The AI Energy Bottleneck: Power, Not Chips, Is the Next Constraint on AI Growth

For the past three years, the AI industry has been obsessed with one constraint: chips.

Nvidia's H100. TSMC's fabrication capacity. Export controls. The narrative was simple — more chips, more compute, more AI. Whoever had the most silicon won.

That era is over.

Microsoft CEO Satya Nadella put it plainly in early 2026: the "chip shortage" has officially ended, replaced by something far more physical and daunting — an "Energy Wall." The primary bottlenecks for AI scaling are no longer the availability of high-end silicon, but the skyrocketing cost of electricity and the lack of advanced cooling infrastructure.

The industry has spent billions securing GPUs. It has not spent nearly enough securing the power to run them. And that gap is about to become the single biggest constraint on AI growth.

The numbers that should terrify every AI executive

Global data center electricity consumption is projected to reach 565 terawatt-hours (TWh) in 2026 — a 26% increase from 447 TWh in 2025. By 2030, Gartner estimates that figure will exceed 1,200 TWh, roughly equivalent to the entire annual power consumption of Japan.

The International Energy Agency projects data center electricity consumption will more than double from 485 TWh in 2025 to 950 TWh in 2030, accounting for around 3% of global electricity demand. The IEA's base case sees global electricity generation to supply data centers growing from 460 TWh in 2024 to over 1,000 TWh in 2030 and 1,300 TWh in 2035.

The driver is AI-optimized servers. Gartner estimates these advanced servers will account for 31% of total data center power consumption in 2026, up from approximately 20% in 2025. Conventional server electricity consumption grew by less than 1% in 2025. AI-optimized servers recorded an 83% consumption increase in 2025 and are forecast to expand by a further 84% in 2026 to reach 175 TWh globally. By 2027, AI hardware will consume more power than all conventional servers combined.

One statistic captures the scale: a single hyperscale AI data center can consume as much electricity as a city of 1 million people. Modern AI workloads now demand 50 to 100 kW per rack, with specialized training configurations pushing beyond 120 kW.

The infrastructure that isn't there

The problem is not just how much power AI needs. It's how fast it needs it — and how slowly the grid can respond.

A modern data center can be permitted and built in 2-3 years. The power to run it can take 5-7 years for natural gas, 10+ years for nuclear, and 2-4 years for solar — while grid interconnection queues stretch beyond five years in many U.S. regions.

Gartner predicts that by 2026, power shortages will delay more than 30% of all data center expansions. By 2028, data centers will account for 10% of all U.S. electricity demand.

The strain is already visible. Half of the 12 GW of U.S. data center capacity planned for 2026 is being delayed or cancelled. Nationwide interconnection requests now total 1.84 terawatts, exceeding total installed U.S. generating capacity.

Goldman Sachs projects U.S. data center power demand will climb from 31 GW in 2025 to 41 GW in 2026 and 66 GW in 2027 — more than doubling in two years. By 2027, data centers would consume 8.5% of total U.S. peak summer power demand, up from 4.1% in 2025. Only 50-60% of capacity scheduled for future years is expected to come online on time.

BlackRock estimates approximately 148 GW of additional power capacity will be needed by the end of the decade to satisfy data center demand — multiples above the roughly 42 GW consumed in 2025. By 2030, planned AI data center capacity is expanding to 109 GW, pushing the projected U.S. data center power deficit to 28 GW.

Google reported that its electricity demand climbed 37% in 2025 — the largest annual increase in the company's history. Since 2019, total electricity demand has grown more than 250%. In its 2026 Environmental Report, Google wrote that its AI infrastructure buildout is "currently accelerating faster than the grid is decarbonizing."

The U.S. problem: speed versus physics

The U.S. is building data centers faster than it can connect them to the grid.

The Department of Energy projects data centers could draw 6.7% to 12% of all U.S. electricity by 2028. The power required would demand building roughly 8,000 km of high-voltage transmission lines annually — around ten times the current pace.

Supply-chain shortages have pushed lead times for critical grid equipment to several years. In early 2026, the DOE invoked emergency powers to shift data centers onto backup generation during peak demand periods. On May 18, 2026, the DOE issued an emergency order authorizing PJM, the largest U.S. regional grid operator, to commandeer backup generators from data centers and large industrial users during extreme emergencies to avoid forced residential blackouts.

The cost is already landing on ratepayers. U.S. residential customers are paying an extra $1.4 billion per year on their electricity bills as a direct result of data center demand. Investor-owned utilities filed $18 billion in rate-increase requests in 2025, the highest since the mid-1980s. States hosting the highest concentration of data centers have seen their historical electricity cost advantage narrow from 5% to 3.7% since 2020, a gap that is set to close further.

China's problem: green ambition meets hard reality

China faces the same bottleneck — but with different constraints.

China's data center power demand is projected to rise by 300 to 500 billion kilowatt-hours between 2026 and 2030, accounting for close to a fifth of the country's total electricity demand growth. State Power Investment Corp. director Pei Shanpeng confirmed these figures, noting data centers will account for 18% of total electricity demand growth over the period.

In 2025, China's national computing centers consumed 196 billion kWh, up 18.1% year-over-year. In Q1 2026 alone, internet data services electricity consumption jumped 44% year-over-year. By 2030, computing center electricity consumption is projected to exceed 500 billion kWh, accounting for over 5% of total social electricity consumption.

The tension is between China's climate goals and its computing ambitions. Beijing wants renewables to supply roughly four-fifths of AI data center power consumption by 2030 — up from around 11% in 2023. But grid operators are wary. A wind or solar farm's output is intermittent. An AI data center demands power 24/7, with no tolerance for interruption.

When the choice is between a clean megawatt that might not be there and a reliable one that will, operators choose reliability — which in China still means a grid with a large coal baseload. The explosive growth of AI will likely accelerate coal power generation and slow the decline of coal's share in China's energy mix. A LEAP model simulation found that AI's exponential compute growth will drive rapid power demand increases, making data centers the third-largest power-consuming sector after industry and residential.

Chinese grid operators have reportedly resisted plans to boost renewables specifically to power AI, citing concerns that intermittent supply cannot meet the relentless, predictable demand of AI training clusters. The country's green-data-center action plan requires new projects in national computing hubs to source most of their electricity from clean sources — but the physical grid hasn't caught up to the policy.

The physical ceiling

The AI industry has spent years assuming that compute scales infinitely. It does not. It scales until it hits the power grid.

Nadella's warning about the "Energy Wall" is not rhetorical. A single training run for a thousand-billion-parameter model now consumes over 1 million kWh — equivalent to a town of 100,000 people's entire daily electricity consumption. AI compute doubles every 18 months. Global power generation capacity grows at roughly 5% per year. The curves are diverging.

Goldman Sachs estimates companies could spend close to $7 trillion building data centers worldwide by 2030. But that spending assumes the power will be there. The data suggests it won't — not fast enough, not reliably enough, and not cleanly enough.

The industry has been fighting a chip war. The next war is for electrons.

What this means for you: If you are investing in or building AI infrastructure, the timeline is no longer driven by GPU availability. It is driven by grid interconnection queues, transformer lead times, and utility rate cases. Companies that secure power capacity early will hold a structural advantage that no amount of chip allocation can overcome. The energy wall is coming. The only question is whether you are on the right side of it.