The 2.5 Cent AI-Power Premium Is a Policy Choice — and China's $945 Billion Is the Proof

The International Energy Agency's World Energy Investment 2026 puts global energy capital formation at $3.4 trillion, a 5% rise on 2025. China leads at $945 billion, ahead of the United States at $615 billion and the European Union at $440 billion; clean energy takes $2.2 trillion of the total, and China, the US and the EU together account for $1.5 trillion, or 66%, of it. China extended a lead it first took in the 2025 edition, when it edged ahead of the United States as the largest national energy investor.

Here is the part the headline misses. China's low-emissions power investment falls from about $315 billion in 2025 to under $290 billion in 2026 — the country is cutting generation capex while grid capex rises to roughly $140 billion. Beijing is spending less on making electrons and more on moving them, and it is doing so through the same procurement machine that supplies the rest of the world. That is why Chinese electrical equipment is the marginal supply in every grid market on earth, including the American one.

China's capex did not just consume hardware; it created the capacity to make it. Chinese transformer exports hit a record 64.6 billion yuan ($9.3 billion) in 2025, up nearly 36% year on year, at an average unit price of 205,000 yuan. Because that supply base exists, American import prices for Chinese electrical equipment sit at 93.8 on an index where June 2012 equals 100 — prices now sit below their 2012 base — and they are still sliding, from 94.5 in July 2026 to 93.8 in September.

On August 31, 2026, US Executive Order 14420 banned imported bulk-power equipment from China and up to 23 other nations. Wood Mackenzie's assessment was blunt: the order could affect more than $22 billion of imports since 2025, with China representing almost all bulk-power imports since the beginning of 2025. At that moment the market was already short: power transformers were 15% short and substations 8% short, with the 100 MVA segment the most acute and the market above 10 MVA having expanded roughly 300% since 2020.

None of this is physics. Transformers are not a scarce natural resource; they are a manufacturable product with a learning curve. The United States had a 15% shortage of a product China shipped $9.3 billion of in 2025, at an average export price of 205,000 yuan a unit across all sizes. The shortage is a capacity, permitting and tariff outcome, and the executive order widens the policy component rather than the physical one.

PJM's capacity market is where the scarcity gets a price, and it has repriced by an order of magnitude. Capacity cleared at $28.92 per MW-day for the 2024/25 delivery year, at $329.17 for 2026/27, at the $333.44 administrative cap for 2027/28 — with that auction 6,831 MW short of its reliability requirement — and at $325 per MW-day for 2028/29. This is not a gradual squeeze; the cap, not the market, has been setting the price for three consecutive delivery years.

Convert that into a per-kilowatt-hour cost and the premium becomes concrete. At $325 per MW-day, one megawatt of firm capacity costs $118,625 a year; at an 85% load factor, or 7,446 MWh, that is $15.93 per MWh, against $1.42 per MWh in 2024/25. The incremental cost is therefore roughly 1.5 cents per kWh at the capped price, and about 3.7 cents had the auction cleared at its uncapped equilibrium above $775 per MW-day. The midpoint of that range sets the US data centre's capacity bill at about 2.5 cents per kWh — before a single dollar of energy-price inflation.

The scarcity rents are landing in two places: Western equipment backlogs and US power sellers. GE Vernova took $29 billion of orders in the third quarter, up 96% organically, booked $12.6 billion of revenue and lifted backlog to $197 billion, guiding to full-year revenue of $46.5-47.5 billion and free cash flow of $16-17 billion. Siemens Energy is running the same playbook in grids: an order backlog of €162 billion, a 1.57 book-to-bill, and a Grid Technologies segment growing revenue 28.6% at a 19.9% margin.

Here is the part that should worry the bull case. GE Vernova's market capitalisation fell from $314.9 billion at the end of June 2026 to $204.9 billion by early November: the market has taken 35% off the stock while backlog grew by $20 billion. It now trades on 18.5 times trailing EV/EBITDA and 4.3 times trailing sales, with trailing earnings of $38.45 a share. The market is not disputing the demand. It is pricing the durability of a premium that Washington created and Washington can withdraw.

Downstream, the money keeps moving. Constellation Energy converts scarcity into contracted cash flow — trailing revenue of $31.27 billion, EBITDA of $7.98 billion and net income of $3.47 billion — while Vertiv sells the boxes that make a megawatt usable, with third-quarter net sales up 61% to $4.31 billion, organic orders up more than 100% and adjusted operating margin of 25.2%. With hyperscaler capital spending running at roughly $725 billion in 2026, the downstream buyers of power are not the ones rationed. The upstream Western manufacturers of grid gear are.

Two clocks matter here, and they are not the same clock. The next four quarters are about converting a policy-made scarcity into cash. The next three years are about whether the policy survives contact with the supply chain.

State the falsifiers plainly. If Chinese transformer export unit prices turn up sharply while volumes hold, the market has genuinely cleared and the scarcity argument weakens. If PJM's next auction clears below the cap, the capacity premium has already begun to unwind. And if Chinese domestic grid spending stalls below the $140 billion the IEA expects, the deflation engine that has held equipment prices down stalls with it. None of those three has happened yet.

Pulling it together. China is spending $945 billion a year, 1.5 times the United States, and rotating that spending into grid hardware, which is why Chinese equipment is the marginal supply in every market and why prices for it sit below 2012 levels. The United States answered a genuine manufacturing bottleneck with tariffs, an executive order banning imported bulk-power equipment, and an administratively capped capacity market that cleared 6,831 MW short. The result is a power premium for US data centres of roughly 1.5 to 3.7 cents per kWh of capacity cost, about 2.5 cents at the midpoint, and a windfall for a narrow set of Western balance sheets. That premium is a policy variable, not a physical one, and it will be repriced by Washington, by PJM, or by the next election long before it is repriced by physics.