Power plays: grid economics and engineering, with Travis Dauwalter

This week I'm joined by Travis Dauwalter. We chatted about electricity grids, across a range of abstractions, like: How the frequency of current on a wire showed operators relative demand/supply conditions before we had computer networks to transfer that information. Why building power transmission infrastructure requires thirty year bets on rates of family formation. What the AI boom means for data center buildouts, and what that implies for energy generation.

This pairs well with previous episodes with Casey Handmer (on solar economics) and Austin Vernon (on fracking).

Sponsors

Check is the leading payroll infrastructure provider and pioneer of embedded payroll. Check makes it easy for any SaaS platform to build a payroll business, and already powers 60+ popular platforms. Head to checkhq.com/complex and tell them patio11 sent you.

Timestamps

(00:00) Intro
(00:28) Understanding the energy grid
(02:22) The complexity of supply and demand
(05:43) Regional differences in energy grids
(08:16) Seasonal and daily energy demand
(11:22) Renewable energy and storage solutions
(17:17) Sponsor: Check
(18:29) Renewable energy and storage solutions (continued)
(24:07) Demand response and time of use rates
(34:08) Bitcoin mining and energy economics
(39:29) Exploring behind the meter arrangements
(40:44) Transmission line challenges and innovations
(45:51) Dynamic line rating and grid efficiency
(50:58) Data centers and energy demand
(58:33) Interconnection queue and grid security
(01:03:38) Understanding the US grid structure
(01:09:46) Wrap

Transcript

Patrick: Hi everybody, and welcome to Complex Systems. I'm here with my buddy, Travis Dauwalter.

Travis: Great to be here. Thank you very much for having me on, Patrick. I'm really excited about the conversation ahead.

Patrick: Thanks very much. We're going to be talking about the energy grid as an institution and our engineering artifact today. You've been in energy for a very long time, I think.

Travis: Yeah, I've been thinking deeply about our electricity grid here in the US for the last 15 years or so. Every time I peel back another layer, I'm shocked at how complicated it indeed is. So it feels like a fitting topic for this podcast.

Understanding the energy grid

Patrick: I feel like I'm something of a talented amateur on this subject—perhaps less on the talent but more on the amateur—but it astounds me that we managed to put it together prior to one, the physical instantiation of any networks to coordinate the installation of the grid. 

[Patrick notes: One of these days I will achieve victory over the childhood impulse to self-minimize, but today is not that day. If you do not already have this impulse, please do not adopt it from me. Many people of good will will be convinced that, if you are not your best advocate, you have an informational advantage over them and therefore your minimization is the ceiling for your actual ability/etc, and of course you may be overstating your abilities, since people do that all the time. Many people of ill will will say that you claimed to be the thing you expressly disclaimed.]

But two, the sort of thinking about formalized thinking of graph theory and networks and similar, which would seem to me to be necessary to string the wires together and have all the math balance. But apparently prior generations were willing to tape things together, deal with the fires, and then eventually get us into some metastable state.

Travis: Yeah, that's kind of how it flowed as well. It got assembled off of choices made here and there. Some of them were stickier than others. We can dig into all of that, but what you're saying really resonates with me in terms of just the sheer physical complexity of the grid.

If I'm looking at a light switch in this room that I'm in right now, and if I were to walk over and flick it on and off, I would know with certainty that light in this room is going to toggle on and off, and there's a wire going from this light switch all the way back to some generating plant, maybe hundreds of miles away. All of our devices, anything plugged in is in some way connected topologically to this grid. It's really amazing.

The complexity of supply and demand

Patrick: The thing that blows my mind even more than the geographical implications and that there physically must be a path hundreds of miles long between you and some physical, chemical, electromagnetic generation apparatus—there's also something that feels a little bit like calculus and a little bit like physics going on. The instantaneous draw on the grid must exactly match the instantaneous production of the grid. As able to, as the universe is capable of perceiving zero difference between those two things, where the actual physical processes involved are not such that one can scale them arbitrarily quickly up or down.

So there is some amount of slack in the system, and implicitly, there is some sort of demand and supply matching apparatus going on, which does possibly humanity's most efficient job of instantaneous demand and supply matching. 

What is that apparatus?

Travis: Yeah, I totally agree with you. It's remarkable that you can turn on that switch with certainty, and that there is this supply and demand match that's happening, not just once a day checking in to make sure we're matching up, but at the second-by-second level.

What ends up happening is, with all the puts and takes of people turning load on in the grid and then taking load off of the grid, it induces incremental change in the frequency of the grid in that local topology. 

[Patrick notes: You can round these jargon words to: “put” is when you produce electricity and add it to the grid, “take” is when you remove electricity from the grid to consume (or store). I would ordinarily have prompted Travis for this, but he was on a roll, and I didn’t want to interrupt.] 

The generators can observe that minor dip in frequency or minor increase in frequency, and they're effectively responding to that. The tolerance band that you're talking about is very narrow but centered on 60 hertz, which is what our grid operates on.

If it starts getting fast, if the frequency starts getting a little fast, that means that we've got an abundant supply compared to demand. That's a signal to generators to start turning things down. And if we're seeing the frequency droop a little, then the supply needs to step in and start injecting a bit more power into the grid

 I totally get what you're saying. It does feel impossible to orchestrate the millions of different users that are on this grid in that way, but there's this really tidy data point that you can observe happening continuously and you're keeping it within that tolerance band around 60 hertz.

Patrick: I suppose this is one of the original operations research problems where in the classical model of operations research modeling a factory, you'd be able to perceive changes in demand in the market in the flow of orders at the factory or changing stock issues at various workstations. And here you are literally using an oscilloscope on the wire and attempting to measure the frequency of the electromagnetic signal.

Regional Differences in Energy Grid

Travis: Yeah, and the alternating current that we operate most of our grid on is ready equipped for that sort of observation so that we can make those fine adjustments. There's also, depending on the market—and we can dig into this a bit more about how regionally we've made different choices across the United States on how to structure those different grids—sometimes generating units that are just sitting there and providing frequency response services for the grid. Their whole job is to just stare very intently at that oscilloscope and make sure that we're doing a good job of keeping within those bands.

Those different policy decisions that were made by the different organizations that oversee our grid have been really effective in keeping the grid as stable as it is. Think about the last time you had a grid outage. Depending on location, but for the most part, across most of America, it's really rare for you to have a grid outage. And if you do, it's often just a flicker. And then the grid is finding its way to its healthy 60 Hertz frequency.

Part of that is the utilities have a lot of arrows in their quiver to help mitigate this supply demand mismatch and ensure that we keep within tolerances. They'll just start running their playbook—first try this, if still having problems then try that—and they just move through the cascade of activities. It will culminate in rolling blackouts if none of them work, but there's a lot of opportunities before you get to that point to cure the problem. They take full advantage of it as they should. We've come to really expect power to just be available whenever we want it. It's almost a feature of living in the modern era, and we have very little patience for when the grid goes down.

Patrick: Yeah, I think the pervasive availability of power along with bandwidth (increasingly), water, et cetera—they're an infrastructural substrate that the rest of society kind of leans on in a load-bearing fashion for doing all the other things that we do with our lives. And so, there's always a background process in my mind when I'm flicking something: wow, this is amazing. And amazing that the vast majority of users never have to think of how amazing it is or to plan their lives around it.

I understand that one of the things that both grid operators and power producers do is to plan around variations in the clock and variations in, I suppose, the annual cycle and the business cycle in terms of power demand. Do you want to sketch out what those curves look like for people who might not have looked at them before?

[Patrick notes: Many grid operators will make the graphs and forecasts available in real time through a web application these days. Tokyo’s is great, though you might fight through a bit of localization engineering to wonder why it quotes things in units of x10 MW. I’ll reproduce a similar graph from Illinois inline so you can see the flavor. Note the daily peak/trough cycle, substantial intra-day variation, and the fact that an alien carefully looking at this data might say “Does your culture do something strange on December 25th? Like, an annual commemorative LLM training run?”]

]

Seasonal and daily energy demand

Travis: Yeah, sure. Let's talk first about seasonality and then we can zoom in onto what's happening hour by hour over the course of a day. It depends a bit on your location—I live in Atlanta, Georgia, for instance, and it's really hot and humid in the summer. Generally speaking, Georgia Power has higher loads that they're facing in the summer and in the winter. We tend to have pretty mild winters here, with the exception of a pretty big snowstorm that came in a couple of weeks ago.

You might see something like a double peak in seasonality in places like the Northeast where there's high energy consumption for air conditioning loads, but then in the winter there's maybe not everyone has gas-fired furnaces, so there's higher electricity consumption in the winter as well as they're trying to heat their homes and their spaces. And you can contrast that with a place like San Diego, notorious for having really lovely weather throughout the year. From a seasonality standpoint, they're going to have less amplitude to that annual cycle.

As we zoom in a little bit to shorter timeframes over the course of a day, put yourself in the shoes of your own energy usage. At midnight, many of us are in bed and the house is kind of shut down and you're not using a lot of electricity. Then you wake up and maybe go down and turn on your coffee maker and throw something in the microwave or turn on your oven, turn up the heat or turn on the air conditioning.

So there's a peak in the morning where the energy consumption goes up as everyone is consuming, waking up and starting their day. There tends to be a leveling off, and that's largely because everyone is moving from their homes to their place of work. That place of work on an energy usage per person basis is a little bit more efficient. Think about a big commercial building with really efficient air conditioning units—the unit amount of energy you need to cool yourself when you're in a commercial building is lower than when you're trying to cool yourself in your home.

[Patrick notes: You’d also see this if society had graphs which broke rent/mortgage expenses down on a per minute level attributed to people based on what building they were physically in. You pay more rent per minute at home than your employer does on your behalf per minute you’re in an office, mostly due to physical and temporal density and utilization.

Mentioning this because your employer’s (or employer’s landlord’s) professional skill with optimization is mentioned below, and that skill is not necessary to achieve relative efficiency over your place of abode.]

The spike happens again as everyone's leaving work and returning to their homes and starting their evenings, turning on televisions and doing their cooking and reengaging with their HVAC system. Then it'll taper off as everyone goes to bed and the cycle repeats. That's sort of what it looks like on a day-to-day basis. And then that little waveform I just described kind of flows up or down as we move through the different seasons of the year.

Renewable energy and storage solutions

Patrick: So one of the things that fascinates me about the economics of this is that, for most grids, there is intentionally a mix of different power generation modalities to hit various points along that waveform. There's a great amount of physical capital in the world that is deployed at various points. At midnight, for example, solar generation—unless you're in a very rare location—you will not be getting sun. But our natural desire for offtake is lower at midnight.

Then solar comes on during the day and towards the peaks, there are higher marginal cost or higher marginal nuisance generators such as—I believe the word is peaker plants—where you might be burning liquid natural gas. From a societal perspective many people would assume, okay, if we could make it 100 percent renewable energy use at some acceptable cost level, we would like to do that. But at certain margins that is cost prohibitive, and so we can cover those margins—and maybe those margins alone—with the use of the peaker plant and then intentionally keep the peaker plant offline for 80 percent of the time, 95 percent of the time, potentially even in some grids perpetually offline except in the case of emergency, which provides that sort of system-wide slack for the extremely disruptive tough to predict events.

I lived in Japan for 20 years, and the only sustained interruption of power that I can remember from my time there through many thousands of earthquakes was the combination earthquake-tsunami event in 2011. 

That would have been a non-event [from a power perspective], but for a power plant being directly impacted by the tsunami. (And several power plants being directly impacted by the earthquake.) 

[Patrick notes: If you trust my recollection of contemporaneous reporting: every nuclear reactor on Honshu failed into quench within seconds, as designed. (And then, a few days later, the site physically struck by the tsunami… had a much worse experience, due to waste heat and other problems, which you can read about in your postmortem of choice.] As I wrote at the time, many engineers were dealing with complex emotions, simultaneously experiencing tragedy with the rest of the nation but also watching their lives’ work deliver for their friends, neighbors, and countrymen.])

A great system-wide wrench thrown in the works in amidst much more direct tragedy, but the infrastructure being down for longer than the period of a few days would also cause its own rolling wave of tragedies.

So, pip pip for grid operators.

Travis: I want to just echo the point that you made about renewables. Being at 100 percent renewable grid is challenging. It's more of a physical problem. The sun is up and available to provide us or to harvest energy for some fraction of the day between one third and one half of the day typically. And wind—you might think, well, let's combine it with wind—and wind tends to do well in the shoulders where the sun is setting or the sun is coming up and you're getting the temperature inversions. But there's still this tricky period where there's really no harvestable, renewable energy source.

So you would have to combine it with some sort of energy storage. What's been great about observing the grid in the last few years, maybe the last decade, is that energy storage is starting to become a more effective way to manage the grid. It's doing two things: First, it's backstopping renewables and removing some of this intermittency that you've been describing. Second, it's allowing us to drive a bit of a wedge between that mandatory supply-demand match that we have.

[Patrick notes: See generally my interview with Casey Handmer, which discusses this subject at some length. One of the opportunities is his startup, Terraform Industries, which intends on storing surplus solar energy as synthetic fossil fuel, to be transported and then combusted at a time/place where energy is more scarce than sunlight was at the time of carbon capture from the air.] 

Instead, we could have the demand be a little bit high and we're using batteries to satisfy that demand load and supply can be lower as a result. Or there could be periods of the day—think about that middle period of the day that I was describing where people have moved to their places of work and the energy consumption has dropped incrementally. That's a good time in places like California, where there's a lot of solar panels on the grid, to take and redirect all of that solar energy that we're harvesting: fulfill the load that is on the grid, but then take any excess and charge these storage units. Then you have energy in reserve that you can deploy later.

Travis: So that's been a technique that we've started to be able to use because chemical batteries and other storage solutions are becoming more financially tractable for these sorts of problems. And then the other point I wanted to respond to was this peaker plant issue that you described.

It's really prevalent in places that have heavy amounts of renewables because you have that sun setting right around the time—this is really prevalent in California, for instance—where the sun is setting and that's about the time that a lot of people are making that transition back to their homes. There's a spike that's coming, but the solar panels aren't there to provide any support against that peak demand, that little spike in demand. And that's where the peakers have to come in and really rapidly respond.

There's this graphic that was created by CAISO, the California ISO, which is the transmission moderator orchestrator in the state of California. CAISO released this amazing chart that tracks year over year what that big spike in late afternoon looks like. When you look at it year over year, it looks a lot like a duck when you're looking at the graph and they've come to dub it the duck curve. It's quite a great illustration of precisely the problem that intermittent renewables can create on the grid that we, when we were designing this grid to begin with, I don't know if they even conceived of the idea of having solar panels harvest energy and inject it onto the grid. And so now here we are, we have this technology, we're using it, but the grid maybe is not quite designed around that sort of scenario.

Patrick: And while we perceive electricity consumption as instantaneous because flip switch light goes on immediately at something slightly under, but imperceptible to our eyes, the speed of light—there's physical processes involved in generation. I saw a graph somewhere, and I will drop in the show notes if I can find it, of like physically, how many minutes does it take to go from "I have a nuclear reactor sitting here, it's currently cold." If we needed gigawatts more energy, making this not cold anymore is an option, but that is not a thing that I can quickly action in the next minute.

[Patrick notes: Per the U.S. Energy Information Administration:

]

How many hours does that take to spin up given that it is otherwise ready to spin up? And then how many minutes does it take to bring water to a boil in a peaker plant given the availability of staff and liquid natural gas on site? There's a similar sort of tiering system of things available. And the more battery capacity we have, the more we can buy some leeway for making decisions that are either costly or consequential or difficult to reverse with respect to usage.

Travis: We can let the batteries do that work or fill in that role. And you mentioned this idea of base load. So if you think about how those waveforms that we were describing earlier look throughout the year, there's a minimum load that will always need to be served in a particular grid. That's a great application for nuclear power plants, for coal power plants, for what we call combined cycle natural gas power plants. What those plants are really good at is running for a very long time, very consistently—they don't want to deviate from their run state and they're very efficient. They tend to be quite large, and so they're really costly to build. But once you get them in place, the marginal costs are quite low and they become an ideal solution for covering that base load.

Then you have these plants that are just—imagine again you're looking at this year-long waveform and the base load is covering up to those lower troughs. And then you have a type of plant called a mid-merit plant, which is meant to run somewhere between half and three quarters of the year. That plant is meant to cover a lot of the rest of that waveform. It's a slightly more flexible type of generating unit, has great economics and we want it to run for a long time. And then those peakers are the last ones in the stack and they're covering all those highest peaks in that wave.

Now, what renewable assets are doing is their marginal costs are effectively zero. So they're kind of the first to take, and states like California have created must-take policies where if there's energy being created by a solar panel or by a wind turbine, we have to make use of that first before we make use of anything else. And you can imagine as more and more solar is coming onto the grid in places like California, then the bottom of that kind of stack that I've just described is no longer the nuclear plant, which we don't want to deviate from its run condition. It's now this slug of renewables that is the must-take.

Then the nuclear plant is sitting on top of that. And if ever the waveform drops into the nuclear power plant's run condition, then the nuclear plant needs to turn down. That's hard on the equipment and it's tough to do. So we're also having to figure out what do you do with all this excess power that's generated by the renewables when this nuclear plant is essentially satisfying all of the base load. The issue we'd run into would mean turning down the nuke plant. So, instead, these batteries become a great solution, like just a dump repository for putting all of that energy until we can use it later.

Patrick: It's also fascinating to me that there are some novel uses of electric power which function as something akin to a synthetic or a natural battery. The classic example of this was aluminum smelting, where aluminum smelting is a relatively simple process in terms of inputs which uses a gargantuan amount of electricity. And so aluminum smelters end up located in the world in places which have sort of structurally low power costs—there's a lot in Iceland, for example.

[Patrick notes: Iceland is naturally blessed with geothermal energy which was conveniently tappable with technology available for the last several decades, and lacking sufficient population to make use of it all, exports much of that energy by turning it into industrial outputs that travel well. Aluminum is the chief example, at about 40% of all exports.

We discuss an easy-to-transport commodity that one cannot conveniently use to wrap soft drinks later in the interview.] 

Every user of electricity in the economy has the notional ability to—magic word for this conversation—curtail their usage, but most of us would find going over to the fridge and unplugging it to be pretty inconvenient.

But if we were professional users of electricity, like if we were the management of an aluminum smelter, we could reasonably make the decision from four o'clock to six o'clock today, just shut it all down. It turns out that there are economic arrangements that a grid operator could make with an aluminum smelter which turned the aluminum smelter from just a sort of passive user of electricity into something akin to its own power generation apparatus. Do you want to sketch out what some of those might look like?

Travis: So just to clarify, are you asking about the onsite generation or taking advantage of the inertia of a really hot molten piece of aluminum that maybe doesn't need heat entered into it?

Demand response and time of use rates

Patrick: I was thinking more along the lines of demand response. [Patrick notes: Broadly, this is a communication and/or economic system designed to incentivize certain grid users to reduce offtake in response to high demand, by making changes to their energy use in substantially real time.] 

Travis: Yeah, you're totally right. Demand response is definitely a technique that can be used there. There's probably two primary signals that industrial facility or commercial building or even your home could use to make demand response a feasible curtailment action. As you mentioned, everyone generally wants their refrigerators plugged in and all else equal, the aluminum smelting plant would much rather have its aluminum smelting production continue on.

But if the utility is willing to make payments through these demand response programs to basically ask the aluminum smelter to reduce their load, and those payments feel attractive to the aluminum plant, then they'll take them up on it. We have the option at the residential level for those who have different smart thermostats. There's often programs that you can sign up for that aggregate all of the neighborhoods together and can effectively act as a virtual demand response asset for the utility.

They'll collect a payment from the utility to curtail our air conditioning units in the peak hours of late afternoon. What ends up happening and how the money flows through that is the orchestrator, this tech company that can coordinate all of the houses and all the smart thermostats—they take a premium or a piece of the total payment that the utility provides and then they pass on savings to the users as well. So there needs to be something in it for us as the resident who doesn't want a really hot home, and we cure that through the economics of this kind of pass-through payment.

The other signal that a utility might be able to send if they don't want to have a demand response program or the aluminum company is not interested in it is some utilities have started to establish time-of-use rates where at different parts of the day the amount of money that they'll charge you for that electricity consumption varies. It tends to be the case that in the middle of the night when no one's using electricity and supply far exceeds the typical demand, the energy prices are quite low. Whereas during that peak period in late afternoon, that's when energy prices tend to be quite high.

So that's another signal that the aluminum smelter can get from the market to curtail their load. And there's been a weird artifact that's come out of that, layered in by all the electrification of vehicles. There's a lot of programs where you can basically avoid plugging in your electric vehicle at that time, or the vehicle itself can become aware or be made aware of those varying time-of-use rates. It will only charge when the time-of-use rate has gone down quite a bit.

It's all designed to save you money, but it's creating this really weird artifact in the grid where if the time-of-use rate trips from a really high rate to a really low rate at 7 PM, then all of the electric vehicles in that area will suddenly flood into the market and demand a lot of electricity to start charging their cars. You get this weird secondary spike because of all the electric vehicles that are beginning their charging process.

So this is something that a lot of utilities are having to start thinking about—do we do tapered time-of-use rates? Do we require that electric vehicles be a little bit less aggressive about charging their vehicles right when it trips over to the lower rate? What do we need to do to solve this problem? Because this is not at all what we had intended when we set these time-of-use rates over the course of a day.

Patrick: This is something where the systems engineering engineer is like, "Finally, I'm professionally relevant." We have so much prior art on this one. But it's an eminently solvable engineering problem.

[Patrick notes: Cron jobs—scheduled processes which run on computer systems—were, once upon a time, manually scheduled by humans. Humans often gravitate to culturally relevant times to schedule things: midnight, for example, or 5 minutes past the hour at a much higher rate than 7 minutes 32 seconds past the hour.

These days, at sophisticated software shops, there is automated tooling that makes it easy for engineers to express the intention “Don’t run this at the same time every other engineer at the company would pick to run this!” One technique is described as jitter; simply adding a short random offset to the start time of a job, to avoid many jobs colliding at exactly the same second.] 

Back when I was living in Japan, they had TEPCO, which is the grid operator and primary power generator in the Tokyo region, but there are a variety of companies which do some amount of generation and a large amount of wholesaling of TEPCO energy to people. The one which I used had a demand response program where TEPCO would use them as essentially a virtual peaker plant—make up a number—contractually promise that TEPCO can give you 24 hours of advance notice and you will provide us 50 megawatts between the hours of 6 PM and 7 PM tomorrow to cover air conditioners coming on as salarymen get home for the day.

[Patrick notes: Softbank Energy, if you’re curious. I’ve written more about this program at Bits about Money and on Twitter.] 

Their way of making that happen was to send out text messages and push notifications to all the customers who had opted into it and say, "We predict that tomorrow you will use four kilowatt hours over this interval. For each kilowatt hour less than four, or fraction thereof, we will pay you three yen," three cents or so. Relatively small individual decisions like, "Okay, maybe I should not be playing the video game on the power consumptive laptop right now because that is worth two cents of utility to me." Aggregate that over several hundred thousand customers in parallel, and then they could fairly reliably promise to TEPCO, "Yep, we're good for 50 megawatts anytime that you make that phone call." And it was typically arranged about a day in advance rather than at the moment. I assume at the moment when you're already looking at a falling signal on the oscilloscope it's far too late to send out text messages.

Travis: Yeah, there's this really great study that some academics pursued—I can't remember the year—they essentially used a light in and around the Chicago area. There was a light that you could plug into a kitchen outlet, it tended to be the kitchen, just some hub in your home. The light would be red if power prices were quite high and they wanted to encourage you to reduce your load, and green when power prices were fine.

They were hoping to observe a lot of response to the light as a signal for what the market prices were, but it turns out that we're not very sensitive to the price of electricity. We just really like it as a modern amenity. If the price is a little high—you mentioned you can save 3 cents times whatever your consumption is—it just ends up not being that much money and most people are willing to pay the premium.

All to say that residential users are not really great targets for demand response type programs and these sorts of mechanisms. But commercial and industrial users, they're often really narrowly focused on their bottom line and they can, if they can save 2 percent of their energy costs or whatever, that's a meaningful number for them. They're willing to engage in those sorts of behaviors.

So a lot of the demand response stuff has been much more effective actually at the commercial building level. Even industrial facilities, they have to weigh the trade-offs of maybe having their aluminum melting stop for a moment and their production slow down or come to a halt—and I don't know, there's a lot of revenue that they're leaving on the table there if they do that. So even industrial facilities tend to be a little bit less responsive to those price signals, whereas commercial buildings are more keenly willing to engage in that sort of behavior.

Patrick: I think one thing we're seeing is simply the agglomeration effects of if there is plus or minus three cents of utility available to you for making a decision, the human cognitive cost of snatching that three cents will swamp it.

And so, residents are unlikely to make that for strictly economically rational reasons.

[Patrick notes: Though, there does exist something of a culture of conspicuous irrationality to save money in certain segments of Japanese society, where that is locally socially esteemed. There are TV shows, and they’re campy, but there is an audience at home eagerly participating in the campiness. You can see quite a lot of this genre on YouTube if you search for e.g. 節約 (setsuyaku).

This culture is established enough that there is a printed warning on the demand response app discouraging users from taking the setsuyaku spirit too far; the worry is, and I am being extremely literal about this, that some vulnerable people would prioritize saving a few pennies over temperature control and die as a direct consequence.]  

It seemed that most people like myself who opted into the push notifications were more doing it out of some sense of social concern or simply the gamification aspect of like, periodically I get a notification on my phone that breaks up the monotony of the working day.

But when you're a large commercial facility, you could reasonably have people or software work on your behalf, which do nothing but power price optimization. And this might not be obvious to people who are not commercial real estate specialists, but if you project out your power costs over the course of a year and they're relatively stable on a year-to-year basis, causing them to go lower does not merely save you an operating expense in year X—that is effectively capitalized into the cost of the property.

Therefore relatively small swings in the operating expenses of a building can cause large valuation changes of a building as capital asset, which do a couple of things that we like. It causes real estate investors, landlords and similar to make capital investment in the present day to quickly sort of structurally change the energy needs of their building. And we get to enjoy those new structural modifications over the course of the next 20, 30 years.

But it also says that, to the extent that there is an obviously incentive-compatible thing where all I have to do is install a piece of software and/or dial my thermostat better in certain hours of the day with someone who probably does not make nuclear engineer money—I will do that immediately. Thank you. And so market works here.

I'd be remiss if I didn't discuss one interesting use for the economics of power because it's a sector of the economy I very rarely say positive things about. And I think this is one of the interesting things they've generally done. The cryptocurrency community is a large user of power essentially because they waste it to produce random numbers that they have a particular emotional or aesthetic attraction to. But the interesting thing is they can titrate up or titrate down on the amount that they are using at any moment in response to demands from grid operators and similar.

[Patrick notes: The single best place to learn about Bitcoin energy usage is probably Nic Carter over at Castle Island Ventures. He’s had a number of great podcast episodes with Bitcoin mining execs, who often come out of an electricity generation background.

I yield to no one in my crypto skepticism, but Nic and those execs are much better calibrated than most people who talk about crypto energy usage because they read that it was really high.]

This got misreported, I think, with respect to some of Texas's grid issues in prior years where Texas was undergoing some amount of grid instability due to a variety of reasons, and it became public knowledge that the grid operators were paying the Bitcoin miners to curtail their usage. And this was like, "Why are we allowing Bitcoin miners to suck up valuable power while we can't necessarily keep the grid on?" et cetera. [Patrick notes: Representative example.]

And in one of the few times in my life where I will defend the honor of Bitcoin miners, what was actually happening there was that the Bitcoin miner was contracting up front to buy a large amount of committed power use at a particular price, which is something that the grid and utility operators want because they have that baseload of energy that they have to sell to someone.

The Bitcoin miner was saying, "Yes, sticking my hand up, I am always good for my slug of that baseload, unless you tell me you need it back. In which case I will turn off my machines in seconds and you will pay me for this service that I am providing in smoothing out the amplitude of your demand curves."

And when this was reported, people who didn't have a great background in energy economics were saying, "Wait, this is the Bitcoin companies extracting money from the state of Texas," et cetera. [Patrick notes: Representative example.] But I don't think people who had that perception of the issue understood the totality of the system where, say what you will about Bitcoin mining—and in other places, I say quite a bit about Bitcoin mining—but in this one instance they were probably good for the economic viability and system-wide stability.

Travis: Yeah, they were taking advantage of a program that was available to all. It's interesting, when I think about tech-oriented consumers, Bitcoin often comes up. And I think I've read in a few locations that these small modular reactors, which is this nuclear power 2.0, is becoming feasible, or there's interest from Bitcoin miners to maybe have an on-site small modular reactor—an SMR—available to them to just have that surety over the energy supply that they're looking for. But like you were mentioning, a Bitcoin miner has the ability to titrate.

And it seems like the trade-offs are okay with them. I can't think of that same sentence being said when it comes to data centers that are trying to support some of the AI work that is going on right now. Those data centers want to run—they're more like a manufacturing facility. They just want to run 24 hours a day, seven days a week. And I would be surprised if data centers would be taking the same sort of bargain with these grid operators, because it's just so valuable for them to be up and running and they just want to run continuously throughout the day.

Patrick: I think there are some interesting margins. Granted there are many engineers that have spent far longer than I have on this particular topic, but in a similar way to the waveform of energy usage tracks humans over the course of their day, the waveform of data center usage largely tracks humans interacting directly with computer systems. But much of the computation which is done on behalf of humans is not done in response to someone sitting at a keyboard—it's done elsewhere in the economy. And some amount of that can be scheduled.

[Patrick notes: This is related to why a decreasing but still maddening number of web applications have office hours. See, the computer program originally running the IRS or Social Security runs on a mainframe. That mainframe has broadly two modes of operation: interactive, where there is an operator typing something into the system, or batch, where we are doing intense calculations to keep the world running. The batches were scheduled overnight, when almost no operators were e.g. keying in tax returns.

The newfangled web interfaces to the mainframe programs use a complex stack of middleware to simulate a helpful IRS telephone agent who has preturnaturally fast and accurate typing. Of course, “she” can only work when it is safe to work. When is it safe to work? During the day, when all the other agents work. Is it safe to work at night? Uncertain; the people who architected those batch programs have long since retired. Do you want to bet the U.S. and global economy that it’s safe to do certain state-mutating operations in the middle of a batch run? Several generations of IRS technical leadership have said “NOPE.” And thus, web apps with office hours.]

And so there are programs at places like AWS and similar where if you were to schedule computation that you have the luxury of scheduling to the middle of the night locally, when we probabilistically think many less people will be watching Netflix and similar, we will sort of cut you in for a break on that computation.

[Patrick notes: The mechanisms here would fill a small book. Most aren’t worth thinking about unless you’re doing very specialized things, but many people in the economy do specialized things for a living, right.

For example, you can get rent servers for really cheap if you’re willing to instantaneously “return” them to AWS any time a higher-paying “on demand” user needs burst capacity. Why would that ever be in your interest? Well maybe you’re doing some image processing or video transcoding where the timeline you need results on is measured in days, and where if an individual server doing an individual video is interrupted in the middle to save 30 cents, your program just shrugs and picks up where it left off the next time it finds itself with a server to run on.

And thus you are providing for Amazon the same service the Bitcoin miner is providing to ERCOT: “Sure, I’ll pre-commit to speaking for a slug of your baseload capacity, in return for you having the right to on-demand curtail my usage, where you pay for that right, to sell that capacity to users with higher instantaneous economic value out of it during particular minutes/hours over a month than I would have.”] 

Patrick: Similarly on AI, there's a difference between inference time compute, which is someone interacting directly with an LLM and/or an LLM being invoked on their behalf, and the training cost.

[Patrick notes: Non-specialists think of LLMs as something you talk to in a chat-like interface. I think non-specialists do not yet understand that there will eventually be very many LLMs doing things on your behalf in the background of life, in the same way that there are non-LLM agents at this very moment e.g. updating their bidding strategies for how much they’re willing to pay for your attention the next time it becomes available for an instantaneous ad auction on e.g. a Google search results page or e.g. an article at the WSJ.]

I think at a lot of margins, it makes sense to run the chips that are doing the training a hundred percent of the time, but one can imagine perhaps that being other than the case at certain margins, which we haven't quite reached yet. And in those hypothetical futures where there is a bounded amount of need for training time, which people are saying might not arrive for a couple of years yet, you could hypothetically do the training overnight, turn off your chips during the middle of the day. They are quite power hungry. 

[Patrick notes: See generally, Situational Awareness.]

Exploring behind the meter arrangements

Patrick: You mentioned one topic which I think is fascinating, which is the co-location of energy production and individual energy users. Because this grid gives us a model where energy just magically teleports from wherever it is produced in the world to me. But that isn't actually true, right? There is some loss over shipping the energy tens or hundreds of miles or kilometers. So, many places to dig into here. One, let's just establish at the outset, it's possible to get electricity without getting it from the grid, right? Can we talk about so-called behind the meter arrangements?

Travis: Yeah, of course. I mean, you said it well, the transmission system is excellent at conveying electricity from the generator to the end user. But if the end user decides that they want generation on site, you bypass that transmission component. A great example of this before we get into some of the bigger applications would just be someone has some solar panels on their rooftop. If I'm sitting, the solar panels sit behind the meter per se. So the meter is out at the street and my solar panels are on my rooftop and they're connected to my house loads.

What the utility observes in my consumption is something less than my true consumption, as long as the solar panels are taking some amount of that load off of the grid. And that's what we mean by behind the meter. As far as the utility is concerned, if I had a big enough solar panel array on my rooftop and maybe even batteries to backstop it, I would maybe have no demand as far as the utility was concerned.

Transmission line challenges and innovations

What you get out of that is, the point you're making, you're avoiding some of the transmission losses. So as a total grid and evaluating grid efficiency, tabbing up the pros and cons, you're also avoiding this effect that's becoming more prevalent on the grid, which is this idea of congestion. What happens in these transmission lines is they were built kind of fit to purpose for some amount of energy usage, but with a timeline of a lifetime of maybe 30 years or 50 years or something like that.

The energy consumption that's happening 50 years from today may—and all odds suggest it will be—far higher than the energy consumption that we're currently experiencing when we build this transmission line. And that transmission line is a lot like a pipe. Think of it as a pipe that's conveying water and it has some diameter and there's only so much throughput that you can put through that transmission line. And then at some point, you're going to just need to build another pipe, another transmission line.

We're fast approaching, and it's being compounded by all of the data center demand that's starting to jump onto the grid, this point where our aging transmission infrastructure is not able to keep up with the amount of throughput that we need to put through those lines. Our choices are build more transmission lines, which are incredibly expensive, or start taking—make it so that the utility doesn't think there's load at the end of that transmission line, and that's that whole behind the meter solution that we're talking about.

Patrick: So, frequent listeners of Complex Systems can probably predict a bit of the answer here, but the physical reality of transmission equipment has not evolved radically since the invention of electricity—or maybe it has.

Well, taking as writ that the physical artifact being shipped out and installed in central Illinois is very similar to the one that we did in 1975, I would expect that the cost of actually installing it on an inflation-adjusted basis is much higher than it was in 1975. Why is that?

Travis: There's land, probably one of the drivers. I don't know what the mix in terms of what percent each of these drivers is, but one of the drivers is these transmission lines have to span a finite and unchanging amount of land and that land continues to be more and more expensive. So as you're trying to get easements through farmlands property, they have more negotiation capability there and can drive the prices up.

I think we've generally tried to get higher throughput lines, especially compared to the lines of the early 1900s. And so there's inevitably more material that we're putting into it and there's more safety mechanisms that we're building in. So I think a combination of those things are making the per unit mile price increase at rates faster than you would expect if you were just tracking inflation.

We're continuing to innovate here though. I think there's a couple of things that we're starting to pursue. There's conducting material that—historically we've used alternating current to transmit over these really long distances because we'll step up the voltage to really high voltages, which draws the current. There's some kind of inverse relationship between voltage and current. And so we can keep our currents a lot lower if we get to really high voltages, and that means that there's less resistance in the line, less heat.

I maybe getting some of the specific details slightly off, but the point is the higher the voltage, the easier it is to transmit over long distances. And so we have these really big step-up transformers that take the voltage that's coming out of the generator and step that voltage up to really high levels, and then send it along the line.

That has been the most effective way to move power historically, but with some of the new conducting materials and new tech, material innovations, we've been able to start doing that with DC lines, direct current lines instead of these alternating current lines. This allows us to harvest that electricity that's coming out of really big solar fields, which are generating electricity as direct current, and then kind of directly inject that into the grid by making use of these high voltage DC lines.

I don't want any transmission operators would find it unfair to characterize them as kind of this stagnant group of technologists, but it's really the case that they tend to be behind the generating assets in the sort of leaps and bounds that they've made, but they continue to try to find ways to improve their transmission.

Dynamic line rating and grid efficiency

The other piece that they're engaging in when it comes to transmission is that the lines tend to be rated for a certain amount of throughput, like the pipe diameter, to borrow that metaphor again, and that pipe diameter rating is set based on that really hot day when the lines tend to sag a bit more and there's more resistance in those lines.

What they're realizing is that actually the pipe diameter kind of changes depending on the weather and some other factors. What we've usually said is a throughput line of, let's say one gigawatt—well, that's like the safest version of throughput. We could expect it every minute of any given year, we will still be able to move one gigawatt through that transmission line, but by being a bit more dynamic in that line rating—dynamic line rating—we are able to say, "Oh, well, when it's cold and it's winter, we can actually move 1.2 gigawatts through that transmission line."

Those choices, that reckoning with those policy choices has allowed us to delay some of the transmission upgrades. The whole idea is, these transmission lines are really expensive. We don't want to build them unless we absolutely have to. So let's pull out every stop and try to make it so that we don't have to spend those dollars yet. Let's do dynamic line rating, let's do behind the meter, let's talk about demand response and battery storage and some of these other techniques and technologies and let them avoid these investments.

Patrick: Dynamic line rating is fascinating to me because presumably that's some combination of work happening in the lab, some combination of regulatory or regulatory-adjacent approval getting, and some combination of abundant cognition available due to the use of software to make math on a day-to-day or minute-to-minute basis much more tractable than it was in those days where we did have electricity but didn't have widely distributed thinking machines and where "calculator" was a person's job title.

We stand on the shoulders of giants in so many ways. I can't imagine doing all this work back in the day.

And again, part of the economic logic here is that there are some margins we are talking about that are effectively continuous but there are some margins where there are great step changes. Let's say hypothetically we can only buy pipes in units of one gigawatt at a time. So, we might buy 10 years of leeway in figuring out, is there really frequently enough a need to ship electricity between location A and location B such that we need to provision one more gigawatt times a hundred miles at a cost of tens of millions of dollars? Or can we install a software upgrade at a cost of hundreds of thousands of dollars and kick that down the line 10 years to see whether demand actually moves in that direction or not?

Travis: Yeah, grid planners have incredible models that try to incorporate population migration. They try to incorporate how much usage is going to happen in a typical household in the future. What about the fact that all the new windows that we're installing in new build homes are much more energy efficient than they used to be? There's a put and take to everything and the models that they develop to try to understand what those investment needs are today to satisfy the demand in the future are so intricate and deep and you're crossing so many different disciplines to try to understand what it might look like 10 years from now.

And so what do we need to do today? A transmission line, for instance, it might take eight to 10 years to get one approved. So we might sit here and kind of mark a line in the sand and say, all right, let's start this whole process of getting a new transmission line. But that transmission line won't be installed for, let's say a decade. We want it to last for another decade or two after that. So we're kind of modeling out 30 years or more to try to understand what the infrastructure need is for this particular investment.

Patrick: Yeah. And there's a fascinating amount of economics and culture and predictions about constantly moving targets that one has to make to do that.

There's a line I like to use about Christmas trees of all things: you need to predict whether children who aren't born yet will believe in Santa Claus in nine years to make Christmas trees. [Patrick notes: I’ve read a few articles about this over the years; this one is pretty decent, but not IIRC the one which opened my eyes to this new facet of Christmas magic.] 

And similarly to model the amount of electricity transfer between Chicago and the various places that supply Chicago, you need to have a prediction in advance of the family sizes of Chicago 30 years from now. And what's the changing mix of demographics and American culture? And is there a resurgence in getting married earlier and having children earlier? Or do we see the continuation of current trends, to have an estimate of that number to inform things that you really need to put in the pipeline now such that they are available in time to meet the demand that you are forecasting.Fascinatingly cross-functional work.

Data centers and energy demand

There's some directionally similar work that gets done with respect to data center planning and siting. And indeed, I think the data center folks, because they have a somewhat direct ability to observe what are the sources of demand in a way that electricity might not have quite as much, get to evaluate that at a faster cycle time than grid operators historically have.

The thing that is really throwing that industry up for its own mini uproar is that many people perceive we're on the edge of a discontinuity with respect to the services that people can be getting from artificial sources of cognition and past that discontinuity is tough to model. Like, are you going to use a standard iPhone and Netflix plus 20 percent of compute demand in the future? Or are you going to need 10 times that provision to maybe like a three-year window? And so that is partly causing the great build out of data centers and also partly causing them to get very creative about energy sources they're tapping.

Travis: Yeah, there's this essay that was written by I think it was Tim Fisk and others, which basically we're brute forcing things. It was called the Bitter Lesson. [Patrick notes: Rich Sutton’s phrasing: essentially, advances in architecture/algorithms over relatively short research timeframes were almost inevitably swamped by upgrades in amount of compute available in those timeframes, and so the only thing smart people in AI should worry about was saturating all the compute. Gwern also a good read on the scaling hypothesis.]

The takeaway that at least I drew from it was there might be more efficient algorithms that we develop and we can find ways to get per unit of inference or per unit usage per inference to be a little bit lower. But really it seems like the most effective ways to move us along this path is to brute force a lot of things, which kind of leads to a bit more of that growing, spiking demand that we're starting to see today.

Would you think, like, my gut tells me that edge data centers would align a bit more with what you're describing as really difficult to predict, but maybe you would argue that all data centers are difficult to model.

Patrick: I think one thing we're seeing is simply the agglomeration effects where grids give us a model. Let's just establish at the outset, it depends, and I'm talking a little bit outside of my direct professional experience here. There have been a number of changes in data center economics over the course of the last few decades. One is the emergence of hyperscalers like AWS—Amazon, Google, Microsoft—there's others but they are increasing share of all the load for compute and related services in the economy.

Partly it's simply that at the scales that they operate and the sophistication that they operate, they can throw better minds and better data at the question of modeling predictive usage patterns.

Where historically data centers were… a fascinating business, we'll probably have an episode entirely on it at some point. But historically data center was run as a subset of real estate where instead of selling people relatively large amounts of square feet, you were subdividing that square feet to rack size or blade size, and then providing some ancillary power and cooling services along with the real estate that you were fundamentally subdividing. And so the sophistication level of data center capacity planning approached the sophistication levels of the rest of the commercial real estate industry, which as someone who grew up hearing a lot of stories from the dinner table—there are some very smart people involved but they're not the team of PhDs that Google would put on the problem.

[Patrick notes: As an example of “No, really, CRE people are actually good at their jobs.”, I offer the ethnographic research (by John Melaniphy et al) cited by my father in how banks make branch siting decisions.]  

And so Google has multiple teams of PhDs on the modeling problem. However, much like there's curves for electricity where we keep finding new things to spend it on as the price goes lower and it gets more abundant, we keep finding things to spend bandwidth and compute and similar on as they become more available. The existence of video streaming as a service, for example, is almost unthinkable from the information superhighway days where a web page using one megabyte of data transfers would be crazy.

And now people routinely blow through multiple gigabytes per month of their residential allocation and multiple tens of gigabytes now, I suppose, due to essentially Netflix and YouTube and all the other things that people don't really think of as engineering marvels. And then throw AI on top of this. Okay, Netflix is very bandwidth hungry, but you can probably predict that there is some asymptote for how many hours per day humans can possibly spend watching video and how beautifully 4K those videos can possibly get. And so there's some asymptote for the bandwidth and compute needed by Netflix.

But there is not necessarily an asymptote for how much human cognition you could use on behalf of one person if the cost of that cognition was low enough. And so can you imagine a world where there are like a thousand PhD equivalents working 24 hours a day, seven days a week on behalf of every fourth grader at my children's school? Yeah, you can imagine that world.

[Patrick notes: I didn’t do Japanese math homework with Lillian today, because I have a podcast coming out. But if I had unconstrained cognitive capacity, perhaps via a clone, clearly I’d have done that homework with her, right?

And if I had a third clone, perhaps I would have carefully kept a list of every problem she got wrong. 

And if I had ten clones, perhaps I would have run a full-dress engineering postmortem on every problem she got wrong. 73 x 20 = 1,560? Oh no, five whys, let’s go. Senior engineer, can you please refresh the sequence of events over the critical 43 seconds?

And if I had a hundred clones, perhaps I would have run a full-on curriculum symposia, discussing a few thousand research papers generated on the subject of bilingual Japanese/English fourth graders, well really one in particular, and what could be done to increase their rate of skills acquisition. Which is totally a rational use of a thousand PhDs if you can conjure them out of the ether for pennies.

The story of humanity is the story of increasing abundance followed by discovering we are unstoppable at finding new things to want.] 

Patrick notes: The goods and services they would be providing are weird ones to describe in language we have currently. But if we find a world where the economics pencil at a thousand, do the economics pencil at 3,000? 7,000? Very plausibly, and then that implies some kind of crazy amount of dynamic range in, okay, so how many data centers need to be within a quick hop of Chicago?

Anyhow, refocusing on the grid topic. A thing that I understand exists in the world is, we have these industrial users of electricity to avoid the cost and losses of transmission and to ensure the availability of electricity. Sometimes they have co-located energy assets. Sometimes the original user of the electricity, the plant out in perhaps somewhere in the Rust Belt goes away, but the electricity asset is still there. Do you want to talk about the downstream economic consequences of that? Because they've been fascinating occasionally.

Travis: Yeah. I think I'm might be oversimplifying a little bit to just kick us off, but really it becomes like a strictly economic question. The work of it is sunk, like building that. And so what we really want to know is suppose it was strictly behind the meter, so it wasn't really connected to the grid in a meaningful way. The question at hand is what will it take from a financial standpoint to connect this asset that's a little bit stranded to the grid? And how does that weigh against the potential revenue upside that we get from starting to sell the power to the grid instead of to this immediate neighbor offtaker.

So I think the choice that ends up being made becomes a relatively simple one for the user or owner of that plant. Where it becomes a bit more complicated is on the utility side, because they may not have, as I mentioned, this is behind the meter, they may not have really ever observed in a meaningful way what that generation was from that plant. Or the true load from that plant and maybe the grid in that area is simply not robust enough to take that full power that's coming out of that plant, or there needs to be some upgrades that the utility will need to make. And the question that's on the top of everyone's mind in that instance is like, who pays for it?

Interconnection queue and grid security

I think the owner of the power plant would say, "Well, you, the utility, you need to pay for it." And the utility is going, "Hold on, wait a second. This isn't our bag to hold." That kind of back and forth can create problems. So what the utilities will often require, and this isn't unique to this behind the meter scenario, but it applies for all generation that's trying to come onto the grid is they'll mandate what's called an interconnect study to understand how that generating unit will interact with the current topology that they have on the grid.

These studies can take a really long time. They need to be sure. It's a lot about safety and making sure that the grid doesn't become unstable by the introduction of this new player on the grid. And the other challenge that the interconnection studies have created is there's this queue, this interconnection queue they call it, that can be quite long with a bunch of generators saying, "Hey, I want to put a generator over here, I want one over here." And there may be dozens and dozens.

I think the last study I saw was that something like 20 percent of the actual assets that go enter the queue actually come out of the queue as real generating plants. So there's this spamming mentality from those independent power producers to just say, "Well, let's tell the utility that we've got six different locations. We really only have financing for one, but one of these will probably make it through the interconnection. And that's the one that of course we'll go and invest in when it finally gets the approvals."

And so this behind the meter installation that is trying to connect to the grid will need to enter the queue, maybe sit there for several years waiting for their turn. And then hopefully through that study, they'll be deemed a worthy asset to add to the grid and then they can join, but all of those different steps need to be followed. And it's in the name of grid security, grid reliability, grid safety. So it's a good reason to have it. It's just that this interconnection queue policy that we've tended to use has gotten a little distorted.

You might hear a lot of people talk about it lately. It's definitely been the thing that the energy industry talks a lot about is interconnection reform, and they're trying to find ways to reduce the queue. There's different methods that they're proposing to make that happen. And you want to reduce the size of the queue and then increase the kind of success of units that are in the queue. Those become good reforms for us to pursue, but it just means more policy, more sausage making, so to speak, and some of the challenges that come—I mean, it's a political economy problem almost at that point.

Patrick: And clearly one reason why the queue is serialized is because if there are 400 people attempting to add locations to the grid, they're not hermetically isolated from each other. Which group you accept determines the dynamic process of what the grid will look like for the marginal person you accept. And so it ends up being exactly a political economy process where you're prioritizing different users' needs against each other, prioritizing locations against each other, probably doing some of the classic work of politics in a democracy of attempting to make the process predictable and transparent and correspond in some way to our intuitions about values and similar.

Travis: And they get—and it's almost in addition to everything you're describing, it's quite literally serial, like there is one unit that will get evaluated. And so a lot of the new—and then once they've evaluated it for the grid and if they approve it, they'll incorporate that new unit into their models and then they'll go and they'll call the next number and "number 101, come on up." And then they'll do their interconnection study.

I think the coolest innovation that the interconnection reform is proposing is that we're now going to do cluster studies. And we'll maybe take like a dozen generating units in different locations and we'll evaluate them wholesale. And that'll increase the throughput considerably. And it also just rightly acknowledges that the grid is complicated and there are interdependencies between generators. So let's study them all together and then bring them in as one big tranche, and then we'll go to the next tranche and do it again. That feels right to me, and I'm glad that those are the sort of policies that are being considered right now.

Patrick: It also seems to be more robust against the statistical realities where not 100 percent of people who attempt to reserve a position in the queue will physically deliver electricity two years after being approved for it. And so, presumably, if you're underwriting on the basis of tranches, you can have things like a statistical model where, okay, recent experience has been that 60 percent plus or minus 20 percent of the people in similarly situated tranches actually deliver on the following timelines. So we can incorporate that into our models versus—well, since we are doing each application serially, we have to assume that they will actually use the—

Oh, there's so many ways we could continue discussing this.

Understanding the U.S. grid structure

One question for you: we have grids plural in the United States. Why is that?

Travis: Yeah, we do. Go at the highest level first. If we just surveyed the whole United States, there's three different, what they call interconnects. There's the Eastern Interconnect, the Western Interconnect, and then there's ERCOT, which is the Electric Reliability Council of Texas, that is just Texas. So you've got Eastern US, Western US, and Texas. And there's a pretty bright demarcation line between these interconnects. So bright, in fact, that if I'm operating on a grid in the Eastern Interconnect, I can't actually export my power directly to someone in the Western interconnect because we're out of phase.

I mentioned before that we have this 60 hertz frequency that we're operating at but not all the troughs and the peaks are in sync between those three different major interconnects. They're out of sync and so you have to transfer the power to this way station, convert it to the right phase and then inject it into the Western Interconnect, for instance, if I was moving from east to west.

The line of demarcation is essentially the Rocky Mountains are in that area. And then, as I mentioned, Texas—why I think it formed in the way it did, my gut tells me it has a lot to do with just how difficult it is to move transmission over the Rocky Mountains. And so it was like, we're going to do our thing on the West side, you guys do your thing on the East side. And then Texas is like, we're going to do our own thing.

And then within those interconnects, there's different transmission organizations that can oversee that local grid. But I just want to make it very clear that grid is all connected to each other. It's much easier to move power from one of these regional subcomponents of the Eastern Interconnect to another regional subcomponent because they're sharing that same phase.

What ends up happening in the Eastern Interconnect, for instance—let's look at where I live here in Atlanta. I'm in a regulated utility geography. I pay my Georgia Power bill. Georgia Power's owned by Southern Company, and Southern Company owns three or four different generating companies, utilities that are vertically integrated. They own generating assets, they own transmission, they own distribution, and the kind of regulatory compact that we've made with Georgia Power is that that's fine.

We'll let you be a monopoly, but in return for that monopoly and the monopoly power, we're going to put a little bit of oversight on it. And we're going to use a public utility commission—or here in Georgia, it's called the public service commission—to basically oversee the utility and its operations, to authorize the prices that utilities are able to charge me when I consume electricity and charge industrial users and commercial users, and then authorize the amount of capital expenditures that Georgia Power wants to spend and then authorize a rate of return that Georgia Power can enjoy beyond their operations.

Travis: That's like one type of grid that's within these interconnects. Another type of grid or kind of regionality are these regional transmission organizations. We call them RTOs for short, and they have—by contrast with the regulated utilities—these are deregulated markets. The transmission and distribution tends to be regulated, but the generation is a competitive environment and they participate in energy auctions in order to sell their wares to the market, which is of course these electrons.

Why different regions form the way they did, I think ends up becoming a bit of a political economy question again. Maybe here in the Southeast, which tends to have more regulated utilities, individuals and legislatures are comfortable with the arrangement and there's some stickiness that's been involved. And then in other areas, there's maybe—one of the downfalls of a regulated utility tends to be, the good thing is it tends to be really consistent in its power delivery. It's often the case that they're doing an excellent job of predicting their supply and then overbuilding because they get a return on the assets that they have.

So there's this—they call it gold plating—but it tends to be the case that a regulated utility plays it safe. You can expect that the utility wants to play it safe. They certainly want more assets in the ground, but also the public service commission, they want to play it safe too, because if they try to thread that needle too closely, they're going to get yelled at and voted out of office or replaced or whatever. So everyone has a tendency to play it safe, but the trouble with that is it gets in the way of maybe some more innovative technologies that might be out there. They're not going to be the first to adopt the newest and latest generating technologies and transmission technologies. That feels too risky.

But in these regional transmission organizations with these competitive markets, I think that it can encourage a bit more competition. And so maybe there's some sentiment at the public level or at the state legislator level that says, "Hey, here in New York, that is under an RTO, we would really like to have a bit more of a competitive marketplace for this energy generation. Let's move in that direction. Let's go the deregulation path."

So the way that you might think about the geography breaking up is starting at these big interconnects that basically make it easy to share power between these sub-sector geographies and then those sub-sector geographies have some autonomy—a fair amount in fact—over what sort of utility system they want to have. Do they want the competitive markets of an RTO or would they prefer the regulated compact that is arranged between a public service commission and a vertically integrated utility?

Patrick: This has been a fascinating topic for me for the last two years. I've been doing a bit of pro bono advisory work with a geothermal nonprofit which is a focused resource organization that is attempting to get geothermal energy used for a wide variety of uses with, you know, ultimate end goal being make it available for baseload. And one sort of shadow purpose of any sort of system that regulates the energy market is to function as effectively industrial policy for power generation.

And so we've had some amount of industrial policy which was pro-solar, pro-wind. We've had effectively an industrial policy which was heavily against nuclear for a variety of reasons, for a number of decades, probably to the world's regret. But no coalition of political interests arose to fix that problem over the course of a couple of decades. And so we made the choices we made.

But you know, an individual polity could choose stability is the only thing that I want for my energy generation policy. Subject to it being on 100 percent of the time, keep the rates as low as feasible, and that is all I want. And a different polity could say, well, okay, there are margins to trade off with here. Like I'm okay with paying a little bit more and maybe accepting—I won't accept seven nines of availability of power, I'll accept six nines and then get X more impact against new technology generation for the purpose of achieving carbon targets or similar.

So, fascinating multi-layered problems, which I imagine we could talk about for another hundred hours, but we are coming up on what is usually people's lengths of time and attention that they have to devote to a similar single podcast. So Travis, where can people find you around the internet?

Travis: Yeah, you can find me on LinkedIn—I'm Travis DeWalter. I also have a Twitter handle at Travis DeWalter. And those are probably the two easiest channels to find me.

Patrick: And if folks are coming to Electricity Grids for the first time, do you have a book or similar that you would point to them as this is the thing that I would give to a young cousin who was getting into the field?

Travis: My favorite resource actually is an online resource called Utility Dive, which does short articles—they tend to not be more than about a seven minute read. That gives you a sense of what's happening on the grid and the sort of puts and takes that people are considering as they're going through it. I would say my best advice for everyone is to just jump in. I would strongly encourage you not to read some 101 primer on the energy grid and instead just start reading and seeing the language that Utility Dive articles are using and then make heavy use of some large language model to fill in the gaps of your understanding. And I think the combination of those two is a really effective way to get up to speed quickly.

Patrick: I would echo that from my own experience as being an engineer in a field that doesn't really have to deal with power draw as a major engineering constraint, at least not in my neck of the woods. Simply reading the work of actual professionals gets you up to speed a lot faster than reading regurgitations in the variety of places in the economy that don't work with people who need to actually care about the thing being on. And LLMs are a wonderful tool for getting up to speed on the lingo and similar in a way that repeated Google searches could duplicate (but not nearly as efficiently) a few years ago.

Travis, thanks so much for coming on.

Travis: Thanks a lot, Patrick. It was great to be here. Can I just say one more thing, which is today is my daughter's birthday. Her name's Madeline and she turned five and I just want to wish her happy birthday. We love you very much. You're the best. Happy birthday, Madeline. Thanks, Patrick.

Patrick: Happy birthday, Madeline. And for the rest of you, we'll see you next week on Complex Systems.