AI’s Next Bottleneck Is Not Intelligence. It Is Electricity.

I remain broadly optimistic about progress in artificial intelligence. The models are improving, adoption is expanding, and the economic potential is still enormous. But recently my concern has shifted. The most serious near-term constraint on AI may not be whether models can become smarter. It may be whether we can supply enough electricity to run them.

The irony is hard to miss. For years, the AI debate focused on algorithms, chips, talent, regulation, and safety. Yet the next major bottleneck may be much more physical: power plants, transformers, transmission lines, substations, permits, and grid connections.

In 2024, Leopold Aschenbrenner’s widely discussed AI forecast imagined a rapid escalation in compute infrastructure: gigawatt-scale clusters by the middle of the decade, 10-gigawatt clusters later in the decade, and eventually 100-gigawatt clusters by around 2030. A 100 GW load, if run continuously, would consume 876 TWh per year—an amount approaching twice Germany’s annual electricity generation. This is not a software deployment problem. It is national-scale industrial infrastructure. (Dwarkesh)

The International Energy Agency’s latest analysis makes the point in less dramatic but still striking terms. It projects global data-centre electricity consumption to roughly double by 2030, reaching around 945 TWh, with demand growing about 15% annually from 2024 to 2030—more than four times faster than overall electricity demand growth from other sectors. (IEA)

The problem is not only generating the electricity. It is delivering it to the right place at the right time. Data centres are concentrated, power-hungry loads. They do not behave like normal incremental demand. One large AI campus can require power on the scale of a city, and it needs that power reliably, continuously, and quickly.

This is where the grid becomes the real constraint. Morgan Stanley has warned that data-centre developers expect power constraints by 2027–2028 because of underinvestment in grids and possible supply-chain disruption. The same report notes that “bring your own power” models—gas, microgrids, batteries, nuclear, and hybrid systems—are gaining momentum because waiting for the grid is becoming too slow. (Morgan Stanley)

The slowdown is already visible. Axios reported in February 2026 that as many as half of the world’s data-centre projects expected to come online this year could face delays. The issue is not lack of ambition: investment is still surging. The issue is that much of the announced capacity is not yet actually under construction, while power constraints, equipment shortages, and community opposition are becoming more serious. (Axios)

The U.S. grid illustrates the scale of the challenge. Berkeley Lab reported that, by the end of 2025, more than 2,060 GW of generation and storage capacity was actively seeking connection to the grid. The queue is enormous, but most projects never get built, and those that do are taking longer to reach commercial operation. For projects completed between 2018 and 2024, the median time from interconnection request to commercial operation had risen to more than four years. (Energy Markets & Planning)

A shortage of transformers adds another layer of friction. Reuters reported in May 2026 that U.S. transformer demand has surged, prices have risen sharply, and lead times for large units can stretch up to four years. That is a serious issue for AI, renewables, industrial projects, and grid modernization at the same time. (Reuters)

This is why the energy problem is not only an AI problem. It is also an energy-transition problem. If economies are to electrify transport, heating, industry, and digital infrastructure simultaneously, the grid must expand much faster than it has historically. The IEA has warned that grid infrastructure can become a bottleneck for clean energy transitions, noting that new grid infrastructure often takes five to fifteen years to plan, permit, and complete. It also estimates that grid investment needs to nearly double by 2030 to more than USD 600 billion per year. (IEA)

Elon Musk put the issue bluntly at Davos in January 2026, arguing that the limiting factor for AI deployment is electrical power, and that chip production may soon exceed the ability to power those chips—except, in his view, in China, where electricity expansion is moving faster. Whether one agrees with Musk’s broader conclusions or not, the basic diagnosis is increasingly shared across the industry: compute is no longer only about semiconductors. It is about energy systems. (The Singju Post)

The obvious response is to build power next to the data centre. That may work in some cases, but it is not simple. Dedicated gas plants, microgrids, nuclear projects, or large renewable-plus-storage systems require land, permits, equipment, fuel supply, interconnection decisions, financing, and operational expertise. They may reduce dependence on the grid, but they do not magically remove the infrastructure problem.

Nuclear power, especially small modular reactors, is often presented as the elegant solution. Politically, it has strong momentum in the United States: the Trump administration has explicitly linked advanced nuclear deployment to AI infrastructure and national security. (The White House) But the technology is not yet a near-term fix. The IEA says the first commercial SMR projects are expected around 2030, while NuScale’s flagship U.S. SMR project was cancelled in 2023 after costs rose and customer support weakened. (IEA)

This leaves the AI sector with a difficult but healthy pressure: efficiency must improve. If chips, model architectures, inference methods, cooling systems, and software optimization can reduce energy intensity significantly, the bottleneck eases. But that creates another risk. If efficiency improves faster than expected, some of the data-centre capacity now being financed may prove excessive. Projects that looked essential under today’s energy assumptions could become stranded or underutilized assets.

That is where the financial risk begins. The AI boom is not only a technology story; it is also an infrastructure and capital markets story. If investors begin to believe that power availability, grid delays, or efficiency gains will reduce the need for today’s massive buildout, the correction could spread beyond data centres into chips, utilities, equipment suppliers, cloud providers, and the wider technology sector.

The future of AI may still be bright. But brightness requires power. The next phase of AI progress will not be determined only by model intelligence, GPU supply, or venture capital enthusiasm. It will be determined by the speed at which societies can build and modernize the electrical infrastructure beneath the digital economy.

In simple terms: the future may be intelligent, but we still need somewhere to plug it in.

---

Originally published in the VerbionPulse newsletter on LinkedIn: AI’s Next Bottleneck Is Not Intelligence. It Is Electricity..