The AI energy bottleneck, with Tim Fist

This week, I'm joined by Tim Fist from the Institute for Progress. We discuss the supply chain and political/policy challenges of the infrastructure buildout required to hit the scaling curves widely assumed by the AI industry's anticipated power usage demands.

This ties together several conversations previously about electricity economics and data centers.

Sponsor: Vanta

Vanta automates security compliance and builds trust, helping companies streamline ISO, SOC 2, and AI framework certifications. Learn more at https://vanta.com/complex

Timestamps

(00:00) Intro
(00:40) Energy bottlenecks in AI development
(02:56) Technical and policy solutions for energy needs
(05:18) Challenges in transmission infrastructure
(12:14) Behind the meter generation explained
(17:50) Solar and storage: The future of energy
(18:47) Sponsor: Vanta
(20:05) Solar and storage: The future of energy (part 2)
(29:07) Power purchase agreements and financing
(33:17) Financing geothermal wells
(33:53) The promise of geothermal energy
(35:25) Challenges in geothermal adoption
(36:59) Industrial applications of geothermal heat
(45:01) Geothermal energy and national security
(49:27) Global investments in AI and energy infrastructure
(56:29) Policy and technical expertise in AI
(01:00:54) The role of government in technological advancements
(01:05:07) Wrap

Transcript

Patrick McKenzie: Hi, my name is Patrick McKenzie, better known as patio11 on the Internet. And I'm here with my buddy, Tim Fist, who is the director of emerging technologies at Institute for Progress.

Tim Fist: Great to be here. Thanks for having me on. Love the show.

Patrick McKenzie: Thanks very much for having me. Wait, that's your line. 

As people might get the impression of during this hour, I'm dealing with something of a cold right now. The good news is we can edit out the coughs. The bad news is we cannot edit up my level of energy. So if you think that I'm experiencing hallucinations, it is because of intermittent availability of energy resources in an unplanned fashion, which will introduce the topic of this talk.

With the AI boom, there's increasing interest, both at IFP, and policy spaces generally in the tech industry and elsewhere, about how the availability of energy and energy economics and the structural factors that impact that are going to likely bottleneck the pace of AI development over the course of the next couple of years. And so I thought I would have an expert to chat about that.

This serves as a companion piece to some of our previous episodes on solar economics with Casey Handmer, fracking with Austin Vernon, Azeem on the energy economics of data centers, and another episode with Travis Dauwalter on the economics of the grid itself.

But starting at a natural point, I think there are some people who might be less convinced that we are looking at an energy bottleneck. Can you sketch out why that might be the case?

Tim Fist: Yeah, I guess if we start with the basics, building AI as well as deploying it widely requires a bunch of specialized chips. So you have Nvidia and AMD with their GPUs, Google with their TPUs, Amazon with the new Tranium. These are installed in data centers to do training runs, but also large scale inference to serve those models to users. And training runs actually now span multiple data centers as well.

Google has pioneered a new deployment topology for AI training/inference: previously there was one physical datacenter, a single big building, which housed the the chips that are involved in training a single AI model. We're now moving to a world where you have multiple big buildings potentially geographically distributed to do that training. And the thing about these chips that are using these data centers is they require a lot of electricity, but actually the main overall cost of building these things on an amortized basis over time is the capital cost of the chips themselves.

[Patrick notes: This is extremely untrue of e.g. servers in regular data centers or, to use an example closer in character to AI chips, the bespoke hardware which dominates Bitcoin mining.] 

Tim continues: So if you assume kind of like a five to six year depreciation time frame, which is what we put into our modeling and seems to map onto sort of general lifetimes for GPUs, the cost of the chips themselves is about 80% of the overall total cost of ownership. So what all this means is that you want 24/7 power to maximize your return on investment in this extremely expensive capital equipment.

[Patrick notes: Depreciation is one place where accounting imposes its lens upon the physical world. In the physical world, different capital investments—from chips to sofas to factories—lose value at different rates, either because they are physically worn down through usage over time or (more to the point in the case of AI chips) they leave their “useful economic life” as improvements in technology make them no longer competitive for, in this case, training frontier models.

Why is it 5-6 years? Why not 3-4? Answer: a combination of a) guessing based on recent history what the actual useful economic life is and b) decisions made by non-technical accountants/etc based on what they think the IRS and other various taxation agencies will sign off on without a fight. If you want some scintillating tables to read, see this publication.]

Technical and policy solutions for energy needs

Tim continues:And so there's this question of how much electricity do you actually need? We put out this big report called Compute in America, which looks at this problem from a few different perspectives, first at the technical bottlenecks to building this stuff, like how much electricity do we actually need and how we do it, and then the policy solutions as well.

We looked at the electricity requirements from two different perspectives. First, the amount of compute that you need to train the most advanced models, which is roughly quintupling each year, so a factor of five each year. This is a pretty significant trend. Second is just the announced plans of the AI industry. So you would have heard of big projects like Stargate that involve hundreds of billions of dollars worth of investment. What is the actual power requirement of those across the timeframe that you're talking about?

Both of these trends pointed us needing at least a gigawatt of power for a training cluster by 2027 and then about five gigawatts by 2030. For comparison, XAI has been in the news a bit recently for rapidly building this huge GPU cluster in Memphis, Tennessee. That cluster is about 200,000 GPUs and about 400 megawatts. So if you look at the needs of the industry, we're talking about 2.5x that size within about a year and a half and then 10x that size within four to five years.

For reference, a single gigawatt is about a very large nuclear plant. And so we're talking about needing multiple large nuclear plants to support single training runs within just a few years. So yeah, our thesis is this is just a substantial amount of electricity that we just don't have on demand. There needs to be solutions to this.

Patrick McKenzie: My napkin math is that we are looking at a future where not just AI usage, but increasingly AI usage at data centers will consume tens of percent of the total US budget for electricity. [Patrick notes: Pip pip for the Situational Awareness essay series for raising my level of situational awareness on this topic.]

And in addition to consuming tens of percent, it will make up a very large portion of the marginal growth. And due to reasons that we discussed in other episodes, we tend to only have just a bit more capacity than we need because the economics of keeping baseline or rather baseload plants offline are very unattractive, particularly for nuclear and similar.

So that delta has to come from somewhere. And it seems unlikely that it's going to come from nuclear if we're unable to get a nuclear plant from the drawing board to production in less than 30 years, when we are talking about three, four, five year timelines. So what looks promising for filling that gap?

Challenges in transmission infrastructure

Tim Fist: Yeah, so it's maybe worth talking about what the bottlenecks to actually building it in the United States are. I suppose the first is just the generation capacity in the first place, as you talked about. So the kind that we need doesn't really exist on demand.

Interestingly, the United States has actually built a lot of new generation capacity over the past few decades, but overall electricity generation has remained flat. And that's because we're essentially taking coal plants offline, but then replacing it with a mix of natural gas, wind, and solar. So we are bringing a lot of generation capacity online, but a lot of it isn't the base load power that you would need to run this 24/7. And we ended up talking about behind the meter generation as a key thing, but I'll talk about the reasons for that.

The second key thing is transmission. Building new transmission lines permitting for that is an absolute nightmare. In 2013, the United States added about 4,000 new miles of new transmission lines. This year, it looks more like 500 miles. On average, it takes about 10 years to build a new line, and some take a lot longer.

One that we've talked about a lot at IFP is the Cardinal-Hickory line in Iowa-Wisconsin, which is going to be connecting, I think, 160-ish new clean energy projects to the grid. It's gone through the strictest form of environmental review, but just keeps getting held up by basically litigation. Environmental groups just keep suing this and these big lines that cross multiple different legal jurisdictions open up a bunch of different stakeholders, each of who can sue to delay the project. This ends up just happening indefinitely. We call this the litigation doom loop.

So it's one thing to be able to bring capacity online, but another thing to actually build the transmission lines to connect that electricity to where you need it.

Patrick McKenzie: We have promiscuously distributed vetoes within the United States political system and that has impacts everywhere, even for building apartment buildings and similar, but the nature of transmission lines is different than the nature of an apartment building in that, as you say, they will cross many jurisdictions, which each get sort of independent bites at the apple from people who are determined to block it. And for various unfortunate reasons, the United States has large and powerful political interests which simply oppose all new net construction at all margins anywhere. [Patrick notes: For much more on this topic, with an unavoidably politicized slant to it, please see Abundance by Ezra Klein and Derek Thompson, among many other places. Also touched on in the Housing Theory of Everything.]

That might be slightly unfair, but rounds to the case when one is talking about a transmission line. You only need one or two determined individuals in a several hundred mile stretch of rights of ways to successfully delay a project by plural years. It's an interesting incentive design and potentially legislative agenda to fix that. What's the name of the landmark environmental legislation that is always used? 

[Tim prompts me with the answer.] NEPA (National Environmental Protection Act), thank you. Can you give people who aren't familiar with NEPA a bit of a background around this because it's one of those things that I never heard about before I got really into energy economics, and then everybody in energy economics is extremely familiar with NEPA.

Tim Fist: Yeah, so there's a bunch of existing environmental regulations that we have that are specific to different forms of environmental harm, related to effects on water and air and environment. And these form substantive requirements around what you're actually allowed to do.

NEPA, on the other hand, is what's known as a procedural law. It basically says when you are building a project that has major impacts, and this can take a number of different definitions, but often it's if you're building on federal lands or engaging in significant environmentally disruptive activity, you need to be following a proper procedure for how you evaluate the environmental impacts. And there's not a clear list of things you need to do, and the interpretation of what you need to do is fairly subjective.

Outside groups can actually sue under NEPA. So even if you do the most strictest form of environmental review, and it takes many years, an outside group can still sue and say, I don't think you went about this procedure the right way. Under NEPA, you can just keep doing that essentially. And so this is this litigation doom loop that we talked about earlier where a project can just be held up indefinitely due to lawsuits under NEPA.

One solution that we talked about a lot at IFP is setting an injunction on how long a project can be subject to these things. So if you say, all you've got four years to lodge as many lawsuits as you want, and then the project has to go ahead no matter what's happened. That's a potential solution here. But in general, in its current instantiation, it just allows this to continue indefinitely.

Patrick McKenzie: So we'll talk about geothermal in a while, but in my work with the geothermal nonprofit [Patrick notes: I am an advisor to Project InnerSpace, a geothermal focused research organization (FRO), and am speaking purely in my own capacity as someone who “got bitten by the geothermal bug” in the course of my engagement with them] over the last couple of years, a thing that was surprising to me was this.

I have a blinkered and somewhat particularized understanding of United States political economy. This would broadly suggest that California puts more calories into green issues than Texas does, inclusive of green power. However, if you take the twin cases of California and Texas, California is very green in a lot of ways, but is lagging on some of the development fronts with regards to renewable energy generation, specifically because California is also home to a large fraction of the US environmentalist movement. They are extremely well capitalized due to wealth generation in California and file a lot of NEPA lawsuits and consume a lot of lawyer-hours.

[Patrick notes: This is a cautionary tale to both the wealthiest tech funders and also the broad “tech middle class”, which are both very deep pockets of non-profit, civic, and political dollars in California specifically and, increasingly, nationally. The “tech middle class” includes people who will make seven figures or more of donations over the course of their working career. You’re welcome to ask your friends in it what the character of their charitable giving is; very little of it would surprise people familiar with this social set.

The rough character of this class is mid-level engineers and managers at large tech companies, startup employees who have had an exit or two, and similar. Every time an IPO happens, about a thousand people quietly join it, and somewhere on the order of 10-20% of what they make goes to charity (eventually), often after a few years cooling off in a donor-advised fund.] 

Patrick continues: This has the probably unintentional consequence that Texas, which many California environmentalists might turn their nose up at, is leapfrogging them with respect to solar deployment, geothermal deployment, and potentially grid improvements as well, although I'm less familiar with grid improvements.

Tim Fist: Yeah, it's deliciously ironic, I think, that Texas is leading the country in terms of new energy build-outs and ease to build new clean energy. [Patrick notes: One reason, covered previously, is that California bet big on rooftop solar and effectively opposed “utility-scale” solar generation. Texas took largely the opposite policy approach. It turns out, I think it is fair to say thoroughly unsurprisingly, that there are returns to scale in power generation, and in important parts of the graph they are accelerating returns to scale.]

Patrick McKenzie: So we've talked about the importance of building additional grid infrastructure because while we have this sort of background model that electricity is fungible once you're on the grid, one, that isn't actually physically true. We do have electricity losses over large transmission lines. But two, it just isn't the case that energy resources are distributed naturally in the fashion where we would like to distribute centers of energy demand. What are the hyperscalers doing to try to co-locate their data centers more closely to natural energy sources?

Tim Fist: Yeah, so there's a couple of obvious things that it makes sense to do. The first is just try and co-locate your load with existing sources of large-scale generation. So we see Amazon doing this in Pennsylvania where they're trying to build data centers right next to existing nuclear plants and suck up spare capacity there. Microsoft also has done the same thing in Pennsylvania. Microsoft actually helped reopen the 800 megawatt Three Mile nuclear plant in Pennsylvania. And you see similar stuff happening in New Jersey as well.

The problem is there's only a finite and I would claim not sufficient amount of spare capacity available from existing plants where you can do this. We see very aggressive bids from hyperscalers now to try and suck up all this additional capacity and then make lots of speculative bids around different places in the grid where new energy projects are coming online.

Behind the meter generation explained

The second step you do after this is look at behind the meter. So rather than connecting to the grid and building new lines or connecting a new load, you instead build your own microgrid. So you're building the data center together with the electricity infrastructure that's required to support that. And this is actually a really promising strategy for AI because a lot of the loads don't need the same sort of low latency connection to fiber backbone cables that you see for more traditional cloud services like those that support video streaming and other kinds of things.

Especially for training, this is most obvious. You have a massive data set of many trillions of words or tokens, and you're basically pushing that data set through many hundreds of thousands of GPUs. And you can essentially do that without being connected to the internet. So you can put the data center that's processing that workload somewhere where you don't even have an internet connection if you really need to. And you can flexibly cite that infrastructure depending on where energy resources are available.

I would additionally make the claim that potentially a lot of deployment or inference workloads might look the same, where right now we're kind of used to this paradigm where you're talking with a chatbot and you're having fairly low latency connections with the chatbot, and this supports a fairly fast back and forth conversation. But a lot of future AI workloads might look a lot more like you give the AI a task and it goes out and does it by itself for a few hours and then comes back to you. And it's only at that point that it actually needs to communicate.

The AI system might need low latency connections to other web services that it's trying to interact with, but you're not needing this fast back and forth with users overall. So in a world where we're deploying many thousands, millions, maybe even billions of AI agents that are all doing work behind the scenes, you might also not have such strong latency requirements. This is a big open question about what is actually the latency requirements and how does that enable you to distribute data center infrastructure in places where you wouldn't normally be able to do it. But overall, this behind the meter path looks more promising in the AI industry compared to other data centers and cloud services.

Patrick McKenzie: The physical limitations of having a human in the loop for various computing tasks are interesting when you think about latency requirements. One, while we are generally averse to wasting humans time, where if one is able to spin up agents on demand, an additional marginal agent having to pause for five seconds or five minutes or five hours is probably pretty close to free.

And then on the flip side, there is basically no way to improve human processing speed and latency between their eyeballs, their brain, and their fingertips, where we have seen great increases in that sort of thing on the computer side of the world. And so for systems where you remove the humans but still have a speed premium, such as automated trading, then you will go to incredible lengths to decrease latency.

For many tasks in the economy that don't have a human involved and that computers are in charge of, things which by computer standards are incredibly languorous are by the rate limiting steps of the process, like moving packages across the country, imperceptible in terms of time.

[Patrick notes: This is an important capacity planning factoid for engineers to know. For example, back in 2021, the White House was very worried that Americans would distributed denial of service attack vaccines dot gov by all hitting it at once, as soon as President Biden got on the news to announce the website. The powers that be directed the technical teams to prepare for 100 million simultaneous users.

This was bonkers prioritization, coming from a misunderstanding of how vaccines dot gov was different in character to healthcare dot gov, which fell over due to a sustained spikey use pattern because of fundamentally different user dynamics. Vaccines dot gov was a read-only website and the queries per second it would need to serve would be limited by people’s abilities to physically key them in. Even assuming 100 million people in America used it in a 48 hour window, another ludicrously unfavorable assumption, the natural spread in access patterns would make actual traffic against the website very manageable over almost all seconds in that window.

(One fun access pattern: VaccinateCA routinely observed a detectable surge in the second after 5:00 PM California time. After talking to users, our best guess was it was largely people checking their personal cell phones “as soon as they got off work” and getting a text from a friend/family member about us. Even with human levels of precision, the bell curve centered on 5:00 PM lead to that second being visible in the graphs.)

But people in the White House didn’t see it that way, apparently, and the tech people they were able to pass marching orders to did not countermand them, and so a very smart team (that I have a lot of regard for) spent a lot of cycles over a few weeks preparing for a traffic event that did not and could never materialize.]

Tim Fist: Yeah, it also seems to be the case that building new fiber backbone infrastructure, either between data centers or from a data center to existing fiber backbone routes, is faster to do than building new electricity transmission infrastructure. So it seems easier to build a new data center and generation infrastructure somewhere fairly remote in the country and then just connect that back with a fiber line. That's easier than trying to connect it with a transmission line.

And even if you do need those low latency connections, building fiber ends up being a lot more tractable overall. This is partly because there's not the same requirements with transmission lines to actually follow a straight line. So if you look at natural gas pipelines, for example, generally easier to build than transmission lines. And this is because you can take a circuitous route to avoid crossing parcels of land where you expect to receive a lot of local pushback or lawsuits for having an environmental disruption there.

Fiber is similar in that it's nice to have these minimum distance straight line connections, but it's not necessary. You can take a more circuitous route as well.

Patrick McKenzie: So you mentioned behind the meter generation, which is a term of art that is well known by people in power circles, but maybe less well known by the typical person who has consumed electricity in their life. Could you just sketch out what that means in practice?

Tim Fist: Yeah, so essentially it's building your own power generation and using it to power a co-located load. So in this case, a data center, but it could also be a factory producing other things, anything that's energy intensive. And we see a lot of increased interest in this over the last couple of years, partly driven by the needs of data centers overall.

It introduces a different set of technical challenges for comparing different types of generation. For example, solar is in part feasible and now cheaper than things like natural gas in certain areas because you can connect it to the grid and not need to have 100% uptime for it to be useful. You're just supplementing existing sources of baseload power.

But if you're connecting solar in a behind the meter configuration, you actually need to really overbuild on the storage side to guarantee that you can have 100% uptime for the power load that you've got. So relative to things like natural gas or geothermal, which I'm sure we'll talk about, or small modular reactors or nuclear plants, solar power looks less promising because you can't get 100% output out of it all of the time. So some of the technologies that have certain economics when you're connecting to the grid have different economics when you think about them in a behind the meter or BTM setting.

Solar and storage

Patrick McKenzie: I also think there's some heady political and unfortunately geopolitical realities which differentially affect different energy types, particularly in the wake of the last couple of weeks. Speaking bluntly, a lot of the progress in solar power over the course of the last decade has come because you can put solar equipment in a shipping container from China and then deploy it in the United States.

It is feasible to deploy relatively large solar plants in relatively remote locations and have just the final bit of the labor and capital investment required to make the plant operational happen in the United States. In a world where access to Chinese goods is curtailed, the relative effectiveness of solar and the relative effectiveness of some other potential technologies, even potentially small modular reactors, becomes maybe a little bit less interesting. It becomes more interesting relatively for forms of power where more of the supply chain sits in the United States. [Patrick notes: Salaryman promotes the subtext to text: geothermal is downstream of the fracking supply chain and the U.S. owns the world market for most of that. This is less true for the parts above the hole, which are under rapid active R&D at present. But e.g. a commanding share of large turbines come from Japan, not from China, though it is not obvious to me that all present or future configurations of U.S. political elites will be able to distinguish between those two nations in trade policy.]

Tim Fist: Yeah. I'd be interested in your perspective on this from previous conversations you had. I'm definitely not an energy industry expert, but through the course of looking into this specifically for addressing the AI data center buildup problem, I have talked to a lot of energy experts.

One perspective is that across a more decade-long timeframe, solar and storage just clearly looks like the best option because you have these really aggressive cost curve reductions. There's this term that I like, I think Sarah Constantine wrote about this, called the "enchipification" of different things where both solar panels and batteries can benefit from Moore's law like trends and the advancement of these technologies where they just quite predictably get exponentially better over time.

Whereas if you look at things like geothermal and nuclear and natural gas, they generally always rely on basically having a turbine. Heating up a big pot of water to spin a turbine, and that generation infrastructure is this big fixed cost that is actually going up over time compared to these other sources of electricity.

So we're kind of at this tipping point where solar and storage is cheaper overall than a lot of alternative energy sources in some cases. Right now, it's cost competitive, but in like 10 years, you just expect it to be way easier to way overbuild your solar and storage compared to trying to do some of these more fixed costs on site generation.

Patrick McKenzie: Also not an energy expert, and to the extent that I have a hardware background as a system engineer, hardware is the part of the system that scares me and I would love to just pretend that power came out of something on the wall and my only interface to it is on the Linux command line. [Patrick notes: I’ve never actually had call to do that, and do not know if it is convenient or technically possible. Every deployment at my companies has had the character “If the computer can execute commands, power/cooling is presumptively adequate. Good job, highly specialized real estate company and/or VPS provider.”] 

Be that as it may, one of the things that I was most surprised by in my adult life was the cost of solar curve, which I will drop a link to this graph in the show notes. It's one of the most important graphs for citizens of the modern world to understand, I think. Up until about the point where I graduated college in 2004, there had been a gradual reduction in the cost of solar energy, and it was not yet price competitive with, for example, coal. If you projected out that gradual reduction, you would think, okay, absent a very large subsidy regime, solar will not be cost competitive with coal or other fossil fuels, and therefore will be a very small part of the energy mix.

That did not happen. And the reason was we had a sharp acceleration of the learning curve, meaning the price of solar got very precipitously cheaper by multiple orders of magnitude in the intervening 20 years. And that curve has not bottomed out yet.

So I think the big question, and I do not have a very developed point of view on this, is what is one's thoughts with respect to batteries over the course of the next 10-20 years. If you think that we're ultimately discussing physical realities, advances in material science, and the chemistry in the universe that we are blessed with or not blessed with as it were, and if you think that batteries become arbitrarily cheap in the same fashion that solar becomes arbitrarily cheap, then I agree it becomes almost irrational to think of much else in the energy mix aside from solar when the sun is on, battery when the sun is off, and we're done. With some amount of redundancy, but it will be a commanding share of the energy mix.

If rates of improvement in batteries decline, then the blended cost of a solar plus battery installation will be dominated by the batteries as the cost of generating solar energy goes to essentially zero. And that would imply a much greater future for other firm energy sources, such as geothermal, next generation nuclear plants, to be a large fraction of the energy mix. So the big bets are on material science, batteries, and what interesting tricks we can come up with in the next couple of years.

Although I do think some hardware problems are actually software problems in disguise and just put a chip on it and have that chip do slightly more intelligent things than we did 20 years ago. It does generate wins at the margin in some infrastructure fields, including very prominently in payments.

[Patrick notes: Although security improvements are probably the most notable win, “Shenzhen eating the world” has caused small businesses worldwide to have much smarter terminals and point-of-sales systems currently than historically prevailed. That allows much better point-of-sale user experiences, such as e.g. tap-to-pay, which would not be feasible without the last two decades of hardware improvements. And indeed in places where small businesses are not able to take advantage of that, you see payment methods evolve to keep more of the transaction on mobile devices, such as e.g. QR code payments in India, which do not require the seller to have a working device or even power/Internet.]

So who knows?

Tim Fist: Yeah, and this idea with enchipification is great because it points out that manufacturing processes where you essentially put down one layer of material and then etch away that material and put another layer of material on top are very susceptible to Moore's law style of improvements. That's a type of manufacturing process that's very easy to make cheaper over time by borrowing techniques from the semiconductor industry. And it seems both batteries and solar benefit from these kinds of cost trends in the way that other technologies don't.

We actually plot out in the report, Compute in America, that I've been talking about (it's a three-parter), in the last part, we plot out costs over time for both batteries and storage. And both seem to just follow this pretty exponential cost reduction trend over time. 

So if you look at the cost per megawatt hour over time, it's pressing levelized cost of electricity, LCOE, a sort of niche term for most people outside the energy industry. But I think in the 1990s for batteries, combined with solar, this was in the thousands of dollars, and now it's more like in the tens of dollars. And there's no strong reasons to think that this won't keep reducing over time.

The upshot of this, if you take this back to comparing solar and batteries versus other kinds of technologies behind the meter, is that if you're building a new natural gas plant now that you want to be powering a co-located load like a data center, the payoff period for that natural gas plant is something like up to 20 years overall. And so you want to have a fair amount of confidence that the electrons that you're generating from that plant are also going to be cost competitive in 20 years time versus alternatives like solar and battery. Otherwise you're incurring this huge, what's known as stranded asset risk.

So if you are an energy developer and you're making this bet that you're going to install all this new behind the meter natural gas capacity around the country, use it to power co-located data centers and then later connect it to the grid so you can continue making money off it, you want to be fairly confident that people are going to actually want to buy the electricity that you're generating from that, even in the case where it might be a lot more expensive than alternative sources.

Patrick McKenzie: I'm glad you brought up stranded assets. As you mentioned earlier, one of the games in power generation the last couple of years has been identifying assets which were stranded, typically natural gas plants or similar that had been powering a co-located industrial use where the industry has moved on, but the physical plant is still there.

For a few years in crypto land, the play was to co-locate a Bitcoin miner directly next to a stranded asset, and get the power for almost free. And since the relative mix of depreciation and energy in Bitcoin mining skews much more heavily to energy than it does in AI, that was a great trade for people while they could make it. But there are finite amounts of stranded assets. We didn't totally denude the United States of manufacturing capacity in the last 40 years, despite people's best impression to the contrary. Most of the low-hanging fruit of the stranded assets has been tapped.

The priority at the moment is avoiding creating stranded assets in the present and projecting that into the future. I think this is complicated by the dynamics of negotiating power purchase agreements (PPAs) with the megascalers and others. Essentially, a thing that the largest cloud providers have been doing the last couple of years is saying, well, we're willing to ink a power purchase agreement with you at some level which is above the typical cost for grid-based electricity. Why is this? One, it's a strategic priority. Two, we feel there's some sort of a communications/marketing tailwind to greening our energy use. And we run enormously profitable businesses and are willing to spend some of our rent on that. And three, they feel like they're in a bit of a foot race right now, where the next three to five years might be the most economically significant ones in history. And so if the price of power is up over the course of next three to five years, well, they'll pay it.

But if you are an energy developer and you assume, contrary to some people in the AI space, that humans are very likely to be around in 20 years - sorry, that is a joke and not a joke at the same time, it's a weird world we live in - if you assume humans are very likely to be around in 20 years and you still want to pay off the new natural gas plant in 20 years, it will not have an advantage of being greener than the rest of the grid. In fact, the opposite is likely to be true.

There is going to be less time pressure. It will be steady state load and based on the overall health of the United States economy and whatever changes in household and industrial energy use over the course of next 20 years. And there won't be huge amounts of time pressure. In fact, there will be a bit of reverse time pressure because 100 energy developers just like you are connecting in the same six month window. And so that does complicate the underwriting quite a bit.

Power purchase agreements and financing

Tim Fist: Yeah, exactly right. The other downside that we've been pointing out to PPAs, power purchase agreements, is that they are well suited for technologies which have known technical risks and cost characteristics. So if I'm trying to sign a PPA for solar or natural gas, I have a pretty good sense of this plant will come online within this amount of time and the electricity will probably cost this much.

But the problem is that as you say, a lot of hyperscalers are wanting to move towards clean firm sources of energy. So things like small modular reactors or geothermal or even fusion. We've seen some companies actually sign PPAs for fusion, which is kind of insane. And for these technologies, there's just a huge amount of uncertainty around, will this actually cost the amount of money that we think it will? And how long will it actually take to come online? And is the person building this going to incur a lot more cost than they would expect trying to build it in the first place and scale it.

So it's nice to have a piece of paper that says, once your power plant has come online and it's generating electricity, I'm going to get this much money off it. But this kind of financial instrument, the PPA, doesn't actually help you scale that up in the first place and incur the technical risk of building this new kind of generation. And so a lot of our policy ideas focus on, if we want to build this with next generation technologies like geothermal and small modular reactors, how can you design a financing regime where you can actually help developers scale that quickly and take on that early technical uncertainty? PPAs can be part of that, but they can't be the only solution because they're not actually helping build stuff a lot faster.

Patrick McKenzie: Mind if I geek out on this a little bit because it's been a focus of my work for last two years on again, off again. I do very part-time work with the geothermal nonprofit.

Tim Fist: Please, please.

Patrick McKenzie: So every project in capitalism has a capital stack associated with it. Capital stacks are varied and different participants in the capital stack have different risk levels and different currencies they want to be rewarded in.

The fundamental problem with geothermal is that there is a tech company which is attempting to create some valuable IP in the world. And perhaps that tech company can raise large rounds from venture capitalists. And the venture capitalists are hoping to get a monopoly on the technology that wins the geothermal race and then take some percentage of the total amount of energy generation over the course of the next 60 years. And that will be a nice payday.

However, to prove that technology out in the physical universe, which is unfortunately the only universe that actually matters at the moment, you have to dig holes. [Patrick notes: Substantial work has gone into producing “digital twins”, but in silico simulation is useful only to the extent it successfully predicts ability to impact the physical universe.]

And the wonderful thing about the United States is we're very good about digging holes due to the oil and gas industry. You call up Halliburton and say, hey, I need a hole. Here's the schematics and the location. Get it done, please. And Halliburton says, absolutely. It will be $20 to $60 million, but that hole, we will drill it. [Patrick notes: Please accept some imprecision here; the actual engineering work being done is quite sophisticated, both above and below the surface, and there are dozens of vendors involved. See the conversation with Austin for more details. But it is really important to understand that the work does involve very deep holes that either a) work or b) are useless very deep holes.]

And the VC rounds run out at a pretty high clip when you are doing $20 to $60 million per hole drilled. And you understand that you are taking an idiosyncratic project risk for each set of pads that you drill beneath. Because perhaps the geography was not quite as amenable to power generation as you expected it to be. Perhaps there was operating or execution error, trying new things and making experiments in the subsurface engineering and the engineering that you do on the top of the hole with dropping down a plant and turbines and similar. And you expect some of those projects to not pan out.

And so the power purchasers want to purchase the power from only the projects that work. They kind of have a bit of an advantage over the financing teams of the project level because the project level gets paid only if their hole is something that can create heat first and then energy later in some net positive way.

Financing geothermal wells

And so we've been attempting to get some capital stacks together, some combination of concessionary capital government grants and loans and maybe creative financial engineering for good to make it sufficiently lucrative such that project financing will finance the digging of the next 20 to 100 holes. And then at some point after the engineering is solved and this is about as deterministic as say a fracking natural gas well is in the United States, the good news is we have, "unlimited" is a funny word but almost rounds to unlimited capacity to finance and actually drill the things. The existing oil and gas industry could plop down 10,000 geothermal wells a year. And some combination of banks and hedge funds would very easily write those checks if the math to writing the checks was as straightforward as it is for a well in a proven geography.

So that is the brief history of two years of discussions of geothermal stuff. And if anyone wants to write a $10 million or $100 million check into drilling wells, hit me up. I know people who could use the money.

The promise of geothermal energy

Tim Fist: Yeah, indeed. I must confess before I started working on this topic, I wasn't really clocking the true promise of geothermal for powering massive point loads using clean firm power and kind of now see it as potentially the most promising, most scalable option across the next five years for bringing new clean firm sources of power in line.

And this is partly due to the Earth's crust having just abundant geothermal energy that's easily enough within the United States to power all our energy needs if we're willing to drill enough holes. But also you can reuse techniques and workforce from the oil and gas fracking industry, which is excellent. So compared to building a bunch of new nuclear plants online where you really need to scale up supply chains that we don't really currently have at the moment, with geothermal, you can just reuse a lot of the existing supply chains and workforce.

Patrick McKenzie: A thing that people might not appreciate about geothermal is we're discussing next generation geothermal, and there's a few physical models to do it under, but people are most familiar with the low-hanging fruit of geothermal in places like Iceland, where the hot water literally bubbles up to the surface of the earth without you doing anything for it. You just skim it off the top or drill a very small amount down. [Patrick notes: This is similar to the history of oil drilling. When we discovered petrochemicals, the first exploited ones were visible on the surface of the earth, and then we drilled hundreds of feet down, and eventually we teched up to fracking and offshore platforms in oceans which have waves taller than early drilling projects.] 

If you're willing to go down six to 10 kilometers, which is very doable with existing fracking techniques, you can get different temperatures of usable hot water essentially everywhere in the continental United States. So when you look at the maps, it feels almost a little bit unfair in that all of the good green energy sources have been very time locked and location locked for a while. And geothermal is just like, you get a geothermal, you get a geothermal, everybody gets a geothermal.

[Patrick notes: Speaking of unfair: game designers have had, since SimCity, a challenge in balancing various power generation methodologies against each other. Frequently that was “coal/gas pollutes, solar/wind produces expensive power intermittently, nuclear produces expensive power and has meltdown risk.” Those are fun, meaningful tradeoffs. But if you update in light of the solar cost curve the last twenty years, the fun gets sucked out of the tradeoffs, because Just Spam Solar Everywhere dominates all other strategies. This is not merely an academic concern: Factorio 2.0 and the Space Exploration mod both made design decisions to nerf solar to keep power management fun. It was just too good.]

Challenges in geothermal adoption

Patrick continues: That's not unalloyed good news for a very funny reason.

Oil resources are distributed in a very lumpy fashion under the ground. Some farmers will wake up one day having a lot of oil under the ground and they will be very happy. Most farmers will not wake up to have a lot of oil. The farmers who wake up to have a lot of oil or natural gas under their property become very wealthy and have an immediate incentive to make sure that someone actually extracts energy from there so that they can start receiving royalty checks.

Every farmer in the United States, to a first approximation, has geothermal resources under their land, which is wonderful, but it means that you will not see an increase in the value of your land as a result of having geothermal, because you have almost an infinite set of competitors. So you are far less incentivized to do the work and accept the hit in your natural business of running a farm. You’d need to take your eye off the ball of that to talk to these new geothermal startups and ask them to please drill some holes in your backyard. You will be dealing with the hole for a while, but who knows if anything useful comes out of it.

Industrial applications of geothermal heat

And then the other interesting thing is informed people in geothermal are also working towards the end of the decade timelines for usable power. But we have to drill some wells. Some of the wells are not going to work. The ones that do work will get hot steam at the top of the hole, or hot water, but not probably hot enough to generate large-scale electricity generation. And so for the first few years of the industry, you're trying to find customers who are using industrial applications of heat. 

Example: milk pasteurization, that sort of thing. [Patrick notes: Low-to-medium grade heat is used in huge amounts in many industrial processes. Paper manufacture, industrial food preparation (prior to baking in some cases), various chemical manufacturing processes, etc. The very vivid examples people have of manufacturing processes like e.g. steel smelting are not coextensive with manufacturing; it would be extremely difficult to do smelting with geothermal energy before we get to the point where geothermal can be used for electricity effectively.] 

A lot of the total energy that we burn or electricity that we use is not used for electricity per se, it's to convert the electricity or the heat energy in a burnable substance into BTUs to generate heat in industrial processes. And so there's something of a game happening right now where people are trying to find plants that are amenable to, okay, you have a perfectly good way to generate hot water right now.

It's that boiler over there that burns natural gas, and that works fine. Can we convince you to re-engineer the plant to work instead with hot water that you get out from underground? We're pretty sure we can generate it, but not positive. And getting people on the gap between “pretty sure” and “positive” is a bit of a challenge for the industry at the moment. [Patrick notes: Note that technical risk here could introduce a bottleneck into a factory at a point it is not designed for. For example, if the heat is ~95% available instead of 99.98% available, you have to redo a lot of the math of downstream processes. Some of that might be adjustable simply with operating changes, but some might imply physically rebuilding the downstream processes, which is something industrial users will be disinclined to go for.]  

But like all things, you can solve challenges and solve risk with sufficient allocation of money. And so we shall see if that happens.

Tim Fist: Do you know how big that demand profile is for customers who would like hot water but not necessarily electricity that you can kind of use as a mechanism to bootstrap the industry or have a backup plan?

Patrick McKenzie: So the good news, bad news. The bad news is the relative amount of greenfield industrial development, which needs industrial heat, is relatively small. The good news is the industry intends to use this as a stepping stone rather than a long-term source of profits or capital or similar. And so when they're thinking, we really care about the next 20 to 100 wells, and then after that the game goes infinite, then you don't need all that many dairy pasteurizers across the United States to count to 20.

However, the number of dairy pasteurizers that we expect this nation to construct in 2026 is far less than the number of dairy pasteurizers that will work in 2026 because they're long-lived assets and the long-term shape of the graph for milk and dairy products in the United States looks like what it looks like. So it's a one step forward, one step back sort of experience in looking into those things. And again, I've spent dozens of hours on this project in the service of people who spent thousands of hours on it. So I apologize in advance for the oversimplification for media use.

Tim Fist: If you'll permit me to nerd out a bit further on the geothermal solution, we did some analysis looking at what happens if you apply historical rates of oil and gas well drilling from the fracking industry to geothermal to see what is the total amount of power generation that you could support with productive wells, just looking at what we've been able to drill historically. Using this assumption that you have an enhanced geothermal well of the kind that you talk about where you essentially have three holes in the ground, one of which is pumping cold water down and two which are sucking hot water back up.

If before I was talking about one gigawatt within a couple of years, five gigawatts by 2030 in terms of a single cluster for training an AI model, but actually we want a whole ecosystem of these things. And if you scale up how much AI power we actually need globally, you can look at production rates from TSMC and say, okay, how many AI chips is TSMC going to be able to produce by the end of the decade? Our estimate is around 130 gigawatts worth of power consumption from AI chips. That sort of could come online by then.

And if you look at the amount of geothermal production you could support just in the United States based on historical rates of drilling across the five top shale plays across the US, it actually exceeds that. So you can get hundreds of gigawatts of energy just from historical drilling rates in the United States alone. Of course, this assumes that all of those wells are productive and not all of them are going to be productive overall. But the potential scalability here is insane. And we've proven that we can dig this many holes in the ground already.

Patrick McKenzie: Supply chains are fractal in nature. Oone of things I love about this podcast is getting to look at new ones and ask the questions of experts on what are the bottlenecks in your industry. My finger to the wind understanding from talking to the fracking people is that they are bottlenecked on the number of drilling rigs that exist in the world rather than say the number of people who have 20 years of experience in fracking.

Nuclear engineering is really rough because if you are bottlenecked on engineers with 20 years of experience, the right time to start educating them was 24 years ago. And given that you didn't do that, then you are sort of stuck.

[Patrick notes: In the wake of the 2011 Tohoku earthquake, there was an outpouring of support for Japan from many places in the world. I ended up writing a short NYT editorial about it, essentially arguing that most humanitarian forms of aid would be superfluous (Japan is perfectly capable of producing hot meals and clothing people, after immediate bubbles in the supply chain were fixed) and that if the U.S. wanted to do anything it should send brains in planes, specifically, nuclear engineers from General Electric.

This occasioned some commentary from people who were not Japanese systems engineers who thought this advice implied that Japan did not have a nuclear engineering industry. sigh 

I sympathize with Mr. Beast, and isn’t that a sentence, when he noted (pg 10) that bottlenecks seem too basic to have to explain to a working adult and yet are absolutely beyond the comprehension of many people who consider themselves functioning professionals.] 

Patrick continues: But if the bottleneck is rigs, then there's a supply chain for rigs. And there's a cost associated with that, there are scaling issues associated with it. But essentially it's like a very large, very complicated industrialized automobile. If we decide we want a fleet that is twice as large as the existing fleet, how many years does it take to get us there? That answer is probably not 30. It's probably closer to five. It is piled upon good news when you look at certain aspects of the geothermal issue and then with great uncertainties as to just a couple of things about it.

Where is the money coming from for the next 100 wells? Then one of the other ones is, I mentioned earlier, the tyrannies of the physical universe. We don't yet know what the curve looks like for productive geothermal wells that are doing electricity generation. We expect there to be a peak of electricity generation immediately after drilling the well, and then it has a decay function associated with it. And there are various mathematical models for this.

But it's all fingers to the wind of like, OK, we've done experiments in lab. We've measured the heat of rocks at various depths via core samples, et cetera. Here's the model. And the model is very sensitive to various parameters that are set by the physical reality that we live in. And if those parameters are set in less attractive portions of parameter space, then this is just a non-starter at anything like the current interest rates, configuration of the economy, and demand for electricity. And if it's in other places of the potential space of parameters and physical reality, then we luck out massively. And we just don't know yet which reality we are in because we need to dig more holes and put more turbines on top of them and see what happens.

Tim Fist: Yeah, we looked at some data from the Department of Energy, the National Renewable Energy Lab, NREL, who plot out potential cost curves for geothermal, enhanced geothermal specifically over the next 20 years. And I think they have three scenarios, conservative, moderate, advanced, and in the moderate, their base case, you reach cost parity with natural gas around 2032. But there's a huge amount of uncertainty around these figures, obviously.

Patrick McKenzie: And without throwing NREL under the bus, one of the important things to retain from the solar experience is that there's a beautiful graph of various credential experts making their best guess as to what the solar curve looks like. And it was not merely me, the untrained college graduate in 2004, who did not successfully predict the next 20 years of technological developments. It almost looks like the credentialed experts - I don't know if "refuse to learn" is not quite the best way to phrase it, but they would get two to three years of additional data, which conclusively disproved the existing analysis that they did. And rather than doing a new analysis, they would say, okay, I update the curve to include the two to three years of observed actuals and then copy paste my analysis from earlier. The curve is going to bend in the unfavorable direction any day now for magical reasons - and we had 20 years of that.

Tim Fist: Yeah, it feels like there's a bunch of government experts both at NREL and elsewhere. The IAEA is another one who's just copped some massive forecasting fails from failing to see what's happening with exponential cost reductions of solar.

Geothermal energy and national security

The other thing I will mention about geothermal that relates back to our core policy proposal that we're making is that most proven geothermal resources are actually in the western half of the United States and these have a huge amount of overlap with federal lands, particularly lands that are owned by the Bureau of Land Management, BLM, that sits under the Department of the Interior. And this allows you to then lease out federal lands for geothermal resources and have kind of an industrial policy around energy where you can create categorical exclusions to NEPA and other burdensome environmental permitting rules on federal land and allow that to happen really quickly.

We talk about this, we introduce this concept that we're working on called special compute zones where the federal government designates regions of the country where you make it really easy to build new AI facilities and associated energy resources in places like this where you can build really quickly on federal lands.

Patrick McKenzie: There's an additional angle which some people have been exploring for a while. One of the ways to get around political economy concerns in DC is to say whatever the good thing that I want for long-term reasons has a short-term positive impact for national security. And so there are airfields and military bases across the United States. And in a situation where the nation was in a true emergency and the power grid was less than reliable, it would be very important that the lights were still on at those airfields.

So I think that the Department of Defense is actively looking into, okay, can we get co-located non-grid dependent sources of base load energy available? Given the political realities of America, when the Department of Defense says, for the purpose of national security, we really need to dig a hole right now, then people across the aisles in the United States of America will say, okay, regardless of generalized impressions with regards to how much level of ceremony we want for doing transmission lines, if you need a hole, you get a hole.

And so again, we're counting to literally the next 100 holes. And if it is possible to get the DoD to sign up for 40 of them, that would be great. But I'm not involved in the day-to-day on discussions over there for all the obvious reasons.

Tim Fist: Yeah, and I think there just is a genuine case from an AI perspective, making the national security case for why it's really important to build this stuff here as opposed to overseas. You can just look at potential near term applications in military autonomy or surveillance or sensing from AI and make a straightforward case of the kinds of systems that the Department of Defense wants to adapt. Are these systems that are being built by industry, these large foundation models that can be applied to a wide range of tasks?

And if that's going to be critical to the defense base over the next few years, that then allows you to unlock a bunch of particular national security exemptions to permitting, but also use the Defense Production Act, which allows you to really resolve a bunch of supply chain issues and solve some of the regulatory barriers to, especially building on federal land, but basically anywhere in the country that you might encounter. And so I think there's just a really clear case for applying these exemptions in this industry in particular.

And then, I personally think that AI will be a really big deal from a national security perspective. But even if you don't, you might support this line of argument because you might think, it's really great for us to bring on huge new sources of clean, firm energy because then that allows us to bootstrap all the other things that we want in our economy, like the mass electrification of manufacturing and transport as well. And so this is a great way to overcome a lot of the extremely stupid regulatory burdens that we've imposed to actually building new clean energy in the United States and get a better long-term outcome for everyone.

Similar to in your podcast with Azim, you talked about that hype cycle around the internet and that leading to a big overbuild of copper and fiber lines, which ended up being this really important infrastructure that was very useful later on, even if the initial cycle was indeed hype around a lot of the startups that were in that dot-com boom.

Patrick McKenzie: Yeah, Byrne and Tobias have a book, Boom, that discusses this thesis in quite a bit of detail.

I personally, fingers to the wind, think is unlikely to be hype, but if the only thing we got out of it was effectively infinite clean energy, shucks. Yeah, we spent a trillion dollars and only got infinite energy for the future.

Tim Fist: Yeah, damn.

Patrick McKenzie: The downside case for the nation is pretty taken care of. [Patrick notes: The downside case to capital allocators in AI labs and startups directly would be dire, but oh well. They’re grown adults, investing risk capital that has been entrusted to them after due deliberation. And, uh, they’re very probably sitting on some of the best investments in the history of capitalism.] 

I think people who have been sort of reading the right papers and plugged into the policy space might understand it, but might not be understood yet by the typical, say, reader of the New York Times.

We discussed having compute in America, the proposed alternative in the world for the compute in America has been in a few geographies. Lay out quickly what the other potential futures look like.

Global investments in AI and energy infrastructure

Tim Fist: Yeah, so we currently see this wave of investment globally into AI data centers and associated energy infrastructure. This is often led by US companies like Microsoft and BlackRock, but we also see large investments coming from foreign companies. So MGX, is this Emirati investment firm, is really key. I think the UAE in general is extremely interesting here. Like they only have I think 1 million citizens as the population of the UAE, with a large immigrant workforce, but only 1 million citizens. But they actually have the world's second largest sovereign wealth fund that's around $2 trillion. So second only to China. And they're now just making a bunch of extremely aggressive bets on AI infrastructure globally with this, not just in the United States, it's everywhere in the world.

There's also Scalar, this Brazilian data center company who's making huge investments in Brazil as well. And on the China side, we've seen a bunch of investment happening there too. So in 2023, they began working on this national computing network, have been making massive subsidies for data centers and chip production domestically, kind of on the order of a chips act every single year. Early last year, I think there was an investment of 150 billion and then 150 billion more recently.

And obviously, if we talk about China in particular, they've over the last 20 years been able to bring on new energy generation way faster than the United States has, so 20 times more than the United States since 2000. And this is particularly true for clean, firm energy and nuclear. Over a period of around five years, 2014 to 2019, they're able to bring 25 gigawatts of new nuclear energy online. So that's like a new gigawatt coming online every few years.

And if you look at sort of recent history in the United States, the last large scale nuclear energy plant we're able to bring online was Plant Vogtle, which I think took 14 years overall, and then was multiple tens of billions of dollars over budget. Our ability to build equivalent infrastructure here is comparatively terrible.

So if you look at places which are going to be more likely to build this stuff over the next few years, China is obviously number one. But then actually, the UAE themselves as well is doing a really good job. I was actually over in the United Arab Emirates recently and did some tours of data center and electricity infrastructure, went to a new few tens of megawatts data center there that was being powered 25% by a nuclear plant that they've just brought online. And they're actually able to bring that online in a new reactor in just four years. So this was their first time doing this. And they're able to do it many times faster than we have been able to despite having a nuclear industry for many decades.

So I think if you look at historical base rates, our ability to bring online energy infrastructure faster than these other countries seems pretty suspect.

Patrick McKenzie: Yep. And geopolitical realities being what they are, we would probably prefer for there to be parity or better with respect to China. And to the extent that American industry is downstream of inputs from the Middle East, that has worked out in a complicated fashion over the last couple of decades and might not be up for a full rerun of those decades of history.

Tim Fist: Yeah, in general, I'd say that the overall thing to know that's driving this is there's kind of a zero sum game around who gets access to high end chips. TSMC is producing the vast majority of the world's high end chips, like five nanometer or better, it's sort of like 85% plus. And for every single AI chip that they're producing, there's far more demand for that AI chip than they're able to produce. So for every single chip they're putting out, you have like five different customers, all who want that.

And so where you put that in the world, there's like a finite number of chips, everyone's vying for where they're actually going to go. And there's an opportunity for the US to, of their 130 gigawatts worth of chips that are going to be produced over the next few years, you ideally want to capture as much of that as you can to get both the economic benefits of this, but also the national security benefits. If you think that you're using these chips to build AI models that could reshape the balance of global economic and military power, ideally you want to be building this stuff within your own borders.

Patrick McKenzie: Yep. So circling back to another facet of the discussion, the cost curves for solar and the batteries are falling precipitously due to the enchipification, as you mentioned, but also because they're sort of intrinsically very large N. We denominate a solar field in square meters of wafers and want to pump out millions upon millions of square meters. 

If you bet heavily on nuclear, we're betting heavily on a bespoke process where we don't get the returns to scale. And just the cycle time through nuclear being measured in decades means that the things that we learn from building a plant that starts its life this year will not be able to impact the cost curve until the mid-20th century, if that.

Tim Fist: Yeah. And the solution that a lot of people talk about is small modular reactors. This is promising because you help solve issues by having modular components. You can do prefabrication in a factory and then send them out to where you're actually installing them. It's more scalable essentially than traditional large scale nuclear plants. But I think a couple of issues here are currently Russia and China are the only countries who have been actually able to bring small modular reactors online and even in those cases, they're just small test units, they're not highly scaled supply chains.

And then you also run into the supply chain scaling issues that you have more broadly with nuclear, so you don't have really a large nuclear workforce that can be easily scaled up in the United States. And then supply chains for things like fuels that are used in nuclear would need significant scaling as well, which is going to take time. Definitely if you're thinking about this as a promising solution over the next five years, I suggest it should be towards the end of your list of promising new clean firm technologies to deploy quickly.

Patrick McKenzie: Yep. And just as a sketch of it, the small modular nuclear reactor is there are variety of companies that are taking a variety of whacks at it, but essentially what if we could make a nuclear reactor self-contained about the size of a shipping container and put it anywhere? And what goes inside the shipping container? Magic. But it'll be magic that is very resistant to the failure modes of existing large scale nuclear reactors.

And I think it's partly a technical/cost innovation and partly regulatory innovation of just for frustrating reasons. Much of the first world has given up on nuclear as a promising technology. And if you re-badge it, maybe if we can get them to un-give up on it, is about 50% of the sales pitch to my understanding.

Tim Fist: Yeah, I think the one design that's been licensed in the US is from Oklo, the SMR provider, correct me if I'm wrong here, and they've signed deals for capacity with a couple of co-location providers, so the companies who sort of sell data center space for hyperscalers to install their chips in. So I think that's kind of promising. But again, signing deals for capacities is not the same as having those plants online and running. And there's huge technical risks involved in scaling this quickly and not losing a bunch of money.

Policy and technical expertise in AI

Patrick McKenzie: So switching gears for a moment, you previously have a machine learning background and hardware background yourself and now find yourself doing policy advocacy. This is a career transition that I've seen a bunch of people around me do over the last few years. And if one has a technology or tech industry background and wants to work in policy as a field or related subfields, what does that transition look like?

Tim Fist: Yeah, great question. My personal story is I was working in the AI industry for about five years and just was looking at scaling laws and trends in the industry and thinking, wow, this is going in a very interesting direction. And I really hope we get this right. Because emerging technologies exhibit this huge amount of path dependence in terms of who builds it and where. It's very important that you get this sort of thing right. I think this is especially true in AI.

I was initially thinking about doing this from a technical perspective, like what are the technical problems that we need to solve to ensure this technology is deployed well or deployed responsibly. And then around 2022, when I was working in the industry, I witnessed the first round of export controls on AI chips come out. So the US basically banned all chips above a certain set of performance thresholds from going to China and I was kind of blown away by how big a deal this was.

In my perspective, this is basically the most important thing that's ever happened in the governance of AI as a technology is this radical reshaping of the global distribution of compute. Like who has the capacity to both build and deploy this technology? And this was done from within the US Department of Commerce with a remarkable amount of foresight for what was coming over the next few years. And I was just thinking, wow, I feel like policy is really where the juice is at the moment in terms of influencing the development of this technology in a positive direction.

So I got really interested in policy in 2022 and decided to make the transition and come to DC as I was previously in Colorado. And my experience was, it was actually fairly straightforward when you have a relevant technical background in an area that's related to AI or energy or other emerging technology that are very hot in DC policy circles at the moment. Just make the case for why your technical background is important for looking at these things.

I think there's a real lack of significant technical expertise in DC, such that if you have a medium level of expertise in one of the industries coming from a company, you can actually add a lot of value in conversations and analysis. And if you're someone who is good at talking to people, is good at pitching ideas and have a technical background, there's just a huge number of opportunities, both within government, there's a range of different fellowships that try and find people who have technical expertise and place them within executive branch agencies or congressional offices. But also think tanks where I've historically worked.

There's just a number of roles where you're doing policy research analysis, you're helping to advocate for different policy ideas to policymakers. There's still a huge gap of accomplished technical people who are willing to work in that environment. So really good opportunities in this space.

Patrick McKenzie: And while you mentioned communication skills verbally focused, I think the ability to write is massively undersupplied in the world, even with machines that are writing for the first time in human history. [Patrick notes: See previous guest Dave Kasten’s essay on this, appropriately titled the Essay Meta.]

And I think there is just a difficult to appreciate level of leverage that is created by being able to write something which is at the 80th percentile for say a post that would end up on Hacker News. I think that many people in our circles don't appreciate that that is where the bar is for a document that materially advances a policy conversation in Washington.

Tim Fist: Yeah, 100% agree with that. I think some of the most popular pieces in policy circles that have come out in the last few years have been people who have worked in the industry just translating fairly technical ideas that are sort of fairly commonplace in the AI industry for a more lay policy audience. That kind of translational writing is actually really undersupplied.

Patrick McKenzie: Yeah. Exporting ideas between communities that have them and ones that have not been exposed to exactly that idea in the local parlance has been a major theme of my career. And if you look at artifacts like the situational awareness paper, the recent AI 2027 paper, et cetera, if your thesis statement was just what would someone in government need to know and how would they get it without having lurked the last 20 years on Less Wrong, you could create an entire career just trying to solve that arbitrage. It will probably not close anytime soon.

The role of government in technological advancements

Tim Fist: Yeah, I agree. And I think furthermore, there's just a sense in which if we think about technical state capacity within the federal government, so the government's ability to actually see what's coming down the line from a technical perspective and technologies like AI and do proactive policymaking with things like export controls, basically the government entirely lacks this. There's no team of really accomplished technical AI experts within government who are thinking about, wow, we see reinforcement learning, test time compute, these other technical trends emerging in the AI industry. What does this actually mean for what policy should look like in a year?

There's actually just basically no one thinking about these things and doing proactive policy making. And it just means that the federal government is extremely reactive to what's happening overall. We see this at the moment with the DeepSeek moment earlier this year where this Chinese company, DeepSeek, came up with these two really powerful open source models in December and January, completely took policymakers by surprise and everyone's sort of freaking out.

And this was pretty observable ahead of time that these trends were happening and that this organization is putting out their research publicly online and releasing things open source that they had approximately reached this capability level. And there's things that you would want to do in response.

So at the moment we have this inference chip, the H20 produced by Nvidia. This is a chip that's specifically designed for the Chinese market to fall within current export control limitations of what chips you're allowed to export to China and which ones you're not. It is extremely good at doing inference workloads. It's actually about 20% better chip for chip than the H100, which is a banned chip.

[Patrick notes: Since we recorded this episode, the U.S. has woken up.]

As I'm sure you know, inference is becoming much more of a bigger deal as we have these paradigms like reinforcement learning, test time compute, synthetic data generation. These all rely on inference compute performance. But at the moment in the news, Nvidia is currently planning to sell about 1.3 million of these chips to a few large Chinese companies. And because export control parameters haven't kept up with where the industry is going, it's somewhat likely that this deal is going to go through and we're going to have many more DeepSeek moments in the future.

Patrick McKenzie: It is enormously frustrating that it is harder for us to write a few paragraphs defining the new version of export controls than it is to speak a new chip into the world.

State capacity: it is what it is.

Tim Fist: Indeed.

Patrick McKenzie: Not to be too pessimistic with respect to the United States because I think there are optimistic and pessimistic takeaways from the experience, but I think many people have a model where the United States government or military or similar has arbitrary burst capacity for doing extremely hard things, which is true to a point in certain configurations of hard things.

We're the most accomplished people in the world with regards to shipping a military significant amount of tonnage to any flat surface on the face of the earth with 48 hours of notice. But the Vaccinate CA experience for me where all the King's horses and all the King's men could not stand up a web app given a year of lead time. And indeed, there was no one in charge of noticing that they were incapable of standing up a web app with a year of lead time. It was enormously frustrating.

The good news is a group of people in civil society was able to do the thing the government couldn't in about 12 hours or so and then get it productionized widely in course of a couple of weeks. So keep an eye on the news and don't assume that just because it seems outrageously improbable that no one is minding the store that there is in fact someone minding the store.

Tim Fist: Yeah, to take a sort of more optimistic lens on this, another thing that happened during the COVID pandemic is an incredible example of state capacity, which is Operation Warp Speed, where the traditional timeline for developing a new vaccine was over 10 years, but the US government proved if it pulled out all the stops, put top people on the problem, used a range of creative mechanisms like advanced market commitments, they brought that time down to just nine months to produce a new vaccine, which was kind of insane. I think this is a huge win and hopefully something that we can do more of in the AI industry as well.

Patrick McKenzie: I think that one might be one of the best single triumphs in the history of Science. We are certainly more capable than we allow ourselves to be on any given Tuesday. Closing that gap is one of the meta problems for our society for the next couple of years. 

That might be the thesis statement for this podcast.]

Tim, where can people follow you online?

Tim Fist: Yeah, so please follow my work at the Institution for Progress (ifp.org). You can also follow me on Twitter. My handle is @TimFiiiiiist.

Patrick McKenzie: For the rest of you folks, thanks much for your time and attention, and we'll be back next week on Complex Systems.

Tim Fist: Thanks very much.