AI's electricity bill: the bottleneck is no longer chips — it's electrons

THESIS. The defining constraint of the AI era has changed. It is no longer silicon: it is electrons. To understand where this industry is heading, watch electricity auctions, grid-connection queues and 20-year nuclear contracts, not benchmarks. Our reading is twofold. In the short term there is a real, badly distributed problem: part of the cost of this build-out is already landing on the bills of households that never asked for it. In the long term, however, AI is becoming the largest buyer of clean, firm energy in history — and that can leave us with a better grid than the one we had. Neither euphoria nor panic: governance.

THE NUMBER THAT MATTERS. The world's data centers consumed about 485 terawatt-hours (TWh) in 2025 and will reach roughly 950 TWh by 2030, according to the International Energy Agency (IEA). One TWh, for scale, powers about 300,000 European homes for a year. Those 950 TWh roughly equal Japan's total electricity consumption today. The US and China account for nearly 80% of that growth, and the AI-specific share will more than quadruple this decade, again per the IEA. Now the context usually missing: even in 2030, data centers will be around 3% of global electricity. That is a lot, it is growing fast, and it is concentrated in a few places — which is the real problem — but it is not "AI drinking the planet dry." The impact is above all local: Virginia, Ireland, Aragon — not evenly spread across the globe.

WHO PAYS THE BILL. Here is the uncomfortable part the industry prefers not to tell. In PJM, America's largest grid (13 states, 65 million people), the capacity auction price — what power plants are paid to be available — jumped from $28.92 per megawatt-day in 2024/25 to $329.17 in 2026/27: more than tenfold in two years. The market's independent monitor attributes $9.3 billion of extra cost to data centers in the 2025/26 auction alone, and $29.4 billion across four auctions. A Fortune analysis puts what has already been passed on to the public at $23 billion. In Washington D.C., Pepco customers pay on average $21 more per month since June 2025, with the local consumer advocate estimating half of that comes from capacity-price spikes. Our reading: this is socializing costs while privatizing benefits, and it is politically unsustainable. The right answer is not to stop data centers, but special large-load tariffs that make them pay for the grid they demand — something several US states are now writing into regulation.

THE BOTTLENECK ISN'T CHIPS. The great paradox of 2026: committed GPUs abound, but there is nowhere to plug them in. Connecting a large project to the grid in markets like Northern Virginia, Phoenix or Dallas now means a 4-to-7-year queue, according to industry data. Large power transformers take 3 to 5 years to deliver. And gas turbines — the favorite shortcut — carry waiting lists of up to 8 years at GE Vernova, Siemens Energy and Mitsubishi, which supply two-thirds of the market. The result: of the ~12 gigawatts of US data-center capacity targeted for 2026, only about a third is actually under construction, per industry estimates. That is why tech companies are bypassing the grid and generating "behind the meter": their own gas plants, batteries and, increasingly, nuclear. All against a staggering backdrop: the big hyperscalers (Amazon, Google, Microsoft, Meta) collectively plan over $600 billion in capital spending for 2026, up 77% from 2025, with Goldman Sachs projecting $5.3 trillion cumulative through 2030. Electricity, not chips, will decide who gets to spend it.

THE NUCLEAR REVIVAL AND THE PRICE OF WATER. The deepest consequence of this race is that AI has revived Western nuclear power. Microsoft will buy the entire output of Three Mile Island (835 MW) for 20 years when it restarts in 2027. Amazon expanded its nuclear contract with Talen Energy to 1,920 MW through 2042 and led a $700 million investment in X-energy's small modular reactors (SMRs — compact, factory-built reactors). Google signed with Kairos Power for 500 MW of SMRs by 2030. And in January, Meta announced deals for up to 6.6 GW of nuclear power by 2035, including over 2,600 MW with Vistra. In total, close to 10 GW of nuclear committed by big tech: nothing has pulled demand for clean, firm energy like this in decades. The other flank is cooling water: US data centers now consume around one trillion liters a year directly, according to academic and industry estimates, and in water-stressed regions that matters. But engineering is moving fast here: Amazon reports an efficiency of 0.12 liters per kWh, seven times better than the industry average of 0.84, and Microsoft is deploying closed-loop designs that eliminate evaporation. Water is a siting-and-design problem, not destiny.

SPAIN, ON THE MAP (AND AT THE LIMIT). This is no longer just about Virginia. Spain ranks sixth worldwide in data-center investment, with roughly €90 billion in announced projects according to DatacenterDynamics. Amazon has raised its bet on Aragon to €33.7 billion, where official scenarios project regional electricity consumption multiplying 4 to 12 times within a decade. Spain's advantage is real: cheap sun, wind and land, with supply already around 95% renewable. So is its limit: about 85% of transmission-grid nodes can no longer accept new high-voltage demand, per the industry. The lesson of the 2024/25 Iberian blackout still stands: without investment in grids and firm backup, cheap renewables alone cannot capture this wave.

OUR READING. The short term demands honesty: the bill is real, it is unfairly split, and public trust is eroding for good reason. If tech companies do not visibly pay for the infrastructure they demand — large-load tariffs, self-supplied power, transparency on water and emissions — the political backlash could halt the build-out cold, and its benefits with it. But the long term tells another story. For the first time there is a buyer with near-bottomless pockets willing to prepay for nuclear, geothermal, storage and grids. Data-center emissions will remain a small fraction of the global total even in 2030, per the IEA, and AI itself is already optimizing grids, materials and drug discovery: the instrument that consumes energy is also the one that can help make it cheaper and cleaner. If this transition is governed well — those who pollute and those who congest the grid pay for it — in 15 years this race will be remembered not for raising power bills, but for financing the abundant, clean grid that electrifying everything else needed anyway. That is the prize. And it is being decided now, in tariffs and interconnection queues, not in benchmarks.