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

Summary

• The headline call: data centers breaking ground by 2027 will be mostly solar. Handmer says “pretty much all the names you’ve heard of” have called him after the Scale Microgrids paper modeled 90% solar / 10% gas — his answer: “You can go all the way 100% solar.” The key insight: hyperscalers are “not power cost sensitive. They are power availability sensitive,” and solar is by far the best option for “firehosing” energy at a given problem.
• Gas wins today, loses at scale. Every turbine before ~2030 is “spoken for,” the Brayton cycle’s spinning Inconel alone costs $35/MWh amortized, and the financing case for GE capacity expansion entails a 25-year payback against solar’s 43% Wright’s Law cost decline per doubling (production doubling every 2-2.5 years, demand elasticity ~6x the marginal capacity increase). “You cannot win.” Dwarkesh’s HBM/CoWoS counter — “we won’t do it” meant “write me a check,” and they did — is the bull case that checks can unlock supply.
• On China: don’t extrapolate the 20x solar-manufacturing lead into AGI victory. China is surrounded by 15 mostly hostile countries, imports oil it can’t defend, and “never underestimate the capacity for an autocratic dictatorship to shoot itself in the foot.” But Handmer concedes his own Terraform synthetic-fuel tech “absolutely asymmetrically helps China” by converting its electricity glut into the two-thirds of energy end-uses electricity can’t currently serve.
• The US could localize solar “from dirt to finished module” in two years or less — it’s ~5 years behind China, not a generation — and even a 200% tariff on Chinese panels “doesn’t matter at all” to data-center economics. The binding constraint is regulatory: NEPA’s four-year reviews mean Texas is out-deploying California 10-to-1; the single highest-leverage policy is a categorical exemption for solar deployment.
• The build math is trivially cheap on land: ~10 acres per megawatt at four nines (1 MW rack + six Tesla Megapacks + 2.5x solar overbuild), so 5 GW = 50,000 acres — comparable to Oak Ridge or Hanford — costing hundreds of millions against $250B of GPUs, i.e. ~0.1% of capex. Modules are 8 cents/watt untariffed; everything around them is the cost problem.
• Structural short on grids, structural long on batteries: batteries do temporal arbitrage where wires do spatial arbitrage, per-capita battery mass has run four-five OOMs (10g phone → 100kg Tesla), and behind-the-meter storage is cannibalizing utilities’ highest-cost peak assets while their opex rises — PG&E as poster child of Baumol cost disease. Handmer flags this as his “most out of the money bet”; every respectable forecaster disagrees.
• AGI valuation frame: $60 trillion in global wages is the lower bound, OpenAI’s $10-20B ARR “sucks” next to McDonald’s, and the value chain today runs 10 cents of electricity → ~$1,000 of user value — so Anthropic could eat a 100x power-cost increase with a $10 surcharge. Long-run, GDP may nominally shrink as AI deflates cognition; “the size of our civilization” is better measured in raw energy use.

Deep dive

1. China’s energy lead is real — but don’t hand it the AGI trophy

• Dwarkesh’s opening challenge: if AI becomes an industrial race — solar panels, batteries, GPUs, transformers — that’s “exactly what China is known for,” with 20x US yearly solar manufacturing and SMIC eventually closing on TSMC. Dwarkesh frames high-speed rail bragging rights as “a sign that you’re really bad at capital allocation” — and the exchange notes China may be “accidentally correct” on the thing that matters most: solar overcapacity.
• Handmer’s geopolitical case: the US is “the luckiest goddamn country on earth” (oceans, friendly neighbors) while China is ringed by 15 mostly hostile countries and imports Middle East oil on tankers its navy can’t defend in the Indian Ocean. Dwarkesh then says, “Never underestimate the capacity for an autocratic dictatorship to shoot itself in the foot.” Handmer’s pushback is to compare Shanghai and Guangdong against the United States, not all of China: “You can have a part of China that is as big as America and as wealthy as America and as innovative as America,” unlike India’s large-but-poor middle class.
• The uncomfortable concession: Terraform’s synthetic fuels — turning electricity (only a third of final energy use) into fuels covering 100% — “absolutely asymmetrically helps China.” Handmer isn’t working with China and doesn’t plan to, “but the physics is very obvious,” and Chinese synfuel projects already exist.

2. America could rebuild the solar stack in two years — if it wanted to

• Handmer rejects the “China is more business-friendly” consensus as “absolutely crazy”: a CCP inspector on your board plus bribes to stay in business, versus US cheap natural gas, financial depth and “world leading automation.” US solar manufacturing is maybe five years behind China’s, and localizing “from dirt to the finished module” — a four-stage process — could take “two years or less if you started today. It’s currently 11 o’clock. So we’re going to start cutting checks by noon.”
• His counterfactual scar: post-Ukraine invasion he expected Europe to localize solar. It didn’t — “They’re still paying Russia a billion dollars a day for the privilege of being invaded.”
• Dwarkesh’s meta-pushback, worth keeping: Handmer’s forecasts are “if Elon ran the government like he ran SpaceX,” not what’s practically likely. Handmer’s partial concession — solar factories will likely never be on the critical path anyway; even a 200% tariff on Chinese solar “doesn’t matter at all” because chips, not power cost, are the bottleneck.

3. Why the smart money is buying gas today — and where that breaks

• The Colossus playbook explains current behavior: xAI in Memphis needed speed, so it bought a building, tapped a gas line (pipeline energy throughput far exceeds overhead electric lines), and rented turbines on trucks. “There was enough available once, maybe twice.”
• But scale breaks it: gas availability, turbine and transformer manufacturing rates, grid capacity, and AI competing with legacy consumers — the PJM forward auction already produced “unsustainably high prices” for households. “What if we’re doing 100 gigawatts per year? You can just break the situation.”
• The Henry Kaiser analogy for the endgame: Henry Kaiser’s Richmond shipyards, bottlenecked on steel, built not just a steel mill but a steel mine. “That’s the same sort of situation” — which is why the speaker is “quite bullish on xAI in particular, because the Elon cinematic universe has just done so much industrial stuff compared to the Googles and Metas of this world.”

4. The turbine-financing trap versus solar’s compounding learning rate

• The physics of cost: the most common thermal-generation route discussed is a Brayton cycle, and “anytime you have a bunch of Inconel spinning at high speed, it’s just going to cost you a bunch of money” — $35/MWh amortized for the spinning components alone, before fuel, heat exchangers or cooling.
• The finance chain: a bank lending GE money for turbine expansion sees benefits in 3-5 years, doesn’t know if the AI bubble bursts or Taiwan is invaded, and must run the plant 20 years to pay back. “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…? You cannot win.”
• Solar’s numbers: Wright’s Law coefficient of 43% per cumulative-production doubling, doublings every 2-2.5 years (~15-20%/year price decline), with demand rising ~6x the marginal capacity increase. “Not only are solar adoption, production, and price decreases continuing, they’re accelerating. And the rate at which they’re accelerating is still accelerating.” His framing: “We’re still in the Apple II computer era of solar.”
• Dwarkesh’s best counter: SK Hynix and Samsung made the same refusal noises on HBM, then ramped anyway once checks were written — “Now Samsung’s coming on board in the States to build AI6 with xAI.” The exchange’s answer: “they all got there in the end” — checks can unlock supply.

5. The dated call: mostly-solar data centers from 2027 groundbreakings

• Today 43% of US data-center power is natural gas; asymptotically, the call is ~100% solar for new load, with coal already costing more to operate than new solar costs to build.
• The near-term call: “by 2027, the majority of new data centers going in… groundbreaking at that point” will be mostly solar — and since 2027 groundbreakings are being planned now, “that’s why they’re calling me. My consulting fees are extremely affordable.” The speaker was the minority voice on the Scale Microgrids 90/10 paper arguing for 100% solar.
• Sizing sanity-check via AI 2027’s compute forecast: ~10M H100-equivalents today → ~100M by 2028 at ~1kW each ≈ 100 GW — “that sounds roughly right.” The interim bridge he flags: a literal “handful” of US sites with dead-smelter grid capacity plus underutilized old plants paid to run harder, with the data center promising curtailment — functionally “a massive captive battery plant.”

6. Ten acres per megawatt, and land is a rounding error

• The unit build: one 1-MW rack, six Tesla Megapacks (~4 MWh each — 24 hours of storage, “two bad nights in a row”), 4 MW nominal solar at 25% Texas utilization, then a 2.5x overbuild to get from one nine to four nines → ~10 acres per MW. At 5 GW: 50,000 acres — versus ~100,000 acres each set aside for Oak Ridge and Hanford.
• The capex answer to Dwarkesh’s “why not just buy 50 gas turbines”: 50,000 Texas acres run “maybe hundreds of millions” against ~$250B of GPUs — ~0.1% of project cost. Modules are 8 cents/watt untariffed; the dollar-a-watt installed figure is everything else, “the thing that breaks my brain at Terraform.”
• Overbuild isn’t waste — Brian Potter’s analogy: a terabyte MacBook you use 100GB of, bought “because it’s cheap enough.” 99.9% of the time the array produces more power than it needs, so the local town can “throw a power cable over the wall” for near-zero marginal cost. And weather forecasting lets you pre-curtail 5% three days early instead of 50% in a crisis. Interconnection barely matters: it’s off-grid, “an optical fiber cable which you can string up on poles.”

7. NEPA, not physics, is the American bottleneck

• The regulatory absurdity, as told: a four-year environmental impact review generates so much paper that “just the environmental impact of producing the report is more than the environmental impact of just deploying the solar. This is bonkers.” The breaking point: a $10,000 biologist finding “a tuft of grass which we believe might be one of the 20 species that this particular species of bee sometimes eats” — on land zoned unrestricted industrial between a rocket test stand and a chemical plant. “I’m going to become the Joker.”
• The perverse detail: desert solar is “arguably positive” (shading, soil moisture retention — it could reverse desertification), yet faces stricter review than grading the lot and covering it in concrete. The ask: a categorical exemption for solar deployment — this is why Texas is out-deploying California 10 to 1.
• The stakes as he frames them: fail to move off fossil fuels in 10-20 years and “we’ll get poor the same way the UK did, because they ran out of coal” — plus flooded coastal cities, requiring solar synthetics and “sulfur injection and a couple of other things.”

8. Batteries do the grid’s job — his most contrarian bet

• The conceptual move: the grid performs spatial arbitrage; batteries perform temporal arbitrage — and since the sun rises daily, the battery’s arbitrage runs ~300 days a year while the grid’s most expensive peak assets almost never hit peak utilization. The speaker flags this: “it’s the one where it’s the most out of the money bet… Everyone else that I consider to be a respectable forecaster in this area disagrees with me.”
• The utility death spiral: behind-the-meter batteries the operator can’t control are cannibalizing exactly the assets it charges top dollar for, while unionized maintenance, eminent-domain litigation and wildfires drive opex up — PG&E “perpetually on the brink of bankruptcy,” “the poster child for Baumol cost disease.” Expect “large-scale pruning” of grids and captive power plants for big loads, as aluminum smelters once had.
• The scaling fact: per-capita lithium-ion has gone from ~10 grams (a phone) to ~100 kilograms (a Tesla) — “four or five OOMs” — and “the average distance the electron is going to travel between generation and consumption is going to decrease… pretty radically.”

9. AGI economics: $60T floor, GDP deflation, energy as the true metric

• The value stack today: a ~$10 subscription delivers more like $1,000 of value, served for ~$1/million tokens of which electricity is ~10% — so “10 cents of electricity is generating $1,000 worth of economic value,” and labs could absorb a 100x power-cost increase as a $10 surcharge, buying turbines “for prices that would make your eyes water.”
• Sizing the prize: OpenAI’s $10-20B ARR — “That sucks. It’s terrible. How can they sleep at night?” — is less than McDonald’s or Kohl’s. But global wages are ~$60 trillion, and that’s a lower bound: estimating AGI from payroll is like capping Caterpillar’s value by “how many men and wheelbarrows to dig a trench.” The frame: every industrial revolution bypasses a bottleneck — first metabolism (“99% of the energy we consume routes around our guts”), now cognition.
• The measurement problem, credited to Gwern and James Bradbury: Dario’s “data center of geniuses” shows up in GDP only as chip and token flows — it “might actually cause a nominal decrease in GDP” while massively increasing real output. Oil per joule is 100x cheaper than food yet ~1% of GDP; elasticity matters more than GDP share (oil shocks cause double-digit GDP drops). Better long-run metric: “the raw energy use that we do rather than GDP.” On AI wages: lab-vs-lab competition pushes prices toward “a small multiple of the marginal production cost of those tokens,” not the $200K a human researcher earns.

10. The attractor state: one human per square meter of silicon

• Strip the stack to essentials: “a big slab of relatively cheap silicon to make the power, and a small slab of relatively expensive silicon to do the thinking” — in space, no battery needed. An H100 matches human-brain flops at ~50x the energy (1,000W vs 20W); at brain efficiency, one acre supports ~50,000 AI souls, making the implicit land value higher than any farmland in history.
• The final form, verbatim: “One human brain can be simulated in roughly a square meter of silicon floating in space… That’s the future human form. That’s my final form” — solar sails with computronium dies, flying “closer to the sun to get more power, right up to the thermal limit.” His speculative coda: 4 billion years of complexification may be “the beginning stages of a collapse back towards the simplest possible thermodynamic-to-cognition stack” — stars, silicon, and lasers to friends: “Hey, I just made up a new meme.”
• The industrial path there: silicon from actual dirt (20 microns suffices for PV), silane-gas purification, ~18 months per refinery — and near-free solar lets you “revisit legacy industrial processes… what if we just use twice as much power and do them faster and cheaper?” Terraform’s pitch: synthetic natural gas from sunlight and air, methanol, ammonia, steel, cement — “I will never hire anyone who can’t do math.”