Artificial intelligence has conditioned the market to think in terms of speed. New models every few months, more powerful chips, data centers announced with numbers that until recently seemed unlikely, and billion-dollar funding rounds. But the emerging limit is not just in software or semiconductors. It’s in the most physical part of the economy: electricity, turbines, transformers, permits, land, water, cooling, and construction.
A phrase attributed to Elon Musk and widely shared online captures this tension well. When asked why big tech companies don’t simply build private power plants next to their data centers to bypass the grid, the answer was straightforward: the bottleneck is in energy plant manufacturing. Beyond the literalness of the quote, the idea aligns with an increasingly visible reality: AI can scale in software much faster than heavy infrastructure can be built.
The paradox is uncomfortable. A team can train or fine-tune a model in weeks, purchase GPU capacity with multi-million dollar agreements, or design a distributed inference architecture in months. But building power generation, reinforcing substations, acquiring gas turbines, expanding transmission lines, or obtaining environmental permits can take years. AI advances as fast as software; energy infrastructure continues to depend on steel, concrete, and regulation.
The new bottleneck is no longer just GPUs
During 2023 and 2024, the AI debate focused on GPU shortages. NVIDIA became the symbol of the race, with each capacity announcement measured in chips, clusters, and gigawatts of compute. While that focus remains relevant, it is no longer enough. Having accelerators is useless without sufficient electricity to power, cool, and connect them.
The International Energy Agency estimates that global data center electricity consumption could double and reach around 945 TWh by 2030—slightly higher than Japan’s current electricity use. AI will be a key driver of this growth, with accelerated server farms expanding at a much faster rate than conventional IT infrastructure. In the U.S., the IEA projects data centers will account for nearly half of the demand increase through the end of the decade.
Latest data points in the same direction. In April 2026, the IEA reported that data center electricity demand grew 17% in 2025, with AI-specific data centers growing even faster—significantly above the overall electric growth of 3%.
In the U.S., the Energy Information Administration predicts new peak electricity demands in 2026 and 2027 driven by data centers, AI, cryptocurrencies, and electrification of other sectors. Total demand would rise from 4,195 billion kWh in 2025 to 4,248 billion in 2026 and 4,379 billion in 2027.
The clear takeaway is that AI has shifted from being just a chip and model industry to becoming an energy industry.
Building your own power generation isn’t so simple either
The seemingly logical solution is for major operators to build their own power near data centers. This approach, known in the industry as “behind the meter” generation, involves dedicated power situated behind the meter, directly connected to an industrial or digital facility. It can reduce reliance on the grid, accelerate deployment timelines, and give greater control over supply.
However, this route also confronts physical limits. Gas turbines, generating engines, transformers, high-voltage systems, industrial cooling equipment, and specialized labor don’t appear overnight. The market already shows clear signs of stress. An analysis published in Engineering indicated that wait times for large gas turbines have reached around five years on average, sometimes up to seven in certain cases, compared to typical cycles of one to three years in the past.
TechCrunch reported in April that the race for new gas plants for data centers was creating a shortage of turbines and significantly increasing costs. Wood Mackenzie estimated delivery times of six years for certain equipment, with no new orders available until 2028.
The case of xAI exemplifies this new phase. The company associated with Musk received approval to install 41 natural gas turbines in Mississippi, with an estimated capacity of 1.2 GW to power its Colossus data centers in the region. The decision was criticized by environmental groups due to the speed of approval and potential emission impacts, but it highlights how industries are willing to seek direct routes to secure power capacity.
The problem is that not all companies can do this. Buying GPUs is one thing; becoming an energy developer, negotiating permits, securing fuel, managing emissions, signing interconnection agreements, and operating critical energy assets is another. AI is pushing tech firms into a realm that resembles heavy industry more than software development.
China, the U.S., and the speed of construction
The comparison with China recurs because the AI race is also a race of industrial capacity. China has demonstrated notable ability in sectors like solar, batteries, power grids, high-speed rail, and electronics manufacturing to build rapidly and scale supply chains. Meanwhile, the U.S. retains leadership in chips, software, capital, and models but faces bottlenecks in permitting, grid infrastructure, equipment manufacturing, and territorial coordination.
The IEA projects that the U.S. and China will account for about 80% of the global growth in data center electricity demand through 2030. Both countries face the core dilemma: it’s not enough to have companies capable of training advanced models; they must also physically power the infrastructure that supports them.
The difference lies in administrative processes, planning capacity, and control over industrial supply chains. China isn’t free of energy, environmental, or grid issues, but its model allows faster execution of strategic projects through state coordination—hard to replicate in more decentralized democracies. The U.S., with its large private market, risks delays due to local permits, litigation, agencies, affected communities, environmental restrictions, and interconnection delays.
Europe views this debate within its own constraints. Data center demand is rising across the continent, with greater regulatory pressure on sustainability, water, land, and energy use. For regions like Madrid, Frankfurt, Dublin, Amsterdam, and Paris, the question isn’t just attracting investment but determining how much capacity the grid can support and under what conditions.
Software runs, infrastructure moves
The key lesson of this phase is that AI has again made physical economics relevant. For decades, much of the technological discourse focused on value moving toward the intangible: code, platforms, data, social networks, and cloud services. All true, but only until compute demand forces gigawatt-scale infrastructure.
Every advancement in models leaves a tangible footprint. More parameters require more training. More inference means more distributed servers. More servers demand more energy, cooling, fiber optic cabling, substations, transformers, land, and construction. The digital world doesn’t float in the air; it exists inside industrial facilities filled with equipment that consumes electricity continuously.
This also shifts power dynamics along the tech supply chain. Turbine manufacturers, transformer suppliers, electric utilities, grid operators, engineering firms, cooling system producers, and data center developers now play more strategic roles. The next competitive edge may not just be in acquiring more GPUs, but in gaining early access to reliable power, permits, and delivered electrical equipment.
Musk’s statement resonates because it touches a simple truth: AI cannot operate without energy. And large-scale energy production requires factories, permits, materials, workers, and years of planning. Software can iterate almost instantly; a power plant cannot.
The myth that AI is purely digital is fading. Moving forward, the race won’t be limited to labs, code repositories, or training rooms. It will extend into industrial zones, transmission lines, ports, turbine plants, copper mines, transformer factories, and municipalities that approve or block projects.
AI has entered its concrete phase—more slow, expensive, and political than the previous era.
Frequently Asked Questions
What is AI’s principal physical bottleneck?
Energy. AI data centers require vast amounts of electricity, along with cooling, grid connection, transformers, turbines, and permits.
Why aren’t private power plants built next to all data centers?
Because equipment shortages, manufacturing lead times, permits, fuel, interconnections, and industrial capacity still pose major constraints at scale.
Has GPU scarcity become less important?
No. Chips remain critical, but electricity availability has become an equally or more important limiting factor for many AI projects.
Which countries have an advantage in this new phase?
Those combining technological capacity with speed in building energy infrastructure, power grids, data centers, and industrial supply chains. The U.S. and China account for much of the expected growth but face different challenges.