Are Exported Tokens Cyber Power Theft? And Why Can't America Build a Grid?

“A blackout is fairly unlikely. But electricity prices going up is one hundred percent certain.” — Ian, on whether AI data centers will crush the North American grid

AI eats software, and energy eats AI. The CEOs of the Magnificent Seven keep repeating the same sentence: we have more GPUs than we can deploy, because there isn’t enough generating capacity to light them up. So the market has started pricing everything electricity-adjacent — a “power shell,” a building already energized, has become hard currency; the previous cycle’s bitcoin mines have been dug up and treated as treasure; Microsoft went and bought a nuclear plant; and OpenAI’s Stargate simply ran its own natural gas pipeline. At the same time, one claim has been spreading wildly across the Chinese internet: openclaw, the “crayfish” agent, sent token consumption through the roof, overseas users discovered American tokens are outrageously expensive, and turned to Kimi, MiniMax and GLM — Chinese electricity, exported in the form of tokens, out-competing the world. Some people call this cyber power theft.

Is it real? If electricity is genuinely the key variable, why is the grid skeleton of the world’s most developed country from the 1960s? Why can’t America assemble a single national grid, while Europe’s sovereign nations are more tightly interconnected than American states? Will AI data centers actually crash the North American grid, or just fatten residential electricity bills?

This episode brings back an old friend, Ian, who writes about power systems on Xiaohongshu under the handle “电力侠 Ian” — Power Ranger Ian. Ian does power systems engineering at a design firm in North America and works day to day on data center interconnection studies and power markets. He compresses the token-export story into one sentence: electricity is at most 30% of a token’s price, and the rest is chips, infrastructure, cooling and whatever the model companies quote — so “exporting electricity can only ever be a gimmick.” But turn the camera back on North America and his judgement gets far sharper: collapse won’t happen, and price increases will, one hundred percent.

What follows is the full conversation, edited and condensed.

1. “Exporting electricity can only ever be a gimmick”: electricity is 30% of a token’s price

Raymond: Today we’re on the hardest bottleneck underneath the whole AI thesis — energy. The Magnificent Seven CEOs keep mentioning that they have enormous numbers of GPUs they can’t deploy, and the underlying reason is a lack of power shell, a lack of generating capacity to keep them running. Layer on the Iran war and whether oil and natural gas will keep rising, and everyone is worried.

Let me start with the hottest item: token export. It originated with openclaw, the crayfish agent — you run one or several agents locally, calling specific large models. When it appeared around Chinese New Year this year, a lot of users’ token consumption rose sharply, and then they found using GPT or Claude too expensive; American token prices are much higher than comparable Chinese vendors’, so they switched to Kimi, MiniMax and Zhipu’s GLM. MiniMax is a newly listed company and its share price rose another 30 to 40% on this wave. A lot of WeChat accounts have now discussed this to an extreme degree. Do you think it’s real?

Ian: Hello everyone, I’m Ian. I normally do power-systems work at a design firm in North America, more on the engineering side, and I also touch system-level analysis for data centers — interconnection, or energy-related aspects of the data center market. I got interested in power markets over the last couple of years and started the Xiaohongshu account “电力侠 Ian,” writing observations and thoughts, though really it’s more like study notes.

Honestly, writing about token export was me riding a wave; a lot of independent media had started writing about it before me, so I essentially stole their idea and added a section from the electricity side. The route by which I noticed it was the same as yours — I saw the crayfish agent, and I’ve written code before, though not professionally, so I wanted to try it. On top of that, North American data centers really are extremely short of power.

But I think that when the data-center power shortage finally shows up in the price of a token, it may account for only 30% to something under 50%, possibly less. The rest is infrastructure and cooling, and those costs are inherently lower in China. So exporting electricity can only ever be a gimmick.

Raymond: It’s certainly part of it, call it 30%, but there’s also the price the model vendors themselves quote, salaries, and the compute needed to develop the model. They use Nvidia chips, domestic firms use different chips, and especially on inference the cost difference is large. So this isn’t just power theft, it’s the whole cost stack.

Ian: Right. Steel is also enormously power-hungry, and exporting steel is likewise using cheap domestic electricity to make things and then sell them abroad. It’s just that tokens happened to catch this data-center wave, which makes it more conspicuous. Many industries are heavy power consumers, not only tokens.

Raymond: Token output is growing extremely fast, so people look at it and feel it’s going to 10x again tomorrow. Fast growth, and gross margins improving too, so it also feels promising.

2. A 1960s skeleton: where exactly is the American grid old

Raymond: So let’s start with grid iteration. China’s electricity is very good, everyone knows that. But a lot of people find it hard to imagine why America, as the world’s most developed country, has an aging grid. We can build EVs and all kinds of electrical things, and the precondition is a good grid. Why is America so far behind?

Ian: Mainly three things. First, equipment age. China’s grid was essentially built out over the last ten or twenty years, with the big infrastructure push around 2000, so substations and transmission lines are ten to twenty years old. America’s transmission lines, substations and switchgear were mostly built between the 1960s and 1980s, and the overall framework was already complete then. So their infrastructure looks rust-streaked; a transmission line goes down and it’s still a wooden pole, possibly forty or fifty years old.

But theoretically these things can last sixty or even seventy years, they just need continuous maintenance. Their logic isn’t tear down and rebuild: grid companies have dedicated capital maintenance teams watching the age of individual pieces of equipment and replacing them when they’re due. Domestically, you might get one budget that replaces all the old equipment in one pass. What that produces is: lots of equipment fifty or sixty years old and still in good condition, but a system that is simply very old.

Second, the path by which new technology gets applied is different. In China, the moment a new technology appears people find a test bed and use it immediately — that’s exactly what happened with renewables; get the equipment out, deploy it, swap it if it doesn’t work. Abroad it’s much more about solving problems with mature, conservative technology. They lead at the university layer, ideas and papers are strong, but actual deployment comes back to infrastructure — and China’s infrastructure genuinely leads. Nothing to be done about that.

Third, US electricity demand has barely grown these past twenty years. This is the most central point: no demand means no growth. Over the past two decades consumption rose only 1% to 2% a year, so the need to renew transmission lines, substations and HVDC wasn’t that dramatic — unlike China, which needs to move all that Xinjiang solar and wind to Guangdong. Their equipment was built forty years ago, and only recently, as data centers multiplied, did they look back and ask whether the capacity is enough and whether the equipment is too old.

Raymond: Sixty-year lifespan, much of it now at forty or fifty years, which maps exactly onto their last big infrastructure cycle in the 1960s and 70s — the US-Soviet Cold War, Vietnam, when America was still a great industrial power.

That said, the 1% to 2% figure sounds off to me; it ought to be negative. Over the past twenty or thirty years America has been deindustrializing, with factories continuously closing, and factory equipment certainly consumes far more power than residential. If factory load was given up and it still held at 1% to 2%, that means residential consumption grew a great deal, which is what kept the total from falling.

Ian: The figure may not be precise, but the conclusion is the same: if the market isn’t growing, then as a utility I won’t invest. In America and in Europe, residential customers ultimately pay for this equipment, and without a business case validating demand nobody invests. Everyone follows the money into infrastructure.

3. “The visiting monk can’t preach in the local temple”: a country of electrical warlord fiefdoms

Raymond: China has one very special condition, which is a single national grid. The greatest need for power is in the east, and the richest energy resources are in the west, so Xinjiang’s renewables can be moved to Guangdong via ultra-high-voltage lines. Our institutions and infrastructure support that kind of national dispatch. I’ve seen you write that America’s grid is divided into several blocks that aren’t interconnected. Beyond the political system, what other economic or design reasons are there?

Ian: Beyond insufficient infrastructure scale, there’s the structure of the North American grid itself. In ISO and RTO regions the generation market is fully wholesale; a power plant has to enter the bidding system and bid for itself. Non-ISO, non-RTO regions resemble China: vertically integrated monopolies where generation through consumption sits with one company, which may be the government or a large utility.

Every region has its own internal rules, there are many incompatibilities between states, and state power is strong while federal power is weak. Above the ISOs and RTOs sits the Federal Energy Regulatory Commission, FERC, responsible for planning the national transmission network — but its authority is limited; it can only recommend. Under the Federal Power Act it cannot compel the construction of transmission lines.

An example. Washington State’s hydropower is very well developed; New York State relies entirely on external natural gas and has no generating resources of its own. If Washington could transmit power to New York, wouldn’t New York residents’ bills fall? Much like that Xinjiang situation. But you’d have to cross many states in between.

Raymond: Washington is on the far west, New York on the far east, spanning the entire country. I just looked at a map and roughly you’d cross eleven states — North Dakota, Wisconsin, all the way through New Jersey before reaching New York.

Ian: Right. Electricity can’t travel through thin air; you have to build cable on land or in the sea, and crossing a state means getting that state’s permits. There’s an interest-allocation problem inside that: some state in the middle may be perfectly self-sufficient, its generators are making a fortune, and those generators have a big voice inside the ISO and can swing the vote. It amounts to asking someone whose interests are harmed to vote for you, which is very hard.

Raymond: Hold on, isn’t an ISO a bidding system? Suppose I can bring cheap hydropower over from Washington — then I’m the most advantaged participant in that system and should be better off.

Ian: Washington’s hydro does have an advantage, and supplying only locally isn’t a big enough volume; if New York also uses its power it earns more. But every state along the way has its own generators, and they don’t have hydro — they may burn natural gas at higher cost than Washington. If cheap external power enters their state, aren’t those generators’ interests harmed?

Also, to clarify, when I say ISO I mean a region. California belongs to CAISO; the eastern US has PJM, which contains a dozen or so states. Washington State is called a non-ISO area.

Raymond: I get it. An ISO is more like a territory-plus-pricing regime: full competition inside the territory, but if you’re an outsider not in the system, I block you at checkpoint after checkpoint. Much of the time it isn’t letting the visiting monk compete with me, it’s that the visiting monk can’t preach in the local temple at all — a Chinese way of saying the outsider is structurally shut out. So it’s essentially a country of electrical warlord fiefdoms. Is that a fair reading?

Ian: That’s exactly right. ISO literally stands for Independent System Operator, an independently operated system region. That line has to cross many states, and if even one state in the middle disagrees it can’t be built; you either detour, or the federal-level national project is simply stuck. And there’s an election every four years — this term the line is well regarded, next term the administration changes. You’ll see plenty of news about lines being shelved for a decade.

There’s also investment. Alternating current can’t travel very far, so you have to use direct current, and DC is more expensive than AC; the economic model only works above a certain distance in kilometers. And after investing you have to allocate cost across states — how much do I pay, how much do you pay — another pile of wrangling.

Raymond: New York is the richest and needs the power most, but it certainly doesn’t think it should bear all the cost; the states in between are all agricultural — Idaho, Montana, North Dakota — and certainly won’t want to shoulder the cost of DC. So that line never gets built.

Ian: And those agricultural states generate a lot from coal, at high cost. You could ask them to contribute, but who contributes how much and who benefits is an extremely complex question. Without a powerful organization above them controlling the planning, it remains very hard to do — in China, State Grid’s design institute might simply do it.

Raymond: Let me add another angle. America elects every four years; this term the Democrats support wind, hydro and solar, next term the Republicans are pro conservative energy — coal, gas and oil. So stakeholders find it very hard to make ten-year investment decisions. The grid you invest in during these four years isn’t even finished when the next administration reverses policy 180 degrees; that may not be illegal, but the economics certainly stop working. Infrastructure runs on multi-decade cycles, and in a place where construction is driven by private business, this is very hard to make happen.

Ian: Which is why they innovate within more independent, smaller scopes. Texas, for instance, is fairly free, so a lot of grid projects and data centers get built faster there.

4. Japan is stuck on frequency, and Europe holds together more tightly than America

Raymond: Before we go down to data centers: America has a state problem and a political problem. Are the same things problems in Europe and Japan?

Ian: Not quite the same. Japan’s issue is that Kanto and Kansai run at different frequencies — Kanto at 50 hertz, Kansai at 60. Frequency has to be fixed within a grid; 60 hertz can only connect to 60 hertz, and stringing them together simply doesn’t work. All of America is one frequency, all of China is one frequency. For Japan to transmit across regions it has to go through DC in the middle, requiring large investment to link the two systems.

Europe’s situation is that the renewable share has grown fairly fast. That’s not really a problem — to decarbonize it has to add renewables. You can count the generation types: natural gas, coal, oil, wind, solar, and nuclear — nuclear is very slow to build, a ten-year cycle. Slightly cleaner options are hydro and gas, but many places in Europe have no large dams. So that leaves gas, and gas has to be shipped from Russia or elsewhere; as far as I know Europe’s gas isn’t especially abundant either, and this Iran war is another problem. So all it can do is add renewables. Renewables are both a blessing and a curse; the downside is that the grid becomes progressively less stable, which is why Europe’s grids need to be strung together so they can help each other out.

Raymond: I understand. Renewables have more complex peaks and troughs, and you need a bigger grid to smooth them. But Europe ought to be worse than America — America is only states, whereas Europe is genuinely sovereign nations, France has its own utility, Germany has its own, and they should be even less willing to accept each other.

Ian: This is a remarkable phenomenon. Europe’s degree of interconnection is actually higher than America’s; they aren’t even one country and yet they huddle together quite tightly, having signed interconnection agreements. Possibly because they depend on each other more — every American state has its own generation methods and can rely on itself, whereas many small European countries can’t generate enough on their own, so some have to use solar and some wind.

5. Collapse won’t happen, price increases will, one hundred percent

Raymond: There’s a lot of talk online that building this many AI data centers will crush the North American grid and cause a collapse. The reasoning is that North America was only growing 1% to 2% a year, and then AI data centers arrive and the gap is enormous. I don’t really understand what “collapse” means. Is it the nationwide blackout you see in films? Am I going to use a model and end up making Americans unable to turn the lights on? From a power expert’s standpoint, what does it actually look like?

Ian: I understand collapse as everyone being unable to use electricity, every location dark, unable to recover in the short term, affecting normal work and life. By that standard, will the US and Canada see a wide-area, long-duration collapse? I think it’s fairly unlikely. But small-scale events — one regional grid unable to supply, or a trip — are very possible, and that recurs continuously everywhere in the world.

Raymond: Domestically it’s very, very rare; basically there are no blackouts. China may have an enormous infrastructure overinvestment, and energy overall is quite good.

Ian: What’s called collapse over here is also the small-scale case. California wildfires burn the transmission lines and tourist areas and residential areas lose power; Texas gets a blizzard, extreme cold outruns supply, and regional rolling blackouts start. Small-scale outage is the lowest tier; the next tier up is the grid running short and taking region after region offline in rotation. Everyone going dark at once — that’s real collapse.

Raymond: So what you think could happen in America is that because there are more AI data centers, they eventually need rolling curtailment?

Ian: In an extreme case it’s possible, but it hasn’t happened; all current outages are from natural causes. Once there are more AI data centers, the most conspicuous situation is this — a blackout is fairly unlikely, but electricity prices going up is one hundred percent certain.

6. “These machines don’t eat and don’t order takeout”: whose bill do AI data centers actually move

Raymond: So let’s talk price. Many AI data centers now claim to generate their own power — OpenAI’s Stargate bought gas and hooked up its own network, Microsoft acquired a nuclear plant, and they don’t even need to interconnect. If they don’t interconnect, doesn’t that mean they won’t push prices up?

Ian: Two cases. First, the data center doesn’t connect to the grid at all and builds enough generation next door to be self-sufficient; that affects nothing on the grid. Second, more common: it builds a plant next door supplying only part of the load, but the data center is still connected to the grid, because sometimes its own power isn’t enough and it needs the grid to supplement.

If you spend without regard to cost, the first is workable. But data centers need to hit 99.99% reliability or they breach contract — for instance Amazon leasing to another company must guarantee uninterrupted service in the agreement, and pays damages if it stops. To self-build, you need more redundant generation to ensure that when one fails another takes over. The somewhat more economical route is connecting to the grid. But the moment you connect, you affect prices.

Raymond: When its own generation isn’t enough and there’s incremental demand, it creates a tighter balance within that local grid and prices are forced up. That’s also what the public minds — your AI data center comes to my state, brings no jobs, and these machines don’t eat and don’t order takeout, and the result is my electricity price goes up. That’s certainly why residents in many states oppose AI data centers.

Ian: Yes. Think of it this way: building one Costco outside my house may not affect traffic; but ten Costcos all built outside your house, and the government has to plan whether to widen the road, and the cost is borne by all the surrounding residents together. There’s a thing in the grid called the rate base, effectively a pool: all invested money goes into it and is then allocated across every customer. Everyone’s price certainly rises; there’s no avoiding it.

Raymond: I’ve heard it said that data center electricity runs on one regime and residential on another. China also has two price regimes, industrial and residential. It seems that isn’t how it works in America?

Ian: Data centers count as industrial, and the tariff structure is completely different from residential. Residential is kilowatt-hours used times a price, plus a fixed monthly charge. Data centers are far more complex: besides daily kilowatt-hours, it also depends on your maximum draw — your annual peak affects grid planning.

There’s also tax. Canada is genuinely going to levy one now: any data center that exceeds a certain megawatt threshold after interconnection pays a 2% tax. That’s the “three birds with one stone” I wrote about before — use that tax revenue to upgrade the grid and residents are happy too. It’s still at the planning stage and hasn’t landed, and I’m not clear on the exact tax base; it’s called something like a Data Center Tax.

Raymond: Data centers actually won’t be unhappy at all; their business model is that they’re happy as long as they can get energized, and this is small money. Capital markets see it clearly: electricity is a very tight balance, and whoever holds power resources, whoever can secure a so-called power shell — a building already energized — the market gives a big premium. Those old bitcoin mines have all had their valuations rise on this AI data center wave; grid connections established earlier are very valuable resources.

7. The cost-causation principle, and a way of drawing power the grid has never seen

Raymond: Grid planning is done to peak. My crude understanding is: exceed the design peak and, without curtailment, the grid may burn out; without curtailment you have to upgrade the whole network. Think of it as a highway — it used to hold ten cars, now it has to widen to twenty, and everyone shares the assessment. But I saw a news item, I think also FERC guidance, saying the cost of widening should be borne entirely by the AI data center: these five lanes are for your use, so it’s all on you. I believe the Magnificent Seven big-tech companies don’t mind paying. But is that workable?

Ian: That’s a sharp question. Some of it is workable, and some genuinely is hard to quantify. Transmission lines are relatively easy to quantify — a line built specifically for you, you certainly fund the investment, and that’s the cost-causation principle: you caused this thing, so you pay for it; you benefited, so you pay for it.

But a lot of things aren’t as direct as transmission lines. With a transmission line the discussion is whether capacity suffices, like being a Costco on a highway with ten trucks that need to get through. But on a smaller time scale there’s also a stability problem. Data center swings are extremely dramatic; GPU load isn’t like ordinary consumption — your air conditioner is a flat line, whereas a data center may go from 0 to 100% in one second. That’s a second layer of shock.

Every load device on the grid draws power differently from this; this way of drawing power has simply never existed before. Perhaps only a small subset of steel mills is similar. The whole grid was not designed for this in the first place. Absorbing that shock requires additional equipment, but it isn’t the data center paying for it — that goes through the ancillary services market: how dramatic the voltage swing is, how stable it needs to be, determines what equipment gets deployed, and companies bid for it. The data center pays nothing at all.

Raymond: So that’s unfair, because it can’t be quantified. Let me explain why the swings are so large. AI data centers have two functions, training and inference. Inference grows fast, but it’s stable demand — the volume isn’t small, but it’s level. The shock comes from training: you start training today, go straight to 100%, train for five days, then shut down to zero. And training a model starts at ten thousand cards, so it’s as though you fired up a giant steel beast that suddenly starts running and then suddenly stops.

Ian: I don’t really understand why everyone has to train together and can’t stagger it. If it could be staggered, that would solve part of the problem.

Raymond: That involves what the large-model design paradigm itself demands of compute. Chinese companies now have some interesting innovations, like using downtime nobody else wants to train small models, or making training more intermittent. But the overall fact of the past two years is: when you run, you run together, you run big, brute force works wonders.

8. Treating data centers as batteries, and queueing in five or six states at once

Raymond: Building a data center in North America is itself easy — put up the building, put the GPUs in. But interconnection reportedly takes four to seven years. Is that true?

Ian: It genuinely happens. Virginia already has a lot of data centers and the queue there is fairly long. But from late 2025 into early 2026 some new policy discussions have emerged; everyone is thinking about how to speed up interconnection and the federal level is pushing too — which is good for America, because data centers certainly have to be built and nobody can wait four to seven years.

One approach is to let data centers interconnect as an interruptible entity. When the grid runs short of generation, cut the data centers off first; one cut is several gigawatts and the grid is immediately fine.

Raymond: What you’re saying is that the grid uses the data center as a battery — a battery that goes offline.

Ian: Right. You want to connect quickly, fine, but you have to accept being cut first in an emergency — don’t cut hospitals, don’t cut schools, cut you first. There’s no consensus yet, but the direction is fairly promising. The other route is combining behind-the-meter generation and self-built plants, which might get you in within three years. The data center side would like to be in production in one to two.

Raymond: Let me explain behind-the-meter generation for listeners. My understanding is that some AI data centers build their own plant without interconnecting, so they aren’t bound by the four-to-seven-year constraint and can start building any time.

Ian: But off-grid means you face environmental permitting. Musk’s xAI is an example: he wanted to go fully off-grid using methane gas generators, but the carbon emissions from that are dramatic, and permitting intensity differs by state. xAI appears to have positions in several states, Minnesota among them; one state’s permitting was very strict so none of the units could be used, and another state was looser so it could. So off-grid isn’t a cure-all either; it’s more that in some states you can find loopholes — put the gas turbine on a truck and you can claim it isn’t primary generating equipment, use it that way for now and move it later.

Raymond: Let me add: there’s an American company called GE Vernova that makes natural gas turbines — essentially gas engines; think of them as large gas canisters. Put them on trucks and they take on the appearance of being ready to leave at any moment, so environmental permitting passes more easily. This also creates large arbitrage between states: if state A lets me in I go, if state B doesn’t I don’t. Which is why AI data centers are mainly in Texas — that’s already established fact.

Ian: There’s another very interesting thing. They now run interconnection processes in parallel, filing applications in five or six states simultaneously, then picking only one at the end and withdrawing the rest. If each takes three years, then I enter the queue in Minnesota, Texas and Ontario, Canada, and go with whichever comes through first.

Raymond: So if I’m an AI data center operator, do I have to build a physical location in all those states simultaneously?

Ian: Ultimately they only build in one state. Each place has a queue of at least two or three years, so I queue in all of them at once, and once I get through I choose where to go — like taking a ticket at ten hospitals and picking which to visit once your number is called.

Raymond: That’s very different from my understanding. I thought you had to physically be at the hospital to queue; what you’re describing is booking ten hospitals on your phone. I’d assumed a lot of people were unwilling to invest precisely because you had to build a real data center there first — buy the land, level it, run the pipes, get the GPUs delivered, and only then queue.

Ian: It’s simultaneous; environmental permitting and interconnection review can both be queued in parallel. That’s also a shock for the grid: if the substation gets built and you ultimately don’t show up, that’s waste, and the money still gets spread across everyone. It isn’t fair to the locality either.

9. Why Texas became the promised land: the bitcoin miners’ legacy

Raymond: We’ll put Ian’s Xiaohongshu account in the show notes; he’s written a great deal about different North American power plants. Take Texas as the example: as an outcome, Texas is where AI data centers most often site. Why?

Ian: On the data it is indeed so, but you have to distinguish. The traditional heavyweight is Virginia, in the PJM region, which has the most data centers. Today’s hyperscale data centers are different, and they more often pick Texas, then California. Why Texas? I can’t say for sure what the most fundamental reason is. Possibly low electricity prices, certainly lower than California; easy land permitting; and freedom on environmental matters. But I haven’t systematically worked out the specific reasons.

Raymond: I think there’s a historical lineage here. Last cycle, in 2020 or 2021, most bitcoin miners withdrew from China and the machines moved all over the world, from Norway and Iceland to Malaysia to hydro stations in Brazil. A mine and an AI data center are in some sense the same thing: a big building with lots of machines computing, needing lots of power. So a lot of mines last cycle were also in Texas. And those miners are now frantically trying to convert into AI data centers.

Ian: Which is why the US energy commission also issued a guideline, which I mentioned in episode 13: even if you use an existing power shell site — it was a mine, there’s a substation next door, convert it to a data center and connect straight in — where previously no process was required, you now have to do a grid feasibility study.

Raymond: What’s the logic? Is it inventing a new administrative step to make life difficult, or are the two genuinely different?

Ian: Because AI data center volatility is too large — I think larger than mining rigs — the shock to the grid needs to be reassessed and can’t be treated as the same as before.

Raymond: Let me explain. A bitcoin mine mines one specific commodity, so it’s extremely price-sensitive. In summer evenings everyone runs air conditioning and pushes prices up, and mining rigs can partially shut down — if it isn’t economic to mine today, don’t mine; you can always mine tomorrow. AI data centers are nothing like that: training starts now, bang, full throttle forward, and even if prices rise the model has to keep running. It’s rigid demand. Bitcoin miners were actually a regulating mechanism for the grid, not a shock. So all of these should be reviewed again; don’t assume it’s merely a change of shell, because ultimately wrecking the grid would be very bad.

10. Storage isn’t the bottleneck; a 200% tariff and a gutted tax credit are

Raymond: I want to ask about storage. I’ll put in the show notes an account I like a lot, 星球研究所 on Bilibili, which did a whole series with State Grid on West-to-East Power Transmission. Solar and wind are so volatile and yet ultimately integrate smoothly into the grid, and storage is a key link. China has all kinds of forms — storing in big batteries, or using gravity to pump water up a mountain and release it. Is there much storage in North America?

Ian: America has plenty of pumped-storage hydro. Virginia has it; Washington State is strong in hydro — it’s the only state in America where it rains half the year; Tennessee is also rich in it. These places all use pumped storage, first because it’s clean and second because it’s very useful to the grid, and it resembles battery technology, just much larger.

But this differs from wind and solar. Wind and solar are mostly invested in by private companies, participating in the wholesale power market for revenue, and then through America’s Investment Tax Credit for renewables — as long as you’re renewable you can get 30% of the tax, 30% of construction cost, credited back — which is a large slice of their annual revenue. Pumped storage, by contrast, is a very large infrastructure-type project with enormous investment, and most of it sits in the vertically integrated non-RTO regions, done through government planning.

Raymond: Wind and solar are inherently very volatile, so if it’s all private operators, how do they connect to the grid? Besides earning government subsidies, shouldn’t they also invest in storage themselves?

Ian: There’s some demand for it. Storage paired with wind and solar gives a better rate of return, because you can buy low and sell high. The US and Canada have energy markets and capacity markets, and price volatility is larger in the energy market.

Raymond: What do energy and capacity mean?

Ian: An energy market, like Texas, settles on instantaneous generation; a capacity market settles on generation over a period.

Raymond: One is a point-in-time figure, the other a period figure.

Ian: Right. So you can pair storage with solar: prices are too low during the day, so I store it and release it in the late afternoon and evening when everyone gets home and demand peaks, and make money.

Raymond: My question is still: China already leads in wind and solar. Is that lead because we do storage better? Is storage actually the bottleneck on America’s clean power development?

Ian: My personal read is that storage isn’t the bottleneck, and it isn’t the reason their wind and solar haven’t grown as fast as China’s.

First, cost. America charges a 200% tariff on Chinese solar panels, so with labor costs on top, panels themselves are expensive, and it has to rely on the local ITC policy, effectively a tax refund: you invest twenty million and I return 30% over ten or five years, so six million. Only with that revenue does the economic model work for many solar and wind projects, and the reason American wind and solar grew rapidly the past two years is that the ITC was still there. But Trump came in and cut the ITC, so the credit keeps shrinking and the economic model keeps deteriorating. A 30% share is too large; electricity prices alone can’t carry it.

Second, America’s natural gas is fairly developed, and with Canada behind it also producing gas, it doesn’t really need renewables to supplement power growth. Whereas China isn’t a gas-producing country.

Raymond: The resource endowments differ on the two sides, so the policy choices differ too. America has gas and oil, so its attitude is different; for China it ultimately comes back to national energy self-sufficiency, so regardless, government subsidies and government guidance will push renewables up.

11. Nuclear is too slow, gas is too fragile, and the legacy burden thirty years out

Raymond: Let’s speculate. America hasn’t built nuclear plants in a very long time; China might build a hundred in ten years and America one in ten years. What role will nuclear play in the AI data center industry? Can anyone go and buy a nuclear plant?

Ian: What Microsoft bought was an old plant, and my understanding is it was near retirement, so they refurbish it and extend its life a few years. Most companies are buying that kind right now — delay the retirement cycle a few years, or rebuild and expand. Building a nuclear plant from scratch specifically to supply an AI data center, I haven’t come across; there may be plans but nothing genuinely under way. They currently feel nuclear is still too slow to build immediately; a ten-year cycle is too slow for data centers — though it may not be slow for the American grid.

Raymond: The Three Mile Island one is an existing plant, not new-build, and even so it’s roughly 2028 before it starts generating for them, with a lot of permitting still needed. But three years against ten is already fast.

Ian: Building from scratch is impossible against their one-to-two-year cycle.

Raymond: Now on Stargate. There’s a drone aerial video that dramatically captures overall project progress, down to what equipment is at which location; I’ll put it in the show notes. Stargate currently runs entirely on natural gas generation — could it be converted to wind or solar in future?

Ian: Theoretically entirely possible. Like Musk’s xAI, which started on gas generation and ultimately wants to take another route; solar or wind can both be converted to, as long as the funding is there, and following environmental policy is fine too. Although I saw news a while ago that Stargate’s Texas project was paused. That was this month’s news; the project was apparently 1.2 GW and wanted to expand to 2.1 GW, and now can’t expand because of funding. Oracle’s share price a while back may have been this too, but I’m not well informed.

Raymond: It certainly has an effect. That said, Stargate’s earlier phases are already properly under construction. If ten were planned and only five get built, that’s still five built, and those five are very large things with strong impact, all running gas engines. Whether the other data centers ahead take greener routes is worth watching. At least in a place as gas-rich as America, getting going first is not unreasonable.

Ian: Gas is a fairly safe energy source for them.

Raymond: A last question with no right answer; let’s just talk reflections. The Iran war has made many people refocus on oil prices, and next a whole series of petroleum derivatives and even fertilizer will get attention. But at the same time, China’s energy situation seems to have no problems at all — decades of frantic infrastructure building is paying off today, and our energy support for industry and production is something other countries can’t imagine. A few months of the Strait of Hormuz being closed seems to have little effect on us; other countries might be destroyed and we’d keep operating. What are your thoughts on energy as a keyword?

Ian: Domestically there’s currently no news at all about power shortages or any energy shortage; even without natural gas, generation wouldn’t be affected. So energy security is a very important subject for a country.

All of North America originally ran on coal as the mainstay, then gradually shifted to natural gas, and behind that was also an energy security consideration — gas is a safer energy source for them, and they’re an exporter. But gas is heavily affected by price volatility. Canada produces gas but the pricing power isn’t Canada’s, so when gas prices rose in 2020 electricity prices spiked. Around three or four years ago, during the pandemic, gas spiked and electricity followed. That’s also why everyone realized they needed to develop renewables more — a well-rounded energy mix is effectively solving energy security. If America still relied entirely on gas, that certainly wouldn’t be the better path.

I think nuclear may need to be developed, and America is thinking along those lines too; it’s just that AI data centers can’t wait. But nuclear is certainly a good technology, and whether from the standpoint of energy security, generation volume, or contribution to grid stability, it’s enormous. Long term, nuclear remains fairly important in North America.

Raymond: But that requires successive American administrations to reach consensus and not reverse course, and ideally for no plant to have an incident in the meantime, running smoothly for ten years. That genuinely isn’t easy.

Ian: Extremely not easy. Actually America’s institutions are fairly advanced; looking at a single state or a single ISO system, it invented a lot of mechanisms, marketized electricity prices, and used price to measure investment. It’s just that applying it nationwide is fairly impossible — and if you did apply it nationwide, it might not still look so advanced.

Raymond: China and America differ. China still has the feeling of playing one large game of chess; America is bottom-up, things formed at state level — this state’s institutions extend as far as they extend, become void a few states over, and another few states start having their own. But over time you find that institutions that look ingenious become, thirty years later, a legacy burden. China happens to still be early, so right now it all looks like upside and the burden isn’t visible — our burden may well show up thirty years from now. Whether there’s better wisdom to handle it then is up to the next generation.

Ian: Yes, exactly. Domestically, power and energy really are developing very well right now.

Recommended by our guest: search for “电力侠 Ian” on Xiaohongshu.

If you're working on this too — or you think we've got it wrong — write to us at [email protected]; if you'd rather not write, just leave your email below.