Below the clouds is electricity.

(Image caption ) Inside the data center, server racks are neatly arranged, and blue indicator lights flash. The new industrial revolution in the AI era does not always occur in roaring factories, but rather in these silently operating buildings.

The New Industrial Revolution in Silent Buildings

Some turning points in history do not begin in parliament, stock exchange, or battlefield, but in a seemingly silent building.

There were no audiences, no applause, and none of the grand, roaring sounds of a traditional factory. Only rows of servers, GPUs, cooling systems, transformers, backup generators, and an ever-flowing electric current. These are called data centers. But in the age of AI, data centers are no longer just the back-end infrastructure of the internet; they are more like new steel mills, new railways, new ports—the underground heart of a nation's future productivity, silent, expensive, and indispensable.

The story of AI, on the surface, is a battle over models; at a deeper level, it's a battle over computing power; and further still, it's actually a battle over energy, land, power grids, equipment, permits, and capital structures. Each layer is heavier than the last, and each layer is more difficult to change.

(Image caption ) AI data centers are not an industry floating in the cloud. They are located on concrete land, connected to concrete power grids, consume concrete resources, and are gradually changing the boundaries of cities, energy, and public costs.

Growth Concentration Behind the AI Boom

After 2025, a signal emerged in the US economy that warrants serious attention. Harvard economist Jason Furman pointed out that while investment in information processing equipment and software accounts for only about 4% of US GDP, it contributed 92% of GDP growth in the first half of 2025; if this portion were excluded, the annualized GDP growth rate during the same period would almost stagnate. Furman himself added that without the AI boom, interest rates and electricity costs might decrease, and other sectors might experience some compensation; therefore, it cannot be simply concluded that the US economy is entirely dependent on AI.

Even so, these figures still serve as a warning: when a country's growth increasingly depends on a few technology companies, a few capital expenditures, a few cloud computing platforms, and a few energy nodes, the economy itself begins to become more concentrated, more vulnerable, and more institutionally dependent on electricity.

This is not your average technology cycle. This is an infrastructure repricing triggered by AI.

(Image caption ) The end of computing power is not the chip, but electricity. No matter how powerful the GPU or how expensive the model cluster, it must eventually return to transformers, power lines and a stable power supply system.

Stargate: Capitalism and the Reality of Electricity

OpenAI's Stargate project is the most typical example of this era, and also the most complex warning.

On January 21, 2025, OpenAI, together with SoftBank, Oracle, and MGX, announced the Stargate Project, a plan to invest $500 billion over four years to build AI infrastructure in the United States, with an initial deployment of $100 billion. In its announcement, OpenAI directly linked this infrastructure to US reindustrialization, national security, and AI leadership. This is not just ordinary corporate expansion, but a public declaration by a technology company rewriting its capital needs in the language of national strategy.

However, between declaration and implementation, real frictions often lie within the system. Subsequent media reports and related cooperation arrangements reveal that large-scale AI infrastructure projects like Stargate inevitably face issues such as the boundaries of responsibility among multiple parties, the pace of capital investment, data center control, and land and power supply. These details are not intended to negate the scale and direction of the project, but rather to remind us that even the greatest capital will must pass through layers of barriers—electricity, land, engineering, financing, permits, and regulations—before it can truly be transformed into electricity on server racks.

(Image caption ) The pace of AI data centers is driven by model competitions and capital expenditures, while the pace of the power system is constrained by planning, approval, equipment delivery, and regulatory procedures. Two sets of time are converging on the same power grid.

The end of computing power is not chips, but electricity.

The problem is that the end of computing power is not chips, but electricity.

The International Energy Agency (IEA), in its 2025 "Energy and AI" report, estimated that global data center electricity consumption was approximately 415 TWh in 2024, accounting for about 1.5% of global electricity consumption. The United States accounted for the largest share of global data center electricity consumption, followed by China and Europe. By 2030, global data center electricity consumption is projected to approach 945 TWh, equivalent to Japan's current annual electricity consumption. More importantly, in mature economies like the United States, where long-term electricity demand is relatively stable, data centers will become one of the most concentrated sources of pressure from increased electricity demand.

This means that the United States needs more than just more chips; it needs to restart an entire power growth system.

(Image caption ) The beneficiaries of the AI industry chain are no longer limited to chip companies. Gas turbines, transformers, copper materials, liquid cooling equipment, power engineering, and on-site power generation systems are being reintegrated into the core narrative of AI infrastructure.

When slow power systems meet the era of fast computing power

For decades, the mature U.S. power system has been accustomed to low growth, slow planning, and lengthy approval processes. This pace is not simply a matter of complacency, but rather a result of institutional design. Utilities have their own investment return cycles, grid interconnection has its own regulatory procedures, new transmission lines in many states require multiple approvals, environmental impact assessments, and legal proceedings, and critical power equipment such as transformers, cables, and gas turbines also face lengthy delivery cycles.

However, the pace of AI data centers is the same as the pace of the Silicon Valley capital and model race. Model iterations are calculated monthly, GPU orders are contested quarterly, and data center site selection is queued by megawatt. The gap between these two paces cannot be bridged by any single company through hard work; it requires a genuine restructuring at the institutional level.

Taking ERCOT in Texas as an example, NERC's long-term reliability assessment has noted that the rapid integration of large loads is changing grid planning assumptions. The overlapping of data centers, AI facilities, crypto mining farms, and large industrial loads is pushing the grid from its past gradual growth to sudden, centralized, megawatt-level queuing stress scenarios. For a power system that still faces extreme weather and reliability risks, this is not just growth, but a stress test.

AI data centers are propelling America's "slow power system" into an "era of fast computing power." This conflict is far deeper and more structural than most market commentaries suggest.

In October 2025, OpenAI submitted a policy document to the White House, summarizing its assessment of electricity in one sentence: "Electronics is the new oil." The document proposed that the United States should significantly increase its annual new power generation capacity, citing a stark comparison: China added approximately 429 GW of power capacity in 2024, while the United States added approximately 51 GW, which OpenAI called the emerging "electron gap."

These figures need to be understood within the context of capacity structure and power quality. A significant portion of China's new installed capacity comes from wind and solar power, whose volatility and dispatchability differ from those of gas-fired, nuclear, or stable baseload power sources. A simple comparison of GW (gigawatt-hours) is not equivalent to a direct comparison of available power.

However, the real issues it reveals cannot be ignored: China adopts an industrial mobilization logic in power construction, while the United States remains constrained by decentralized governance, local approvals, litigation procedures, and capital recovery mechanisms. This is not a simple comparison of which is more advanced, but rather a difference in the pace of two systems. The competition for computing power in AI is forcing this difference in pace to be clearly measured.

China's advantages lie in its construction speed, supply chain completeness, and large-scale manufacturing capabilities of renewable energy. The US's advantages lie in cloud computing, chip design, capital markets, AI models, and its global technology ecosystem. However, AI infrastructure isn't about individual strengths, but rather the system's ability to operate synchronously. GPUs can be delivered in batches, but electricity cannot be generated immediately upon order; capital can be priced in a day, but power lines cannot cross state borders in a day.

This makes the AI infrastructure development a competition of energy at the institutional level, rather than just a competition of capital and technology.

(Image caption ) Behind every large data center is not just corporate investment and technological competition, but may also affect residents' electricity bills, local tax revenue, land use, water resources and community affordability.

Power First: The Repricing of the AI Industry Chain

From an industry chain perspective, the beneficiaries are no longer limited to chip companies. Gas turbines, transformers, silicon steel, copper materials, liquid cooling equipment, power engineering, natural gas pipelines, nuclear power, small modular reactors, geothermal energy, battery energy storage, grid software, and on-site power generation systems—all of these have been reintegrated into the core of the AI narrative.

In the past, energy companies viewed technology companies as stable electricity customers; today, technology companies are beginning to behave like energy developers. They are not just buying electricity, but scrambling for land, grid connection slots, turbines, and transformers, and some are even considering configuring or building their own power sources directly next to data centers.

This is a logic known in the industry as "Power First," which is becoming the primary principle for site selection in the AI era.

In the era of traditional cloud computing, data center site selection could prioritize factors such as network connectivity, tax policies, land prices, customer distance, and climate conditions. In the AI era, the first question becomes: Where can we find sufficient, cheap, stable, and quickly grid-connected electricity?

Without electricity, even the best land is just empty land; without electricity, even the most numerous GPUs are just expensive inventory; without electricity, even the grandest model vision is just a line of numbers on a PowerPoint presentation.

(Image caption ) The AI era is not an era of floating in the clouds. Only those who can integrate energy, computing power, capital, and public governance into a single, trustworthy framework will truly be qualified to build the next generation of intelligent civilization.

Public costs, institutional boundaries, and the next generation of intelligent civilization

GFM is more concerned with how the expansion of tech giants will change the boundaries of public costs than with how this expansion will affect them.

Data centers are not an industry floating in the cloud. They are situated on concrete land, connected to concrete power grids, consume concrete water resources, and alter the electricity pricing structure, noise pollution, and tax ecosystem of concrete communities. Community backlash against large data centers has emerged in several parts of the United States. This is not simply anti-technology sentiment, but rather residents questioning a systemic issue:

When a few tech companies require massive amounts of electricity for AI competition, who bears the cost of upgrading the power grid? Are residential electricity bills silently paying for corporate infrastructure? What local governments are gaining in exchange for tax breaks: long-term, high-quality employment, or high-energy-consuming, low-labor-density asset-heavy data centers?

These issues cannot be completely swallowed up by the phrase "national competition." Every local resident's electricity bill is a real-world manifestation of this grand narrative.

The IEA's assessment is relatively restrained: data centers are not the only source of global electricity demand growth; electric vehicles, air conditioning, industrial motors, and the resurgence of manufacturing are also changing the power system. However, in mature economies like the United States, where electricity demand has historically been relatively stable, the impact of data centers is more concentrated and sudden. In other words, while AI may not be the only major issue affecting global electricity, it is becoming one of the most acute new pressure points on the US power grid.

To complicate matters further, the energy requirements of AI itself continue to evolve.

Early competition in large-scale model development focused primarily on training, pursuing larger clusters, faster interconnections, and higher density. In the future, inference needs may account for a larger proportion. Inference can be more decentralized, but it could also create a massive long-term infrastructure load due to the continuous expansion of the user base. Training is like an intensive battle, while inference is like a city's daily water supply—the former is intense, the latter is continuous.

The real and lasting energy pressure often comes not from a single peak, but from the constant and continuous consumption.

Therefore, the energy problem in AI cannot be solved simply by "building more power plants." It requires several systems to be implemented almost simultaneously.

Power grid planning must shift from a low-growth assumption to a high-growth assumption, treating data centers as a new type of industrial load, rather than ordinary commercial customers.

The cost-sharing of power grid expansion must be priced publicly, and who benefits, who pays, and who bears the risk cannot be left to tacit conventions.

Energy procurement must move beyond simply "purchasing green electricity certificates" to the level of "adding real supply" and "dispatchable supply." Carbon offsetting on paper cannot replace real system reliability.

Computing power itself should also be designed as an adjustable load. Non-real-time inference tasks, backup capacity, local energy storage, and portable workloads can provide flexibility to the grid, rather than requiring uninterrupted power supply indefinitely.

Finally, AI companies need a more sophisticated public narrative. They can't just portray data centers as a means of innovation, employment, and national competition; they must also honestly explain their impact on electricity prices, water resources, land, noise, taxes, and communities. Infrastructure lacking transparency will ultimately erode public trust in technology.

The real contradiction of AI lies here: it enters the world in the name of "intelligence," but first exposes the cumbersome nature of the energy system; it promises to improve efficiency, but in the initial stage, it requires a huge amount of material input; it seems intangible, but it relies more on steel, cement, copper wire, natural gas, cooling water, and land than many traditional industries.

(Image caption ) After 2025, although investment in information processing equipment will only account for about 4% of GDP, it will contribute 92% of GDP growth in the first half of the year, highlighting the phenomenon of concentrated growth under the boom of AI.

This is not irony, but the true nature of modern civilization.

Every information revolution ultimately impacts the physical world. Printing required paper, ink, and transportation; the telegraph required copper wire; the internet required fiber optics and submarine cables; and AI requires chips, electricity, and data centers. Humanity always believes it is becoming lighter, but each new form of information buries a heavier infrastructure beneath the earth.

Therefore, the era of trillion-dollar infrastructure investment in AI data centers cannot be understood merely as a round of technological investment. It is a new industrialization and also a new institutional test.

It tests whether the United States can find a truly working coordination mechanism between local governance and national strategy; whether technology companies can establish a credible boundary between private gains and public costs; whether capital markets can distinguish between genuine productivity expansion and capital expenditure bubbles; and whether the energy industry can transform from a conservative public utility into a provider of new infrastructure supporting the AI era.

In the coming years, the market will continue to ask who has the strongest model, the most users, and the largest GPU cluster. But what will truly determine the AI landscape may be a few more fundamental questions:

Who can get electricity? Who can get a stable supply of electricity? Who can get electricity at an affordable price? Who can get electricity without disrupting the community, passing on costs, or sacrificing system reliability?

The AI era is not an era floating in the clouds. It is pulling us back to the earth, back to power plants, transformers, transmission lines, rivers, land, and the electricity bill that every ordinary family receives at the end of the month.

This is GFM's basic assessment of AI energy:

Computing power is not the new oil. Electricity is the new institutional foundation.

Whoever controls electricity growth controls the boundaries of AI; whoever can integrate energy, computing power, capital, and public governance into a single, trustworthy framework is truly qualified to build the next generation of intelligent civilization.

⸻

Disclaimer:

This article is an analysis of energy and industry trends and does not constitute investment advice, legal advice, securities recommendations or any basis for trading.