China is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer

Summary

Casey Handmer, Caltech PhD and founder of Terraform Industries, joins Dwarkesh Patel to discuss the energy landscape powering the AI revolution. Handmer argues that solar energy, not natural gas, will ultimately power the hundreds of gigawatts of new data centers required for AI — and that the US can compete with China on energy if it acts decisively. His central claim is contrarian: solar panels follow a Wright’s Law learning curve of 43% cost reduction per doubling of production, a rate that will make solar unbeatable within years.

The conversation covers why hyperscalers currently choose natural gas (speed and convenience), why that approach hits scaling walls at 100+ GW, and how off-grid solar-powered data centers of 50,000 acres could bypass the broken electrical grid entirely. Handmer paints a vivid picture of an “energy singularity” where solar becomes so cheap that data centers produce excess power, reversing the grid relationship. He also discusses China’s geopolitical energy vulnerabilities, environmental regulations blocking clean energy in the US, and the philosophical endgame of turning solar energy into silicon computation — eventually culminating in space-based “computronium” wafers.

The discussion also touches on Terraform Industries’ work on synthetic natural gas from solar power, the economics of AI data centers, and why GDP is a poor measure of AI’s true economic value.

Highlights

”Solar’s 43% learning rate is astonishing”

yt-dlp --download-sections "*18:00-19:00" "https://www.youtube.com/watch?v=3cDHx2_QbPE" --force-keyframes-at-cuts --merge-output-format mp4 -o "3cDHx2_QbPE-18m00s.mp4"

“The Wright’s Law coefficient for solar is 43%. For every doubling of cumulative production, we get a 43% reduction in cost. And roughly every two to two and a half years, we’re doubling production and dropping the price by 43%. So it’s roughly 15-20% cost reduction per year.” — Casey Handmer, 18:00

”The grid is the poster child for Baumol cost disease”

yt-dlp --download-sections "*26:00-27:00" "https://www.youtube.com/watch?v=3cDHx2_QbPE" --force-keyframes-at-cuts --merge-output-format mp4 -o "3cDHx2_QbPE-26m00s.mp4"

“The delivery cost is perpetually on the brink of bankruptcy. Every additional solar panel, wind turbine, or gas plant you add — getting that power to your house is really expensive. The grid is the poster child for Baumol cost disease.” — Casey Handmer, 26:00

”50,000 acres of solar-powered data centers”

yt-dlp --download-sections "*27:00-28:00" "https://www.youtube.com/watch?v=3cDHx2_QbPE" --force-keyframes-at-cuts --merge-output-format mp4 -o "3cDHx2_QbPE-27m00s.mp4"

“If you want five gigawatts, that’s 50,000 acres. And 50,000 acres sounds like a lot, but the amount of land put aside for the Hanford site to produce the plutonium in the Manhattan Project is comparable. Nevada alone is 80 million acres.” — Casey Handmer, 27:00

”One human brain simulated on a sheet of paper floating in space”

yt-dlp --download-sections "*58:50-59:50" "https://www.youtube.com/watch?v=3cDHx2_QbPE" --force-keyframes-at-cuts --merge-output-format mp4 -o "3cDHx2_QbPE-58m50s.mp4"

“One human brain can be simulated in roughly the area of a sheet of paper, floating in space. That’s the attractor state. Computronium at the center of a solar cell.” — Casey Handmer, 58:50

”China’s energy vulnerability is its Achilles’ heel”

yt-dlp --download-sections "*1:37-2:25" "https://www.youtube.com/watch?v=3cDHx2_QbPE" --force-keyframes-at-cuts --merge-output-format mp4 -o "3cDHx2_QbPE-1m37s.mp4"

“China’s surrounded by 15 countries with no good mountain ranges or rivers to protect it. They get almost all their oil from countries that they don’t control, on fleets of oil tankers that they can’t defend. Synthetic fuels could level the playing field.” — Casey Handmer, 1:37

Key Points

• China’s energy race (0:00) - China has 20x the yearly solar panel production but depends on imported oil from countries it cannot control
• Synthetic fuels as geopolitical equalizer (2:12) - Electricity is only 20% of final energy use; synthetic fuels could supply the other 80% and neutralize China’s energy advantage
• US manufacturing capacity (5:28) - The idea that the US cannot compete in solar panel manufacturing is wrong; it could scale up in about two years
• Why hyperscalers choose gas (8:28) - Natural gas is fast to deploy, delivered by pipe, but runs into constraints at scale
• Solar’s 43% Wright’s Law learning curve (18:00) - Every doubling of production yields 43% cost reduction, happening every 2-2.5 years
• Still in the Apple II era of solar (19:30) - Solar adoption, production, and price reduction are all accelerating, not slowing
• Grid as Baumol cost disease (26:00) - Transmission infrastructure costs keep rising while generation gets cheaper
• Off-grid data centers (27:00) - AI data centers should have captive power plants, bypassing the grid entirely
• 50,000 acre solar data center layout (28:30) - 10 acres of solar, some batteries, one truckload of racks per megawatt
• NEPA blocking clean energy (40:24) - Environmental regulations designed for industrial pollution are being applied to solar farms in deserts
• Batteries replacing temporal arbitrage (44:04) - Battery storage is growing exponentially, performing the temporal shifting the grid used to do
• GDP is broken for measuring AI value (49:14) - Human labor generates $60 trillion in payroll; that is what AGI could ultimately replace
• AI’s value exceeds its GDP contribution (52:00) - Like oil, AI could be insanely valuable while appearing cheap in GDP terms
• One acre per 1,000 AI souls (58:00) - At current hardware efficiencies, one acre of solar supports roughly 1,000 H100-equivalent AI workloads
• Computronium endgame (58:45) - The ultimate state is integrated solar arrays with compute on the same silicon wafer, floating in space

Mentions

Companies

• Terraform Industries (0:14) - Casey Handmer’s company making synthetic natural gas from solar power
• xAI (7:30) - Building data centers in Texas; Handmer is bullish on their industrial approach
• Meta (8:28) - Planning 5 gigawatt data centers
• GE (14:00) - Makes gas turbines for power generation
• Tesla (28:30) - Megapacks referenced as battery storage solution
• TSMC (0:51) - Referenced for leading-edge chip manufacturing
• Samsung (20:00) - Coming on board for HBM production
• Kaiser Industries (12:00) - WWII analogy for vertical integration under supply constraints

Products & Technologies

• Brayton cycle turbines (14:00) - Core technology in natural gas power generation
• Starlink (30:00) - Mentioned as potential connectivity for off-grid data centers
• H100 GPUs (56:00) - Referenced for compute-to-energy calculations
• Silane (SiH4) (1:02:00) - Silicon tetrahydride gas used in purifying silicon for solar cells

People

• Elon Musk (7:19) - Referenced for xAI’s industrial approach and agreement on solar’s potential
• Dario Amodei (53:00) - Referenced for the thought experiment about AI data center economics

Surprising Quotes

“We’re still in the Apple II computer era of solar. The rate at which it’s getting cheaper is because there’s a lot of innovation left, not because we’ve exhausted the easy gains.” — Casey Handmer, 19:30

“I would never hold up a flag saying ‘I’m pro-China.’ That is just a sign that you’re not thinking clearly. China called the most important thing correct — that solar energy is the key to future industrial growth.” — Casey Handmer, 1:16

“The environmental impact statement for producing the paper the environmental impact is written on is more than the environmental impact of the actual solar project itself. It’s crazy town.” — Casey Handmer, 41:30

“Humans are glacially slow compared to computers. It takes longer for one neuron to fire than it takes for the internet to reach you to sleep between you tapping out ‘hello.’ It’s kind of nuts.” — Casey Handmer, 59:30

Transcript

0:00 Today I’m interviewing Casey Handmer. Caltech PhD on gravitational wave black hole physics, formerly at the Jet Propulsion Laboratory at NASA. Now founder and CEO of Terraform Industries. Casey, welcome. Big picture question I’m interested in. There’s this big industrial race — who’s winning between the US and China? Who can build the most batteries, transmission lines and transformers? China has been winning at least in recent decades. They have 20x the amount of yearly solar panel production. Obviously we have export controls on chips right now, cutting them off from TSMC’s leading edge. What’s your view of how the United States wins this?

1:00 Do you think that China is better at building stuff? Do you think the Chinese business environment is more favorable? I feel you can make these first principles arguments that they’re killing it, but it doesn’t mean they’ll win. People say they’re so much better at building infrastructure. I would never hold up a flag saying “I’m pro-China.” That is just a sign that you’re not thinking clearly. Why would you devote, in 2025, so much energy to defending China’s model? They’re devoting a lot to solar because they’ve correctly identified that solar energy is the key to future industrial growth. They called the most important thing correct.

1:37 Well, they’re in a similar situation to Japan before World War II. The United States is the luckiest country geographically — it’s surrounded on two sides by oceans and friendly neighbors. China’s surrounded by 15 countries with no good mountain ranges or rivers to protect it. They get almost all their oil from countries that they don’t control, don’t have leverage over, on fleets of oil tankers that they can’t defend because they can’t operate effectively in the Indian Ocean. If you get synthetic fuels working at scale, that could dramatically help China.

2:32 China has all this electricity production. But electricity is only about 20% of final energy use in a modern economy. The rest, you need gas and liquid fuels. Or coal — they use a lot of coal in China. The ability to turn that electricity into synthetic fuels which can supply 100% of energy needs could neutralize China’s energy advantage. This technology levels the playing field. But at the end of the day, China still contains an autocratic government. Never underestimate the capacity for an autocratic regime to mess things up.

3:28 There are parts of China which are hugely productive and innovative. You have to compare not all of China but Guangdong against the United States. Some of these regions are as big as America and as wealthy. At the same time, we want to find the truth here. We shouldn’t just drop out of the battle and just give up. Provided we don’t take extra effort to shoot ourselves in the foot, we can compete. Right now we are export controlling chips because we recognize AI is a key strategic advantage. Energy is also a key input in this AI race, and yet we’re not doing to them with these cheap energy technologies what we’re doing with chips. We should export control solar and batteries.

5:28 The idea that the United States cannot compete on solar panel manufacturing is wrong. We have abundant human resources, great financial markets. We could literally copy-paste their manufacturing processes. We could probably do that in about two years if we started with real urgency. A lot of technology already exists here. It’s mostly a case of putting in phone calls and saying “We need you to 10x the size of your operation.”

8:28 Let’s talk about the data centers they’re building — 5 gigawatts in Meta’s case. Right now the hyperscalers are choosing natural gas. Why are they wrong and you’re right? If you’re like xAI right now trying to build a data center, you want to get it done super fast. What are the factors of production? You need a building, you need chips, you need power. You can tap into a gas line and the energy transmission capacity of gas pipelines is way higher than electricity overhead lines.

12:00 As you grow, you start to run into constraints. Some of those constraints are around turbine manufacturing rate, around grid capacity, and running into the interests of existing consumers who depend on legacy electricity production. We just saw a recent forward auction in PJM where prices spiked. Not to reach prematurely for analogies, but this is like Henry Kaiser in World War II who was making ships and eventually had to vertically integrate all the way down to mining iron ore.

18:00 The Wright’s Law coefficient for solar is 43%. For every doubling of cumulative production, we get a 43% reduction in cost. Roughly every 2 to 2.5 years, we’re doubling production and dropping the price by about 43%. So it’s roughly 15-20% per year. And demand isn’t saturating — for every 15-20% price reduction, demand skyrockets by probably an even greater amount.

19:30 The so-called pros have been consistently wrong. They keep saying solar is going to saturate this week. Instead, not only are solar adoption, production, and price reduction all continuing — the rate at which they’re accelerating is also increasing. We’re still in the Apple II computer era of solar.

27:00 For really large captive loads like AI data centers, you’re going to have to build your own power plant. It might seem inefficient, but if you’re dealing with massive supply constraints on the grid, there’s no two ways about it. 10 acres of solar, some batteries, one truckload of data center racks — that gives you about one megawatt. If you want five gigawatts, that’s 5,000 of those, so about 50,000 acres. For context, the Hanford nuclear site in the Manhattan Project was comparable in size.

40:24 The environmental regulations that are blocking renewable energy in the United States are insane. Texas is out-deploying California 10 to 1 on solar. If you want to put solar panels on empty land in the middle of nowhere, you’ll probably trigger NEPA, which means a four-year environmental impact study. The environmental impact of producing the paper the study is written on is more than the environmental impact of the actual solar project itself.

49:14 GDP is broken as a measure of AI’s value. Human labor generates on the order of $60 trillion in payroll globally. That’s what AGI can potentially replace. The current AI companies are generating on the order of $10-20 billion in annual revenue. How can they sleep at night? Kohl’s generates more yearly revenue than that. But once we have models that are actually replacing human labor, the value is enormous.

58:45 An H100 has roughly the same number of floating-point operations as a human brain. It uses 50x more energy — 20 watts versus 1,000 watts. We know hardware can improve to close that gap. The ultimate attractor state is integrated solar arrays with compute on the same silicon wafer, floating in space. One human brain can be simulated in roughly the area of a sheet of paper. That’s the post-human state — computronium at the center of a solar cell.

1:05:00 Terraform is my day job. We are making synthetic natural gas from sunlight and air. We also have a methanol process which is a precursor to every hydrocarbon you could possibly want. Our jobs are available at terraformindustries.com. If you’re an ambitious hardware person looking to work on the most important industrial challenge, come work for us. Eventually we get to go and build it all on Mars as well.