Compute in America: Playbook for Secure Clusters at Home
Summary
Arnab Datta proposes "special compute zones" as a solution to expedite the construction of the five-gigawatt AI compute clusters in the US.
SESSION Transcript
Hi everyone, I'm Arnab Datta. As you mentioned, I work at the Institute for Progress and Employ America. Earlier this year at the Institute for Progress we did a series called Compute in America. So I'd like to be talking about that today, essentially to establish a policy playbook for how to get to secure five-gigawatt clusters here at home in the next three to five years.
So yeah, this is the question that was really animating us. How can we build these here in the US and how can we make sure they're secure? So you all know the industrial age has begun. It's a race for infrastructure.
Leading firms are planning five-gigawatt scale clusters. This is equivalent to several large nuclear plants. AP1000s are just over a gigawatt. So we really haven't built this type of infrastructure in a very long time.
As Charles outlined, we're entering a period of load growth and so we really need to build this incredibly fast. And global competition is fierce. The Middle East is offering cheap power capital from its sovereign wealth funds. China is doing similar things with their state-backed financing agencies and they're also building energy infrastructure incredibly quickly.
And why does this matter? If you believe that AI is going to be at the center of scientific and economic progress over the coming decades, it's really important that we maintain that competitive frontier edge here at home and also that we ensure that our adversaries can't misuse it.
So Charles kind of stole a couple of my slides earlier, but did a much better job of talking about this. So I'll be quick here. But one of the challenges that we have, we just don't have enough energy build-out and we can't really necessarily expect a lot of the major hyperscalers to take this on on their own, even though energy is a major bottleneck for them.
And building these data centers here, it's only 10, you know, it's less than 10% of the total cost, including operational costs over the period of operating for the data center. And so for them to put the capital up front necessary to build the energy infrastructure as well would just blow their costs up. And as Charles mentioned, we need firm power, so that means 24/7. And we have challenges across the stack.
Again, Charles went through quite well the different technology problems. I think I'm indicative of the problem. I didn't even put hydro up there, so it's a good point.
I'm going to focus on this last part here which is permitting and litigation risk, because that's a really, really important challenge that we don't have a good solution for right now. A lot of the conversations around permitting reform center on the point of how long our permitting timelines are. And that's really important. But the why is not really understood as much.
And a huge piece of that is because of the litigation risk associated with our permitting processes. So agencies from the federal government all the way down to local and state basically have to expand these litigation-proofing exercises in order to get a project permitted.
And that is what really extends the time. And that creates just an enormous amount of uncertainty. Because if you're putting capital into a project and you don't have certainty around the timeline to get it permitted, it opens you up to a significant amount of risk on the cost side.
So what's the solution? Our thesis is this idea we call the special compute zone. So there's a market failure around building secure AI data centers here at home. Companies would love to do it and would love to invest more, but it requires a lot of capital.
Energy policy to us is the opportunity. If the government can offer more certainty, speedier energy development for AI data centers, it can require adequate security in return. So that's really what we tried to operationalize. And I'll talk a little bit more about the policy solutions we mentioned in the report.
Just before I do that though, I want to talk about what uncertainty looks like. This is a diagram from the 80s from Chevron on its stage-gating process for new technologies. So you can see there's all these different things that happen when you get from inventing a new technology to commercializing it and getting wide adoption.
And so early on you have a lot of the research and development stages, opportunity identification to start. And then you have these different stages. Commercialization and broad application are where we still have a lot of uncertainty in terms of the litigation risk that I talked about.
So if you want to build a transmission line which is a mature technology, we know how to do this. It can take years, even more than a decade, to build the transmission line necessary to connect some of the new energy projects to the grid. So there's a project that I've written a lot about called the Cardinal Hickory transmission line, which began its public regulatory processes in 2014. It got its first approvals from the local governments and three separate federal agencies in 2019 and 2020.
That project, the construction did not complete until 2024 because of all the litigation that occurred during that period. So it took 10 years and that project connected ultimately nearly 200 emissions-free energy sources to the grid. And we lost a lot of carbon reduction simply from that. So if you care about the carbon reduction aspect of it too, this is a huge cost.
So what are the opportunities for reducing uncertainty? There's the financing side so you can speed up federally available financing. One example I will talk about is the Department of Energy's Loan Programs Office. The Loan Programs Office for now has a lot of capital that it can deploy to energy projects.
But one challenge of that is that the regulatory process, particularly in the nuclear space for next-generation technologies take a long time because these projects also need to be licensed under the Nuclear Regulatory Commission. And so because those regulatory processes take so long, they can't deploy capital even for activities that don't require any groundbreaking but still need to be completed in the first years of the project.
So things like contract acquisition, you know, there's long lead times for the nuclear supply chain and so you need to get capital out. All of that risk basically sits on the developers. So that's one example. Another example, very hot live issue right now is preserving the tech-neutral tax credits from the Inflation Reduction Act.
So these are mobilizing just a tremendous amount of capital into energy generation. And if they go away, we could lose over tens of gigawatts of energy that's planned in the pipeline right now. So that's really, really important. There's also on the supply chain we could utilize the Defense Production Act for the gas supply chain.
So there's a lot of challenges right now. GE Vernova, for example, one of the major CCGT producers says that they're sold out by 2030. You could use the Defense Production Act to prioritize orders for AI data centers. And then finally permitting.
So if you could utilize categorical exclusions which are essentially a lower threshold, lower level of review under the National Environmental Policy Act, you could make that land, federal land available and you could reduce the uncertainty associated with the permitting and litigation challenges that I mentioned before. So these are just a few things that you could utilize and tie to conditions for security that would make the build-out possible here at home and more safe.
This picture didn't turn out really very nicely on this slide, but we proposed essentially an all-of-the-above energy approach, including hydropower. And this has a lot of different assets. It just kind of, I think as a visual, hopefully you can get a sense we have a lot of assets here in the US. There's these big black dots that are retiring coal plants.
There's a program at the Loan Programs Office called 1706 that allows you to deploy capital, lending capital, to convert those coal plants into other forms of energy infrastructure, including natural gas, because it's lower carbon intensity than coal. And those are connected to transmission lines already. So we have opportunities there to potentially just convert those assets.
You can see in the west there, there's lands, significant lands in that yellow that are owned by the Bureau of Land Management. Those map quite well with geothermal resources. So we should make those lands available for geothermal. Thank you, everybody.
I'm not going to be here at the one o'clock discussion, but I'm happy to stick around if anyone has any questions.
So yeah, this is the question that was really animating us. How can we build these here in the US and how can we make sure they're secure? So you all know the industrial age has begun. It's a race for infrastructure.
Leading firms are planning five-gigawatt scale clusters. This is equivalent to several large nuclear plants. AP1000s are just over a gigawatt. So we really haven't built this type of infrastructure in a very long time.
As Charles outlined, we're entering a period of load growth and so we really need to build this incredibly fast. And global competition is fierce. The Middle East is offering cheap power capital from its sovereign wealth funds. China is doing similar things with their state-backed financing agencies and they're also building energy infrastructure incredibly quickly.
And why does this matter? If you believe that AI is going to be at the center of scientific and economic progress over the coming decades, it's really important that we maintain that competitive frontier edge here at home and also that we ensure that our adversaries can't misuse it.
So Charles kind of stole a couple of my slides earlier, but did a much better job of talking about this. So I'll be quick here. But one of the challenges that we have, we just don't have enough energy build-out and we can't really necessarily expect a lot of the major hyperscalers to take this on on their own, even though energy is a major bottleneck for them.
And building these data centers here, it's only 10, you know, it's less than 10% of the total cost, including operational costs over the period of operating for the data center. And so for them to put the capital up front necessary to build the energy infrastructure as well would just blow their costs up. And as Charles mentioned, we need firm power, so that means 24/7. And we have challenges across the stack.
Again, Charles went through quite well the different technology problems. I think I'm indicative of the problem. I didn't even put hydro up there, so it's a good point.
I'm going to focus on this last part here which is permitting and litigation risk, because that's a really, really important challenge that we don't have a good solution for right now. A lot of the conversations around permitting reform center on the point of how long our permitting timelines are. And that's really important. But the why is not really understood as much.
And a huge piece of that is because of the litigation risk associated with our permitting processes. So agencies from the federal government all the way down to local and state basically have to expand these litigation-proofing exercises in order to get a project permitted.
And that is what really extends the time. And that creates just an enormous amount of uncertainty. Because if you're putting capital into a project and you don't have certainty around the timeline to get it permitted, it opens you up to a significant amount of risk on the cost side.
So what's the solution? Our thesis is this idea we call the special compute zone. So there's a market failure around building secure AI data centers here at home. Companies would love to do it and would love to invest more, but it requires a lot of capital.
Energy policy to us is the opportunity. If the government can offer more certainty, speedier energy development for AI data centers, it can require adequate security in return. So that's really what we tried to operationalize. And I'll talk a little bit more about the policy solutions we mentioned in the report.
Just before I do that though, I want to talk about what uncertainty looks like. This is a diagram from the 80s from Chevron on its stage-gating process for new technologies. So you can see there's all these different things that happen when you get from inventing a new technology to commercializing it and getting wide adoption.
And so early on you have a lot of the research and development stages, opportunity identification to start. And then you have these different stages. Commercialization and broad application are where we still have a lot of uncertainty in terms of the litigation risk that I talked about.
So if you want to build a transmission line which is a mature technology, we know how to do this. It can take years, even more than a decade, to build the transmission line necessary to connect some of the new energy projects to the grid. So there's a project that I've written a lot about called the Cardinal Hickory transmission line, which began its public regulatory processes in 2014. It got its first approvals from the local governments and three separate federal agencies in 2019 and 2020.
That project, the construction did not complete until 2024 because of all the litigation that occurred during that period. So it took 10 years and that project connected ultimately nearly 200 emissions-free energy sources to the grid. And we lost a lot of carbon reduction simply from that. So if you care about the carbon reduction aspect of it too, this is a huge cost.
So what are the opportunities for reducing uncertainty? There's the financing side so you can speed up federally available financing. One example I will talk about is the Department of Energy's Loan Programs Office. The Loan Programs Office for now has a lot of capital that it can deploy to energy projects.
But one challenge of that is that the regulatory process, particularly in the nuclear space for next-generation technologies take a long time because these projects also need to be licensed under the Nuclear Regulatory Commission. And so because those regulatory processes take so long, they can't deploy capital even for activities that don't require any groundbreaking but still need to be completed in the first years of the project.
So things like contract acquisition, you know, there's long lead times for the nuclear supply chain and so you need to get capital out. All of that risk basically sits on the developers. So that's one example. Another example, very hot live issue right now is preserving the tech-neutral tax credits from the Inflation Reduction Act.
So these are mobilizing just a tremendous amount of capital into energy generation. And if they go away, we could lose over tens of gigawatts of energy that's planned in the pipeline right now. So that's really, really important. There's also on the supply chain we could utilize the Defense Production Act for the gas supply chain.
So there's a lot of challenges right now. GE Vernova, for example, one of the major CCGT producers says that they're sold out by 2030. You could use the Defense Production Act to prioritize orders for AI data centers. And then finally permitting.
So if you could utilize categorical exclusions which are essentially a lower threshold, lower level of review under the National Environmental Policy Act, you could make that land, federal land available and you could reduce the uncertainty associated with the permitting and litigation challenges that I mentioned before. So these are just a few things that you could utilize and tie to conditions for security that would make the build-out possible here at home and more safe.
This picture didn't turn out really very nicely on this slide, but we proposed essentially an all-of-the-above energy approach, including hydropower. And this has a lot of different assets. It just kind of, I think as a visual, hopefully you can get a sense we have a lot of assets here in the US. There's these big black dots that are retiring coal plants.
There's a program at the Loan Programs Office called 1706 that allows you to deploy capital, lending capital, to convert those coal plants into other forms of energy infrastructure, including natural gas, because it's lower carbon intensity than coal. And those are connected to transmission lines already. So we have opportunities there to potentially just convert those assets.
You can see in the west there, there's lands, significant lands in that yellow that are owned by the Bureau of Land Management. Those map quite well with geothermal resources. So we should make those lands available for geothermal. Thank you, everybody.
I'm not going to be here at the one o'clock discussion, but I'm happy to stick around if anyone has any questions.
