The building isn’t the first thing that catches your eye when you pass one of the new AI campuses being built in rural Mississippi. The substations are to blame. They are arranged in rows. Fenced off and humming under floodlights, transformers wait for turbines that haven’t even arrived yet.
In a way, the easy part is the chips. They are transported in crates. Pipelines, river water, uranium pellets, and steel towers spanning half a continent must be used to generate the electricity. And more and more, no one can rush that part.
| Topic Brief: AI’s Energy Bottleneck (2026) | Details |
|---|---|
| Subject | The shift in AI’s primary constraint from chip supply to power supply |
| Estimated global data-centre electricity use (2024) | Around 415 TWh |
| Projected global data volume by year-end 2025 | 213.56 zettabytes, growing at 25.4% CAGR |
| Major US private power moves | xAI’s ~1.9 GW Mississippi gas-turbine project; Microsoft’s restart of Three Mile Island in 2028; Meta’s 6+ GW nuclear contracts |
| China’s grid investment (2026–2030) | Roughly CNY 5 trillion across State Grid and Southern Power Grid, a 40% jump over the prior cycle |
| Ageing US grid statistic | Around 70% of transmission lines and large transformers are over 25 years old |
| China’s 2030 west-to-east transmission goal | Over 420 GW, up from ~340 GW in 2025 |
| Key emerging concept | “Energy islands” — private, purpose-built power adjacent to AI clusters |
Silicon has dominated discussions about AI for the majority of the past two years: who produces the best GPU, who is hoarding Nvidia allocations, and whether export controls will be a problem. That conversation has been subtly moved, but it hasn’t vanished. The bottleneck has moved, according to anyone who is actually developing these systems. It’s no longer chips. They are powered by electricity, and the wires that transport it are becoming more and more important.
The explanation is hidden in the way contemporary AI functions. Quick bursts of inference were performed by earlier chatbot-style models, which would respond to a question, pause, and then repeat. The so-called agentic systems, the younger generation, exhibit distinct behaviors. They spend minutes or hours reasoning, dissecting issues, reviewing their own work, and utilizing outside resources.
Because of this, GPUs remain pinned at full draw for much longer than data centers were intended to handle. Global data-center electricity consumption was estimated by the International Energy Agency to be around 415 TWh in 2024 and to double by 2030. However, when compared to the surge in agentic workloads, this doubling appears to be closer to 2027.
The United States, where the grid was primarily constructed during Eisenhower’s presidency, is where the strain is most evident. Most circuit breakers have been in use for at least thirty years, and about seventy percent of transmission lines and large transformers are older than twenty-five. Backlogs of several years have resulted from interconnection queues. Lead times for high-voltage equipment are expressed in years rather than months. Additionally, utilities are hesitant to invest in long-term projects that may become outdated by the next election due to regulatory whiplash, including the planned reversal of the 2009 endangerment finding.
The hyperscalers ceased to wait as a result. In Mississippi, xAI is assembling nearly two gigawatts of private gas turbines. In addition to reviving Three Mile Island, Microsoft is concurrently developing more effective chips like the Maia 200. Google has inked a 1.2 GW contract to provide carbon-free energy around-the-clock. More than six gigawatts of future nuclear capacity have been secured by Meta. What is emerging resembles self-contained energy islands—power plants with computers attached, rather than the other way around—rather than data centers linked to the grid. All of this is a kind of silent admission that the grid as it is now will not survive.
Naturally, China is acting in the opposite way. Beijing is stepping up its use of the public grid rather than avoiding it. In order to expand the ultra-high voltage network that transports renewable energy from the windy, sunny west to the data-hungry east, State Grid and Southern Power Grid have committed approximately five trillion yuan over the next five-year plan, a 40% increase over the previous cycle. The “Eastern Data, Western Compute” approach locates new facilities in inland, colder provinces where cheaper electricity and liquid cooling extend each megawatt. Ironically, it’s a centralized, slower-feeling strategy that might ultimately be faster overall.
The philosophical differences between the two are difficult to ignore. Americans are encircling their computers with moats. Beijing is constructing roads. Time horizons and whether the next US administration views grid investment as infrastructure or as ideology will likely determine whether the road strategy or the moat strategy prevails, rather than engineering.
However, there is a counterargument that is worth considering. The computation per watt is constantly getting better. Ten years ago, Germany’s weather service spent more than $20 million on two Cray supercomputers that filled rooms and used two megawatts to produce roughly a petaflop. At less than $5 million, a single Nvidia Blackwell rack can now surpass that performance at 120 kilowatts. In less than two years, Google reportedly increased its monthly token throughput from ten trillion to over thirteen hundred trillion, a hundredfold increase, without experiencing even a hundredfold increase in electricity consumption. Hardware continues to improve in efficiency, sometimes more quickly than demand.
Efficiency, however, only purchases time. Whether the wires, turbines, and cooling towers can be constructed quickly enough to feed an industry that doubles its appetite every few years is the deeper question that no one on either side of the Pacific has fully addressed. The chips will continue to come in. They do it every time. Who actually gets to use them will depend on power, whether it be political, physical, or slow.
