Casey Handmer on Energy, China, and Why He Thinks Solar Will Power the AI Build-Out
Dwarkesh Patel
Dwarkesh Patel frames the conversation with a worry. If AI becomes an industrial race over who can build the most solar panels, batteries, GPUs, transmission lines, and transformers, that is exactly what China is good at and what the United States has not been known for in recent decades. China, Patel notes, has about 20 times the yearly solar manufacturing of the US, and he expects SMIC to eventually catch up with TSMC's leading edge despite export controls. So why doesn't China simply win by default?
Casey Handmer is the founder and CEO of Terraform Industries. He previously did a Caltech PhD on gravitational waves and black holes, then worked at Hyperloop and at NASA's Jet Propulsion Laboratory. His answer runs through the whole conversation. The US can still compete if it stops obstructing itself, and solar plus batteries, not natural gas, will supply most of the energy for AI. He openly calls this a contrarian position.
China's bet on solar: "accidentally correct"
Patel pushes on the idea that China is simply better at capital allocation, pointing to the success of BYD and CATL. Handmer is dismissive of some of China's industrial showpieces. Being very good at building high-speed trains in 2025, in his view, signals bad capital allocation rather than good. On solar overcapacity, though, he concedes that China "might be accidentally correct." They picked the most important thing, and that should count for something.
He attributes this to geography rather than wisdom. The United States, he says, is "the luckiest goddamn country on earth," bordered by two oceans and two friendly neighbors. China is surrounded by about 15 mostly hostile countries with no strong natural barriers. It gets almost all its oil from the Middle East, from countries it doesn't control, carried on tankers its navy cannot defend in the Indian Ocean. Energy independence is therefore a strategic necessity for China, a situation Handmer compares to Europe's.
Patel points out that this makes Terraform's own work awkward. Terraform turns electricity into synthetic fuels. Patel explains that only about a third of final energy use in a modern economy is electricity, so a technology that converts cheap electricity into fuel would let China's electricity advantage spread across the rest of its economy. Handmer agrees without hedging that it "absolutely asymmetrically helps China." Terraform does not work with China and has no plans to, he says, but the physics is obvious, synthetic fuels are a century old, and there are already projects in China on them. He would not be surprised if China were taking the idea seriously.
He still does not concede the race. Autocracies, he says, have a large capacity to "shoot [themselves] in the foot." Patel pushes back. Even if China's governance is flawed, parts of the country such as Shanghai and Guangdong are as large, wealthy, and innovative as the US, unlike India, whose middle class is larger than America's but far poorer. Handmer accepts the caution and restates his point: the US is still in the race as long as it doesn't go out of its way to hurt itself.
What an energy embargo would actually do
Patel sets up a mirror image of US chip export controls. If energy is a key AI input, China could retaliate by export-controlling solar panels and batteries. Handmer thinks that would hurt China more than the US. China depends on the US export market. More importantly, China's ability to make advanced chips is "basically not there," while American solar manufacturing, though "embryonic," is by his estimate only about five years behind China's. He says the US is already on track toward about 100 gigawatts of solar capacity per year.
On cost, he rejects the mainstream explanations for China's edge one by one. Cheaper labor is no longer true, he says, since the comparison should be with Mexico. Laxer environmental regulation is true. But the claim that China is more business-friendly he calls "absolutely crazy," citing CCP inspectors on company boards and the need to pay bribes where the rule of law is weak. Against that, he lists what the US has: much cheaper natural gas, abundant oil, human capital, financial capacity, and world-leading automation. Mostly or fully automated American solar factories should be able to compete, and the US "could literally copy paste" existing factory designs.
On speed, he recalls expecting Europe, after Russia invaded Ukraine, to localize solar production "from the dirt to the finished module," which he describes as roughly a four-stage process. He thought Europe could do it in about two years. It didn't, and he says Europe is still paying Russia "a billion dollars a day for the privilege of being invaded." He believes the US could do the same in two years or less by calling manufacturers in the US, Germany, and elsewhere and offering a blank check to grow their factories tenfold.
Patel objects that many of Handmer's forecasts are really answers to "what could happen with World War II–level motivation," or "if Elon ran the government like he ran SpaceX," rather than what is likely to happen. He notes that xAI is focused on chips, not solar, because chips are the real bottleneck. Handmer agrees about the bottleneck and uses it to make his point. Even a 200% tariff on Chinese solar, with no workaround through countries like Vietnam, would still leave solar "a bargain" relative to the total cost of a data center. What matters is having competitive chips installed, cooled, and running. With dozens of solar manufacturers worldwide competing, he "very much doubt[s]" solar panel supply will ever be on the critical path.
Why hyperscalers are choosing gas today
Patel then puts the hardest challenge. The people with real money at stake, the hyperscalers building 1–2 GW sites (5 GW in Meta's case) for 2028–2030, are choosing natural gas. They can see solar's learning curve too. Why are they wrong?
Handmer's answer is that they are not wrong for their current circumstances, but those circumstances will not scale. He uses xAI's Colossus data center in Memphis as the example. Racing to get online, xAI bought an existing building instead of constructing one, and brought in power and cooling equipment on trucks. Crucially, it could tap a local gas line. Gas pipelines carry far more energy than overhead power lines and are easy to upgrade, so access to gas generally means you can get enough power. There were enough rentable gas turbines "once, maybe twice."
As demand grows, he argues, new constraints appear. Gas availability is one. He mentions discussion of building in Pennsylvania and parts of Texas where gas is stranded, but notes that growing US export capacity means gas prices will not stay very low forever. Then come turbine manufacturing rates, transformer production, grid capacity, and competition between AI loads and ordinary consumers. He cites a recent PJM forward capacity auction that, in his description, produced "unsustainably high prices" for households. A gigawatt a year can be met with turbines indefinitely, he says, but at 5, 50, or 100 gigawatts a year "you can just break the situation."
He reaches for a wartime analogy. Henry Kaiser's Richmond shipyards began building ships for the British and eventually ran four yards in parallel, until steel became the bottleneck. Kaiser Industries responded by building its own steel mill and even its own mine. Handmer sees the same dynamic of massive vertical integration ahead, and says he is particularly bullish on xAI because "the Elon cinematic universe" has done far more industrial work than Google or Meta and can reach down into primary material supply if necessary.
The grid is the expensive part
On the PJM prices, Handmer argues the increase is probably driven more by delivery costs than generation costs. Adding a solar panel, wind turbine, or gas turbine is relatively cheap. Getting the power to a house is expensive. He cites PG&E, which he describes as "perpetually on the brink of bankruptcy." Power lines are built and maintained by generally unionized labor in already built-up areas. New transmission requires eminent domain and years of litigation in which public money is spent on both sides. Then there are wildfires. He calls this "the poster child for Baumol cost disease."
As a result, he expects "large-scale pruning" of grids that can't be afforded under the current regulatory regime. For large captive loads such as AI data centers or aluminum refineries, he thinks you will have to build your own power plant, as aluminum plants historically did. Patel suggests that redundant plants at every site seem inefficient. Handmer replies that if you're sensitive to power supply, "you just have to do it," and that Colossus's truck-mounted power plant in a parking lot is not inefficient but simply the cheapest way to get power.
Turbines, the Brayton cycle, and $35 per megawatt-hour
On how soon turbines run out, Handmer says everything before about 2030 "is spoken for." More could be made, but spinning up additional turbine production is relatively expensive. He walks through conventional generation: burning a fuel out of chemical equilibrium with the atmosphere, making heat, and turning that heat into motion and then electricity. The common conversion, he says, is a Brayton cycle, the same kind of machine as a jet engine. Anything with Inconel spinning at high speed is inherently expensive to build. He doesn't know GE's current retail prices for its roughly 100 MW turbines and suspects they've risen a lot if flexible. His recollection is that about $35 per megawatt-hour covers only the amortized cost of the high-speed, high-temperature spinning components, before fuel, heat exchangers, or cooling.
Hyperscalers care about availability, not price
Handmer's central and "very counterintuitive" claim is that hyperscalers don't care about the cost of power. He contrasts them with a Pennsylvania retiree who is highly sensitive to her electricity bill. Patel estimates that an AI subscription on the order of $10 delivers perhaps $100 to $1,000 of value to him. Handmer estimates that serving tokens costs a lab around a dollar per million tokens, with electricity about 10% of that. In his framing, about 10 cents of electricity generates $1,000 of economic value. A lab could absorb a 100-fold increase in its electricity costs by adding about $10 to a subscription, and could buy turbines "for prices that would make your eyes water."
So why solar in 2032, when Patel expects hundreds of gigawatts of new data center demand? Handmer says there simply aren't enough turbines. Turbine production has only ramped back to roughly early-2000s rates.
Solar's learning rate
Patel argues that both solar and gas supply chains will respond to demand. Handmer says the difference is the learning rate. By his account, solar has a Wright's Law coefficient of 43%: every doubling of cumulative production cuts cost by 43%. Production doubles roughly every 2 to 2.5 years, and prices fall about 40% per doubling, which he puts at roughly 15–20% per year. He attributes the improvements partly to roughly 10,000 manufacturing process engineers working on it full time. When Patel notes that this could be true of any process, Handmer says the learning rate can only be sustained if demand elasticity outpaces it, which it does for solar. Each price drop, he says, causes demand to grow by probably six times more than the added production capacity.
This is where he says "the so-called pros are definitely wrong." Forecasters have repeatedly predicted that solar demand would saturate, and instead it keeps "blasting out the top of the graph." He claims adoption, production, and price declines are accelerating, and that the rate of acceleration is itself accelerating, in the sense that solar's fitness for its markets keeps improving. We are, he says, "still in the Apple II computer era of solar."
Asked why the same demand-driven logic doesn't apply to gas turbines and transformers, Handmer frames it as a lending decision. A bank financing GE's turbine expansion would see results in three to five years. It would not know whether the AI bubble had burst, whether China had invaded Taiwan, whether Siemens or others had outcompeted GE, or whether GE's structural problems had resurfaced. It would also need the plant to operate at capacity for about 20 years. Looking at current solar and battery prices, he asks what the odds are that turbines will still be price-relevant in 25 years, and concludes "you cannot win."
Patel offers a counterexample. Memory makers like SK Hynix and Samsung initially hesitated to expand HBM production for fear AI demand wouldn't last, and CoWoS packaging was another bottleneck, yet capacity did eventually expand. Handmer's reply is that "we can't do it" usually means "write me a check," and those checks were written. He mentions, with an "I think," that Samsung is now coming to the US to build AI6 chips with xAI.
How fast new load goes solar
Patel cites a figure that 43% of US data center power currently comes from natural gas. Handmer agrees that new load will asymptotically approach 100% solar by around 2040. Gas plants that still make money will keep running, but he says operating an existing coal plant already costs more than building new solar. Exponentially growing AI demand also means new capacity will dwarf existing stock.
He believes the binding constraint will always be the chips, so the pace depends on how fast TSMC ramps GPU production. Patel uses the AI 2027 compute forecast: about 10 million H100-equivalents today and about 100 million by 2028, at roughly a kilowatt each, or about 100 GW. Handmer finds that plausible.
He says "pretty much all the names you've heard of" have called him about this. They reference a recent paper with Scale Microgrids arguing for 90% solar and 10% gas, on which Handmer was the minority voice arguing for 100% solar, a case he has also made on his blog. Many of these callers are discussing about 5 GW over the next few years at over 90% solar. His forecast is that by 2027, the majority of new data centers breaking ground will be mostly solar. When Patel notes that 2027 groundbreakings are being planned now, Handmer says, "That's why they're calling me." He adds that without visibility into companies like Meta, he doesn't know when they'll hit limits on transformers or on "municipal peak load" shaving.
He describes that latest idea in detail. In a handful of US locations, perhaps where an aluminum smelter once stood, there is latent grid capacity, along with old generators running at 40–50% capacity factor that could reach about 80%. A data center could pay to run such a plant harder, locate at the old smelter site, and promise to curtail when the grid needs power. That effectively requires a large captive battery plant, which Handmer says "arrives on a truck." The advantages are that the power already exists and little land is needed. Solar's drawback, he admits, is that it is "a farming operation" requiring huge amounts of land, less than 1% of which goes to batteries, roads, and data center buildings.
Land: is there enough, and where?
Patel relays what someone in the industry told him. Energy is a small share of data center cost, but the real difficulty is securing tens of thousands of contiguous acres and getting permits and interconnection. Handmer calls the idea that there isn't enough land "garbage." Anyone who has looked out an airplane window over the US, especially west of about 110° longitude, can see how much there is. The land doesn't need to be flat. Nevada alone is roughly 80 million acres, about 90% of it federal. He wouldn't pave all of Nevada, he jokes, but people might visit to see the solar.
He also rejects the idea that Europe lacks solar. He says it's sunny for up to 20 hours a day in summer, though it's seasonal, and suggests southern Europe, especially sparsely populated Spain, as the place for a hypothetical 100 GW European AI build. Seasonality matters because an expensive GPU fleet needs about "four nines" of uptime to maximize tokens per dollar spent on the whole project, which means heavy solar overbuild to cover winter. He doesn't see overbuild as waste. We produce about 40% more food than we need, and that is much better than producing 40% less. A data center whose solar plant produces more than it needs 99% of the time could supply the surrounding town at essentially zero marginal cost, reversing today's dynamic where utilities tell data centers to disconnect. Patel adds Brian Potter's analogy of buying a 1 TB laptop while using 100 GB, because extra capacity is cheap.
Contiguity isn't required either, Handmer says. Parcels can be wired together. His ideal is solar arrays "as far as the eye can see" with a central node. By his estimate, that node's floor area is about 10% racks, 10% access space, perhaps 50% stacked batteries, plus cooling, at anywhere from 100 MW to 10 GW. The only external connection needed is optical fiber, or microwave or laser links, though he doubts Starlink has enough capacity. The result is a self-contained, off-grid "world of computation" on private land somewhere in Texas "where no one lives."
The arithmetic of a 5 GW solar data center
Patel notes that rack power density is heading toward a megawatt per rack. Handmer uses that as his unit. One 1 MW rack, with cooling left to air-conditioning specialists, needs about 24 hours of battery storage, enough to get through two bad nights. He says even less is needed in South Texas, and that power draw can be cut substantially during a run of bad days with only a small loss of compute. At about 4 MWh per Tesla Megapack, that's about six Megapacks, each roughly a truckload, next to one truckload of rack.
Texas solar runs at roughly 25% utilization, so averaging 1 MW would need about 4 MW of panels, roughly 4 acres, if every day were identical. Hitting four nines instead of one requires about 2.5x overbuild, or about 10 acres per megawatt. A 5 GW campus therefore needs about 50,000 acres, perhaps 10–20% less at scale. For comparison, he says Oak Ridge and Hanford were each set aside at about 100,000 acres during the Manhattan Project. Hanford was planned with space for four plutonium-producing piles, spaced out in case they exploded, and in the end only two were needed.
Patel brings up an Austin Vernon post: if diesel generators could cover about 10% of winter generation, the required solar could fall by about 60%. Handmer agrees there's a trade-off and calls it an easy optimization. Take NREL solar data for a location, simulate a year, and add panels and batteries until you hit your target uptime. The optimum is broad, and a third source such as diesel or a gas turbine can be added.
Patel asks why a company in a hurry would hire an army of perhaps 30,000 workers to cover 50,000 acres of desert rather than outbid rivals for the last 50 turbines. Handmer says Meta, in particular, has realized that Zuckerberg "is running out of time to spend his money to win," but insists the capex isn't crazy. Five gigawatts of GPUs is around $250 billion by their shared estimate, while 50,000 acres of Texas land might cost hundreds of millions of dollars, about 0.1% of the total. The "usual suspects" quote about $1 million per megawatt of installed solar, but Handmer says untariffed modules cost about 8 cents per watt, or $80,000 per megawatt. Since the panels are "the magic part" that converts sunlight to electricity at 25% efficiency, he argues everything else should cost less than they do, and calls this a cost problem Terraform works on. His takeaway: hyperscalers are "not power cost sensitive," they are "power availability sensitive," and solar is the best way to "firehose" energy at a problem because "it rains down from the sky."
Regulation, not tariffs, is the real handicap
Handmer expects electricity prices to rise, but calls that a reflection of "regulatory irrationality," which he says is also true in Europe and Australia. Prices will keep rising until people demand access to power technology invented over the last 50 years. Tariffs don't matter much given cost-insensitivity. What does matter are environmental rules that block renewable deployment. He credits these for why "Texas is out deploying California 10 to 1."
He acknowledges that the early-1970s environmental laws had sensible goals. In practice, though, putting solar on private land in the middle of nowhere can trigger NEPA and a four-year environmental review. He says, with evident sarcasm, that the review generates enough paper that producing it has a bigger environmental impact than the solar project. In Southern California, because off-grid solar is new, projects can end up regulated like chemical plants. He argues that solar on desert land is arguably positive because shading improves soil moisture retention, and that deploying panels could even help reverse desertification. Yet solar faces stricter review than grading the land and pouring concrete, or parking rusting, oil-leaking cars on it, which in many cases needs no permit.
His main policy ask is a categorical exemption for solar. He would gladly post money in escrow guaranteeing that the panels are removed after 20 years and the land returns to desert. What drives him "to become the Joker" is paying a biologist $10,000 to report a tuft of grass that might be food for a bee species that isn't endangered but might become so, on land zoned for unrestricted industrial use between a rocket test stand and a chemical plant. He adds that he doesn't want to drive species extinct. But he argues that failing to move industry off fossil fuels within 10–20 years will make the US poor, the way he says Britain became poor after running out of coal, and will flood coastal cities through climate change. Addressing that, he says, needs solar synthetics and also sulfur injection, among other things.
Batteries as a substitute for the grid
Patel raises the long stagnation in transmission construction and the shortages of substations and transformers. Handmer points to his blog posts, one prompted by a conversation with Patel about two years ago and the more recent "How to feed the AIs." He says he's dead serious about the latter even though it is "the most out of the money bet": everyone he considers a respectable forecaster in this area disagrees with him.
His argument starts by conceding that the grid won't get cheaper or easier to build. He says DOE projections of needed grid construction over the next decade and actual construction are "not even in the same order of magnitude." But batteries, he argues, do the grid's job in a different dimension. The grid performs a spatial arbitrage, moving power from where it's cheap, next to a power plant, to where it's expensive, your house. Batteries perform a temporal arbitrage, storing power at one time and releasing it at another. Before batteries, the only meaningful storage was pumped hydro, limited in location, capacity, and efficiency. He expects batteries everywhere: at solar farms, at substations, at retired power plant sites, and in homes. Per-person battery ownership has gone from perhaps 10 grams in a phone to around 100 kg for a Tesla owner, which he calls four or five orders of magnitude, and he expects the trend to continue.
The economics then turn against grid operators, in his view. Solar's daily swing from midday surplus to evening demand is predictable, so batteries are used perhaps 300 days a year. Expensive high-voltage lines and substations that exist for bad-weather contingencies see peak use almost never. Batteries installed behind the meter, outside operators' control, keep eroding the use of those assets while their operating costs rise. He concludes that the average distance an electron travels from generation to use is already shrinking and will shrink "pretty radically."
Weather forecasting helps too. Patel notes you can predict solar output three days ahead but can't change how many batteries you have. Handmer says you technically could by trucking batteries around, but doubling battery size will be cheaper as prices fall. Forecasts let you curtail 5% now rather than 50% in three days, keeping total annual curtailment to about five hours instead of about 24, and staying at four nines.
Why GDP may undercount AI
Patel then turns to what an AI-heavy economy would look like. Hardware value depends on software, and today's models aren't hugely valuable in pure economic terms. He says OpenAI's roughly $10–20 billion in annual recurring revenue is less than McDonald's or Kohl's yearly revenue. But human labor earns roughly $60 trillion a year in wages, and human-level AI would be worth at least that. Handmer agrees it's a lower bound and compares it to estimating Caterpillar's maximum market cap by counting men with wheelbarrows.
Handmer frames industrial revolutions as ways to bypass bottlenecks. Before the Industrial Revolution, the bottleneck was metabolism: how many oats humans and horses could digest into mechanical work. Today about 99% of the energy we use bypasses our guts. AI, he says, routes around cognitive constraints, continuing what writing, printing, computers, and the Internet began. His example is the credit card, which replaced the cognitive work of building a trust network with centralized trust.
Patel credits James Bradbury and Gwern with the point that GDP may make AI look underwhelming, much as the Internet's consumer surplus is hard to measure because much of it is free. Handmer says the same is true of oil. By his account, oil is about $8 trillion a year, yet per unit of energy gasoline is roughly 100 times cheaper than the cheapest food. Valuing oil at food prices would put it around $800 trillion, and cheapness obviously didn't hurt us. Patel adds that oil is a small share of GDP, yet oil shocks have caused double-digit GDP drops, so demand elasticity matters more than share. He paraphrases Gwern: a "data center of geniuses" would register in GDP only through its inputs (chips, energy) and its outputs (tokens), neither huge relative to the value it produces. If it automated human work, it might even lower nominal GDP. Patel suggests that energy use may eventually be a better measure of a civilization's size than GDP.
Handmer largely agrees on pricing. Where AI competes with humans, it keeps some pricing power. But in new kinds of work where AI mostly competes with other labs, he guesses prices would fall to a small multiple of the marginal cost of tokens. An AI as capable as a $200,000-a-year researcher, one who by Handmer's self-deprecating estimate produces maybe 10 hours of top-level work a week, would be worth far more because it can be copied, and far less at the margin because it costs only the H100s to run.
Patel carries this into a land-based valuation of cognition. Assuming 10 acres of solar per megawatt of H100s and roughly a megawatt equaling a thousand human-equivalents, an acre of desert becomes worth far more than the best farmland ever farmed, at current hardware efficiency. He adds that an H100 is said to match a brain's flops but uses about 50 times more power (around 1,000 W versus 20 W). If hardware reached brain-level efficiency, he figures that one acre could support something like 50,000 "AI souls." Handmer says it could be much more, because neurons are much slower than transistors. He recalls that a phone saves power by sleeping thousands of cycles between keystrokes, since everything humans do is "glacially slow" to a computer.
Stripping the stack down to silicon
For 2035, with AGI bottlenecked only on deployment, Handmer asks what minimum matter the computation needs. Racks, grid, transmission, and ISOs aren't necessary. He credits Elon Musk's habit of deleting anything not absolutely needed. What's left 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." In space, that's all, since the sun always shines. On Earth, you add batteries and interconnects, but not transformers or even DC-to-DC converters. A buck converter or relays can match array output, battery charge, and GPU draw. A solar panel about the size of a desk produces about 500 W in full sun, he says. He imagines aliens with a different silicon stack integrating the solar cell and a patch of "computronium" on the same wafer, since it is "all silicon all the way down."
Pushing the thought further, he describes a future TSMC manufacturing integrated solar dies that act as solar sails. They fly closer to the sun for more power, up to their thermal limit, or farther out to explore, and steer with LCD panels built into the wafer. When Patel asks what the post-human state is, Handmer answers that one human brain could be simulated with roughly a square meter of paper-thin silicon floating in space. He calls this "the future human form" and "the attractor state," assuming some software improvement. Patel wonders whether that's what a Dyson sphere would be made of.
Materials, Handmer says, aren't the limit. Chipmaking and solar both start with silicate rock, which is chemically reduced, purified to about six nines for solar or nine nines for high-end computing, then grown into crystals and cut into wafers. He says a PV cell needs only about 20 microns of silicon, ordinary dirt contains plenty, and a new silicon refinery takes about 18 months to set up with current technology. Nearly free solar power could let industry revisit processes optimized for energy efficiency, trading more power for less capex and shorter lead times, such as less efficient electrolytic reduction. Noting that he's "not really a chemist," he describes the silane (SiH₄) route: turn silicon into a gas, separate out contaminants, then heat it to recover pure silicon.
An energy singularity, and a possible collapse in complexity
Patel says discussions of the AI singularity usually focus on cognition, but unbounded cognition would both supply and demand more energy, so what does an energy singularity look like? Handmer offers a speculation he says he has no strong reason to believe either way. Evolution has produced ever more complex ways of using thermodynamic gradients, from RNA-based organisms to industrial economies. We may now be seeing the start of a collapse toward the simplest possible thermodynamic-to-cognition stack: stellar fusion against cold space as the gradient, and silicon as the converter. Electrons are pushed across a gap in a solar cell and return through logic gates, "making decisions about things and then beaming lasers to their friends," announcing a new meme. Patel calls this the possibility that after four billion years of increasing complexity, there may be a big collapse.
At the end, Handmer describes Terraform, founded almost four years ago. It makes synthetic natural gas from sunlight and air, has a methanol process, and is working on ammonia, steel processing, desalination, cement, and, he says, essentially everything primary industry does except glass and paper. He says the team is mostly mechanical engineers and that the company is still small. He notes he expected competitors by now, but the only one doing similar work is a small UK startup. He ends with the aim of eventually building the same systems on Mars, helping "our robot overlords make more of themselves out of dirt." The core forecast he leaves on the table is that most new AI power will come from off-grid solar and batteries within a few years, which he acknowledges nearly every forecaster he respects disagrees with.
Today I'm interviewing Casey Handmer. Casey has worked on a bunch of cool things. Caltech PhD on some gravitational wave black hole gimmick stuff, then Hyperloop, then the Jet Propulsion Laboratory at NASA. Now he is founder and CEO of Terraform Industries. Casey, welcome. Thank you. It's great to be here finally.
Big picture question I'm interested in. To the extent that AI just ends up being this big industrial race - who can build the most solar panels? Who can build the most batteries? Who can build the most GPUs and transmission lines and transformers? This is not what the US is known for, at least in recent decades. This is exactly what China is known for. They have 20x the amount of yearly solar manufacturing the US has. Obviously we have export controls right now, but over time SMIC will catch up to TSMC's leading edge. What is the story of how the United States wins this? Why does China just not win by default?
Do you think that China is better at capital allocation than the United States? Do you think the Chinese business environment is better for business than in the United States? I feel you can make these first principles arguments about these other industries where they're killing it, but it doesn't seem to have hampered BYD or CATL.
People say they're so much better at building high-speed trains than the United States. 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. Why would you devote, in 2025, so much industrial effort and money…
They're devoting a lot to solar overcapacity, which in your opinion is the key to future industrial growth. I think they might be accidentally correct. They called the most important thing correct, right? Which should count for something.
Well, they're in a similar situation to Europe, but unlike the United States. 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, with no good mountain ranges or rivers or anything to really separate them. They get almost all their oil from the Middle East. From countries that they don't control, don't have strong diplomatic relationships with, on fleets of oil tankers that they can't defend because their navy doesn't have the ability to operate effectively in the Indian Ocean.
But you're working on this, right? If you get synthetic fuels working at Terraform. Doesn't that asymmetrically help China? Which might be fine. 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. There are projects in China right now working on synthetic fuels. It would not surprise me if they were thinking pretty seriously about this.
Just to spell out for the audience, China has all this electricity production. And the bottleneck is that only a third of final energy use in a modern economy comes from electricity. The rest, you need gas and whatever to transport things... Or coal. They use a lot of coal in China. Right. What Casey is inventing is a technology to turn that electricity, which only can supply a third of end-uses right now, into synthetic fuels which can supply 100% of the electricity your civilization needs. China's energy advantage then becomes overwhelming. This technology levels the playing field.
It levels the playing field a lot. But at the end of the day, China still contains the poorest Chinese people anywhere on earth. Never underestimate the capacity for an autocratic dictatorship to shoot itself in the foot.
I don't know. I agree that they've obviously made bad decisions, but even if you have the poorest Chinese people anywhere in the world, they can still be quite rich. Like Singapore is richer, whatever. Also, there are parts of China which actually contain quite rich Chinese people. You have to compare not all of China against the US, but Shanghai and Guangdong against the United States. You can have a part of China that is as big as America and as wealthy as America and as innovative as America. Like the Indian middle class is larger than the US middle class. But also it's nowhere near as wealthy. Whereas there are parts of China which are humongous, which are actually as wealthy as the United States, and in many cases as innovative, etc.
Yeah. I'm saying don't underestimate it, but at the same time, we want to find the truth here. The truth is we should not count the United States out of the battle and just give up. We're very much still in the race now, provided we don't take extra effort to shoot ourselves in the foot.
Right now we are export controlling chips for the purpose of keeping our AI lead and we recognize this is a key input in our ability to compete in AI. So we are going to export control China's ability to have these chips. Energy is also a key input in this AI race, and if China wanted to do the converse of what we're doing to them with these cheap imports, what they would do to us is to export control solar and batteries.
It would be asymmetrical, it would hurt them worse than us. If they did tariffs? China obviously depends upon the US export market for its economic dynamism. It's going to hurt both parties to sever the link. But if you sever the link completely, China's ability to make advanced chips right now is basically not there, whereas the United States can make them. 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. If we decided we want to produce 100 gigawatts of solar capacity every single year… We’re already on track to do that.
Is it going to be as cheap as it is to do in China? My views on this are somewhat different from the mainstream, which is great because this is a podcast. The mainstream view would say China has cheaper labor, which is no longer true, because they compare to Mexico. And it's got lower environmental regulations, which is true. And that it is more business friendly, which is absolutely crazy. There's no way you could justify that your company having to have an inspector from the CCP on its board who harasses you about Xi Jinping every day helps you do your business. And also the rule of law is not great. So you're constantly having to pay bribes to people in order to stay in business.
The idea that the United States cannot compete against that with mostly or fully automated solar panel manufacturing in the United States—which has cheaper natural gas by far, abundant oil, abundant human resources, great financial capacity, world leading automation, etc.—is crazy. We could literally copy paste solar manufacturing factories.
How much additional solar power capacity do you think we could be putting on, that's manufactured in the US, by 2028?
This is a good question. When Russia invaded Ukraine, I thought, finally the Europeans will see sense and they'll pull the trigger on, "We need to localize production of solar panels from dirt to the finished module," which is roughly a four-stage process. They didn't. They're still paying Russia a billion dollars a day for the privilege of being invaded. But at the time I thought they could probably do that in about two years. I think the United States could probably do that in two years or less if you started today. It's currently 11 o'clock. So we're going to start cutting checks by noon.
You could ramp up pretty quickly. A lot of technology already exists here. It's not like it has to be invented from scratch. It's mostly a case of putting in a phone call to all the different manufacturers here, in Germany, and so on and saying, "We need you to 10x the size of your factory, starting today, blank check, go."
A lot of your predictions seem to be not predictions, but more like, "If we had World War II-levels of motivation, if we had Manhattan Project-level intensity around doing a specific thing, how fast could we do it? Like if Elon was running the government, how fast could it happen?" He was for a brief period. Maybe then we should put it like, "If Elon ran the government like he ran SpaceX." As opposed to the question of "What is actually practically likely to happen, given that we are not treating it with World War II-level intensity."
If you look at xAI, which Elon is involved in obviously, what are they actually focused on right now? They're focused on the chips because they understand the key bottleneck is the chips, not the solar power. 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. If you need solar to run your data center, it doesn't hurt in terms of the overall cost picture. It doesn't matter at all.
What matters is having the chips at competitive capabilities per chip, and enough of them installed in your PCBs, in your data centers, hooked up to your liquid cooling, ready to go. That's actually something that Elon and his companies are great at. It's figuring out this mass production, semi-automated mass production. They've got this facility in Texas which is making the Starlink receivers, completely automated. At what point does, "Oh, we don't have a solar panel factory," become on the critical path? I very much doubt it's ever going to be on the critical path. There's dozens and dozens of manufacturers of solar panels worldwide that are all competing against each other.
So you're a big solar bull. Yeah. Right now the hyperscalers are making decisions about the data centers that they're building. They're going to be 1-2 gigawatts, 5 gigawatts in Meta's case. They’re making decisions about how they're going to be actually powered. The people with actual money on the line are choosing natural gas. It's not like they can't see the learning rate. They're building things which will be online in 2028 or 2030. Why are they wrong and you're right?
It's their job. They probably know more about it than I do. But in all seriousness, if you're like xAI right now trying to build the Colossus data center in downtown Memphis, you want to get it done super fast. "What are all the different things we need? What are the factors of production to build this? We need a building. We don't have time to build a building so we'll buy a building. Okay. We'll adapt it. We need power, we need thermal cooling." That stuff you can deliver on a truck so that's what they did. You need access to gas. They had access to gas there. They could tap into a local gas line.
If you can tap into a gas line, generally speaking, you can get enough power. The energy transmission capacity of your regular gas delivery pipelines is way, way higher than electricity overhead lines, and it's easy to upgrade. So if you're in this situation right now, you say, "Are we constrained by our ability to go and rent gas turbines?" No, they're not, because there was enough available once, maybe twice. But at a certain point, you realize as you grow, you start to touch all these additional constraints.
Some of those constraints include gas availability. So there's a lot of chat about doing this in Pennsylvania where there's quite a lot of stranded gas, and in parts of Texas. But at the same time, the United States is gearing up in its ability to export natural gas overseas, so the price will not be infinitely low forever. You start to run into constraints around turbine manufacturing rate, around transformer production rate, around grid capacity, and also running into problems where the AIs and the humans who depend on legacy electricity production and delivery utilities are competing with each other. We just saw this recent forward auction in PJM result in very high, unsustainably high prices for consumers who depend on cheap electricity to heat and cool their houses and have general prosperity.
If you look far enough in the future, you can just turn up the dial arbitrarily high. You can say we're going to put in a gigawatt a year. Well, we can meet that constraint with gas turbines. We're not going to run out of natural gas at 1 gigawatt per year indefinitely. 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.
Not to reach prematurely for analogies, but Henry Kaiser set up the shipyard in Richmond, just down the road here in San Francisco near Berkeley. He was initially making ships for the British and by the end of the war he had four separate shipyards operating in parallel, to the point where he was bottlenecked on his supply of steel. Steel was rare enough in the war, because everyone was using it for different things, that 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.
That's the same sort of situation you have here where you have these massive industrial verticals. Here I'm 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. They can reach all the way through down into primary material supply if they need to. The reason that these current plans are being done based on natural gas is that this is the sort of…
PJM has all kinds of different sources of power. They have nuclear as well, they have gas, they have coal, all kinds of stuff. This price here is probably driven more by the delivery cost growth than by the generation cost growth, if that makes sense. When you pay your utility bill, the cost is sometimes broken down into a delivery cost and a generation cost, sometimes importation costs and other things. The delivery cost is what it costs the utility to build and maintain all the power lines that connect all the houses to all the power plants in some gigantic area divided by your marginal usage, with all kinds of other complicated rules designed to make it fairer.
The problem that we see—and the reason that PG&E here in California, for example, is perpetually on the brink of bankruptcy—is that 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. Why? Because you've got generally unionized labor that has to build and maintain power lines in areas that already have built up infrastructure. You have multiple collisions, whether this is a power pole on your own street or building a new transmission line which requires you to eminent domain land. So you're in court for years and years and years spending public money litigating against other people who are also spending public money to litigate against you on behalf of other interest groups and so on and so forth. Then you've got wildfires. It's just the poster child for Baumol cost disease.
One of the reasons that we're going to see large-scale pruning of these grids is that we just can't afford under our current regulatory regime to maintain. When you say pruning, will everything just go off-grid? It's fairly clear to me that 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. If you had an aluminum plant back in the day, you would be building your own power plant for it as well.
It seems inefficient to have redundant power plants at every single industrial site. Let me paint a grand vision for you. It would seem inefficient, but if you are sensitive to the cost of power expressed in supply elasticity or something like that, you just have to do it. There's no two ways about it. Is it inefficient for the xAI Colossus
data center to have its own captive power plant, which it does on the backs of a bunch of trucks in the parking lot? No, it's not inefficient. It's the cheapest way for them to get power.
Okay, AI might be a special case. But big picture question. Across different kinds of ISOs, from Texas to Pennsylvania to whatever, people are building data centers which will not be online for many years. They're choosing natural gas. What's going on?
We haven't completely exhausted the supply of turbines relative to GPUs.
Do you have some estimate of when we'll run out of them? Because we can also make more.
Everything before about 2030 is spoken for at this point. Yeah, you could make more turbines. The funny thing is that it's actually relatively expensive to spool up additional production of these turbines.
Here's one thing you have to grapple with sooner or later. Conventional power generation is a steam engine. You have some kind of chemical that you find inside the earth that is out of chemical equilibrium with the atmosphere and you burn it and it makes heat. It could be coal, could be gas, oil.
You're giving me the true birds and bees here.
Yeah, exactly. And it makes heat and you boil water. The water goes through some kind of mechanical contrivance that creates motion. That motion twists a magnet and generates an electrical field which then pushes electrons down wires, which then push electrons through a series of gates that then approximate thinking.
It's kind of complicated. But the key step in this is converting heat into electricity in the most efficient way. 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. The jet engine on an aircraft is a Brayton cycle as well. Anytime you have a Brayton cycle with a bunch of Inconel spinning at high speed, it's just going to cost you a bunch of money.
Because it's inherently inefficient or what? It's just inherently expensive to build.
Okay. What is the cost of… GE makes these 100 megawatt gas turbines, right?
I don't actually know what the retail price is. I would suspect that if their price is flexible, it would have gone up a lot. But if I recall correctly, $35 a megawatt hour is just the floor cost for...
How much, sorry?
$35 a megawatt hour just for the Brayton cycle. We're not talking about the fuel, we're not talking about the heat exchangers, we're not talking about the cooling ponds or anything like that. Just the amortized cost of the high-speed, high-temperature spinning components is $35 a megawatt hour.
Do you think the hyperscalers are being irrational, or do they have some reason?
To be clear, they don't care about the cost of power. This is very counterintuitive. For Grandma Kettle in Pennsylvania, she's very sensitive to electricity costs. We don't really want her to suffer in her retirement from unaffordable electricity costs and having to sit there shivering. That's not the image that we want.
At the same time, what is the economic value to you of using Claude or Grok or whatever you use on a monthly basis?
A lot.
It's obviously much more than the subscription, but is it 10 times more than the subscription maybe?
Yeah, easily.
Let's say the subscription is on the order of $10. The value is on the order of $100.
No, it's probably more like $100 and $1,000.
How much does it cost xAI or Anthropic or whatever to serve your usage? The marginal variable cost of serving it, in electricity, is less than 10% of the actual cost of… Well, their cost of serving it is maybe a buck per million tokens or something like that. The cost of electricity is about 10% of that. So 10 cents of electricity is generating $1,000 worth of economic value.
It's very obvious that Anthropic could be like, "Our electricity cost basis has increased by a factor of 100. Now instead of paying 10 cents on your bill, on your $100 bill for power you're paying $10. So we're putting your subscription up to $110 for an electricity capacity charge." Then they could go out and buy turbines for prices that would make your eyes water.
Okay, so then why are we going to get the solar future? In 2032, we're going to have hundreds of gigawatts of extra demand for data centers and at that point, most of it is coming from solar? Why is that?
There aren't enough turbines being manufactured. But also, I think in the early 2000s… We can probably overlay the graph of how many turbines were being manufactured. Right now, we're at historical—
They've ramped up basically to the early 2000s rate again. But I don't know, you have to make more solar panels as well, right? There will be supply elasticities for both solar and natural gas. Is there some reason to think that it's worse for the supply chain involved in having a natural gas-powered data center than a solar one?
Yeah I do. The learning rate for natural gas is nowhere near as steep as solar. It just tells you that it's easy to make solar panels, much easier to make solar panels. There are very few manufactured products which are easier to make. The Wright's Law coefficient is 43%. So every time we double cumulative production, we get a 43% reduction in cost.
What is the basis of that? Why are we finding 43% worth of things that can be made cheaper or more efficient every single year?
Roughly speaking, there's 10,000 manufacturing process engineers working on this full-time.
That could be true of any process, but no other process sees the kinds of learning rates that solar is seeing.
That's not strictly true. In order to sustain this over a long period of time, you obviously need to have demand elasticity that exceeds your learning rate. Otherwise, you would, after a couple of OOMs, saturate your market at the current price and you'd have no additional growth. But in this case, roughly every two years we're doubling production. Every 2-2.5 years, we're doubling production and the price is coming down by a factor of ~40%. So it's roughly 15-20% per year. Then just as a result of that price reduction, demand skyrockets by probably six times more than that additional marginal production capacity increase.
This is one point where I'll say the so-called pros are definitely wrong. Conventional wisdom is that solar demand is going to saturate this week. It's going to saturate. We've got a graph here somewhere that's like, "This year is it, it's never going to grow anymore." Instead, it's just blasting out the top of the graph. This conventional wisdom is wrong. Not only are solar adoption, production, and price decreases continuing, they're accelerating. And the rate at which they're accelerating is still accelerating.
The rate at which it's accelerating is accelerating?
Yes.
As measured in the total fraction of energy that's coming from solar?
In the sense that its fitness for the markets that it is being produced for is increasing over time. So it's still extremely early. We're still in the Apple II computer era of solar.
Backing up, if the story is that the reason solar is getting cheaper is because there's a lot of demand for more solar, and that demand can sustain economies of scale or whatever is going on…
Yes. I'm going to go on a limb here and agree with Elon Musk on this.
Then shouldn't that also be true of gas turbines and transformers and power stations and whatever else that's required for the non-solar future? We're expecting AI to drive up demand for power regardless of the source. To the extent the story for solar becoming cheaper over time is just that demand will go up and that will drive efficiencies, why isn't that true for…?
Let's say you're a bank, and you're trying to decide whether to lend GE a bunch of money to expand production of their gas turbines. You can write them the check today. They'll start scaling up their factories. They'll start to see the benefits of that in three or four or five years. You don't know if the AI bubble will have burst by then. You don't know if China will have invaded Taiwan by then. You don't know if Siemens or Philips or someone will have outcompeted you. You don't know if GE's major looming structural problems will cause it to be unable to compete, as it has in the past.
In order to make that money back, you also have to then operate that plant at that capacity for 20 years. If I was looking at the same charts as they're looking at right now, I'd say, "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.
I feel like there was actually a similar discussion a year back when AI people were like, "No, AI is real. This is going to happen." Then SK Hynix, Samsung, etc., were like, "We're not ramping up HBM production because HBM is used largely for AI workloads, and if this demand doesn't continue, then our additional manufacturing capacity for HBM will not have been worth it." Then there was another bottleneck with CoWoS. What happened after that? Did they end up indeed ramping up their production?
I think so. Well, so when someone says, "We can't do it, we won't do it, no way, no how," what they're saying is, "Write me a check." And they did. Now Samsung's coming on board in the States to build AI6 with xAI, I think. So they all got there in the end.
Maybe it's worth going into the numbers. Right now, 43% of US data center power consumption is from natural gas. Basically, you think asymptotically that it will be 100% solar if you go to 2040?
Yeah. Obviously legacy production, coal and stuff, is going to retire over time. If a gas plant is still making money, people will keep operating it. But at a certain point… 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. Also, capacity is going to increase a lot so that helps to dilute the existing production.
And the amount of use is going to increase a bunch. The amount of data center use of energy will just be exponentially higher. So the new stock matters a lot as compared to the existing stock.
Anyway, I want to know in 2027, what fraction is natural gas? In 2030, what fraction is natural gas versus solar?
For new load?
Let's say new load. For new load, 2035, etc. If eventually you're right that we'll pave the earth in solar panels to sustain quadrillions of AI souls, what is the pace of that?
The question to ask is, what is the major constraint on that ramp up? Then everything else will just draft in behind. I suspect that the hardest thing to make will always be the silicon, like the GPUs. So the question is really, "How quickly does TSMC ramp up its production of GPUs?" That's a question for you, not for me.
I'll use some numbers that AI 2027 used for their compute forecast. Even if you don't buy their singularity thing, I think they did a reasonably good job with crunching the numbers on their compute forecast. I think they said there's on the order of 10 million H100 equivalents in the world today. I think they said by 2028 there'd be 100 million, so basically 10x more H100 equivalents in the world.
About a kilowatt each, something like that.
Okay, so that's like 100 gigawatts. That sounds roughly right. You're not the first person to give me a call and ask me about this. I'll put it that way. I'm not going to name names. Pretty much all the names you've heard of have given me a call and said, "We know that you're a minority voice on the paper that came out recently with Scale Microgrids talking about how you could do 90% solar, 10% gas." I said, "You can go all the way 100% solar." I wrote a blog post about it. So they always call me up and say, "What about this?"
They're all talking like 5 gigawatts in the next few years. That's just like 90+% solar for just those. So within a few years, we'll probably see that the majority of new data centers that are going in will be mostly solar.
Within how long?
Let's say by 2027, the majority of new data centers going in at that point would be mostly solar.
Going in as in…?
Groundbreaking at that point.
But if you're groundbreaking in 2027, you're probably planning it now, right?
That's why they're calling me. My consulting fees are extremely affordable. But I don't have deep visibility—because I'm not in the same room with the Meta people—as to when we're going to hit the wall on transformers and when we're going to hit the wall on just how much municipal peak load we can shave off, which is the latest thing that's been doing the rounds.
It turns out there's a handful of places in the United States—and by handful, I mean literally a handful—where there might have used to be an aluminum smelter. There's a bunch of latent capacity in the grid. And there's also a bunch of generators on the grid that are notionally turned down. They operate at, say, 40-50% capacity factor, but they max out at about 80% capacity factor because you've got to bring them down for maintenance pretty often, especially if they're old.
So they're saying, "Well, you know, we could pay you just to operate this old coal plant or something at higher capacity. It'll go down this power line to this place where the smelter used to be. We'll set up there, and then we promise to curtail when you need the power." That basically means they just have a massive captive battery plant as well. Which is fine, you just buy that and it arrives on a truck. The major advantage to doing that over the pure solar play is that the power is already there, so there's no risk there. And you don't need a massive amount of land.
The problem with the solar approach is that there's no two ways about it. It's a farming operation. You need a huge amount of land. The total amount of land that you're using, less than 1% is under batteries, under roads, under data center structures, et cetera, etc. It's mostly solar.
Let's get into what this looks like. If you've got a 5-gigawatt plant you want to build, break down the numbers for me in how much land in terms of solar you need to farm this out.
I was talking to somebody in this space and they said, "Obviously, the cost of energy for these data centers is a small fraction of the total cost. Most of the cost is going towards chips. So then the issue is just, can you make the energy available?" They were saying that even though solar panels themselves you can acquire, the issue is getting that much contiguous land and getting the permitting to interconnect it. That's apparently a big hassle.
It's kind of a nightmare.
So they're like, "Well, at that point, is it actually easier than just getting on the grid or…?" Anyway, if you need tens of thousands of acres of solar, where can you do that?
Basically in Texas. There's this very popular misconception that there's not enough land to do solar. This is garbage. If you've ever flown in an aircraft in the United States and you've ever looked out the window, you'd be like, "Oh, wow, look, there's a lot of land you could put solar on." Especially west of like 110°.
Does it need to be flat, or no?
No. Doesn't matter. Do trees grow on mountain slopes? So it doesn't matter. For reference, Nevada is something like 80 million acres. Just Nevada, which is like 90% federal land, is 80 million acres.
I would never say that we should sacrifice Nevada to the AI and pave the entirety of Nevada from one wall to the other. But I just saw a bunch of things in my feed the last couple of days that Vegas is falling apart. The boomers are retiring. No one goes there anymore. People would go to see the 100 million acres of solar.
Even if you did it in Nevada—
We can do it anywhere. You can do it anywhere you can find the land. People say that you can't do this in Europe because Europe doesn't have solar power. Europe has solar power. I've been to Europe in the summer. It's sunny for 20 hours of the day. It's a bit seasonal. But that's not a big deal.
But I mean it is. Because energy is a small fraction of the cost, you care more about making sure the chips are running all the time, right? In practice, what happens is… Let's say Europe hypothetically awakens from its slumber and decides it wants to participate in AI. I hope it does. They say, "Well, we're going to have to put 100 gigawatts of solar down at some point to build these data centers. It will most likely be in southern Europe. Spain is not particularly heavily populated. That's a great place to start. So we put in 100 gigawatts of solar data centers in Spain."
Basically, if you're spending AI hyperscaling money on your GPUs, you want to have four nines of uptime in order to maximize your tokens per dollar spent on the entire project, not just on that. This is a very subtle point. I can go into vast detail on it later on maybe. Let's just say you need four nines of uptime. In order to achieve four nines of uptime in the middle of winter, you need to have a lot of solar overbuilt.
Is solar overbuilt a bad thing? No. Is the fact that we produce 40% more food than we need a bad thing? No. It's much better than producing 40% less than we need. It just means that effectively, you have a giant captive power plant attached to a data center that 99.9% of the time produces more power than it needs. 99% of the time it produces much more power than it needs.
That can now actually be the source of power for the local utility, which, instead of being like, "Naughty, naughty data center, you must disconnect when we tell you to", they say, "Hey, data center, I noticed you've got a bunch of power you're not using 360 days of the year. Would you mind ever so much if we threw a power cable over the wall and we powered our entire town off your spare power at essentially zero marginal cost, plus whatever residential batteries that we need in addition to local power supply."
Brian Potter had a good analogy in his blog post about this. He's like, "My MacBook has a terabyte of storage and I use 100 gigabytes. I just got the terabyte version because it's cheap enough and I might need it at some point that it's worth it." You're saying solar gets so cheap that it's the way we'll treat hard drive space. We get a bunch of excess.
Also the market will be made at the new marginal consumption and production. All the people who are working in the space right now are like, "Oh, I'm in the business of delivering power or storing power. I'm going to serve the AI market because that's where all the growth is occurring." That's where all of US GDP growth is occurring right now.
I guess you didn't answer the question of, yes, theoretically we could do this, but is it going to be possible to get the permitting to have tens of thousands of acres of contiguous land?
It doesn't need to be contiguous. It helps if it's contiguous. It doesn't need to be convex. You can have a bit over here and a bit over there and you can wire them together relatively easily. In fact, in the limit, you have fields upon fields of solar arrays with…
Tell me your dream, Casey.
Fields, just solar arrays as far as the eye can see. Then within the solar arrays, roughly in the middle of them, you have your batteries and your…
I've played Factorio. I remember this optimal layout of batteries and solar.
You've got your batteries and you've got your data centers. So in terms of ground floor area, it's roughly 10% racks, 10% access to the racks, maybe 50% batteries stacked up on top of each other, and there's also cooling, something like that. That's in terms of what sits in the centralized node. That could be 100 megawatts or it could be 10 gigawatts, depending on how you want to scale this.
But then all you need to connect that to the outside world is an optical fiber cable which you can string up on poles, you can run it underground. You could even use microwave links if you really wanted to. You could use Starlink if you really wanted to. I don't know if Starlink would be fast enough. I'm not sure if its capacity is high enough. You could use laser links if you really needed to. That's it. It's this completely self-contained world of computation.
Because it's off-grid.
Yeah, it occurs off-grid, on private land somewhere in the backwoods of Texas where no one lives and no one will ever live because it's completely inhospitable to humans.
In terms of the ratios, one trend that was impressed upon me is that the power density of racks is increasing a lot as the flops per GPU are increasing. A megawatt per rack is what they're heading to now, which just seems bananas to me.
I think it was even more than that. Let's get concrete here for a second. Let's say you've got one rack and it's 1 megawatt. I'll leave the cooling to someone who specializes in air conditioners, but it's basically throwing air conditioners at the problem. Then you have batteries. So in order to get four nines of uptime on this… In South Texas, you actually need less than this. But let's just say it's 24 hours worth of battery storage. That means it'll get you through two bad nights in a row, basically.
Actually, it turns out that you can significantly decrease power consumption with a very small reduction in overall compute. So if you've got like three really bad days in a row or something, you can dial back your power usage quite a lot without compromising your inference or training.
Okay, so you've got, say, a Tesla Megapack, something like four megawatt hours. So one megawatt rack, and then six Tesla Megapacks, each of which is roughly one truckload worth of stuff. So one truckload worth of rack, and then like six truckloads worth of batteries.
Then in order to operate this at an average power of 1 megawatt, your solar arrays in Texas will be something like 25% utilization. So on average, if the sun came up every day and the day was the same length all the time, you would need 4 megawatts of solar arrays, which is about 4 acres of land. But 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. So 10 acres of solar, six truckloads of batteries, one truckload of data center, and some cooling stuff.
For how big of a data center?
One megawatt. That's just for one megawatt. So 10 acres, one megawatt kind of situation at four nines. If you want five gigawatts, then that's 5,000 times 10. So 50,000 acres. At a larger scale, you can probably cut all those numbers down by 10-20%, but it's on that order.
And 50,000 acres sounds like a lot.
It does sound like a lot, is it not? The amount of land put aside for Oak Ridge was 100,000 acres. The amount of land put aside for Hanford was about 100,000 acres.
What's Hanford?
Hanford was where they made the plutonium in the Manhattan Project.
But I don't know how big that was. Is it like, "Oh, this is so small," and then you're like, "Oh, but it's 100,000 acres," or…?
It's still largely unpopulated now because it's a National Laboratory. The reason they did that was they thought, "Oh, we're going to need four piles to produce plutonium." These are not nuclear reactors that produce exothermal energy, so you can't actually make nuclear power with them, but you're making plutonium with them. In the end, they only needed two. They wanted them spaced out because they thought they might just spontaneously explode, and a bunch of other facilities and plants and stuff as well.
Austin Vernon had an interesting blog post where he said that 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 because you don't need to plan for that contingency.
Yeah, there's a balance here. This is not a very complicated optimization problem. For people who do optimization problems for fun, this is how you do it. You start off with a bunch of NREL data on what your solar abundance is in this particular part of the world, and then you just start throwing solar panels and batteries at it over the course of a one-year simulation until you hit the number of nines you want. To an extent, you can trade the amount of panels and the amount of batteries you've got back and forth, and there's a very broad optimum. Or you can throw in a third thing like a diesel backup or a gas turbine.
The issue here is—if Meta or Microsoft or whoever just wants to get something off the ground—this might be low opex to have this huge solar farm, but it's high capex, where you need to hire 30,000 people to go in the middle of a desert and install 50,000 acres' worth of solar panels. They're like, "Why would I not just buy 50 gas turbines instead?" Why not just outbid Microsoft, or Meta outbids Google or something, for the last gas turbine that's available that year?
Totally. The thing that Meta has realized is that Zuck is running out of time to spend his money to win. The capex is not crazy high, just to be clear. The capex is still dominated by just the GPUs. How much does five gigawatts' worth of GPUs cost?
I don't know if my numbers will be wrong but $250 billion or something?
$250 billion sounds about right. Is 50,000 acres going to cost $250 billion in Texas?
That's so much money. Wait, I did the math in my head and like, that's a lot of money.
We're talking maybe hundreds of millions of dollars, something like that. So it's like 0.1% of the cost is land. How much does a megawatt of solar cost? If you go and ask the usual suspects, they'll tell you a million dollars. But this is one of the things that breaks my brain at Terraform, which is my day job. The modules themselves, without tariffs, would be 8 cents a watt, so that's $80,000…
8 cents a watt? But they're like a dollar a watt, including installation and everything.
Including installation and everything. But the panels are the magic part. They're the thing that turns sunlight into pure electrical energy at 25% efficiency. Everything else should be less than that. If you want to work on that project, come and work with us at Terraform because we're very cost-sensitive.
We'll give you an opportunity to shill, don't worry.
In all seriousness, the central takeaway is that the hyperscalers are not power cost sensitive. They are power availability sensitive. For all these things, you just run into this supply elasticity wall at the rates of increase that we're talking about. Solar is by far the best option for firehosing energy at a given problem because it rains down from the sky.
Between the fact that maybe solar prices will go down and the fact that demand is going to go up, do you think electricity prices are likely to rise?
Yes, but electricity prices at this point are a reflection of a regulatory irrationality. This is the same situation in Europe and Australia for that matter. Your prices will rise until you've had enough and you say, "No, we demand that you allow us to take advantage of power technology that's been invented in the last 50 years."
In terms of things that are causing us to lose to China, tariffs are neither here nor there because as we've discussed, we're not sensitive to cost on power. But the environmental regulations that are actively preventing us from deploying renewable energy in the United States… This is the reason Texas is winning. Texas is outdeploying California 10 to 1.
The regulatory environment around solar is just insane. It's insane. In the United States, part of the reason that solar has not been deployed at massive scale yet is that a bunch of laws went into action in the early 1970s that were intended to protect our environment. And that makes a lot of sense. And our environment's a great thing we should protect.
I think people will be familiar with NEPA and whatever, but how is it especially impacting solar?
Let's say you've got a bunch of private land out in the middle of nowhere, and you want to build solar on it. You'll probably end up triggering NEPA, at which point you now have to do what is not in the law but considered necessary under current regulations. 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. This is bonkers. It is crazy town.
The thing that drives me particularly crazy in Southern California is that just because solar is kind of new, and off-grid solar is very new, unless you're very, very careful you end up getting regulated as though you're trying to build a chemical plant even though it's a solar array. The impact of solar arrays on desert is arguably positive because it shades the ground and improves soil moisture retention. If you wanted to reverse desertification, you would basically just deploy solar panels on it and that would pay for the process.
But you end up having to go through more stringent environmental review process than if you just wanted to grade the whole thing and cover it in concrete, or if you grade it and then park a bunch of old rusting cars that are dropping oil into the aquifer, which in many cases you don't need a permit for at all. But to build solar, you have to go through this whole process.
If there's one thing that anyone listening to this can do, it would be to have a categorical exemption for solar deployment. Or if I put money in an escrow account that says after 20 years we have to pull this out—we'll pull all the solar out of the desert and it goes back to being desert—I will do that in a heartbeat.
But if I have to hire another biologist for $10,000 to be like, "Well, on that 40-acre plot we found a tuft of grass which we believe might be one of the 20 species that this particular species of bee sometimes eats, and this species of bee is not technically endangered but it might be at some point in the future… Therefore, you can't deploy there." Even though it's zoned unrestricted industrial and it's sandwiched between a rocket test stand and a chemical plant, for example, in an industrial part of the desert… I'm going to become the Joker. It is insane.
We need to be a bit balanced about this. I don't want to drive species into extinction. But the meta problem here is if we don't move our industrial stack off fossil fuels in 10 or 20 years… First of all, we'll get poor the same way the UK did, because they ran out of coal, basically. The second thing is we'll get poor because we'll flood our coastal cities than Florida underneath climate change. We need solar synthetics for that part. We also need to do sulfur injection and a couple of other things.
People will point out that transmission line growth has been stuck in a rut for decades. We have all these bottlenecks in terms of substations and transformers, etc. Why will this not hamper this abundant solar future?
That's a really great question. You and I had a conversation along these lines almost two years ago when we first met. It caused me to go and write a blog post.
This is a good way of thinking about it. There's another blog post you wrote, which was also related to a conversation we had, which is "How to feed the AIs."
That's much more recent. That was after dinner, I think. To be fair, I usually am fairly clear in my blog posts if I'm shitposting or if I'm serious, but this one actually I'm dead serious on. It's actually the one where it's the most out of the money bet as well. Everyone else that I consider to be a respectable forecaster in this area disagrees with me on it. That to one side.
We know why the grid is expensive. It's a lot of wires strung up in hard to reach places that are hard to maintain, especially as the workforce ages, with regulations and all the rest and eminent domain and so on and so forth. So the grid's not going to get cheaper anytime soon or easier to build. If you look at the projections of how much grid the DOE would have us needing to build in the next 10 years versus how much's actually being built,
it's not even in the same order of magnitude. You say, "Are we totally screwed?" The answer is, "No, we're not totally screwed" because batteries actually do the same job that the grid does. This is kind of weird. Hear me out. The grid transports power from one place to another. It transports almost instantaneously at the speed of light, so it's actually performing a spatial arbitrage.
The idea being that right outside the local nuclear power plant, power is really cheap because they make a lot of it. And in your house, power is really expensive because you don't have a power plant in your house. You pay the intermediary a small fee and they allow this trade to take place. That's basically how the grid works.
Until quite recently, the only way we had of meaningfully storing energy, storing electricity on the grid, was pumped hydro. That only works in a handful of places and with limited capacity. It doesn't work all that well either. The efficiency is not great. Now we have batteries. Batteries store power at one time of day and they release it at another time of day. Batteries are performing a temporal arbitrage, an arbitrage over time.
But they can be local or they can be more remote. I think we'll end up seeing batteries next to the solar arrays, and batteries in the middle of the grid at substations, and batteries on the sites of existing power plants that get turned off, and batteries in your house, and batteries everywhere in between.
One way of thinking of this is, what is your per capita allocation of batteries in kilograms per head? When you and I were much younger, the lithium ion battery was just in your cell phone. So let's say it's 10 grams per person or something. Nowadays half the people in this town drive Teslas, so your per capita allocation of lithium ion batteries is 100 kilograms or something like that. We're talking four or five OOMs of increase of total battery per person. That trend is only going to continue.
We've got batteries that are performing this temporal arbitrage. The sun comes up every day, right? So the power swings from midday—you're otherwise curtailing the solar array—to dusk when everyone's watching TV and cooking dinner or running the air conditioners to cool off in the evening. It's very predictable. Whereas, "Oh, we had really bad weather, so we had to use the power line that runs to the extra power plants over by Hoover Dam or something." It doesn't get used nearly as much. Its peak utilization happens almost never, which means that 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. That includes the substations and transformers and stuff that serve that.
So it's a really bad position to be in if you're a grid operator. You've got this aging existing thing that the batteries are cannibalizing. The batteries are being installed behind the meter. You don't have a say in whether they're being installed and how they're being used. All you know is that your utilization of your asset where you get to charge top dollar for it is just dropping year after year at the same time as your operating costs are increasing year after year.
So it's just very clear that the average distance the electron is going to travel between generation and consumption is going to decrease in the future pretty radically. It's already decreasing. It's going to continue to decrease.
It's especially helpful for solar, but solar is the one that's most intermittent. You can predict the amount of solar power you're going to get in three days pretty accurately because of weather prediction. But you can't change the amount of batteries you have. Well, actually in the limit you can because you can put them on trucks and drive them around. There could be a capacity market for batteries where you drive them around to people who need them. In practice, it's going to be cheaper just to double the size of your battery because batteries are going to keep getting cheaper and cheaper. But what it does mean is you can say, "Well, I know that I'm going to have three low days, so 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, so I'm still at four nines instead of having to curtail 24 hours because I can't predict the weather."
Okay, let's assume you're right. I think at some point, you will be right. Maybe we disagree about—sorry, I'm not qualified to disagree. Maybe you and some other person disagree about what year it happens. But it's hard to deny that in the asymptote, our civilization is headed towards lots of energy use for AI and a lot of that coming from solar. In that asymptote, I want to get to the crazy nerd sci-fi… What does our civilization look like? What is happening?
Kardashev Level 1.
Let's wait to get to turning the entire earth into an AI factory. But let's say in the 2030s, where you've gotten multiple people who are building sites on the order of 5 gigawatts or 10 gigawatts. The value of the hardware is dependent on its complement, which is the software. Right now, AI models are fine. The hardware they're running on, the economic value they can generate, is sort of bottlenecked by how good the software is. But if you actually had AGI, if you had human-level intelligence or maybe even better, running on an H100, that H100 is worth a lot. We're paying a lot for humans to do work.
Right now, I don't think AI is that valuable. The models themselves aren't super, super valuable in terms of just pure economic value. OpenAI is generating on the order of $10-20 billion ARR.
That sucks. It's terrible. How can they sleep at night?
But for context, McDonald's and Kohl's generate more yearly revenue than that. But the promise of AGI is to automate human labor. Human labor generates on the order of $60 trillion of economic value. That's how much is paid out in wages to labor around the world. So that's what AGI can do. Even if you curtail it to just white-collar work, that's still tens of trillions of dollars of value. So once we have models which are actually human-level, they will be worth at least that, pending the fact that you can build them or you can run them.
I don't think we should constrain ourselves to being like, "Oh, well, maybe it'll be some fraction of current payroll," because that's very contingent on humans being humans.
That's a lower bound, to be clear.
Oh yeah, lower bound for sure. But if you think about someone trying to estimate the upper bound for the market cap of Caterpillar based on, "Well, it takes this many men and wheelbarrows to dig a trench. So it couldn't be more than that."
One way to think about the industrial revolutions is 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. How much oats can a human or a horse physically digest and then convert into useful mechanical output for their peasant overlord or whatever? Nowadays we would giggle to think that the amount of food we produce is meaningful in the context of the economic power of a particular country. Because 99% of the energy that we consume routes around our guts, through the gas tanks of our cars and through our aircraft and in our grids and stuff like that.
Right now, the AI revolution is about routing around cognitive constraints, that in some ways writing, the printing press, computers, the Internet have already allowed us to do to some extent. A credit card is a good example of something that routes around a cognitive constraint of building a network of trust. It's a centralized trust.
That's interesting. I want to credit James Bradbury and Gwern with making this interesting point when I was talking with them a couple of days ago. If you measure it by GDP, AI's outputs might be underwhelming. One of the complaints that economists have about the Internet is that it's hard to measure the consumer surplus that's created by the Internet because a lot of the goods and services that are made available, you pay zero for them. They don't show up in GDP.
Well, it's the same with oil.
In the sense that energy's like only 1% of GDP?
Well, oil is like $8 trillion a year or something, right?
Yeah.
But if you said, "Well, one day we're going to consume 100 times more energy in the form of oil than in the form of food—and the per joule cost of food is whatever it is, the cost of a Big Mac—then oil should be like $800 trillion a year." Per unit energy, oil, like gasoline, is 100 times cheaper than the cheapest food that humans can digest. Does that mean that we've shot ourselves in the foot by using oil to run our economy because it's so cheap? No.
Right. Also its fraction of GDP also doesn't correspond to how important it is. For example, oil is like 1% of GDP or something. But 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.
Anyways, on the original point about AI, you're going to have this huge deflation. Gwern put it this way. He's like, "If you imagine Dario's data center of geniuses, how is that showing up in GDP? Well, it would be the inputs which are the chips, the energy, etc., and the outputs which are just the tokens. Neither of those is going to be that astronomical in comparison to the value that data center of geniuses is producing."
In terms of GDP numbers, if that data center of geniuses automates or complements a bunch of human work, 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. In the long run, it might make more sense to think of the size of our economy, or the size of our civilization, as the raw energy use that we do rather than GDP. Again, GDP will see this huge deflation because the variable cost of running AI will just be pretty cheap as compared to paying humans wages.
At the point where you've got a mixed economy with an AI doing my job and also a human doing my job…
I love how this is the new way we use the phrase "mixed economy".
Obviously, I still have some pricing power relative to humans, and the AI thus has pricing power. But if it were the case that a new kind of job emerges that AI is really well adapted to, because it's not competing against humans for most of those roles, 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. That would be my guess.
It might be a mistake to assume that if we're going to pay a top AI researcher $200,000 a year—Lol. Let's say for the sort of AI researcher that I could be, $200,000 a year—that if an AI comes along that's as good as me, even taking into account the fact that realistically speaking, I only get maybe 10 hours of really top cognitive work done a week, that it would also be worth $200,000. Obviously, it'd be worth much more than that in the sense that you can copy-paste its output and much less than that in the sense of whatever the marginal additional cost of spooling up H100s is. If some kind of role comes along that the AIs are really well specialized at and outcompete the humans quickly, then we'd also expect to see that both the cost of providing that service would drop drastically, at the same time as the overall value generated in the economy by that service would increase a lot.
Exactly, if we think that the value of cognition is going to be unbounded, and the way to derive cognition—to the extent you think solar will eventually win—you can derive it from how much land it takes to power an H100 using solar panels.
That is a very interesting derivation. At a minimum we're going to just fill up all the land. At some point you might have a declining marginal value of cognition or something. We kind of discussed this earlier. If you have 10 acres of land feeding one megawatt of H100s or something, let's say a megawatt is 1,000 humans. So one acre is a thousand humans' worth of cognition. The implicit land value there is a lot higher than it is as undeveloped desert. It's also a lot higher than it is as the most productive farmland that humanity has ever had.
At current hardware efficiencies. I don't know if it's worth spelling out. Basically, 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.
Is that right?
20 watts vs. 1000 watts? We know hardware can be at least as efficient as the human brain. The human brain can generate this many flops on 20 watts. If you do that calculation, that's 50x1000, so 50,000 AI souls off of one acre?
It could easily be much more than that because neurons are much slower than transistors, obviously. Probably 10 years ago, one of my friends reminded me, the way your phone saves power is it goes to sleep between you tapping out "hello". H, it takes a nap, like 10,000 cycles. E… It's kind of nuts. I think Elon's talked about this in the context of self-driving cars as well. Anything humans do is glacially slow from the perspective of a computer.
Let's go back to the original point. I was explaining why I think it's plausible that there could be more than hundreds of gigawatts of extra demand from AI in the 2030s. I want to understand what that looks like in the real world. At that point, it has become basically this industrial problem. Can you generate enough solar panels and solar modules and batteries, and not to mention the chips themselves? That's the industrial point, and then there's a cultural point as well. Let's start with the industrial point. I want to know what the year 2035 looks like, if we've got AGI and we're just bottlenecked by the ability to deploy it.
What do you need in order to run? What is the minimum amount of matter that you need in order to perform these calculations? Right now we're talking about AI racking and grid and transmission and a bunch of ISOs and all the rest. You don't need any of that stuff.
Obviously, xAI is on top of this because the first thing that Elon will always ask is, "delete anything you don't absolutely need." 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, because it's in the sun all the time, so you don't need a battery. If you're on the Earth, you need a battery as well, so you need some interconnects. You don't need a transformer. You don't even need a DC-to-DC converter. You can actually make do with a buck converter or with relays or whatever to match the current output of your solar array with the charge state of your batteries and the power consumption of your GPU or something.
But a solar array about the size of this desk, for example, will generate about 500 watts in full sun. So you can actually imagine aliens who have different silicon technology stacks building their systems as an integrated solar array with a bit of computronium in the middle, for example, on the same wafer. But that's basically all you need.
On the same wafer? Because it's all silicon?
It's all silicon all the way down. What's silicon made of? It's an element. It's chemically in the crust. There's no shortage of it.
This is a great prompt for a sci-fi exercise, because especially in space, you don't need batteries. The future TSMC just manufactures integrated solar dies.
And they can fly around. They're solar sails, and they're relatively dense so they don't fly crazy fast, but they don't need to because they're immortal.
Is this what the Dyson sphere will be made of, Casey? Is it just going
to be computronium at the center of a solar cell? They can fly closer to the sun to get more power, right up to the thermal limit, and they can fly further from the sun to go and explore or fly to other planets or something. They can adjust the orientation of the solar sail with LCD panels that could be integrated into the wafer itself. What's the post-human state? That's it. A solar sail with a silicon die in the middle for compute? One human's worth of computation.
One human brain can be simulated in roughly a square meter of silicon floating in space. How much, sorry? 1-square-meter of silicon, like the thickness of a sheet of paper, floating in space. That's the future human form. That's my final form. That's the attractor state. That's assuming a little bit of software improvement, but I don't think that's… All that's assuming is software improvement. The Dyson sphere just needs a little bit of tweaking of the algorithm.
The area of the panel is the variable there. What do you need in order to make the silicon? Making solar arrays, making chips is a multi-stage process. Basically you start off with silicates, which are rocks ideally in a relatively pure form. You chemically reduce them. A couple of different processes can do that. Then you purify them into, ideally six nines of purity for solar arrays, maybe nine nines for really nice computers, and grow crystals, cut wafers, etc.
So then the constraint is, well, how quickly can you convert the crust into enough silicon to support silicon thought? What does the silicon ecosystem look like? Any thoughts? Well, it's pretty quick. 1 kilowatt per square meter and then you use that just to rip oxygens off the underlying dirt, it doesn't take all that long too. You only need about 20 microns of silicon to make a solar PV array. You mean like actual dirt?
Yeah. Actual dirt has plenty of silicon in it. For example, setting up a brand new silicon refinery takes about 18 months. But that's just with the current technology, I actually think we may find new ways. One of the nice things about having infinite free solar power, approximately free solar power, is that you can revisit a bunch of legacy industrial processes that have been optimized for efficiency and say, "Well, what if we just use twice as much power and we just want to do them faster and cheaper?" Less capex, less lead time, more power. Well, you can start solving problems. It turns out that if you want to chemically reduce silicon, you can do it electrolytically with less efficiency and under a hydrogen-rich atmosphere or something.
One of the ways that silicon can be refined is by turning it into silane, which is a silicon tetrahydride. I'm not really a chemist, but I think that's right. So SiH4, which is a gas, it's actually like methane, but one down on the periodic table. Don't breathe it though. Once it's a gas, you can filter it from all the contaminants which don't form gases or can be separated by density, much like how uranium is sometimes enriched, but much, much less difficult. You then heat it up to separate it back into pure silicon where you can then precipitate out a crystal. The reason I think this is interesting is because whenever people are talking about the AI singularity, often their expertise is not in energy or physics or whatever. They focus only on the cognitive elements of the singularity, which is like how much faster can we make AI smarter, etc.
I think this is really interesting. If we have unbounded cognition, which sets up both the ability to supply and to demand more energy, I'm very curious, what does the energy singularity look like? We're just trying to saturate as much energy that the earth receives and turn it into cognition. I hadn't thought about that before, but there's this idea that evolution resulted in this continual ramification and complexification of the thermodynamic gradient. You start with very simple RNA-based organisms. Now you get this industrial economy. But it may be the case—I don't have a strong reason to suspect one way or the other—that what we're seeing is the beginning stages of a collapse back towards the simplest possible thermodynamic-to-cognition stack.
We have fusion in stars and the inky blackness of space and that provides our temperature gradient. Then the most efficient way to convert that into usable cognition is silicon. Literally electrons being pushed across the Fermi gap in a solar array and then taking the return path through some set of gates, making decisions about things and then beaming lasers to their friends, saying, "Hey, I just made up a new meme." That is an interesting concept. For 4 billion years we've been increasing the variance in complexity of creatures and then you might see this big collapse.
Should I give you the opportunity to plug why people should work for Terraform? Just to give you an introduction, Terraform is my day job. It's a company I founded almost four years ago. We are making synthetic natural gas from sunlight and air. We are also working on other core primary materials stuff. We also have a methanol process. Methanol and methane together are precursors to every hydrocarbon you could possibly want. Another chemical, ammonia, processed steel, desalination. We can also make cement and a few other things. Basically everything that the primary industry does, except for glass and paper. We are hiring.
Our jobs are available at terraformindustries.com. Yes, the website's meant to look like that because we're very cool. We are some very special people. I know a lot of smart people and I'm privileged to work with some of the smartest people I know. We are mostly mechanical engineers. I will never hire anyone who can't do math. I will never have the problem at Astronomer because we don't have a head of HR.
Also the CEO is not having an affair. Yeah, step one. I think that was a more crucial issue, Casey. Heads of HR can get into trouble. I'm just saying, everyone does math. It's very important to me that Terraform is the place that ambitious hardware people go to become the best they can be. That is really important. It's not here to check in and get your paycheck and optimize some shiny widget.
It's still a small team. It's still like a one project per person kind of situation. And I will level you up—maybe not quite like a Jensen "torture you into greatness" kind of situation, but at times it's going to feel that way.
You get to work with the best people that there are, at least on the West Coast of the United States, on this sort of thing. It's also a unique company. I thought years ago, by now we'll have competition. We don't. No one else is doing this except for a small startup in the UK.
So you get in on the ground floor and it's going to be super cool technology. Eventually we get to go and build it all on Mars as well and help our robot overlords make more of themselves out of dirt. It's pretty cool. Nice. Come work for us. Casey, thank you so much for coming on the podcast. Thanks for having me. This was fun. Yeah.
Article published