AI's Power Problem: Why Electricity, Not Chips, Is the Real Bottleneck - Part 2

Every AI cluster starts life as a request for megawatts, not GPUs. Chip supply is no longer what decides whether a data centre gets built — power does, and in most of the world's biggest AI markets, power simply isn't there yet.

And once that power arrives, nearly all of it leaves again as heat. This chapter follows AI's physical footprint through its first two hard constraints: getting electricity to the rack, and getting rid of what the rack turns that electricity into.

Behind-the-meter gas turbines are becoming standard equipment at AI campuses that can't wait years for a grid connection. Source credit to be added on upload.

Electricity Becomes the Bottleneck

For the last two years, the AI industry's binding constraint was chips. That's no longer true. Across the biggest data-centre markets — Northern Virginia, Dublin, Phoenix, Dallas, Singapore, Amsterdam — new high-capacity grid connections now face waits of four to seven years, and the US interconnection queue alone has swelled past 1,500 gigawatts of proposed projects.

Hyperscalers have responded by going around the queue rather than waiting in it. As of 2026, US developers have announced roughly 101 gigawatts of behind-the-meter natural gas generation — power plants built on-site, dedicated to a single campus, with no dependence on the public grid at all. A dedicated gas plant can be commissioned in twelve to eighteen months, versus five to seven years for a utility connection, and for a company burning through capital on an AI build, that gap in timeline outweighs almost any other consideration.

The scale of what's queued up is striking on its own. Natural gas capacity planned for interconnection between 2026 and 2030 has grown from roughly 24 gigawatts at the start of 2025 to about 64 gigawatts today, with AI-driven gas demand alone projected to reach 6 billion cubic feet a day by 2030. Nearly 7 gigawatts of already-planned US data-centre capacity has already been delayed or cancelled this year for one reason: the power wasn't there when it was needed.

↑ Back to Table of Contents

Where the Power Actually Comes From

"Behind-the-meter" power has to come from somewhere, and mostly that somewhere is gas — dispatchable, quick to build, and able to ramp up and down to match the sharp, spiky load profile of AI training runs in a way that solar and wind alone can't, unless paired with a lot of battery storage most sites don't yet have.

That creates a genuine tension for operators who've made public net-zero commitments. Investors underwriting those pledges are increasingly reluctant to count uncaptured gas generation as compatible with them, and enterprise or sovereign clients are starting to write emissions language directly into hosting contracts. Nuclear, including a new generation of small modular reactors, is being signed at scale by the largest hyperscalers, but those projects run on a construction timeline measured in years, not months — too slow to solve the immediate power crunch even as they reshape the picture for the early 2030s.

All of which sharpens the question this series keeps returning to: does building generation capacity for AI add genuinely new clean electricity to the grid, or does it simply divert electricity, and public goodwill, that would otherwise have gone to households and other industries? A gas turbine built specifically to bypass the interconnection queue isn't green accounting — it's a new, direct emissions source, and it should be counted as one.

↑ Back to Table of Contents

Nearly All of It Becomes Heat

Here's the physics that makes this a two-part problem, not one. With very few exceptions, essentially all the electricity a data centre consumes for computation is eventually converted into heat. A single megawatt of IT load produces on the order of 8,760 megawatt-hours of thermal energy a year at full utilisation — the entire electricity bill, reappearing as something that then has to be removed from the building.

The chips driving AI workloads are making this harder every generation. Liquid-cooled GPU racks running at 120 to 150 kilowatts are now common in new AI campuses, densities that air cooling simply cannot handle. The upside buried inside that density increase is that modern liquid cooling systems reject heat at 55°C to 65°C — hot enough to be genuinely useful elsewhere, rather than low-grade warmth with nowhere to go.

↑ Back to Table of Contents

Turning a Liability Into a Resource

This is where the story stops being purely about consumption. Waste-heat recovery is moving from pilot project to standard infrastructure, especially in Europe, and the economics are already competitive: recovered data-centre heat is landing at roughly €12–30 per megawatt-hour in 2026 benchmarks, against €35–55 for a conventional gas boiler, with recovery infrastructure costing a fraction of what a new combined-heat-and-power plant would.

The clearest working examples are in the Nordics. A major hyperscaler's heat-recovery partnership in Finland now delivers up to 350 megawatts of combined thermal power, covering roughly 40% of district heating demand for a metropolitan area of 250,000 people. In Greater Copenhagen, a long-running data centre now supplies heat, free of charge, to 6,000 local homes — about 80% of that district's heating demand.

Policy is starting to catch up to the opportunity. Germany's Energy Efficiency Act now mandates a minimum heat-reuse rate for larger new facilities, rising from 10% in 2026 to 20% by 2028, and the EU's recast Energy Efficiency Directive requires any facility above 1 megawatt to formally assess waste-heat recovery before it's ever allowed to switch on. Where the infrastructure exists to use it, AI's waste heat is starting to look less like an externality and more like a second product line.

↑ Back to Table of Contents

What Power and Heat Add to the Ledger

Two things are true at once here, and the tension between them is the point. Electricity is now the single hardest physical constraint on how fast AI can be built — hard enough that companies are willing to build their own gas plants rather than wait for a cleaner grid connection. And the heat that same electricity turns into is not automatically waste; in the right location, with the right policy behind it, it becomes district heating, greenhouse warmth, or industrial process energy instead.

What determines which outcome you get isn't the technology — liquid cooling and heat-recovery equipment already exist and work. It's geography, grid maturity, and whether the regulatory environment treats waste heat as a compliance requirement or an afterthought. That same geography-first pattern is about to matter even more in Part 3, where the resource in question isn't power that can be piped from a neighbouring plant — it's water and minerals that can't simply be trucked in from wherever they're abundant.

↑ Back to Table of Contents

K M Manohar’s Insight

AI’s power problem is not simply an energy problem.

It is a location problem, a timing problem, and ultimately a systems-design problem.

Building more data centres without building the right power infrastructure first will only move the bottleneck from chips to electricity.

But there is another side to the ledger.

The electricity consumed by AI does not disappear — it becomes heat.

That means the same infrastructure creating a sustainability challenge can also become a source of district heating, industrial energy and greenhouse warmth.

The real breakthrough, therefore, is not merely making AI consume less.

It is learning how to connect power, cooling, heat, geography and policy into one system.

For India, this distinction could become strategically important: where future AI campuses are built may matter almost as much as how powerful their GPUs become.

My view: The smartest AI data centre of the future will not simply be the one with the fastest compute. It will be the one that turns every unit of energy into the maximum possible economic and social value.