Elon 的前电池负责人谈如何将变压器缩小 100 倍 | Drew Baglino、Heron Power

That's actually what a data center is. You take a gigawatt of power and convert it into 700 megawatts of heat. You spend 300 megawatts to get that heat into the atmosphere.

I presume your transformer is going to be more efficient than the 10% loss.

We can reduce the loss from grid to chip by about a factor of 2. Our core first product at HeronLink, the transformer part is 100 times smaller. You still need to go through an isolation stage when you're going from a high-voltage grid to a computer. You need to galvanically isolate, but rather than doing it at 60 hertz, which is how we do it today, we can do it at hundreds of kilohertz because these devices can switch so fast.

It's a materials science thing. You're trying to find materials that can block more voltage and are more thermally conductive, because a semiconductor is either conducting or not. There are a bajillion transistors in a GPU, and a power device has 1.

Right.

But it's 1 highly engineered device. What silicon carbide is doing is allowing you not to make a 1-volt transistor like you have in a GPU, but to make a 10,000-volt transistor, and it is actually smaller than the GPU. So we halve the loss. For somebody building a gigawatt data center, that means they can get 35 megawatts more useful compute out.

Okay, so why did California power rates go up?

There are a lot of weird reasons.

You're listening to "Gradient Dissent", a show about making machine learning work in the real world, and I'm your host, Lukas Biewald. Today, I'm talking with Drew Baglino, an old friend of mine who started as a junior engineer at Tesla and ended up running all of powertrain and energy, which is most of the company, and reporting directly to Elon. He then went and started Heron Power, which has raised hundreds of millions of dollars to revolutionize the energy grid. This is a really interesting conversation on topics I know a lot less about than AI. I hope you enjoy it. I'm super excited to talk to you about your company, but I haven't known you for a long time. I wanted to go back and start with your career at Tesla, because that was a really interesting ride. I think you were there for 18 years. Can you talk about that experience?

1. Stanford Led Drew To Tesla

Sure. Maybe I'll say first that we knew each other in college and went to Stanford together. We may have lived in some funny off-campus housing together—more than 1 of those off-campus houses, I think. I remember even back then, you and I would talk about different ways software would change the world, and you were always better at the software classes than I was. But I took them too, so I have a little bit of a software background.

That's actually a little bit of what brought me to Tesla: some connections back at Stanford. I ended up signing up for an energy field trips class that Gil Masters was running, and it sounded like the coolest class ever. We were going to visit zero-emissions buildings, wind-turbine manufacturing facilities, solar-manufacturing facilities, and large installations. We went to the Hoover Dam and some large wind farms and solar installations, including the SEGS parabolic-trough mirror solar installations in the Mojave, which were pretty cool to check out. They're still running today.

Uh-huh.

The power-tower one is nuts. No, you also see these. They're on I-15, I think.

I think I see them from the airplane sometimes on the way to San Francisco.

I signed up for that class because it sounded really fun, and that's where I met JB Straubel, one of the early Tesla founders. He and I kept in touch, and even after we lived together in Dire Wolf—was that what it was called? I don't remember what it was called, to be honest—but that house on Embarcadero.

Some Grateful Dead song.

Yeah.

Yeah.

That house on Embarcadero. I left that house and went to D.C. to work at a think tank called Resources for the Future, where I was basically doing programming. I was coding up this crazy model in some bizarre database language that I can't even remember—maybe it was called Stata or R, or maybe it was both of those languages.

Those are bizarre languages, by the way. Classic stats languages.

Yeah.

Yeah.

I can't remember which one it was, but I was working on this really interesting economic model of the Washington, D.C., metro area, and I was bored, kind of out of my mind, mostly because of the lack of impact. It was interesting work, but it was translating into nothing that I could see in the physical world, at least not on a time horizon that was interesting to me.

I kept in touch with JB, and at some point he said, “You know, I think we might have a role for you.” I had a background in embedded DSP because I took some graduate courses at Stanford in robotics and communications that involved embedded DSPs. He said, “We need somebody who knows embedded DSPs.” I said, “Oh, I can do that.” And that's how I ended up there.

The first project was, “We licensed all this analog control design for the motor control, the charge control, and the battery management from AC Propulsion. That's how Tesla got a lot of its early technology: through licenses from AC Propulsion, including the motor design. We want to replace all that analog stuff controlling the car. It doesn't work well. We want to do it in a DSP.” I said, “Sweet. That sounds like an amazing project.”

That was really the first project I was responsible for from beginning to end at Tesla. It meant I spent a lot of time in electric cars, driving around these streets and parking Roadsters in the Mission District. There was also a really fun project involving converting an internal-combustion-engine smart car that we bought in Mexico to an electric smart car.

I remember parking it at 22nd and Bryant. Do you remember that house I lived in on Bryant Street?

Totally.

I would just park it on the sidewalk. It fit on the sidewalk right next to the staircase, and it didn't even block the sidewalk. That's how small the smart car was.

It was great going from working on code in Stata to model the economy and transportation system of Washington, D.C., to embedded software, where you could immediately see the result in the vehicle you were driving yourself. I would be driving it with the laptop out, showing me the signals of what was going on. It couldn't be more real-time feedback.

That's what got me hooked on what we were doing at Tesla and why I stayed there so long: seeing the work you do translate into the real world so quickly.

I feel like you also had this awesome experience of working really closely with Elon before it was clear what a successful entrepreneur he was going to become. Do you have any fun stories about Elon back then?

2. Elon Sets Impossible Targets

I can talk through a really early story—my first exposure to Elon, actually.

I worked for JB, the CTO, directly throughout my early years. One of the things I was responsible for, beyond all this embedded software that made the car go and things related to that, was modeling the physics of our products. How far could they go? How much energy would they consume per mile? Would the car overheat when it was towing a boat? These were all the physical aspects of the electric powertrain and the systems attached to it. How quickly could it charge, and things like this?

During the period when we were conceptualizing what Model S could be—our first, let's say, mass-market EV. I mean, the Model 3 was truly the mass-market EV, but Model S was way more mass-market than the Roadster was—I was in that architectural role, defining what the car needed to look like. If we wanted to make the car do this, that meant the battery needed to look like this. That meant the motor needed to look like this. This was what an optimization of that system would be.

We got some guidance from JB and others about what kind of car we were trying to make. We brought our presentation to Elon: “This is what we think the battery needs to be”—how big it needed to be and how much energy needed to be in it.

We went through the assumptions: We thought the car would weigh this much. We thought it would be this slippery. We thought we could get tires that were this efficient. We knew we wanted to hit this EPA range, and there were a couple of other dimensioning load cases. We went through most of that with him, and he was following most of it.

Then we get to the conclusion: we think we need somewhere between 8,000 and 8,800 cells. Ultimately, it’s complicated math, but it is just math. You have some sensitivities on how good all of these numbers could be, and this is the range: 8,000 to 8,800 cells.

This is the first time I’ve ever interacted with Elon. He just says, “It can’t be more than 7,200. It’s 7,200.” And we’re like, “Oh, okay. Do you want to go through any of the assumptions? Where do you think we’re wrong?” He says, “It can’t be more than 7,200,” and the meeting’s over. Everybody is like, “What just happened?”

For me, it was also, “Did we do something wrong?” We didn’t really understand how to take it. Looking back at it, he had information that I didn’t have access to, like how much the cells would cost. He was in the background pushing Panasonic, our cell partner, way out of their comfort zone on how much energy density they could get per cell. I didn’t know any of that.

I don’t actually think he did either. He was just doing the math in his head: “Oh my God, if it is 8,000 cells, how do I fit that many cells? How much is the car still going to be affordable and accessible?” He didn’t communicate any of that.

But that’s part of his genius, I would say. By putting out what seem like impossible targets—“You said it has to be 10% to 20% more than this number, but it can’t be. I’m telling you, it can’t be more than 7,200”—what it forces, and we did it, is for everyone to say, “Okay, we’ve got to make the motor more efficient. We’ve got to figure out how to make the car more slippery, use less energy on road cooling itself, and condition the cabin. We do have to push Panasonic really hard to make the cell more energy-dense.”

You’re breaking everybody’s back at the same time. I mean, you’re trying not to. Let’s say you’re loading everybody’s back equally to achieve this goal, and doing it in a way that forces the A game out of everyone. That’s a classic example of how I think Elon operates.

At the time, I was like, “I never want to be in a meeting with him again,” because it was such an intense interaction with not a lot of context. But when you step back and appreciate why that’s the way it’s communicated, and how that drives results, it certainly drove the result we needed. Everything that I just said happened: we did get a more efficient powertrain, and the battery energy density was much higher than Panasonic was initially proposing.

It’s almost good that we did that, because other things changed along the way, like the way range was measured by the EPA changed in a not-so-favorable way. You kind of needed everything to go this way just to have a compelling product anyway.

It’s one thing that I’ve observed from the outside, watching you and a lot of my other friends or mutual friends work for Elon: people seem to burn out pretty quickly for this reason. Everyone’s got a story like that, and most people seem to walk away from that story being like, “This is ridiculous. I told him the truth, and he just didn’t listen, and then he fired me when it turned out to be the truth.” I hear that all the time.

But you seem to enjoy the experience, and you managed to work for him for quite a long time and really rise up inside Tesla. Do you think there’s something about you that’s different from most people?

You need to have a thick skin. You need to decouple your work product from your ego. Not everybody can do that. I think that’s probably one of the hardest things for engineers, especially in high-paced companies, to do: to accept that what they just spent the last 3 months on might get torched, and for a really good reason.

It may not initially be obvious to them why, but in retrospect, if they think back on it, whether they’re at the company or not, they’ll be like, “Oh, yeah, that was the right decision.” I think that’s just hard for engineers to do. But you need to be able to do that if you’re going to be nimble, if you’re going to respond to the newest information, and if you’re going to find the failures in what you develop, whether in software or in hardware.

You can’t attach your ego to your work product. With JB’s support, JB helped me through that. Having a great manager early on was important; I had 2 amazing managers early on. One was JB Straubel, and the other was Craig Carlson, who was a highly contributing engineering executive at Tesla from around 2007 or 2008 until 2013. He ran software at Tesla.

Because I was responsible for a lot of this powertrain software, I worked for Craig for a while. Both of those leaders really taught me how to check your ego at the door. If you’re going to do a good job in a hardcore engineering organization, I think that is hard for any young engineer to do. But if you can do it, you can go far. I think that’s true in business generally, not just in an Elon company.

What was it like when JB left and you stepped into his role?

3. Drew Takes On Executive Duties

It was fulfilling to be exposed to so many new areas of technology. I think that was really what retained me at the company: all of these new problems.

When JB left, I was responsible at the time for powertrain architecture. That’s the physical modeling of all of the systems in the vehicle and how you optimize them to deliver the best results. That was a team of around 20 people that did what we would call “model versus actual.” You’re responsible for predicting what the range of the vehicle will be, what the watt-hours per mile will be, or whether it’ll overheat on a grade going up to Tahoe.

But you’re not just predicting it. You’re also going out and running the test, and if it doesn’t match, figuring out why. You use the gap between model and actual to make both the actual and the model better.

I was also leading the energy engineering organization—the team that was building, designing, and installing the Powerwall and the Powerpack. We were already starting to work on the next versions of that stuff at that point, as well as the Superchargers. We were responsible for engineering the Superchargers.

When JB left, I took on the battery pack—how batteries are made—and the cell roadmap inside those batteries. How do we improve the cell energy density and the cell affordability? I also took on the propulsion system: how do you physically design and validate the motors, the gearboxes, and all the power electronics in these systems?

I took on all of these new areas of technology that were really fun. I also started to play a more significant role in the business aspect of energy: how do you make money with our energy products? That was something JB had also been the executive steward of.

The last thing I became was one of the officers of the company. I joined all the board meetings and the earnings calls, which was fascinating.

Give us a story. Tell us a story, if people might not have heard about that experience. I feel like you got this inside look at a really iconic company at a really interesting time.

Yeah, okay. Look, I don’t think it will be a surprise to you at all that Elon leads a very busy life. What was always impressive to me was that, ahead of an earnings call, I worked closely with Zach Kirkhorn and the other members of the leadership team, like Jerome and others. We had a lot of anxiety about these calls and prepared a lot.

Frequently, Elon would be coming from probably Neuralink or maybe even OpenAI at the time, to be honest, because he was still heavily involved with OpenAI. Then he’d just show up, maybe have 5 minutes in the room to chat with us, and the recording button would start for the earnings call.

That’s—I don’t know how he did it. To be honest, I don’t know how he did it. That’s just part of his genius: that he can do that. But it was certainly a high-anxiety time for us.

Are we aligned going into this? Are we going to say the same things? Those were difficult times.

But we, of course, maintained good alignment with Elon through other avenues, not just what comes up ahead of one of those earnings calls. I’m trying to think if there was any specific event. But, I mean, you can imagine. It’s not quite like a live-recorded podcast in a studio like this. Hello. But you have the people who are going to move the stock listening to every word you say. If you’re not prepared and you put your foot in your mouth, you’re probably not going to be at the company the next day. So, probably some of the more stressful moments of my life.

Anything else in your experience worth sharing?

4. Tesla Simplifies Production Hell

Yeah, I’ve got some other good stories. I think there’s been a lot made of “production hell.” Ramping the Model 3 was very intense, and there were so many reasons why. The company was in a major growth period in so many facets of how to operate itself—not just how to produce that many vehicles, but how to get the parts in, get them in at quality, deliver the vehicles to customers once they’ve been produced, and keep them serviced and operating well once they’ve been delivered. Every aspect of the company needed to grow massively with Model 3, by orders of magnitude.

There’s been a lot made of the production-hell part of it, and I think this is another area where a lot of people, as you would say, were like, “Oh, I said the truth, and then I was fired.” There are always 2 sides to the story, and I’m not here to defend Elon. I’m just here to provide another layer of context to all of that.

You have to step back and think about how to run this company that is trying to do probably one of the hardest things any company will ever do: go from delivering 10,000 to 20,000 cars a year—maybe 50,000 was what we were doing at that point. I don’t know. Maybe it was 100,000 Model Xs and Ss that we did in that time frame per year, I think it was—to doing millions of cars a year. When you think about the sheer tonnage of material, the millions of dollars and billions of dollars of work in progress, it’s an intense time.

Any given interaction that an executive would have with anybody at that time—you, as the executive, are trying to drive the end outcome, right? While, maybe, in an interaction on a production line, an executive might say, “Just delete that thing,” or, “We don’t need that,” is that actually what they really mean? What they’re trying to say is, “We need to produce more units.”

With any individual bottleneck, you don’t want to over-index on, “He said we shouldn’t do that here, so we shouldn’t do it anywhere.” Actually, the point really was: How do we make more product holistically? I don’t want to imply that anybody wanted to make bad product. You just wanted to make more product.

I’ll actually get into the details on this one. We had this robot that was trying to install a moisture-absorbing bag. It was kind of like this flimsy bag filled with desiccant material or something, to try to avoid condensation building up between the top of the battery pack and the underfloor, where the seats are and where your feet are. It was partially to act as a desiccant and partially to be a noise barrier, but it was this super-flimsy object.

There was this robot trying to get it onto the top of the battery pack. I think Elon may have coined this, or somebody else did, but we called it the “flufferbot” or something, because it looked like this ridiculous thing and was always erroring out. When you stepped back and looked at this, you said, “Why do we have this part?” Second, “Why do we have a robot installing it?” It had to be taped down. It was really complicated.

In classic outcome-oriented thinking—which is true of so many things in life—first, we got rid of this robot. The line was stopped because the thing wouldn’t get installed correctly. You don’t need to stop the line; a person can fix it. There’s space for a person to go and correct it. Don’t stop the line.

Then it was, “Why are we even using a robot at all?” If you’re going to have a person fixing the robot, just get rid of the robot. And then it was, “Wait a minute, this part should just be deleted. Just get rid of this part.” You go through that process.

I think in this case we went through that process in 2 or 3 days. By taking that one flufferbot—or I can’t remember the name of it. It was something like that. Somebody is going to respond, comment on this, and say, “This is what it was actually called.” Just by going through that example with the team, they found 7 other versions like that—7 other similar opportunities where the part could be deleted, or the process or the way we were automating it could be changed.

Sometimes those interactions are really painful, especially if the person in that interaction with Elon is the person who spent the last 2 weeks coding up this complicated robot and defending their work, and then maybe wasn’t there the next day.

Yeah.

But the broader team was like, “Oh, yeah, I get it. I get it.” It’s not really about the flufferbot or the part. It’s about the overall system. There are other examples where, if you viewed them from another lens, you could delete, delete, delete, or simplify, automate, whatever.

So, now that you’re running a company yourself and you’ve seen this one example in a CEO of this really iconic, inimitable CEO, what do you take from it? Are there things that you leave behind? How do you find your own style?

5. Drew Builds Teams With Humility

Yeah. Even working within Tesla, I learned a lot from Elon, Zach, Jerome, Doug, and Craig. I think I was developing my own mixture of all of these individuals who were my role models, influencers, and peers.

What I value more than anything else in the organizations I build is humility and transparent communication, and a commitment to trying to fail fast. This is another reason why ego attached to your work product is a bad thing. If you attach a lot of ego to your work product, you’re more likely to take the positive side of any finding.

“This test result is a little funny. It’s not like it passed or failed; it’s funny. Maybe we need to retest it.”

Okay, I’ve got to stop you here, though, because—

You know what I mean?

But, yeah, totally. The stories you’ve told, Elon has been right. But I guess my experience of Elon is mostly from Twitter, where this doesn’t seem like a guy with zero ego who’s sort of dispassionately looking at the world.

I’m not talking about Elon.

I see.

Yeah, I’m not talking about Elon here.

Okay.

I’m talking about how I try to run my companies and why I reinforce the divorcement of ego from your work product.

I see. Okay.

I’m talking about what’s important for the people on my team and what I ask of them.

Uh-huh.

And that’s not a question of my style. I’m not talking about my style necessarily.

I see.

I’m talking about what I want to see in my team members. And the reason is—and I’m just providing another analogy—if you put a lot of ego into your work, you’re not going to find the problems in your work.

If you want your product to get to market and be the best in the market, and you want it to get into the market quickly and iterate quickly, you need to find the problems quickly. That’s where trying to detach yourself from your work product and being able to attack it as openly as anybody else could—as openly as Elon undoubtedly would, to bring Elon back into the conversation—is key.

That aspect of what I learned at Tesla, I try to bring to Heron Power, where I’m the founder and CEO now. Anyway, so that’s key. The other things that I value are a commitment to hiring people who are better than you, which I think also doesn’t always come naturally to people.

Oh, yeah.

People have a habit of wanting to throw their weight over their direct reports, and one of the ways in which they do that is by feeling better than them. One of the ways they feel better than them is by thinking they're smarter than them or whatever. But actually, you want to do the exact opposite.

You want to bring on people who are smarter than you, have more experience than you do, fill in the blind spots that you have, and really elevate them, because that's how you grow, too. I have a whole management philosophy that I've kind of developed over time. I'll send it to you. I'm forgetting—there's a bunch more, but those are...

Humility and hiring people who are better than you are 2 key ones. I'll think of some more as we keep chatting.

What about really aggressively pushing? Is that—I mean, that's something I've actually noticed that people who come out of working for Elon often have: a style where they push really hard. Do you also do that?

I do. I try to do it pragmatically, but I do. I think it's something I learned from my father. My father was an electrical engineer.

Uh-huh. I've met him, yeah.

Yeah, yeah, yeah. But—

Awesome guy.

Was your father—

My father, actually—it's funny—does consulting on electricity grids.

Yeah, yeah.

Yeah.

Yeah, I remember that.

My father was an electrical engineer and was also an engineering manager. He led divisions within Teradyne that delivered the products. He was always telling me, even before I was at Tesla, that when you develop a timeline, you basically have to do the bottoms-up thing, where the team says this is how long they think it's going to take, and then you have to bring your top-down view of what the business need is or what you think the team can do.

You sort of push as hard as you can for something in between, because if you give people 12 months, they'll take 15 months. If you give people 9 months, they'll take 10 or whatever. That's just human nature.

I try to bring pragmatism to it. I don't think it's worth denying some reality. I think that's where you run into problems if you're like, "Okay, well, we've got this large stamping tool, and it's going to take, whatever, 40 weeks to get it cut, delivered, and ramped up as a tool." I've been through enough programs to know that's probably not that compressible. Maybe you can make it 38 or something.

But then, when the team brings that as the reason why the project is 2 years long, you have to say, "Well, we could soft-tool it. We can work around that part for a lot of the early prototyping. We can find ways to compress the program that maybe add a little bit of cost, but don't require the project to be this long." I think people have this habit of long-pole-in-the-tent Whac-A-Mole, where they're trying to put up a pole to make the schedule be what they want the schedule to be. It's sort of your job as an engineering executive, or an executive of a company, to just keep cutting the poles down.

Wait, the long pole, though, right? Don't you want to cut down the long pole?

Yeah, yeah. But people just keep putting up new examples. It's like, "Oh, okay, so we figured out how to make the stamping not be your problem. Now it's something else." You just have to keep going through that a couple of times.

Mm-hmm.

You have to sort of train the team on how you're willing to let the schedule be developed. I think when people land on a long pole, anybody that isn't the long pole just rests and is like, "Well, we've got to wait for that stamping to show up, so I don't need my part to be done until 36 weeks from now, because my lead time is only 4."

It's like, "Uh-uh. No. If your lead time's only 4, do your work now, get it here in 4 weeks, rev it till it's perfect, and go help the guy with the 36-week lead time get it done sooner."

You know, it's funny—my experience with software is different, where I feel like what usually happens with software projects is, if you ask people how long each component is going to take, they way underestimate it. They're talking about the first 90%, the second 90%, the third 90%.

If you really do a bottom-up analysis and let a team do it, especially a junior team, you'll end up with this ludicrously short timeline, really needing to pad it, and then watching it expand over time. It's not usually taking a longer timeline and pushing them to compress it. It's more like trying to get people to be realistic.

I mean, yeah. With software, I totally agree. Software's totally different from hardware. I think with software, it's all about the list of integration milestones or functionality-demonstration milestones that you want to rally around, and architecting that list and sequencing it in a way where you're kind of getting the nights-and-weekends work done, but you're also focusing folks.

Software has a lot of hazards. You can basically fill time endlessly with more code or rearchitecting over and over again.

And that's why I think short deadlines are really important in software, right? You get people to cut scope and do the thing that actually matters.

Yes.

Because I started writing software in physical systems, the physical systems had real deadlines. The car needs to be driving for this event. We're going to do this.

With the Roadster, it was crazy. We had a 400-person event at an airport in LA. I remember Arnold went, and lots of the celebs went, and every single one of them was going to do a 0-to-60 drive. They were all going to experience the 4-second 0-to-60 in the Roadster. The date was fixed, and the software—everything—needed to work. It's incompressible.

Right.

The only thing we couldn't figure out how to deal with was keeping the whole system cool during all the back-to-back 0-to-60s. The vehicle technicians came up with this dry-ice rapid-chiller setup. The car would go behind the curtain, and they were doing this thing with dry ice to cool everything down, and then they'd send it back out.

We did all those 0-to-60s, and everything was fine except the magnesium-cast motor mount cracked at 300 or something. Luckily, the motor was pinned between the back of the battery pack and the rear trunk thing. When you'd floor it, the motor would slam into the battery pack, so you heard a thump. But it still worked.

Nice.

When you slowed down and went on regen, it would thunk into the trunk.

Uh-huh.

But everything was fine. Anyway, that was a fun example. But, yeah, you need integration milestones if that date is 10 weeks from now. This is something Craig Carlson taught me, and he was amazing at it. It's like, "Okay, what are we doing this week? What are we doing next week? What are we doing after that?"—just laser-focusing the team on that stuff.

Yeah, this is a total aside, and I do want to get to Heron Power. But I had this interaction with you, I think, 10 years ago, where you were talking about luxury cars and what makes a luxury car. It was things that were so stupid, like having a bigger piece of metal or something, that it completely made me not ever want to buy a luxury car and always drive cheap cars. Did I take the right takeaway from that conversation?

I'm trying to remember the conversation.

Well, what makes a car fancy? Are they actually faster?

Well, okay, what makes a luxury car? I think some of it is the choice of metal. There are some interesting examples, like the NVH package—the noise, vibration, and harshness package—which is basically the content that you add to absorb the sound that the tires make on the road. They're the primary sound generator. And then, in an internal-combustion-engine car, the engine makes noise.

So you basically spend money to absorb that sound. In a budget car, you spend not a lot of money, and in a luxury car, you'll spend, I don't know, $400 or $500. The difference between the regular sedan, where you spend $50 on NVH material, and the luxury sedan, where you spend $500 on NVH material, is the difference between a quiet cabin and an acceptably noisy one. That's an example.

But that's pretty good, yeah.

Yeah. But that's the kind of thing you're doing with a luxury car. Another example would be closing the door. A luxury car sounds nice when you close the door. It's got this satisfying thunk. It's like, "Ah, this is a well-built car. I'm happy I bought this car."

Yeah.

Right? And a not-luxury car is like, the car door closed and stayed closed when it was supposed to stay closed in this accident and that accident. That's what it did.

Uh-huh.

The difference between the satisfying sound and the not-so-satisfying sound is the stiffness of the structure of the door and how many lips are on the seals. All of these things contribute to that thunk and the feeling of solidity in the door. Again, it's basically weight—materials that you've added. You can always optimize it, but the thing is, if you take the optimized luxury car, it just means the not-luxury car will be lighter. That is kind of the thing you're trading. You're trading some mass. I think it is mass.

Uh-huh.

And some cost for these creature comforts or perception aspects of the vehicle.

And are you a car guy after all these years of working on cars? Do you look at cars and think about—

No. I never was much of a car guy. I love—I will never drive a non-electric car, or at least not happily.

Yeah.

I love how much quieter they are and the responsiveness. I can understand the linearity of steering and what good damping feels like on the suspension system. I can appreciate all of these things. But I was never going on the weekends and rally driving on a racetrack or really following that closely. Of course, when I was at Tesla, I was following all of the competitors, but as soon as I left Tesla, I really wasn't.

Although you do kind of drive like a maniac, is that because you're confident in the specs of the car, or was that before Tesla?

Yes.

Tesla definitely had an influence. I spent a lot of time on racetracks at Tesla developing the traction control and the stability control. I was responsible for that aspect—how the car stays on the road from the perspective of braking, steering, and the propulsion system.

I've spent time on dry racetracks and on slick tracks where you put water down to understand what driving in the rain is like. I've been in Alaska, Minnesota, and Sweden driving on snow and ice. You learn how to interpret where you are relative to the edge of stability on a vehicle just from the feedback you get through the steering wheel and the sound of the tires. You learn these things.

Since I did it for so much time, and I'm still driving those same cars today—I drove a Model 3 here today—you know the car really well. You can understand its limits, and you can have fun.

Nice.

And I'm always chronically late, so—

I think so.

That's another reason why I drive a little bit faster than most people find comfortable.

What about autonomy? You were one of the first people I knew talking about autonomous vehicles, and then it got really hot, and you were right in the middle of it. How has that—tell me about what you were thinking when it started, and then how has the trajectory surprised you?

6. Tesla Takes Autonomy Beyond Physics

Yeah. One of the things that JB asked me to do in 2012 or 2013 was start working on this problem statement. We needed to be competitive. The Model S needed to be competitive with all these German cars and have adaptive cruise control and a lot of these active safety systems.

I went out and met with Bosch to see their radars, talked to Mobileye, understood what was out there, and hired some people. I hired some controls folks to start developing the autonomy stack. At the time, we didn't really call it the autonomy stack. We called it adaptive cruise control. But I was developing the car-driving-itself functionality and was responsible for Autopilot in the first year or so as we were putting the team together.

I didn't have your background. I didn't have any exposure to machine learning at that point, and I didn't have any exposure to what computer vision could be. I was, and still am, very much a physics-first control-systems person. I was an electrical engineer as an undergraduate, and I had a concentration in signals and systems, which is basically control theory in electrical circuits and software.

I came at it from a physics-based perspective: If I'm going to have a car drive itself, I'm going to create a physical model of the world around me. I already can fully model the car and plan optimal control for the car from here to there if I understand where the car needs to go. That worked really well for adaptive cruise control, but it pretty quickly became clear that it wasn't going to work well all the way to the end, even back then.

What was interesting to me in that whole period of time—this was 2014, or I think it was 2014—was that Mobileye had a demonstrator car to sell its cameras. Their demonstrator car was an Audi A8 or something like that, and they would have the car drive itself using their single-camera, monocular image-sensing stack.

Wow. Just one camera?

Just one camera.

Wow.

They'd have the car drive itself. I think they also did some fusion with the radar or something, so they had a camera and a radar, and they would have the car drive itself. It was a really impressive demo.

Elon got a demo from the Mobileye CEO, and we were doing adaptive cruise control. We weren't talking about the car driving itself, nothing like that. We were developing the platform of the vehicle to meet the Euro NCAP crash-safety standards, or active-safety standards—detect pedestrians and cyclists, automatic emergency braking, and all this stuff with some radar and some camera.

Satish and I were developing this platform that could meet all the European requirements and also have the car do adaptive cruise control. This was early 2014. Elon comes out of this Mobileye drive and says, "We're going to have the car drive itself." Then it's, "Forget about meeting the European requirements. We're just going to have the car drive itself fully."

We're like, "This is the hardware we got. We're going to use this hardware?" And he's like, "Yeah, that hardware can do it." It was like, "Okay." That's when the idea started that people have 2 eyes, they drive, so why can't we drive the car with just vision? It really came out of that Mobileye drive.

I was in the position to meet with Elon every week and hear, "When's the car going to drive itself from California to Boston?" That was his question in every meeting to me and the early Autopilot team. He'd ask, "Can it be next week?" That was the question.

This is the amazing value of the distortion, the reality distortion field, that folks like Elon and others bring to a conversation. It's not, "So how many years is it going to take?" It's, "Can it be tomorrow? Why can't it be tomorrow?" That's the question.

The team was ramping up. We were getting the computer-vision people in and all the people who could start thinking about how the car could drive itself. Every week, that was the question. At some point, I'm like, "I'm an energy nerd. I'm a physical-systems person. I think I'm going to go do something different."

That's when I started working on Tesla Energy. When the first Superpack, Powerwall 1 came along, the team was 3 engineers. I started doing that, and Autopilot moved on to a series of leaders who tried to work with Elon through the "Why can't it be tomorrow?" reality. We're still in the "Why can't it be tomorrow?" phase a little bit, although it's really impressive.

Really awesome, yeah.

FSD is incredible.

Totally.

The team is amazing. I actually interviewed Ashok when he joined, because I was part of building that early team. I mean, Ashok is amazing.

The team is amazing. The Tesla team is incredible. What they've done to go from heuristics and physics-based optimal-control planning to nothing but nets—to go from pixels to actuators—is incredible.

Could I have predicted that? No. I just didn't have the context at all. I knew how to do estimation and Kalman filters and all of these things that are in control systems, where you're working with situations where you don't have all the information, and you do have to do, effectively, recursive, on-the-fly state estimation. I was familiar with all of that type of work, but I just had—

I had taken an AI class in college. I knew about Bayesian trees and all of that. I think we took the same classes. I knew about it, but I had no idea where it would go. I could never have predicted it. Kudos to the team for sticking with it and getting to where they've gotten. It's incredible.

Totally. All right, so let's talk about the electricity grid.

Sure.

I was going to say, you're 2 sentences in, and you haven't said data center.

I know, I know, I know. Sometimes I need to stress the other things—

Nice.

Because AI is said so often. But yeah, we have load growth in AI, which, by the way, when I started Heron, wasn't as obvious as it is now.

Totally.

But—

But okay, maybe backing up, because I'm actually really not an expert on this, and probably people watching or listening are even less expert. What's the problem here? I turn the lights on, and the lights consistently go on. The electricity grid seems pretty stable.

We complain about power prices in California for some reason, but of all the things in my life, the electricity grid seems pretty stable and okay. What's the issue?

7. The Grid Needs New Capacity

Yeah. Just an interesting fact about California: The retail rates are really high in California, but the average household electricity bill in California is middle of the pack in terms of average electricity bills in the country. That's because people don't actually use a lot of electricity in California. Very few people have air conditioning, everybody's on fancy LED lighting, and it's a moderate climate, right? It's not like Texas.

While the retail rates are high, the average cost of electricity to a homeowner is middle of the pack, which is just a fascinating thing. It actually means that if you go back 10 years, before the retail rates were high, electricity was super affordable in California, which is contrary to the thesis.

I'm not here to defend the retail rates of California. I think they need to go down, and they are going down as we bring more load on. Maybe that's the point. From the 1970s, maybe 1980s, until just a couple of years ago, there really wasn't electricity load growth in the United States and in the Western world. That was because of energy efficiency in all of these technologies: lighting, air conditioning, and computing as well.

Data center load growth was not very much, and actually, as a percentage of total end-use electricity, it did not go up very much from the 1990s all the way until the 2010s, because the efficiency of compute kept improving dramatically. Not just compute, but storage—everything: networking, how to cool them, everything just kept getting more efficient. You kept adding more compute, but you kept making it more efficient at the same time.

So the overall load of the United States hasn't grown since the 1980s?

Yeah. From the 1980s, the 1990s, the 2000s, and the 2010s—through that whole period of time—average load growth was 1% or less. So there was very little growth in the system.

What that did was, I would say, almost calcify the industry that supports the utility sector—the supply chain to the utility sector—because if you're not in a growth industry, what are you doing? A lot of the talent left the States and went to areas that were growing, like Japan, Korea, China, and India, because that was where demand was going gangbusters, where new things needed to be deployed.

The new energy technologies were being developed and improved in other places. The other thing that added to this calcification is that the utility incentive model was built around the fact that electricity use isn't always going to go up.

The way utilities make money is not by increasing the sales of electricity, but by getting a guaranteed rate of return on the CapEx they deploy. Now imagine this period of time when there's no load growth. The utilities don't have a lot of reason to deploy a lot of CapEx. What's going to happen? The suppliers are going to charge more money for what they are deploying, because that's how the utilities get a return on investment. This is all natural incentives. It's all just incentives.

So the utility is incentivized to spend more money, then?

Yeah. The utilities have an incentive to deploy more CapEx year over year.

So I would think we would maybe make some awesome projects and get all kinds of wild new technology.

I mean, they're regulated, so they have to justify their rate case to the public utilities commissions, and the public utilities commissions are going to say, "Hey, why are your expenditures outpacing inflation?" They have some pushback.

Okay, so I have to ask—

But—

Sitting here in California, why are the rates going up, then? It sounds like nothing's changing. What happened?

Yeah, yeah. All right, let me just finish this thought—

Okay.

So you're in this environment where it's stagnant. The amount of hardware being deployed is almost maintenance CapEx. The utilities are incentivized to spend more money on that maintenance CapEx because that's how they make their return.

Then the suppliers that are shipping a flat number of units are just getting fat and happy, charging more for that same unit over time because that's actually what the utility wants to do: spend more money on it. That whole supply base became siloed and a little bit uncompetitive.

If we're going to have electrification growing at the pace that it's now growing—3%, 4%, 5% per year—I think we need to triple the total amount of electricity generated and consumed if we want to have an all-electric, sustainable energy economy, which I'm really passionate about.

We need new suppliers. At Tesla, we used to call it "towing a dinosaur," which is kind of what's happening. Or a boat anchor. The supply base is almost like a boat anchor on the path toward electrification.

Wait, sorry. I keep interrupting, but why would electrification need triple the amount of electricity? Isn't the main way we consume power through electricity? Are there so many gas cars that that's going to be the main thing?

Well, electricity represents a third of end-use energy consumption right now, which is the simplest way to think about it.

Really? So gas is more, then?

Yeah. Well, you have fossil fuels in industry and buildings. You have fossil fuels in transportation. You have fossil fuels in the whole chemical sector, including fertilizers and building materials, plastics, and composites.

I worked on a paper called "Master Plan Part 3," which was effectively about whether you can build a sustainable energy economy. What is involved? Do the resources exist? Do the technologies exist? Is it an economical thing to do? Short answer: yes, you can. But it ultimately means that we will produce 3 times as much electricity as we do today.

Wow.

But in doing so, we will have, effectively, a fully sustainable and more affordable energy economy in the future, when we do that.

Okay, so why did California power rates go up?

There are a lot of weird reasons. One is the cost of wildfire abatement and the liability for prior wildfires. It's probably the single biggest individual adder in California relative to other places.

The other is a little bit what I'm talking about: the incentive structure and the cost of the components that go into maintaining and building out the grid inflating over time. The other is that California is actually kind of a big state, and it's pretty rural on average, so there's actually a lot of distribution lines relative to the number of people.

And we do have, and the utilities have a mandate to serve the rural folks in northern Mendocino County as much as they do in downtown San Francisco. You compare that to Delaware or Maryland, and the ratio is not in favor of the California utilities.

California also got built out long enough ago that a lot of stuff needs to be repaired and replaced, so that's a unique thing in California. Another one, and this is probably one of the most controversial, is that for a long time California had net energy metering, or NEM, for solar. That meant that you, as an end customer, got compensated at the retail price of electricity for putting your solar onto the grid.

You don't get compensated at the same price that the utility pays to send you electricity from the grid. So, conceptually, when the sun is not out, the utility pays the spot-market price of electricity, which is whatever it is, plus it has to pay for transmitting the power from where it's generated to you and then distributing the power to you. The utility pays for all those things, and then you pay a retail price of electricity for the power that you consume. But when you go the other way, the utility has to buy it back at the retail price, not at the spot price, whereas it could have bought power from the market at the spot price. So effectively, they're paying you way more than they would pay a generator anywhere else.

So they're not making money they would have otherwise made.

Well, the simple way to think about it is this: If the way that you amortize the cost of your infrastructure is all of your costs divided by all the kilowatt-hours that you serve, the kilowatt-hours served by rooftop solar don't get put in the denominator. Those fixed costs get divided over fewer kilowatt-hours served, and so the fixed costs go up.

Rooftop solar is so popular in California that it has a significant impact on the denominator.

Mm-hmm.

The denominator got a lot smaller, or didn't grow. So the numerator, even if it stayed fixed, means the per-kilowatt-hour cost has to go up for the same infrastructure. The thing that's changing that now is that there's a high penetration of rooftop solar, so fewer people want to have it. NEM has been changed so that you don't get compensated in the same way, and we have load growth again.

We have electric cars in homes. We have heat pumps in homes. We have data centers, maybe even in homes—Xfra [?] is this idea of data centers in homes. All of those things have meant that, for the first time in many decades, PG&E actually had the numerator go up.

Mm-hmm.

The numerator of that equation went up, and so that's why the rates didn't go up in California. They think the numerator is going to go up next year, and they think the rates will even go down.

If you listen to Patty Poppe, CEO of PG&E, whom I know, Patty is great. Hi, Patty, if you're listening. That's what she's saying: If we keep having this load growth and the regulatory structure doesn't change, we're actually going to reduce rates in California, which would be great.

So what are you going to make? I would have thought maybe you look at this and think, “Okay, I should make solar panels or some kind of new energy generation,” but that's not what you're doing.

8. Power Electronics Rebuild The Grid

Yeah. What we're trying to do is provide alternative, let's say, software- and silicon-based solutions to the largely mechanical, passive solutions that utilities, people building power plants, and people building factories have to choose from today.

Going back to why this opportunity exists, there wasn't a lot of load growth, so there wasn't a lot of need for innovation in that supply base. I saw it directly. I was responsible for deploying Megapacks, and over 10 years, we did 4 different versions of utility-scale storage: Powerpack 1, Powerpack 2, Megapack 1, and Megapack 2.

In the course of 8 to 10 years, we iterated and dramatically improved the cost, made it easier to install, and did all of these things. But over that same period, we were still interconnecting to the same annoying stuff that was hard to get from the same suppliers—stuff that was largely made abroad and wasn't improving.

So what are these things? What are we talking about?

Yes. It's switchgear that controls the flow of electricity or interrupts the flow of electricity. Think large mechanical switches that take hundreds of milliseconds or maybe even seconds to open. So, switchgear is one of those things.

Is this a physical thing?

A physical object.

Is it like those blade switches from Frankenstein?

Yeah, blade switches from Frankenstein. But imagine they're motorized.

Nice.

Sometimes they're inside a vacuum, or maybe they're in some other inert gas that helps suppress the arc.

But I guess I might look at that and say, how could you make a motorized blade switch better? What more could it do? Does it let power through or not?

Well, yeah. I'll get there.

Or pick any of these.

No, I'll get there. So switchgear is one. If you go to somebody building a data center or a factory, one of the longest-lead-time things they have to buy is the switchgear.

So why? That's maybe the question, because it is just simple mechanical stuff. And then transformers, and all the things that are on the switchgear and the transformers, which are meters and sensors for voltage and current. All of that gear is kind of this cottage industry, partially because the utility industry is really balkanized. There are 3,000 utilities in the United States, and they all have their own rules and requirements.

When they were all formed, there wasn't a federal highway system. There wasn't the internet. So they all have their down-the-road supplier, and they all develop their own requirements and their own custom stuff.

That is the ecosystem we're looking at: this world of passive steel and oil, copper-wound transformers, mechanical switches, and ancillary components that have a little bit of software written around them to control them. That is what we're trying to provide alternative solutions for, and how we're doing that is by using wide-bandgap power semiconductor material in the form of really robust, capable, proven devices that have been deployed in electric vehicles and grid-scale batteries to replace these mechanical objects.

Sorry, what is a wide-bandgap? Is that a wire?

A wide-bandgap semiconductor. You've seen an integrated circuit. You've seen a GPU, right? Now imagine that highly engineered, multilayered device isn't engineered to store 1s and 0s, move 1s and 0s really quickly, and do math and all that kind of stuff, but instead is designed to optimally conduct and interrupt the flow of electricity and block voltage. That is what power semiconductor materials do.

Wait, let's talk about that, because it makes sense to me that compute would need a really complicated system, but I feel like interrupting power and letting power through feels very simple. What are you actually doing?

Did you take semiconductor physics at Stanford?

I didn't, and I'm sure people watching this have, so you might need to back up.

Yeah. It's pretty awesome. The way all semiconductor devices work is that you're selectively choosing or doping different regions of the base material, which is the semiconductor of choice.

In GPUs and CPUs, it's silicon. In some power devices, it's not. It's silicon carbide or GaN or other materials. You start with some substrate, and then you dope different regions and etch different regions to selectively channel or control the flow of electrons, or even enhance the number of electrons in different regions of the material.

Ultimately, it's a very mechanical process for how you make these semiconductor devices. You have masks for photolithography, and you're doing vapor deposition or atomic layer deposition of atoms—doping atoms, atom by atom, into the matrix of the substrate. You're actually—

And that's, again, like on GPUs, the flow of actual electrons is really light, right? These are really small currents at the level of the chip. I mean, if you add them all up—

Because it's 1 volt and it's like a kilowatt, so it's 1,000 amps. Nuts.

But, I think on each individual part of the chip, it's a pretty small flow, right?

No. That's actually how the power density and the compute performance over time have improved: you're driving higher and higher current density in the chips and figuring out more effectively how to pull the heat out, the heat flux out. Effectively, all of the operations happening in the chip—especially at high frequency—are, every clock cycle, charging a capacitor and discharging a capacitor. It's all of these parasitic capacitances in the device, and then you're actually moving electrons through a resistive medium, the resistor, which generates heat.

You have capacitive discharge and charge, and then you have the actual current flow. All of those things are generating massive amounts of heat. That's what data centers basically are: you take a gigawatt of power and convert it into something like 700 megawatts of heat, and you spend 300 megawatts to get that heat into the atmosphere. That's actually what a data center is.

The 700 megawatts of heat is almost all inside the chips. Maybe 8% is lost in power conversion on the way to the chip, maybe 10%, depending on the architecture. So you get a gigawatt at the grid, you get 250 megawatts to 300 megawatts of cooling power, and maybe 650 megawatts of useful compute.

I guess your transformer is going to be more efficient than the 10% loss. Is that—?

Yeah. We can reduce the loss from grid to chip by about a factor of 2, so we halve the loss. For somebody building a gigawatt data center, that means they can get 35 megawatts more useful compute out.

I see.

Something like that.

Cool.

Yeah.

And the switches—presumably, you can switch faster, or what's the—?

So, back to what's happening at the semiconductor device: wide bandgap means that, intrinsically, because of the way the material works, you can block more voltage per unit of material, per micron or nanometer. You can also be more efficient; there's less loss. The highway that the electrons travel through when the semiconductor is acting as a conductor is more efficient.

Because you can block more voltage per unit of material, you have less parasitic capacitance between where the electrons are flowing and whatever your cooling device or the bottom of the device is. So you can actually switch the device faster, because that parasitic capacitance—that charge—has to charge.

Every time you imagine lifting the output of the power semiconductor, what you're doing is—this is the top voltage that's connected to the device, and this is the bottom of the voltage that's connected to the device. The output of the device needs to swing from the lowest voltage potential to the highest voltage potential, and there are all these parasitic capacitances in the material, which are physical. When you go to these wide-bandgap materials, you can make them physically smaller, and that means they can switch faster. Or, for the same loss, they can switch much faster.

That's ultimately what silicon carbide and these other materials enable. Gallium nitride is even better at that. So it's a materials-science thing. You're trying to find materials that can block more voltage, are more thermally conductive, and are low-loss when they're in the on state. They can conduct electrons really well when they're trying to conduct.

A semiconductor is either conducting or not. That's what makes it a semiconductor. You can control that with the gate voltage on a transistor, right? You can control whether the channel in the device is conducting electrons or not. You turn that on and off with the gate voltage. There's a bajillion transistors in a GPU—I don't even know how many, trillions or whatever—and in a power device, there's 1.

Right. Right.

But it's 1 highly engineered device. What silicon carbide is doing is allowing you not to make a 1-volt transistor like you have in a GPU, but to make a 10,000-volt transistor.

Wow.

In fact, it's actually smaller than the GPU, and it's able to block 10,000 volts. That's what materials science has enabled over the past 3 or 4 decades.

So now, when you're thinking about this mechanical switch—

Mm-hmm.

—that we were using to control the flow of electricity, which literally physically needs to move the distance of my arm—

Yeah.

—because otherwise we're going to have an arc in free air—

Yeah.

—you can now do that inside a material, a solid-state material. You can start and stop the flow of electricity across 10,000 volts, all within a solid-state material, with nothing moving. That is what power electronics is. That's the power of power electronics.

So are you imagining that these big transformers that freak people out—

Yeah.

—that we see around San Francisco and around the world, do you imagine those become tiny?

Yeah. Our core first product, the Heron Link, has a transformer that's 100 times smaller volumetrically per unit of power. The way we're doing that is by switching the— You still need to go through an isolation stage when you're going from a high-power or high-voltage grid to a computer, because otherwise, if you had lightning or something like that, you might see the voltage from the grid side come over to your computer. That would be very bad.

So you need to galvanically isolate the high side from the low side. But rather than doing it at 60 hertz, which is how we do it today in these giant, oil-filled gray boxes, we can do it at hundreds of kilohertz because these devices can switch so fast.

You take that 60-hertz power and use these power devices to chop it up to a super-high frequency, hundreds of kilohertz. You run it through a high-frequency transformer that, because of the way the physics works, can be much, much, much smaller. It's actually inversely proportional: as the frequency goes up, you can make the transformer that much smaller. It's really an energy-product thing.

The way you store the energy in a transformer is— A simple way to think about it is that at 60 times a second, you can move a packet of energy through it, versus 100,000 times a second. That's the simple way to think about it. If the voltage is the same and the energy is basically current at that voltage moving up and down, if it moves up and down 200,000 times a second instead of 60, it could be 200,000 times physically smaller.

That's the principle at work. The objective is to make the transformers way smaller. These giant switchgear objects that are interrupting the flow of electricity now have power electronics on either side of the transformer, and they can natively interrupt the flow of electricity. You don't need another mechanical switch, because the passive transformer can't do that. It's passive.

It can't change the voltage from one side to the other except in a static ratio. It can't stop the flow of electricity; it can't do anything. So you need to have these large mechanical arms and levers do that work for you.

So do you imagine replacing all the infrastructure with these new devices?

Yeah. Over time, I think as we've got 50-year-old stuff out there and it needs to be replaced, you're going to replace it with active, solid-state transformers instead of the passive ones, because you can get all of these interesting functions integrated into 1 object, and they're smaller. They're easier to service and easier to maintain. They're not flammable, like the oil in those transformers. That oil lights on fire.

I don't know if you've seen it, but when I lived in Bernal, more than 1—

Wow.

—transformer lit on fire down the road. You have a power outage. It's just flammable oil, and you've got this thing that's getting hot inside. If you don't maintain it—and even if you do maintain it—sometimes these things happen.

There are lots of reasons why you'll use it to replace the existing infrastructure, but our primary focus right now is that the grid is growing really quickly. There are people building solar facilities, people building battery facilities, people building factories, and people building data centers, and they need and want a more scalable, better solution. So that's our first focus: the folks building the new stuff.

Do you have a point of view on data centers and their impact on the grid? I feel like that's become pretty controversial—

topic lately.

9. Data Centers Can Support Grids

I think properly designed data centers can be grid-positive because data centers are the best utility customer. They are baseload. They consume a ton of kilowatt-hours per kilowatt. If you look at your house, where you have a 200-amp service, which is like 40 kilowatts, you're probably never using more than 10 kilowatts peak, and on average, you're probably using 1 or 2 kilowatts, maybe even less. So the utility has invested in the infrastructure to serve you 50 kilowatts, and you use 1 kilowatt on average.

A data center—you invest in all the capacity to deliver a gigawatt, and the data center uses 800 megawatts on average. They're a way better customer. If you think about the simple equation I described before—the numerator is the total cost to keep the grid healthy, and the denominator is the total kilowatt-hours served—the data centers are the best customers. They're basically like aluminum smelters or steel mills; they are the customers that make the grid affordable for everybody else.

The way that could not end up being true is if the utilities, because of their incentive structure, are building more than they need to serve that data center, or they're amortizing the costs of connecting that data center more on the residential ratepayers than they are on the data center. I actually think the data center should just pay for the infrastructure that comes to them because the return on investment of these data centers is so high, and that infrastructure actually isn't that expensive compared to the GPUs. But if you think of data centers as a utility customer, they're the best ones, and if you actually look at the numbers, the states with the highest penetration of data centers overwhelmingly have had the lowest electricity rates and actually have had rates reduced.

So you actually think data centers will cause electricity rates to go down?

I do.

Interesting.

And the other thing is, because of the uniqueness of training, data centers need to have some energy storage to participate in absorbing all of that training ripple. Once you put energy storage at a data center, the data center can support the grid through peak periods of high load. They can stabilize the grid frequency and voltage rather than disconnecting whenever there's a problem. So they can do all these things to be grid-supportive, in addition to being an awesome customer.

That should also drive down the cost of the utility. Whereas historically, they've actually done some things on the opposite side, the rectifiers they used have had flicker problems. Utilities have had to clean that up with extra hardware. Just this week, 3 gigawatts of data centers all turned off at the same time because they didn't have storage set up to do that. There was a little bit of a grid wobble, and they all disconnected, and that can cause a liability concern. So historically, data centers were maybe a grid liability from a stability perspective. In the future, I think they can be an asset, and they can drive rates down.

What do you think about nuclear power?

I'm down for nukes. I don't really have any aversion to nuclear power. I'm happy that Diablo Canyon was recently extended here in California. I think that with proper engineering controls, nuclear power plants can be safe, and of course, you need to maintain them and do all those things.

I would love to see the US focus on its waste plan, because there are plans in place in other countries, and we don't really have one, but we should have one. The history of nuclear in Asia has been a history of nuclear power plants continuing to become more affordable to build as countries get reps. Costs, in thousands of dollars per kilowatt, have progressively come down. I think the US hasn't seen that trend because we stopped building nuclear, but I'm confident that if we start building nuclear again, we can find ways to do it more affordably, like has happened in Asia. They did that without any safety problems or anything like that, so it's not like you're trading safety for cost.

I think nuclear is needed as part of the mix. You need some baseload renewable power. I'm actually a big geothermal fan. I think geothermal is going to give nuclear a run for its money in terms of affordable baseload, 24/7 renewable power.

What about just solar plus batteries? When I look—

Yeah.

—at the simple math, it seems like that's actually the cheapest way to get baseload. Why isn't there more focus on that?

Solar plus batteries can absolutely provide baseload power. One of the things we did was do it on islands—a lot. We took some islands in American Samoa fully renewable with solar plus storage. You can definitely do it.

But when you think about the LCOE—the levelized cost of electricity—with solar plus storage, when you consider that worst-case period, the winter with that bad storm or the number of cloudy days, if you really want to have 100% solar plus storage, the life-cycle cost of energy is still basically 10 cents. Maybe a little bit more, depending on how good the solar asset is. Maybe you can get it to 8 cents if it's an incredible solar asset, like a desert, and you don't have to worry about the winter. So in the Middle East or something like that, you get to 7 or 8 cents or something like that.

But geothermal can beat that, I think. That's definitely what Fervo's trying to do. They're trying to get down to 6 or 5. And when you look at the nuclear plants we have today that are fully depreciated, they're 2 or 3 cents. The question is, can you build new nuclear and beat that? The Georgia facility we just turned on was, I think, more than 10. Could we find affordable nuclear? Could we develop more affordable nuclear that gets down below solar plus storage?

Is that happening in China, or is the government subsidizing it? China's actually ramping up nuclear quickly, right?

Yeah. South Korea has built a lot of nukes, and they've been able to get to this LCOE; China, too. Whether they subsidize it or not—it's such a tricky question, right? There's so much government support in the raw-material industries in China. So you go and build a nuclear facility, you're benefiting from government support in the concrete, in the steel, in the fancy metals. Whether they directly subsidize the nuke or not, they've subsidized everything that went into the nuke in some way or another.

Okay. Your new transformer is, like, an order of magnitude, maybe—or you're saying double the efficiency?

Mm-hmm.

Do you feel like that enables new uses of transformers that people aren't using today?

Um.

Or, if that's not the intention, that's fine. I was just curious.

10. One Transformer Replaces Three

Well, no. I like the question. What we're really doing is changing the thought process of how you'd go from medium voltage down to the load.

If you're building a factory right now, you might actually have a number of transformers in series. You probably have some main interconnection to the transmission grid that's 300 kilovolts or something like that. There's a transformer there that goes from 300 kilovolts to something lower, like 13 kilovolts or 34 kilovolts. You'll distribute that medium voltage around, and then you'll have another transformer, maybe to 480, and then you'll have another transformer from 480 to 208. You kind of end up with these cascading series of transformers and protection switchgear and all this kind of stuff in a row.

With this power-electronics approach, you can elegantly do it in one stage and also have multiple different output voltages. So you're just simplifying the build-out of factory infrastructure, electrical infrastructure, energy-facility infrastructure, or data-center infrastructure. I think that's a good thing because ultimately, if we are going to triple the electricity sector, we need to be able to kill three birds with one stone, right? We're just doing more with less.

Mm-hmm. I've been playing around with building little robots in my garage.

Yeah.

The motors you need actually have a significant enough power draw that they can melt the little wires and cause problems. I've been using a little transformer off the battery to get down from the 24 volts that the LiPo battery puts out to 5 volts.

Yeah.

And that thing is actually surprisingly heavy. Are there mini versions of what you have that could be used in consumer electronics?

Yes.

Definitely. I'm looking at one right now that's probably, like, a 200-watt, 120-volt—actually, it's a 240-volt AC-to-15-volt output. It's powering the laptop over there. That's an example of a solid-state transformer, and that's an Apple AC-to-DC adapter for a MacBook, right? That thing is probably switching 800,000 times a second, power semiconductor devices inside of it.

But that's, like, $60 probably, even if you bought an off-brand one.

Yeah.

Shouldn't it be cheaper?

It's definitely cheaper than if you were doing it at 60 hertz. That same thing, if it were 60 hertz, would be, I don't know, 10 times bigger, way, way more—

Uh-huh.

—and cost a lot more.

Interesting.

It's sort of a grass-is-greener phenomenon. That is much smaller than it would've been for the same power 5 years ago, also with power electronics. I like to pull out this 70-watt, tiny GaN adapter.

Oh yeah, let's see it.

It's 70 watts, with 2 outputs, USB-C. It does 5 or 6 different voltages, and it power-shares from one to the other.

Wow.

It's tiny.

How many amps can it do? Oh, it plugs right into the wall. Cool.

Basically, what we're doing at Heron is, in the same way that this technology has made these consumer-electronics chargers tiny and super capable, we're doing that but for industrial electronics. Megawatts instead of 70 watts. So 5 million watts is the rating of our product instead of 70 watts.

Awesome.

Same concept, though.

Anyway, what's inside this?

So, yeah. You've got an isolation transformer that's probably the size of my pinky. That is isolating the 240 volts AC or 120 volts AC from the 5 to 15 volts DC.

On either side of that isolation transformer, you have these power semiconductor devices. The die area of each is probably, like, 10 millimeters by 10 millimeters or maybe 20 millimeters by 20 millimeters—probably more like 5 millimeters by 5 millimeters.

I want to take it apart and look at it.

Yeah. You have, like, 5-by-5-millimeter die-area power semiconductor devices that are GaN, switching the voltage waveform from the grid 1 million times a second to make the DC across that transformer.

One of the things I've noticed lately is that my USB-C cables power at wildly different amps. So I got little meters to measure how much power I'm getting from my USB-C cables. Everyone should buy them. They're amazing.

You realize that some of the cables are actually limiting the amount of current that you get. Then I realized it's actually upstream: that converter. Different ones will put out different amounts of amps. So I really wanted to get good ones where I could—

Juice it.

—put a ton of charge into my laptop and charge it super fast. But I couldn't find ones that actually got to the top of the USB-C rating. Should we expect better converters in the near future?

Yeah, I think it's actually happening in tandem. The batteries also have to become more capable because they're also a bottleneck. I like to think of it this way: Have you heard the parking-lot problem?

No.

Lithium-ion batteries can be thought of as a giant parking lot at a stadium. If you imagine that the parking lot at a stadium has exactly the number of spots as there are people going to the stadium—it's got 60,000 spots—if the parking lot were set up as a square, it would take forever for the 60,000th person to figure out where to park.

Totally.

But if the parking lot were set up as a line and the person could literally just drive along until they saw the hole, they could find their spot really quickly. Now, the line doesn't really help you get to the stadium. You need something to get them from the line to the stadium. But that's effectively the battery problem.

There are parking spots for lithium ions in the anode and the cathode, and they can only get to their parking space so quickly. Making a linear parking lot in a battery is actually super expensive because it's a ton of surface area. So you're always trading off this 2D/3D problem.

Is that why the charging slows down so much?

Yes, exactly. You can find the first spot easily in the parking lot. You can find the last spots really hard.

Wow, that's a great analogy.

It's the best analogy for lithium-ion batteries.

All right. Maybe a good spot to end. Thank you so much, Gabe. Thanks so much for listening to this episode of Gradient Dissent. Please stay tuned for future episodes.