China is killing the US on energy. Does that mean they’ll win AGI? — Casey Handmer
What this covers
Casey Handmer and Dwarkesh Patel examine whether energy abundance will determine who wins the race to AGI, with Handmer arguing that solar power's economics make it the inevitable energy supply for AI data centers, and that this trajectory leaves the US still positioned to compete with China despite its current solar manufacturing deficit. The conversation moves through the physics and economics of power generation, the structure of hyperscaler incentives, and the geopolitical implications of energy-intensive AI systems. Handmer contends that power-availability vs power-cost sensitivity is the key insight: because electricity represents only roughly 10% of the marginal cost of serving tokens, while users extract $100–$1000 of value per serving, hyperscalers are indifferent to electricity costs and care only about supply. This flips the traditional energy economics.
The argument proceeds through several distinct technical and policy domains. On the physics side, Handmer explains why Wright's Law—solar's steep 43% cost reduction per production doubling—outpaces gas turbines' shallow learning curve, and how Spatial vs Temporal Arbitrage through batteries can substitute for grid infrastructure, enabling self-contained off-grid data centers. On the industrial side, he draws parallels to Henry Kaiser's vertical integration during shipbuilding to argue why scaled AI energy deployment will require building down to primary materials. The conversation identifies regulatory barriers—particularly NEPA environmental reviews—as the actual constraint on US solar deployment, not resource or cost, pointing to Texas outpacing California by tenfold as evidence. Handmer also addresses the geopolitical asymmetry: China's manufacturing lead in solar panels may matter less than the US's geography and chip supply chains when competing for total AI dominance. The discussion concludes by sketching how Solar Overbuild plus batteries can achieve four-nines uptime, inverting the relationship between data centers and utilities.
Handmer argues that solar power, due to its uniquely steep learning rate and the fact that AI hyperscalers are power-availability-sensitive rather than power-cost-sensitive, will inevitably dominate AI energy supply via off-grid solar-plus-battery data centers, and that the US can still beat China in this industrial race despite China's solar manufacturing lead.
- Hyperscalers don't care about the cost of power, only its availability, because electricity is ~10% of the marginal cost of serving AI tokens
- Solar's Wright's Law learning rate of 43% per doubling outpaces gas turbines, which face slow scaling and 20-year payback risk
- Batteries perform temporal arbitrage that substitutes for the grid's spatial arbitrage, enabling self-contained off-grid data centers
This asset isn't compiled yet
You're seeing its claims, ranked. Compile it to build the argument threads, weight them, and check each claim against your library — the full view.
The grid performs spatial arbitrage—transporting power near-instantaneously from where it's cheap to where it's expensive—while batteries perform temporal arbitrage, storing power at one time of day and releasing it at another; batteries can therefore substitute for the grid's function, and being installed behind the meter they cannibalize the grid operator's most expensive, lowest-utilization assets, shrinking the average distance electrons travel between generation and consumption.
“The grid transports power from one place to another... so it's actually performing a spatial arbitrage.”
Each industrial revolution bypasses a binding constraint: the pre-Industrial Revolution bottleneck was metabolism (how much food a human or horse could digest into mechanical work), now 99% of consumed energy routes around our guts through cars, aircraft and grids; the AI revolution analogously routes around cognitive constraints, just as the printing press, computers, and credit cards did.
“every time you figure out the industrial revolution, what you're doing is you're finding some way of bypassing a constraint or bypassing a bottleneck. The bottleneck prior to what we call the Industrial Revolution was metabolism.”
Getting power to your house is expensive not because generation is costly but because the grid requires unionized labor to build and maintain power lines in built-up areas, eminent domain litigation that drags on for years, and wildfire liability—making it the poster child for Baumol cost disease and forcing large-scale pruning of grids that we can no longer afford to maintain.
“even though the cost of an additional solar panel or additional wind turbine or additional gas turbine or whatever is relatively cheap, getting that power to your house is really expensive.”
Like the internet's consumer surplus and oil's cheapness, AGI's enormous value may not register in GDP—a 'data center of geniuses' would show only modest inputs (chips, energy) and outputs (tokens), and by automating human labor it could cause nominal GDP to decrease even as it massively increases the valuable output of civilization; therefore civilization's true scale is better measured by total energy use than by GDP.
“it might actually cause a nominal decrease in GDP while at the same time contributing massively to what we might think of as the valuable stuff human civilization can produce.”
Gas-powered AI scaling works at 1 GW/year but breaks down as you turn the dial higher—at 5, 50, or 100 GW/year you hit constraints on gas availability, turbine manufacturing rate, transformer production, and grid capacity, while AI loads compete with residential consumers for legacy electricity, as shown by the PJM forward auction producing unsustainably high consumer prices.
“What if we're doing 5 gigawatts per year? What if we're doing 50 gigawatts per year? What if we're doing 100 gigawatts per year? You can just break the situation.”
Hyperscalers don't care about the cost of power because electricity is only about 10% of their roughly $1-per-million-tokens serving cost, while the economic value to the user is on the order of $100-1000—so 10 cents of electricity generates $1000 of value; Anthropic could raise electricity costs 100x and still profitably buy turbines at any price, meaning the real constraint is availability, not cost.
“So 10 cents of electricity is generating $1,000 worth of economic value.”
Solar panels have a Wright's Law coefficient of 43%, meaning every time cumulative production doubles, cost drops 43%; with production doubling roughly every two years, this means ~15-20% annual price reduction, and crucially demand elasticity exceeds the learning rate so each price cut drives demand up about six times more than the marginal capacity increase, sustaining the loop—unlike natural gas, whose learning rate is far shallower.
“The Wright's Law coefficient is 43%. So every time we double cumulative production, we get a 43% reduction in cost.”
A bank deciding whether to finance GE's gas turbine expansion faces the risk of operating that plant for 20 years to recoup investment, but can't know whether the AI bubble will burst, whether China invades Taiwan, or whether competitors win—and crucially, in 25 years gas turbines cannot be produced at a price relevant in a world where solar and batteries are already at today's prices, so the financing case for new gas turbine capacity fails.
“What are the odds that in 25 years' time we can produce gas turbines at a price that is relevant in a world where solar is already at its current price and batteries are at the price where they're already? You cannot win.”
Batteries cycle predictably nearly every day (~300 days/year) because the sun rises daily and demand peaks at dusk, giving high utilization, whereas the most expensive high-voltage grid assets like substations and transformers exist for rare peak events and have very low utilization—so grid operators face an aging asset base being cannibalized by behind-the-meter batteries they cannot control, with rising operating costs and falling utilization.
“the utilization of the batteries is on average, let's say 300 days a year. The utilization of your most expensive, highest voltage grid assets is much, much lower.”
When a role emerges that AIs are well-suited to and don't compete with humans for, AIs compete only against other labs, pushing the price down to a small multiple of the marginal cost of producing the tokens—so an AI as good as a $200,000 researcher won't be worth $200,000 but rather worth much more in copy-pasteable output and much less in marginal compute cost.
“it'd be competing against the other labs. You'd actually see the cost pushed down to a small multiple of whatever the marginal production cost of those tokens is.”
When Henry Kaiser scaled WWII shipbuilding to four parallel shipyards, he became bottlenecked on steel and so built his own steel mill and even a steel mine—illustrating that massive industrial verticals reach all the way down into primary material supply when scaling, which is why Elon's vertically-integrated companies are well positioned for AI energy scaling.
“Kaiser Industries went off and built not only a steel mill, but also a steel mine. They went and started digging rocks out of the ground to turn into ships.”
Conventional power generation is fundamentally a steam engine: you burn a chemical out of equilibrium with the atmosphere to make heat, boil water, spin a magnet, and generate electricity; the key expensive step is converting heat to electricity via a Brayton cycle, and anytime you have a Brayton cycle with Inconel spinning at high speed it inherently costs a lot to build.
“The most common way is the same for a nuclear plant or a gas plant, combined cycle plant or a coal plant or whatever. It’s what's called a Brayton cycle.”
Hyperscalers like xAI chose natural gas for speed: gas turbines arrive on a truck, and the energy transmission capacity of regular gas delivery pipelines is far higher than electricity overhead lines and easier to upgrade, so if you can tap a gas line you can generally get enough power quickly.
“The energy transmission capacity of your regular gas delivery pipelines is way, way higher than electricity overhead lines, and it's easy to upgrade.”
The solar panel factory will never be on the critical path for AI; even a 200% tariff on Chinese solar wouldn't matter to the overall cost picture, because the real bottleneck is having enough chips at competitive capability installed in data centers—which is exactly what xAI and Elon's companies focus on.
“Even if Trump puts in a 200% tariff on Chinese solar and we're not able to bypass it via Vietnam or something, it's still a bargain. It doesn't matter.”
The US is geographically the luckiest country on earth—bounded by oceans on two sides and friendly allies on the others—while China is surrounded by 15 mostly hostile countries with no natural barriers, gets almost all its oil from the Middle East via tanker fleets it cannot defend because its navy can't operate effectively in the Indian Ocean.
“The United States is the luckiest goddamn country on earth because it's surrounded on two sides by oceans and on the other two sides by friendly allies. China's surrounded by 15 countries who are mostly hostile to it”
Really large captive loads like AI data centers or aluminum refineries will have to build their own dedicated power plants, reverting to how it used to work—this is not inefficient if you are sensitive to power supply elasticity, because it is the cheapest way for them to actually get power at scale.
“for really large captive loads, like AI data centers or aluminum refineries or whatever, you're going to have to build your own power plant for them, which is how it used to work.”
Synthetic fuels technology—turning electricity into liquid fuels—would asymmetrically benefit China because it converts China's electricity abundance into the transport and industrial fuels that currently require imported oil, neutralizing China's energy-security disadvantage.
“It does. It absolutely asymmetrically helps China. We're not currently working with China, we don't plan to, but the physics is very obvious. Synthetic fuels have been around for 100 years.”
Environmental regulations from the early 1970s, especially NEPA, irrationally block solar deployment: a four-year environmental impact review generates so much paper that the environmental impact of producing the report exceeds that of the solar itself, and off-grid solar gets regulated like a chemical plant—even though solar arrays arguably benefit desert by improving soil moisture, while grading land and dumping rusting oil-leaking cars often needs no permit at all.
“That’s your four-year environmental impact review, which generates so much paper that just the environmental impact of producing the report—because you have to cut down trees to make paper—is more than the environmental impact of just deploying the solar.”
An H100 has roughly the same flops as a human brain but uses ~1000 watts versus the brain's ~20 watts (50x more energy), so since the brain proves hardware can run that cognition on 20 watts, with software/hardware efficiency gains one acre of solar feeding a megawatt could host on the order of 50,000 'AI souls'—and likely far more since transistors are vastly faster than neurons, making the land value of solar-for-cognition far above the best farmland.
“an H100 has the same amount of flops as a human brain, but also uses way more energy than a human brain. It uses 50x more energy.”
The minimum matter needed to run AI computation is a big slab of cheap silicon for power plus a small slab of expensive silicon for thinking; in space you need no battery (always sunlit) and no transformer or DC-DC converter—just a buck converter or relays—so aliens or post-humans could build integrated solar-plus-computronium on a single silicon wafer, with one human's worth of computation fitting in roughly one square meter of silicon the thickness of paper.
“What you actually need is a big slab of relatively cheap silicon to make the power, and then a small slab of relatively expensive silicon to do the thinking. If it's in space, that's all you need”
To achieve four nines of uptime for AI data centers you need solar overbuild of about 2.5x: in Texas a 1 MW average load needs ~4 MW of panels (25% capacity factor) plus 2.5x overbuild yielding ~10 acres of solar, ~24 hours of battery (six Tesla Megapacks), with the overbuilt power 99.9% of the time exceeding need—which can then be sold to local utilities at near-zero marginal cost, inverting the data-center-must-curtail relationship.
“in practice, because you're aiming for four nines instead of one nine, you need an overbuild of about 2.5x. So you've got about 10 acres of solar.”
Oil's small fraction of GDP (~1%) does not correspond to its importance, because oil shocks cause double-digit GDP decreases—so the elasticity of demand often matters more than a good's raw fraction of GDP, paralleling how AI and energy may be undervalued by GDP metrics.
“if you don't have oil, then you have these oil shocks, which cause double digit decreases in GDP. So the elasticity of demand often matters more than its raw fraction contribution to GDP.”
Because solar output can be predicted accurately three days ahead via weather forecasting, an operator can preemptively curtail slightly (e.g., 5% now) to avoid a deep curtailment (50% in three days), keeping total annual curtailment to only about five hours and maintaining four nines uptime rather than failing because of unpredictable weather.
“I will start curtailing now by 5% so I don't have to curtail by 50% in three days. Then overall for the whole year I'll only curtail five hours”
Texas is winning the solar race against California by a factor of 10 to 1 in deployment, demonstrating that the regulatory environment—not climate, resources, or cost—is the decisive variable in whether solar gets built.
“This is the reason Texas is winning. Texas is out deploying California 10 to 1.”
OpenAI generates only on the order of $10-20 billion ARR—less than McDonald's or Kohl's yearly revenue—which shows current AI isn't very economically valuable yet; but AGI's promise is to automate human labor, which generates on the order of $60 trillion in global wages, so human-level models will eventually be worth at least that, with white-collar work alone in the tens of trillions.
“OpenAI is generating on the order of $10-20 billion ARR... for context, McDonald's and Kohl's generate more yearly revenue than that.”
The conventional wisdom from forecasters that solar demand will saturate each year is consistently wrong—solar adoption, production, and price decreases are not just continuing but accelerating, and the rate of acceleration is itself accelerating, meaning we are still in the 'Apple II computer era' of solar.
“Conventional wisdom is that solar demand is going to saturate this week... Instead, it's just blasting out the top of the graph. This conventional wisdom is wrong.”
Just the amortized cost of the high-speed, high-temperature spinning Brayton-cycle components of a gas turbine is about $35 per megawatt hour—excluding fuel, heat exchangers, and cooling ponds—setting a hard floor on conventional generation cost.
“$35 a megawatt hour just for the Brayton cycle. We're not talking about the fuel, we're not talking about the heat exchanges, we're not talking about the cooling ponds or anything like that.”
If you allow diesel generators or a gas turbine to cover ~10% of generation during winter, you can reduce the required solar panel installation by about 60% because you no longer need to overbuild solar for the worst-case winter contingency—making it a broad optimization trade-off between panels, batteries, and backup.
“if you have diesel generators or something which can take over 10% of the generation during winter, then you can have a 60% reduction in the amount of solar panels you need to install”
The belief that there isn't enough land for solar is garbage; Nevada alone is 80 million acres of mostly federal land, terrain need not be flat (trees grow on mountain slopes), and a 5-gigawatt data center needs about 50,000 acres—comparable to the 100,000 acres each set aside for Oak Ridge and Hanford in the Manhattan Project—so land is a trivial fraction (~0.1%) of total project cost.
“There's this very popular misconception that there's not enough land to do solar. This is garbage.”
Evolution has continually ramified and complexified the thermodynamic gradient from simple RNA organisms to the industrial economy, but we may be witnessing the beginning of a collapse back toward the simplest possible thermodynamic-to-cognition stack: fusion in stars and cold space provide the temperature gradient, and silicon—electrons crossing the Fermi gap in a solar array and returning through logic gates—is the most efficient way to convert that gradient into usable cognition.
“what we're seeing is the beginning stages of a collapse back towards the simplest possible thermodynamic-to-cognition stack.”
Solar modules without tariffs are 8 cents a watt ($80,000/MW) versus the ~$1/watt people quote including installation; the panels are the 'magic part' that turn sunlight into electricity at 25% efficiency, and everything else in the system should cost less than that—a major cost-reduction opportunity.
“The modules themselves, without tariffs, would be 8 cents a watt, so that's $80,000…”
Per capita lithium-ion battery allocation has risen four to five orders of magnitude from when it was just a cell phone (~10 grams/person) to now when many people drive Teslas (~100 kg/person), and this trend will only continue, distributing batteries across solar arrays, substations, retired power plant sites, and homes.
“When you and I were much younger, the lithium ion battery was just in your cell phone... Nowadays half the people in this town drive Teslas, so your per capita allocation of lithium ion batteries is 100 kilograms”
It currently costs more to operate an existing coal plant than to build a new solar plant, so coal becomes cheaper to simply turn off, and as overall capacity increases it dilutes legacy fossil generation toward an asymptote of 100% solar.
“It is the case right now that operating a coal plant costs more than building a new solar plant. So it's just cheaper to turn off.”
Only about a third of final energy use in a modern economy comes from electricity; the rest requires gas, coal, and liquid fuels for transport and industry, which is why converting electricity into synthetic fuels would dramatically amplify the value of a country's electricity production.
“only a third of final energy use in a modern economy comes from electricity. The rest, you need gas and whatever to transport things”
If trade links were fully severed, China's ability to make advanced chips is basically nonexistent while the US can make them; conversely the US ability to make solar arrays is embryonic but only about five years behind China's, so an export-control war on solar would hurt China worse than the US.
“The United States’ ability to make solar arrays is embryonic, but it's actually not that far behind China's. It's maybe five years behind.”
Never underestimate the capacity for an autocratic dictatorship to shoot itself in the foot; China still contains the poorest Chinese people anywhere on earth, which is a structural weakness even though parts of China are as wealthy and innovative as the US.
“Never underestimate the capacity for an autocratic dictatorship to shoot itself in the foot.”
Setting up a brand new silicon refinery takes about 18 months with current technology, but with approximately free abundant solar power one could revisit legacy industrial processes optimized for efficiency and instead use twice the power to do them faster and cheaper with less capex and lead time—for example reducing silicon electrolytically or via silane gas.
“setting up a brand new silicon refinery takes about 18 months. But that's just with the current technology”
By 2027, the majority of new data centers breaking ground will be mostly solar-powered (90%+ solar), based on conversations Handmer has had with major players planning 5-gigawatt projects and his published analysis showing you can go all the way to 100% solar rather than the 90% solar / 10% gas in the Scale Microgrids paper.
“Let's say by 2027, the majority of new data centers going in at that point would be mostly solar.”
The US could build out a full four-stage domestic solar panel supply chain (from dirt to finished module) in about two years or less if it started with full intensity, because the technology already exists and it is mostly a matter of telling existing manufacturers to 10x their factories with blank-check funding.
“I think the United States could probably do that in two years or less if you started today.”
Devoting enormous industrial effort to building high-speed trains in 2025 is a sign of bad capital allocation rather than something to boast about, undercutting the assumption that China's manufacturing prowess reflects superior capital allocation.
“I would never hold up a flag saying, “I'm really good at building high-speed trains.” That is just a sign that you're really bad at capital allocation.”
It's currently 11 o'clock so we're going to start cutting checks by noon.
“It's currently 11 o'clock. So we're going to start cutting checks by noon.”