder if there’s already a huge wait for turbines.
Tobias Mann (22:44)
Right.
Brandon (22:44)
That the queue is getting bigger.
So I can’t imagine that everyone’s gonna wanna burn this gas responsibly. So we’re gonna get in the longer line for cleaner turbines that are now necessarily not gonna be in as high demand because there’s less requirements that we have to do this. So they’ve been actually repurposing old jet engines to turn them into the kind of turbines that can be used to generate energy.
Yeah, and as we all know, jet engines are incredibly environmentally efficient.
Tobias Mann (23:18)
There’s also an interesting play here. The race to get compute capacity up is so intense right now that the cost of choosing the most efficient fuel or energy generation mechanism is in some cases secondary to getting it deployed in the first place. And so we’ve also seen adoption of natural gas fuel cells. So similar to what you see and have seen in NASA spacecraft, but instead of hydrogen and oxygen running on these things, it’s predominantly natural gas reacting with oxygen to form CO2. The greenhouse gas emissions are roughly similar to gas turbines, but the pollutants tend to be fairly smaller because you’re not combusting anything.
So you don’t necessarily have as much kind of smog producing output from these systems. The downside is that they’re not particularly power dense and so you need a ton of these fuel cells and they’re expensive.
Brandon (24:28)
Right.
Tobias Mann (24:28)
But if you can’t get turbines and you want your capacity now, that’s what people are doing.
Brandon (24:33)
Yeah. That would be nice if more of those became more of an option, right? How are they to produce compared to turbines? I mean the turbine queue is big, but are these fuel cells, can they ramp up production quick enough that they’re a better option?
Tobias Mann (24:47)
This is an open question right now. There’s I think what one or two companies in the US that actually make these things.
Brandon (24:54)
Okay.
Tobias Mann (24:55)
Bloom Energy is the big one and big neoclouds like Nebius are deploying them in pretty large volumes in order to to supply power to their datacenters because they wanna deploy in areas that have good connectivity but are constrained on power. It’s oftentimes in those situations easier to get a natural gas line out to the building, than it is to get, you know, transmission lines hooked up.
Brandon (25:22)
Sure. Yeah.
Dan (25:23)
Yeah, yeah, in fact there was a report out just this week that the ability to build extra infrastructure, the grid, is actually holding back on datacenters.
It’s outpacing the rate at which the utility companies and the grid operators can actually add and build new lines to supply the power.
Brandon (25:52)
Yeah, I believe it. I believe it. I mean there’s a lot of projects happening in a lot of very remote areas, you know, that require a lot of transmission. Up in my part of Michigan, no one is actually discussing building a datacenter, but a lot of people are worried about it. And those of us who know this stuff are pretty much like, you need to relax.
There’s no transmission out here. And also beyond that, there’s no space for someone to build their own co-located power facility. You know, there’s a lot of protected land, a lot of occupied space and the zoning and everything, you know, doesn’t allow it. But there are a lot of areas like mine where there is space, but no transmission lines. And yeah, those areas are ripe for the picking, but you’re gonna need to supply your own power.
Tobias Mann (26:39)
Hermantown, Minnesota, right outside of Duluth on Lake Superior. Lots of water, plenty of power, cheap land – and not popular at all.
Brandon (26:51)
With datacenter operators? Yeah. Yeah, totally. And that’s the thing is I think a lot of people they
Tobias Mann (26:52)
No, no, not popular with the residents of that area. So yeah.
Brandon (27:00)
They don’t want them here. I mean, people don’t want them here. People don’t want them there. No one wants these things in their backyard, but increasingly I feel like we’re running out of space to do that. I don’t know. We’ll see what happens. Dan, I guess you’re just lucky you’re not here dealing with all this deregulation. Hopefully the air stays a little bit cleaner over over where you’re at.
You know, and I hope Tobias is right that we don’t have a whole bunch of coal plants going up in the near future.
Dan (27:26)
Yeah, we’ve got rid of all our coal plants over here, but there is a shortage of energy and we’ve got the most expensive energy in Western Europe, if not the whole G7 in the UK. And there’s no shortage of datacenters.
Tobias Mann (27:46)
No.
Brandon (27:46)
Yep.
Dan (27:46)
Plenty of datacenters and I know of at least three new ones that are being built within about a 20-mile radius of where I live, which is just to the north of London. So they’re coming. And people are worried about where the energy is going to come from to supply them.
Brandon (28:09)
Totally.
Tobias Mann (28:10)
I guess here’s a good place for us to place the bet. Which happens first? Small modular reactors or space datacenters at scale?
Dan (28:22)
Hahaha
Brandon (28:23)
My god. Yeah. I don’t want to take a bet on either side of that equation because I don’t know that I’ll be alive to see either one happen. I would love it if small modular reactors won out because that would mean that there would be more safe nuclear in more places. But I jeez, I don’t see it happening anytime soon.
Tobias Mann (28:45)
If you thought the NIMBY for datacenters was bad, just wait till you attach nukes to the sides of them.
Brandon (28:52)
I can’t imagine people will like it even if it is relatively safe compared to you know, we’re not in the Three Mile Island-Chernobyl era of nuclear power anymore. But I don’t think very many people have really bothered to learn to understand that.
So, well, no matter which way it swings, we will be here at the Kettle to tell you all about it, and we hope you’ll tune in and not be too miserable by the end of our future episodes. Thanks for listening. ®
