Topics Pioneers Insight Method Author
Back

55. From Power Shortages in China to Power Shortages in the U.S.: An AI Energy Bottleneck? — 唐奕波

·苔藓之火

Rendering the long image…

Topics AI Data Center Power Bottlenecks
Pioneers 唐奕波

Summary

  • The core issue in a power shortage is how much capacity can be dispatched when needed, not how much electricity is generated over the year. 唐奕波 uses rough dispatchability factors of about 80% for coal, 5% for wind and 1% for solar to explain China’s 2021 problem: load growth outpaced the buildout of firm power, a gap later addressed through the “three 80s” coal approvals. Storage operates on a different scale: a 2-hour battery shifts power within the day; it does not turn solar into an around-the-clock source.
  • The U.S. power shortage began with manufacturing reshoring and was then amplified by AI demand. 唐奕波 recalls that power prices and generator profits improved first: CEG and VST shares rose about 3–5x in 2023, while gas-turbine equipment orders did not turn meaningfully higher until mid-2024. Market-based adjustment takes years to feed through. Raymond’s summary of the process: “Their shortage is demand-led; ours was purely about suppressing supply.”
  • The gas-turbine oligopoly and its service revenue cannot prevent new equipment demand from collapsing. Raymond estimated that GE Vernova, Siemens Energy and Mitsubishi together held about 85% of the market; equipment sales also come with roughly 15 years of service contracts, carrying operating margins of about 40%. Annual demand peaked at around 100GW in the previous cycle; 唐奕波 says that when power-demand growth slows, new equipment demand could fall to 10–20GW. As he puts it, the industry’s fortunes are “determined by supply and demand, not by how good the business model is or how high the barriers are.”
  • Constraints on heavy-duty turbine deliveries are pushing demand toward more expensive, more marginal sources of supply. The 3 heavy-duty turbine makers are planning roughly 25% capacity expansion over 3 years, while lead times have stretched to 5 years. Buyers are moving down the stack from light-duty turbines to fuel cells, gas engines and aero-derivative equipment. 唐奕波 puts heavy-duty turbine costs at about $0.06/kWh, light-duty turbines at about $0.08/kWh and fuel cells at at least $0.12/kWh at the time; the cycle will only truly end if cheaper heavy-duty turbines can scale substantially.
  • The opportunity for domestic blade makers comes from extreme scarcity, but entering the supply chain takes years of validation and sustained investment. Yingliu began making blades in 2015; its first sample took about a year, and getting from drawings to a paid small order takes at least 3–4 years. 唐奕波 recalls that the company continued investing about RMB200M in capex and RMB200M–300M in R&D even when net profit was around RMB100M. “Either exceptionally forward-looking or exceptionally committed,” he says; the stock has already risen severalfold.
  • The next opportunity may shift with data centers toward merchant power assets. 唐奕波’s conditional view is that if resident opposition, state-level restrictions and midterm elections slow U.S. construction, 2027 training demand could move overseas. He is watching a Nordic power producer with a dividend yield of about 5%: unlike regulated assets, which earn an allowed return on new investment, merchant generators can capture the upside from higher power prices directly.
  • Europe has not yet developed a clear aggregate power shortage, but new load and grid access rights are worth tracking. 唐奕波 estimates that data-center grid-connection applications rose from more than 200GW in early 2025 to more than 300GW in October 2025, and could reach about 450GW in June or July 2026; these are early-stage applications, and actual construction still needs to be verified. With connection waits of 8–10 years, existing interconnections at retired factories and power plants have value. He cites a retired RWE coal plant sold for about €400M: buying its connection avoids starting the queue again.
  • 唐奕波 puts crowding in the current equipment trade at 8–9 out of 10 and sees no more left-side opportunities in new equipment. Since 2023, the trade has spread from overseas to China and from reliable generation solutions to marginal supply. In March and April 2026, constant WeChat Channels recommendations for “gas turbine exports” became his signal that he did not want to keep playing these stocks: “It’s hard to make money now.”

Deep dive

1. Solar’s latecomer advantage made leadership hard to monetize

  • 唐奕波 entered solar in 2018, just as the “531” subsidy cut hit. He recalls the sector losing 30%–40% of its market cap in a week: “Once the subsidies were gone, there was effectively no demand.”
  • His bearish case rested on “leaders die, latecomers win.” Looking back over the industry’s first decade-plus, he saw each generation of leaders go bust. Technology shifts can overturn the incumbent route, while later entrants on the same route can use equipment that is one-third cheaper to make more efficient products, with unit capex potentially only one-fifth to one-quarter as high. From a 5- or 10-year end-state perspective, he saw an exceptionally poor manufacturing business model.

2. China’s shortage was not of electricity, but of dispatchable capacity at specific moments

  • 唐奕波 separates power shortages into kilowatt-hours and kilowatts: enough electricity over a full year does not mean the system has enough capacity to meet immediate demand at 5 p.m., when the sun is down and wind is weak.
  • His rough dispatchability factors are about 80% for coal, 5% for wind and 1% for solar; 100GW of solar therefore contributes only about 1GW of dispatchable capacity. Asked by Raymond why coal is not counted at 100%, he says the system needs reserve capacity and plants take time to start up and shut down.
  • In his framework, dispatchable capacity usually needs to exceed peak load by a factor of 1.2, leaving 20% headroom. China’s load grows about 5%–6% a year, and in 2021 the ratio fell below 1.2 for the first time.
  • Residents may not have noticed much because “China prioritizes residential use.” Factories bore the brunt through power cuts and staggered production: one group gets a few hours to run, then hands over to the next.

3. The “three 80s” quickly restored China’s firm power supply

  • 唐奕波 recalls that the “three 80s,” roughly from 2022 to 2024, meant approving 80GW of coal power each year—“three Three Gorges projects every year.” China had workers and equipment ready, so projects could be built in a little over a year after approval. “After 2023, this stopped being a problem.”
  • Thermal power construction may previously have been only about 10GW a year; it is now around 40–50GW. Nuclear and gas-fired power are also expanding. Asked whether China could again face a baseload shortage, he says, “I don’t think so”: the transition had been too aggressive, and the priority of stability and security is now understood.

4. Two-hour storage does not turn solar into an around-the-clock power source

  • A common configuration pairs 1GW of solar with 0.2–0.4GW of storage for about 2 hours. That shifts midday surplus into the evening and helps balance the day, but does not carry power across days or weeks.
  • To make 1GW of solar behave like a source with a dispatchable-capacity factor of about 0.6–0.7, 唐奕波 says the scale would be 1GW of solar paired with 1GW of storage for 10 hours—10GWh of batteries. The cost “would explode.”
  • He points to off-grid power at African mines: meeting demand 24 hours a day might require a 1:16 configuration, yet outages would still account for about 10% of the time, for only 90% power reliability. Raymond admits he had assumed that adding batteries next to solar “felt like it would solve the problem.”

5. U.S. power prices rose first; equipment orders followed later

  • U.S. power demand had been flat for about 10 years and Europe’s for about 20. Existing gas plants could generate less and make room for solar. The initial break in that balance came not from AI, but from manufacturing reshoring after the pandemic, along with subsidies for semiconductor, automotive and battery plants.
  • 唐奕波 recalls manufacturing investment at one point “possibly growing about 10x year on year.” Demand recovered, but no firm power had been built for years. Raymond summed up the difference between China and the U.S.: “Their shortage is demand-led; ours was purely about suppressing supply.”
  • The first effect came through power prices: in 唐奕波’s account, CEG and VST shares rose about 3–5x in 2023. Improved generator earnings then created an incentive to invest, but gas-turbine equipment orders did not turn meaningfully higher until mid-2024. After years of losses, manufacturers do not expand capacity the moment conditions improve.

6. The three heavy-duty turbine makers have both resource and technology barriers

  • Raymond estimated that GE Vernova, Siemens Energy and Mitsubishi together account for about 85% of the global gas-turbine market. He put their cumulative shares at roughly 40%–50% for GE, about 20% for Siemens Energy and the low teens for Mitsubishi. 唐奕波 says China has not produced a strong competitor because of both a lack of incentive—“rich in coal, short on gas, poor in oil,” with insufficient domestic demand—and a technology gap.
  • 唐奕波 compares gas turbines with aircraft engines, calling both “the crown jewels of industrial manufacturing.” Combustion temperatures can reach about 1,700–1,800°C, and the equipment must run for 200,000 hours without failure—a very high bar.

7. The real money in gas turbines is in OEM parts and long-term service

  • 唐奕波 illustrates the economics with a hypothetical: sell equipment for $10M and bundle a service contract worth $20M–$25M over about 15 years. In the previous cycle, equipment was sold at close to zero profit; service operating margins were about 40%.
  • “The most profitable part is spare parts,” not simply sending technicians to a site. Power plants can rarely maintain these highly technical components themselves, making OEM contracts close to mandatory. Parts margins may be 40%–50%, while the incremental margin on service itself is about 20%.
  • Asked by Raymond why GE could not win customers with a 40% price cut, 唐奕波 clarifies that GE was trying to win other companies’ service business. Making a rival’s spare parts is uneconomic; buying them from the OEM hands the profit back to the OEM. GE once set a target for servicing competitors’ equipment, but never met it and abandoned the business after a little over a year.

8. When incremental power demand goes to zero, equipment orders can fall off a cliff

  • Raymond asks why the industry can lose money for years despite long contracts and high barriers. 唐奕波’s answer: supply and demand determine the industry’s condition, “not how good the business model is or how high the barriers are.”
  • During the expansion around 2008, annual gas-turbine demand peaked at close to 100GW. 唐奕波 says slower growth in power demand can sharply reduce new equipment demand: if growth falls to zero, demand may be only 10–20GW, because the installed fleet is still young and replacement demand is limited. He calls the sensitivity of incremental power-equipment demand to power-demand growth a “second derivative.”
  • The downturn accelerated in 2015–2016. By 2018, all 3 companies had each laid off about one-third of their workforce and closed about one-third of their factories; the industry did not recover until 2024. Pressure from clean-energy policy, emissions rules and uncertainty around changes in political parties further discouraged capacity expansion.

9. Five-year lead times reflect both limited appetite to expand and a blade bottleneck

  • AI power-demand estimates keep rising, but the 3 heavy-duty turbine makers are planning only about 25% capacity expansion over 3 years, with orders already booked 5 years out. Raymond thinks professional managers are happy to maintain strong conditions; 唐奕波 adds that blade supply is also constraining turbine output.
  • 唐奕波 says the cycle will only truly end when heavy-duty turbine capacity expands: if it doubles in 3 years, more expensive alternatives will no longer make economic sense. But almost the entire supply chain is outside China, so an OEM deciding to expand cannot solve the problem on its own.

10. The supply gap is pulling increasingly marginal sources into the market

  • A roughly 500MW heavy-duty turbine may generate power at about $0.06/kWh. Baker Hughes light-duty turbines are about 50MW and cost about $0.08/kWh; they were bought up in 2025: “If you can’t get one 500MW unit, buy 10 50MW units and make do.”
  • Elon Musk even had equipment removed from abandoned gas-fired power plants in Europe and shipped back to the U.S. 唐奕波 calls this “very strange supply” emerging after prices rose.
  • Bloom Energy fuel cells cost at least $0.12/kWh at the time, but can be delivered in 3–6 months. 唐奕波 stresses that when they were in demand in 2024–2025, it had not been proven they could supply a 500MW-to-1GW power plant all at once. Existing projects were mostly supplementary power for hospitals, schools and small campuses, though Oracle was still willing to buy.
  • Further out, Caterpillar, Cummins and Weichai are converting diesel engines to run on natural gas, while FTAI is repurposing retired aircraft engines for power generation. Costs are higher and efficiency lower, but these options can fill the gap when heavy-duty turbines cannot be delivered.

11. Blade shortages are what made customers willing to try Chinese suppliers

  • Siemens Energy’s CEO said on an earnings call that blades were the bottleneck to expansion, prompting 唐奕波 to trace the supply chain. Major suppliers include Howmet Aerospace and PCC, which together produce about 80% of blades. He also notes that related stocks have risen 4–5x and trade at P/E ratios above 40.
  • High-temperature blades depend on skilled workers. 唐奕波 observes that Howmet’s headcount had still not recovered to its 2019 level in 2024–2025. Only extreme scarcity created an opening for domestic suppliers; otherwise, for such a critical component, “no one would even consider trying them.”

12. Yingliu’s commercial validation was built on a decade of investment against the cycle

  • Among domestic companies, 唐奕波 looked at Yingliu and Haimai Technology, focusing on Yingliu. The company started making blades in 2015; later, GE gave it a drawing, and the first sample took about a year. That proved it could make the part, not that it could produce at scale or get it into a turbine.
  • Moving from small batches and low yields to running thousands of hours in a turbine, then winning a paid small order, takes at least 3–4 years. Switching to Siemens Energy would require starting validation again. Raymond’s questions bring out the distinction: making a blade is a long way from commercial validation.
  • 唐奕波 recalls that when Yingliu’s net profit was about RMB100M, it was still investing about RMB200M in capex and RMB200M–300M in R&D every year, and kept that up for 10 years: “Either exceptionally forward-looking or exceptionally committed.” The stock has risen as much as 5x; at the time of the conversation, it was up about 3x.

13. U.S. construction hurdles could push training demand overseas

  • 唐奕波 sees the next constraint beyond equipment: residents are opposing data centers as power prices rise, and midterm elections are turning the issue into a political one. New York and Texas have passed relevant laws. The federal government wants construction to speed up, while states may impose limits, delaying approvals and construction.
  • His view is conditional: if these hurdles prevent the U.S. from building its planned data centers by 2027, training could move overseas while inference stays near customers. Japan, South Korea and Southeast Asia already have related plans.
  • Europe is also seeing migration within the region: grid connections are full in traditional hubs such as Germany, France, the Netherlands and Ireland, pushing projects toward solar in Spain and Portugal and hydro and nuclear power in the Nordics. 唐奕波 sees potential in places with more power and grid capacity.

14. Merchant generators offer price upside; regulated assets offer certainty

  • 唐奕波 is researching an unnamed Nordic power producer with a dividend yield of about 5%. It is a merchant generator, so higher demand and power prices can flow through to profit. But data centers have to be built before higher prices feed into EPS, leaving room for a lag in market recognition.
  • Raymond illustrates the upside with a hypothetical: if power prices rise from RMB0.02/kWh to RMB0.20/kWh, a 5% dividend yield could become 50% a year later. 唐奕波 agrees with “the broad logic,” but the discussion does not give a specific earnings forecast for the company.
  • Regulated assets earn an allowed return on investment—for example, invest RMB10B and receive an allowed return of 8%—so growth mainly comes from additional capex. Overseas brokers may recommend these assets because five-year plans make EPS easier to forecast. 唐奕波 is more interested in power-price upside after a long period of stagnation.
  • To guard against blind spots, 唐奕波 recalls traditional auto analysts who missed BYD and Tesla: not having seen change in an industry for a long time does not mean it cannot happen next year. He reminds himself, “Don’t become an old fogey” (不要做个老登).

15. Europe is seeing structural tensions first; an aggregate shortage depends on demand materializing

  • European power consumption still fell in 2024 and only grew in 2025; it has not returned to its previous equilibrium. 唐奕波 sees no clear aggregate supply shortage yet, but solar is depressing prices at midday while prices rise sharply in the evening or in certain seasons, already driving storage development.
  • A conservative forecast is annual demand growth of about 2% over the next few years; an optimistic case that includes data centers could exceed 3%. 唐奕波 worries that new power supply and policy adjustments move more slowly in Europe than in the U.S. If demand materializes, the resulting gap could be large.
  • Raymond suggests that reindustrialization could add 1 or 2 percentage points of growth. 唐奕波 still sees data centers as the key driver.

16. Grid-connection applications are an early signal; existing connections sell time directly

  • By 唐奕波’s recollection, European data-center grid-connection applications rose from more than 200GW in early 2025 to more than 300GW in October that year. He estimates they could reach about 450GW in June or July 2026. These are early applications and expressions of interest, not a final buildout forecast.
  • He also cautions that U.S. interconnection queues in earlier years included many solar applications, like “claiming the number for a Beijing new-energy vehicle license plate” just to hold a place. Applications need to be tracked alongside actual construction.
  • In many parts of Japan, South Korea and Europe, grid connections now take 8–10 years. Existing connections are therefore valuable: 唐奕波 cites a retired RWE coal plant that sold for about €400M, with the buyer paying for the interconnection and avoiding a new application.
  • Raymond understands this as discounting the “time value” of 10 years into today.

17. The equipment trade has spread from core scarcity to crowded, marginal supply

  • Asked whether there are still left-side opportunities in the equipment chain, 唐奕波 answers, “No more.” Raymond points out that small engines still represent relatively few GW in absolute terms; 唐奕波 responds that they do not face the same clear blade bottleneck and can expand “very aggressively,” suggesting the supply gap has been filled.
  • 唐奕波 puts crowding at 8–9 out of 10: since 2023, the trade has spread overseas to China and from reliable solutions to marginal supply over 3 years. In March and April 2026, repeated WeChat Channels recommendations for “gas turbine exports” became his signal that he did not want to keep playing these stocks: “It’s hard to make money now” (现在很难挣钱了).