AI’s New Bottleneck: Power Supply, Not Chips

Thanks to my ICIS colleague, Aliena Huang, for this important article on AI and electricity supply versus the importance of computer chips. It could be, as I discuss below in my take on these issues and as Aliena concludes, that sufficient electricity supply will be the biggest factor in winning the AI race between the US and China. The question then becomes which country has the best approach to providing the power needed to achieve success. My guess is that will probably be China.

AS GLOBAL DATA volumes continue to rise and artificial intelligence moves into a new phase of “agentic” systems in 2026, the limiting factor for further AI expansion has shifted.

The challenge is no longer the supply of advanced chips, but rather the availability of reliable, large‑scale electricity.

The next generation of AI models—designed to reason, plan and operate over extended periods—requires continuous, high-intensity computing. This is pushing data centres into sustained peak power usage, far beyond the short bursts typical of earlier chatbot-style AI.

This shift has exposed the constraints of ageing power infrastructure, particularly in the United States, and is driving a divergence in strategy between the US and China. My view is that will probably be China who wins because of is centralised and long-term state-driven investment model.

US technology companies are increasingly bypassing the public grid to secure their own private energy sources. By contrast, China is tackling the issue at the national level, investing heavily in strengthening and modernising its entire grid system.

Why AI Is Now an Energy Problem

Global data is expanding at an extraordinary pace. International Data Corporation – an information technology market intelligence and analysis provider  – forecasts that worldwide data volume will more than double from 2025 to 2029.

They predict that global data volume will reach 213.56 zettabytes (ZB) by 2025, growing at a compound annual growth rate (CAGR) of 25.4%, doubling approximately every three years.

While storing data accounts for only a small portion of data‑centre energy use, the explosion in demand for AI processing—especially retrieval and inference—points towards far higher consumption levels.

Retrieval means fetching the right information from a database, a search index, or a knowledge store so the AI can use it.

Examples include searching documents to find relevant paragraphs, scanning emails to pull out the one you’re asking about. Checking a company database for product specs and finding notes, reports or numbers before responding.

Retrieval is fast, lightweight, and mostly about finding the right text or data.

Older AI systems mostly worked by predicting the next word, with no access to external facts. Modern systems retrieve information first, so answers are more accurate and less “hallucinated”.

But retrieval is relatively low on energy consumption compared to reasoning.

Inference involves the processes of, for example, understanding and summarising a document, planning a multi-step task, analysing data, writing code, making recommendations, breaking down a problem and deciding what to do next.  These are the reasoning steps of AI.

Inference is what happens every time an AI responds to you, even in a casual chat.

Inference is computationally heavy — especially in modern “agentic” systems that think for longer, run in the background, carry out multi-step plans, evaluate their own actions and revise answers. Hence, the high energy use.

Older AIs mostly did quick inference (seconds). New “agentic” AIs do long, continuous inference (minutes to hours). This pushes GPUs to run at full power for long periods, creating a huge demand for new energy. Agentic AI is artificial intelligence that can independently plan, reason, and take multi‑step actions to achieve a goal, rather than just responding to individual prompts.

The International Energy Agency in April last year estimated that global data‑centre electricity use reached around 415TWh in 2024, and earlier projections suggested this could double by 2030. When overlaid with data‑growth forecasts, however, the doubling point may arrive even earlier—potentially by 2027.

This acceleration is tied to the rise of “agentic ecosystems”: Advanced AI systems capable of autonomously handling multi-step tasks, analysing large quantities of previously underused “dark data” and delivering complex decisions.

Dark data is information that an organisation collects and stores but never actually uses for analysis, decision‑making or any meaningful purpose.

These agents perform deep reasoning, break down tasks, assess errors and draw on multiple tools—often running in the background for long periods.

These workloads keep Graphis Processing Units (GPUs) and accelerators operating continuously at high utilisation, unlike the quick-response pattern of older AI models.

Graphics Processing Units are types of computer chips that are very good at doing many small calculations at the same time.

Originally built to render video‑game graphics, GPUs turned out to be perfect for AI, because modern AI models require huge amounts of parallel maths.

The result is a surge in energy intensity as companies build big new computing centres.

New facilities announced in early 2026 include a Microsoft site in Wisconsin using vast numbers of GB200 processors; Meta’s “Prometheus” facility in Ohio consuming roughly 1GW; and xAI’s Mississippi base approaching 2GW—making it one of the most power‑hungry AI campuses ever built.

As these clusters operate at consistently high load, the industry’s attention has shifted. Chip performance still matters, but the overriding constraint is now the ability to secure and sustain enough power.

The US Response: Private Power as the Grid Falters

The American grid, much of which dates from the mid‑20th century, is struggling to keep up.

According to the American Society of Civil Engineers, around 70% of US transmission lines and large power transformers are over 25 years old, and 60% of circuit breakers have been in service for more than 30 years. This ageing infrastructure is poorly suited to the constant thermal stress generated by advanced AI facilities.

Circuit breakers are safety devices that automatically stop electrical flow if currents become too high, protecting against overloads and faults. Thermal stress refers to the heat-induced strain on electrical equipment when components are run continuously at very high power, increasing the risk of failure.

A further problem is grid congestion. While new power plants are being built, the transmission infrastructure needed to deliver that power is not expanding fast enough.

Interconnection queues have grown to multi‑year delays and supply of high‑voltage equipment involves long lead times.

On top of this, regulatory uncertainty—particularly around emissions standards—has deterred investment from utilities.

For example, utilities are experiencing mixed signals as the EPA plans to roll back the foundational 2009 “endangerment finding,” which legally recognised that six greenhouse gases endanger human health and welfare.

According to the Wall Street Journal in this 10 February article, the Trump administration plans to repeal this finding, described by EPA Administrator Lee Zeldin as “the largest act of deregulation in the history of the United States”.

The rule—expected to be finalised shortly—would remove requirements for measuring, reporting, certifying and complying with federal greenhouse-gas standards for motor vehicles, while suspending many industry compliance and credit schemes.

Although these changes would not immediately apply to emissions from power plants, repealing the endangerment finding could pave the way for broader rollbacks, undermining long-term planning.

Industry groups and utilities already face uncertainty as they weigh fuel-switching, carbon capture investments, or deploying advanced emissions controls—especially when today’s rules might be reversed tomorrow.

In response, major American technology firms have begun to secure power independently of the public grid:

• xAI. Elon Musk’s company is building an almost fully off‑grid solution by purchasing large-scale gas turbines. With around 1.9GW of private capacity planned, xAI aims to avoid grid‑connection delays entirely.
• Microsoft is pursuing a mixed strategy through partnering with grid operators, restarting the Three Mile Island nuclear plant in 2028 and designing more energy‑efficient AI chips, such as the Maia 200, to reduce consumption at the hardware level.
• Google is adhering to strict sustainability goals, Google has signed a 1.2GW agreement focused on delivering round‑the‑clock carbon‑free energy, combining renewable generation with storage to meet the continuous demands of AI systems.
• Meta is locking in future nuclear capacity through long-term agreements with developers, securing more than 6GW of potential output later in the decade.

Together, these approaches indicate a shift towards “energy islands”: Private, purpose‑built sources of power erected beside AI clusters. This reflects a loss of confidence in the public grid’s ability to expand fast enough.

The China Response: A National Infrastructure Solution

China is taking a very different path. Rather than bypassing existing infrastructure, the government is strengthening and re‑engineering the grid itself.

Backed by a dominant domestic electrical manufacturing base—particularly in transformers and transmission equipment—China is rolling out a systemic, long-term upgrade.

The State Grid Corporation of China and China Southern Power Grid have together pledged approximately CNY 5 trillion in investment during China’s 15th Five-Year Plan (2026–2030)—representing a 40% increase over the previous cycle. State Grid alone plans CNY 4 trillion, while Southern Power Grid has earmarked around CNY 1 trillion across the next five years.

A core focus of this budget is the expansion and modernisation of the Ultra-High Voltage transmission network, ensuring large-scale delivery of renewable energy from western regions to eastern demand centres. The National Energy Administration aims to increase “west-to-east” grid capacity to over 420 GW by 2030, up from about 340 GW in 2025

Beyond physical transmission, intelligent grid despatch systems and virtual power plant designs are being integrated to allow data centres to  adjust energy use in response to supply conditions.

Cooling efficiency is also being improved. Under the “Eastern Data, Western Compute” strategy, new data centres are being located in cooler western provinces, while liquid cooling systems help push power usage effectiveness closer to the ideal minimum.

The net effect is a grid designed not only to support AI, but to do so at scale, cost‑effectively and with lower energy waste.

The Emerging Global Picture

By early 2026, the central challenge in the AI race has become unmistakable: Power, not silicon, is the limiting factor. While American technology companies build private nuclear and gas supplies to escape a constrained grid, China is upgrading its national power system to deliver energy wherever it is needed.

The long-term competitiveness of AI ecosystems may depend less on breakthroughs in AI technology and more on which country can ensure the most abundant, reliable and efficient electricity.

The post AI’s New Bottleneck: Power Supply, Not Chips appeared first on Asian Chemical Connections.