NextFin News - The world's climate diplomats gathered in New York this week for the annual UN Climate Summit, and the most contentious item on the agenda was not coal, oil, or a national emissions target. It was the electricity bill for artificial intelligence. As the UN Climate Summit convened at headquarters on 23 September alongside Climate Week NYC, the breakneck growth of AI data centres collided with rising power prices to produce a new and uncomfortable question for the energy transition: who pays when a single industry's power demand doubles within five years?
The answer being forged in New York is reshaping climate diplomacy itself. Turkey's incoming COP31 presidency used the sidelines of the UN General Assembly to launch a voluntary pledge demanding that AI companies disclose their energy use and power their operations with clean energy. Cities led by C40 pushed a Global Urban Data Centre Pact after polling showed nearly eight in ten Americans are skeptical or critical of data centres. And the UN Secretary-General's response was not another emissions target but the Global Grids Accelerator, a new facility aimed at unlocking electricity infrastructure investment in Africa and South-East Asia. The transition, in short, has moved from generation to connectivity — and from pledges to power-purchase agreements.
The Numbers Behind the Pressure
The scale of the demand shock is what makes this different from previous industrial load growth. The International Energy Agency projects global data centre electricity consumption will more than double to around 945 terawatt-hours by 2030 — slightly more than Japan's total electricity consumption today — with AI the single largest driver. By the end of the decade, data centres in the United States alone are set to consume more electricity than the country's aluminium, steel, cement, chemicals and all other energy-intensive industries combined. In the US, data centres account for nearly half of electricity demand growth between now and 2030.
The domestic picture is already visible in official data. The US Energy Information Administration expects American electricity use to grow 1% in 2026 and 3% in 2027 — the strongest four-year growth stretch since 2000 and the first time demand has risen for four consecutive years since 2007. The EIA is explicit about the cause: "increasing demand from large computing facilities, including data centers." Administrator Tristan Abbey tied the two energy systems together, noting that natural gas output is expected to reach nearly 109 billion cubic feet per day this year and that "natural gas supply is critical as we forecast that US liquefied natural gas exports expand and electricity demand rises through 2027."
Prices are moving with demand. Total average US retail electricity revenue reached 13.83 cents per kilowatt-hour in May 2026, up 5.3% from a year earlier, with residential rates at 18.44 cents, up 6.2%. Natural gas, the marginal fuel across much of the US power system, has been climbing from multi-year lows: the Henry Hub spot price was $2.97 per million British thermal units on 15 September, up from $2.71 on 11 September, though still below the EIA's full-year 2026 average forecast of $3.80.
The political temperature is rising alongside the physical one. In January, Microsoft pledged that its data centres would not drive up local electricity rates — the first of a series of commitments that culminated in March, when Amazon, Google, Meta, Microsoft, OpenAI, Oracle and xAI signed a White House Ratepayer Protection Pledge agreeing to build, bring or buy their own generation and cover the cost of power-delivery upgrades so expenses are not passed to households. The administration also pressed the PJM Interconnection, the nation's largest grid, to hold an emergency capacity auction after data centres were attributed to $23 billion in capacity costs that watchdogs called a wealth transfer to consumers.
Why This Is Structural, Not Cyclical
It is tempting to read the current squeeze as a cyclical spike — high gas prices, a hot summer, tight grids — that will fade once supply catches up. That reading is wrong on the demand side. Three features mark this as a structural regime shift rather than a mean-reverting cycle.
First, the demand floor keeps rising with no plateau in sight. The IEA's base case has data centre consumption reaching 1,200 TWh by 2035, and its uncertainty range spans 700 to 1,700 TWh. A cyclical load would show elasticity to price; AI compute demand is tied to model training and inference volumes that grow independently of the power price. The UN's economic commission for Europe warned in New York that data-intensive technologies are growing faster than electricity infrastructure can handle, raising concerns about voltage oscillations, unintended disconnections and cascading failures in systems that depend on renewable energy. Because AI workloads are volatile and can spike unexpectedly, grids built around variable renewables cannot adjust in real time — a technical mismatch, not a temporary shortage.
Second, the infrastructure response is mismatched by an order of magnitude. Investment in data centre infrastructure worldwide is forecast to roughly double only by 2050 — from about $800 billion a year in 2026 to $1.8 trillion in 2050 — while consumption doubles before 2030. On the grid side, the International Renewable Energy Agency puts annual investment at roughly $0.5 trillion in 2025 and says it must reach about $1 trillion a year between 2026 and 2035. More than 2,500 gigawatts of wind, solar and storage projects — equivalent to nearly twice the entire generating capacity of the United States — are currently stalled in grid-connection queues worldwide. The IEA estimates that grid-enhancing technologies and regulatory fixes could free enough hosting capacity to connect between 1,200 and 1,600 GW of advanced-stage projects, which means the queue is partly a policy failure rather than a pure build problem.
Third, the policy response is structural, not temporary. Moratorium legislation on data centres has been introduced in at least 15 states, and more than 100 local governments have enacted outright bans or restrictions that are blocking or delaying an estimated $64 billion of projects. New York became the first state to impose a statewide moratorium on hyperscale data centres; Maine's legislature passed its own before the governor vetoed it; Virginia lawmakers from both parties are now calling for one. These are not short-term grid fixes. They are a renegotiation of where and how a whole industry is allowed to build.
The cyclical overlay is real but secondary. Natural gas prices will mean-revert — the Henry Hub has already traded back toward $3 from winter storm spikes — and electricity price spikes tied to weather and fuel costs will ease. But they are now oscillating around a permanently higher demand base. The right framing is a structural demand shift with a cyclical price wave riding on top of it.
The Second-Order Consequence: Big Tech Becomes an Energy Company
The first-order effect of AI power demand is obvious: more electricity, higher bills. The second-order effect is what the New York talks revealed — the technology sector is being pulled into energy markets it never planned to enter, and that is changing both its cost structure and its political exposure.
Hyperscalers have responded to grid constraints by signing up for dedicated, always-on generation. Microsoft agreed to a 20-year power-purchase agreement with Constellation Energy to buy the output of the restarted Three Mile Island Unit 1 in Pennsylvania — 835 megawatts of nuclear baseload that Constellation plans to bring back online in 2028, with the company investing $1.6 billion to revive the reactor. Google signed a 25-year PPA with NextEra Energy to support restarting Iowa's 615 MW Duane Arnold plant, with power expected by early 2029. Amazon has pursued co-located generation at the Susquehanna site in Pennsylvania. On the fossil side, Babcock & Wilcox received notice in March 2026 to proceed on a $2.4 billion design-build contract to deliver 1.2 gigawatts of natural gas-fired generation for Applied Digital's AI campuses — four 300 MW boiler and steam-turbine units serving a single industry's data centres. And in the largest single fossil commitment, Meta agreed with Entergy to build and finance 10 gas-fired power plants in Louisiana totalling 7.5 gigawatts — more than a 30% increase to the state's entire grid capacity — to serve its Hyperion AI campus.
This has a direct financial logic, and a direct risk. Fuel represents about half the cost of electricity from a large power plant, so a doubling or tripling of natural gas prices could make "bring your own power" data centres materially more expensive to run. The exposure runs both ways: if hyperscalers instead lean on the shared grid, they push wholesale and retail prices higher for everyone else. Either route ties the economics of AI tokens and cloud contracts to commodity markets that these companies have never had to hedge at scale.
"We must openly discuss AI's growing energy consumption and it is time for governments to start setting the terms," said Murat Kurum, Turkey's environment minister and COP31 president-designate, launching the Antalya AI pledge in New York. "We expect companies to be transparent about their energy use and to power their operations with clean energy."
The political logic is equally direct. Household electricity bills are rising faster than overall inflation, and they have become a live political issue ahead of the US midterm elections. When energy costs become a voter issue, the AI industry's social licence to build becomes conditional — which is precisely why the industry showed up in New York with transparency pledges rather than waiting to be regulated. Cities feel this first: C40 polling found that 79% of Americans are skeptical or critical of data centres, yet if clean energy is integrated into their design, net support swings from -33% to +16%. That 49-point gap is the entire political story in one number.
The Counter-Thesis: AI Could Cut More Than It Consumes
The strongest argument against treating AI power demand as a climate setback is that AI itself may be a net decarbonisation tool. Proponents note that widespread application of the technology could drive large efficiency gains across the economy and optimise renewable-energy integration. The IEA estimates that existing AI applications could cut more emissions than data centres add. On this view, the New York hand-wringing mistakes a temporary infrastructure bottleneck for a permanent climate problem; once grids are built and models become more efficient, the net effect is positive.
This argument deserves weight, but it rests on two assumptions that the New York evidence does not yet support. First, the efficiency gains are economy-wide and diffuse, while the power demand is concentrated, localised, and immediate. A gas plant approved in Louisiana this year does not wait for a future software optimisation somewhere else in the economy. Second, the "net-benefit" calculation requires transparent energy-use data that the industry does not currently disclose, which is exactly the gap the Antalya pledge is trying to close. Until disclosure is real, the optimisation argument is an assertion, not a measured outcome.
The falsifying signal is specific. If US data centre electricity consumption exceeds roughly 1,000 TWh by 2030 — above the IEA base case — while AI-enabled efficiency gains remain unmeasured, the net-benefit thesis fails. A second check is the fuel mix: the IEA's Lift-Off Case already has nearly half of the additional electricity generated for data centres between 2024 and 2030 coming from fossil fuels. If gas-fired capacity added specifically for data centres exceeds 20 gigawatts per year through 2028, the clean-power alignment pledge is not holding.
What Comes Next
The New York talks set up three watch points for the rest of 2026 and into COP31 in Antalya this November.
Short term (months): natural gas prices and the quarterly EIA Short-Term Energy Outlook revisions to the 2027 data centre load forecast. A sustained move above $4.50 per MMBtu at Henry Hub would stress the economics of gas-backed "bring your own power" campuses and could force a rethink of projects like Meta's Louisiana build-out. An upward revision of more than 5% to the 2027 load forecast would confirm the structural pressure is intensifying, not easing.
Medium term (through COP31 in November): whether the Antalya AI pledge converts from a New York announcement into signed commitments with binding transparency requirements. The pledge sits alongside Turkey's broader 35x35 initiative to lift electricity's share of final energy consumption to 35% by 2035, from about 20% today — a target that only works if the new demand is clean. The Global Grids Accelerator, launched in New York for Africa and South-East Asia, will be judged on whether it converts grid priorities into financed, operational projects rather than another announcement layer.
Long term (to 2030 and beyond): the build-out of transmission. The beneficiaries of the current regime are clear: generators with dispatchable capacity, grid-equipment suppliers, and jurisdictions with permitting speed. The exposed are households facing rising bills, utilities caught between reliability mandates and decarbonisation goals, and tech companies whose sustainability pledges now depend on power markets they are only beginning to understand.
The central judgment from New York is plain: the energy transition is no longer a story about replacing fossil-fuel generation with renewables. It is a story about whether the grid can be built fast enough to carry both the climate agenda and the AI boom — and who is billed when the two collide.
Explore more exclusive insights at nextfin.ai.
