Energy as the ultimate governor of intelligence
We measure a nation’s AI capability the way that we rank model releases. Benchmarks, parameter counts, and token cost all serve as our leading indicators of a country’s competitiveness relative to peers. That’s a pitifully incomplete scoreboard that emphasizes today’s capabilities at the expense of tomorrow’s diffusion.
CEO of Nvidia and leather-jacket-clad AI industry narrator, Jensen Huang, has popularized a “five-layer cake” AI framework that demonstrates the true profundity of AI capability dimensions. It reveals an ecosystem of dependencies, in which each layer is capped by the layer that precedes it.
At the base sits energy, which, as Huang notes, “sets the ceiling on how much intelligence can be produced at all”.
If energy is the fundamental governor of the rate at which AI capability can accumulate, then the compounding national advantage is not model quality, but rather the physical and institutional capacity to add power quickly. The former is an output that cannot materialize in the absence of the latter.
The scale of AI’s energy needs is substantial. In its base case scenario, the IEA projects that data center electricity consumption will nearly double from 485 TWh to at least 945 TWh by 2030. Because data centers are built where power is reliable, capital is cheap, and digital infrastructure already exists, the US, EU, and China sit squarely in the nexus of demand.
However, these nations possess very different capacities for adapting to AI’s surging energy demands. With no nation willing to concede the AI race, those differences have never loomed so large. What follows is their genesis and an exploration of three dynamics born of the collision between AI demand and Western grids: demand exiting the grid entirely, costs spilling onto ratepayers, and a buildout defaulting to whatever can be delivered fastest.
Why China, the US, and the EU adapt at different speeds
China’s economic development is well-documented. Over the past 40 years, per capita GDP has exploded by nearly 50x in nominal dollar terms. Unsurprisingly, over the same 40-year period, China’s per capita energy consumption grew by 495%, spurred by both a richer population and booming industrial manufacturing. Meanwhile, in the US and EU, consumption has been basically flat.
These drivers have created two divergent electricity generation trends. China, confronted with a rapidly developing economy and surging energy consumption, has added enormous amounts of electricity generation infrastructure, growing capacity at nearly 8.5% per year over the past 40 years. Total US and EU generation logged annual growth of just 1.3% and 0.8%, respectively.
Total Electricity Generation: US, EU, and China
Annual electricity generation (TWh) | 1985–2025
None of this is an indictment of the Western strategy or evidence of a gap in ambition. This is merely a story of different pressures giving rise to different systems. China’s growth, driven by energy-intensive industrialization, has been contingent upon massive electricity consumption. The US and EU, meanwhile, have historically hosted much more diversified economies, steadily shifting from industrial manufacturing to technology and services, which are far less energy-intensive. The flat lines in the graphic above show economic growth decoupling from electricity generation in motion.
2000–2025, Real GDP against electricity generation
| GDP Total | GDP CAGR | Electricity Total | Electricity CAGR | Elasticity1 | |
|---|---|---|---|---|---|
| United States | +69% | 2.13% | +19% | 0.69% | 0.33 |
| European Union | +41% | 1.38% | +7% | 0.26% | 0.19 |
| China | +587% | 8.01% | +681% | 8.57% | 1.07 |
In short, China built capacity ahead of demand at a fervent pace due to energy-intensive industrialization. It built institutional capacities to support that goal: top-down build quotas, state-owned grid operators, policy-bank financing, vertically integrated domestic manufacturing. The US and EU imposed efficiency measures and shifted to tech and services-based economies. They built institutions that prevent unnecessary construction across decades when construction was, in fact, unnecessary: prudence review and the “used and useful” doctrine, distributed authority, permitting as a binding step.
In a world in which electricity constrains AI capability and diffusion, China’s recent history endows it with advantageous muscle memory. Adding grid capacity is a familiar endeavor. The West, however, must refamiliarize itself with this art.
Where AI demand collides with a stagnant grid
This refamiliarization is not optional. The data centers are not waiting. They are drawn to a brew of skilled labor, local IT demand, capital, and preexisting infrastructure. The US and Europe, despite not presently possessing the latent power of China, host those ingredients in spades. They are consequently the world’s two largest regional data center by count, and rank first and third by capacity (China, of course, places second).
Thus, the data centers have arrived, and all indications are that they will continue to proliferate. The result is a collision between soaring demand and a stagnant grid. Three dynamics, flagged earlier, will define the outcomes of this collision: demand exiting the grid, costs spilling onto ratepayers, and the buildout itself defaulting to whatever the supply chain can build fastest.
Behind-the-meter: Data centers exiting the grid
The first dynamic is exit from highly congested systems altogether. Grid connection queues in the US and EU now run five to ten years. As idle compute clusters carry extreme opportunity costs, developers are increasingly electing to go “behind-the-meter” (BTM). BTM usually takes one of two forms.
The first is the data center developer building out its own onsite energy generation. This is how xAI was able to build a 100,000-GPU cluster in just four months, deploying hundreds of megawatts of onsite gas turbines.
Notably, xAI’s speed came at a cost. The turbines were unpermitted and have attracted an NAACP-led Clean Air Act suit.
The second BTM option is co-location, which, while also adopted in the US, is the dominant BTM solution in Europe due to the prohibitive cost of gas. Co-location consists of building a data center in close proximity to an existing generation or storage asset, directly drawing from that asset rather than the grid itself.
Who pays: Ratepayer burden and rising bills
The second dynamic is cost; specifically, who bears it. In the US, the answer is increasingly ratepayers themselves. Meeting the energy demands of the AI boom will require substantial investments in the electrical grid. Barring rate structures that assign those costs to the loads driving them, this buildout will ultimately be subsidized by US ratepayers via 15% to 40% higher electricity bills by 2030. In some areas, current conditions are even more dire than those future forecasts.
Because data centers are attracted to unique ecosystems of which there are a finite number (again, defined by the presence of skilled labor, IT demand, capital, and preexisting infrastructure), they tend to cluster in specific geographies. For example, Europe’s “FLAP-D” markets (Frankfurt, London, Amsterdam, Paris, and Dublin) account for 62% of Europe’s total data center capacity. Meanwhile, a relatively small corridor in northern Virginia, known as “Data Center Alley,” represents the largest data center market in the world.
That data centers are not evenly distributed means that their effects are not either. In cluster regions, data center load dwarfs national averages. Across the whole of Virginia, data centers account for 25% of total electricity demand. In some data center-adjacent areas, wholesale electricity prices have increased by as much as 267% over the past five years.
Even if not indicative of the aggregate reality, these headline figures have put politicians in a difficult position: support the development of AI, which is increasingly seen as a strategic geopolitical and economic necessity, or endorse ratepayer protections and moratoriums.
Supply chains, not preferences, decide the generation mix
The third dynamic is simple pragmatism: build anyway and build whatever the supply chain can reasonably deliver. Here, lead time has a far greater influence on the generation mix than ideology or preference.
Every project awaits a similar suite of grid equipment. Large power transformers have a lead time of about 128 weeks. Generator step-up units clock in at roughly 143 weeks. That floor is common to all projects. Differentiated lead time comes from what stacks atop it.
Three manufacturers control roughly three-quarters of the large-frame gas turbine market, and they are effectively sold out. GE Vernova, for example, closed 2025 with an 80 GW backlog and expects to be booked solid through 2030. Manufacturers are advising would-be customers to plan for seven- to eight-year timelines. Batteries face tight supplies as well, but a greater diversity of manufacturers means that the pipeline is far less constrained.
The consequence is visible in the predicted constituents of the near-term energy buildout. For example, in the US, the EIA anticipates a record 86 GW of new US utility-scale capacity in 2026, of which 51% is solar, 28% battery storage, and 14% wind. Natural gas accounts for roughly 7%.
New US Utility-Scale Capacity Additions: 2026
Planned utility-scale generating capacity additions (GW) | 86 GW total — largest single-year addition since 2002
- Solar 43.4 GW·51%
- Battery storage 24.3 GW·28%
- Wind 11.8 GW·14%
- Natural gas 6.3 GW·7%
GW
In the fast-moving world of AI, in which every month lost represents millions in deferred revenue, the most attractive megawatt is that which can be built the fastest. Increasingly, that implicates solar and storage.
What AI energy demand means for investors
For decades, Western grid modernization has been a perpetual bridesmaid. It has been endorsed by everyone, funded by no one, urgent to no one in particular (or at least no one with a sufficiently authoritative pulpit). AI changes that dynamic. When electricity becomes the binding constraint on a technology that governments treat as strategic infrastructure, the institutions built to constrain construction acquire a reason to permit it. Interconnection queue reform in the US, the EU’s Grids Package, and transmission spending at record highs all indicate that the hulking machinery of Western grids is beginning to lurch into motion. Crucially, the primary catalyst was not the long-awaited landing of the case for climate, but rather the sudden case for competitiveness.
The composition of the buildout and the speed at which it needs to occur cannot be teased apart. They run along an intertwining path towards a green transition. With gas turbine supplies exhausted into the next decade, the fastest, and therefore most valuable, megawatt is a renewable one. AI demand, so often cast as the energy transition’s primary antagonist, may prove to be its most powerful accelerant.
This transition, and the race driving it, will demand tremendous capital. In deploying it, investors will need to answer a long list of questions that straddle social, governance, supply chain, and environmental issues. Which utilities and grid-equipment suppliers sit on the right side of a decade-long capex cycle? How exposed are data center-heavy regions (and the companies that operate within them) to ratepayer backlash and potential regulatory whiplash? What are companies doing to reduce that exposure?
Answers will not come easily, but they will decide who generates a return on this energy revolution.
Huang’s cake has five layers, but only one of them is poured in concrete and laced in copper. Benchmarks measure the intelligence a nation has; grids determine the intelligence it can add. The nations that build will set their own ceiling on the latter.
Stay updated
Author Information
