The grid has a problem. Not a future problem — a right-now problem. In the first four months of 2026, the United States poured $49.5 billion into data center construction alone. That is nearly four times the pace of a year ago. Global investment in data centers is on track to cross $1 trillion this year. The grid infrastructure those buildings will eventually plug into? Much of it was designed in the 1970s and 1980s, before a single teraflop was ever computed on commercial silicon. Something has to give. Right now, the grid is giving first.
A Power Surge Unlike Anything in Grid History
US data center power demand stood at 31 gigawatts in 2025. By 2027 — just two years from now — that figure is projected to reach 66 GW. To put that in perspective: 35 GW is roughly the total generating capacity of Spain. We are adding, in two years, an entire Spain worth of power consumption from one sector, in one country. Anthropic has estimated that training a single frontier AI model will require 5 GW of dedicated power by 2027 — more than most mid-sized European nations consume in total.
I spent years at CERN, where the LHC draws around 200 MW during full-energy run periods. You develop a visceral respect for what serious power density looks like. Every substation, every transformer, every cable run is accounted for. But even the LHC is a rounding error compared to what the hyperscalers are now demanding at grid scale. This is a genuinely new category of infrastructure stress — and we are not remotely prepared for it.
Source: Goldman Sachs, EPRI, IEA (2025-2026 reports)
Why the Grid Cannot Simply Scale Up
The data center power demand crisis is not a political problem, though politicians are certainly attempting to make it one. It is a fundamental infrastructure timing mismatch. Transmission lines take 10 to 15 years to permit and build. High-voltage transformers — the critical link between generation and consumption — now carry lead times of 2 to 3 years due to global supply chain constraints. A data center campus can be designed, built, and ready to energise in 18 to 24 months. The maths does not work.
The US Energy Information Administration projects annual electricity demand will grow roughly 2% per year through 2027, driven overwhelmingly by AI infrastructure. That sounds modest until you realise that 2% of a very large number, concentrated in a handful of counties, creates acute local stress. Northern Virginia — which hosts more data center capacity than any other region on Earth — is already straining its regional grid. Texas, Louisiana, and Illinois are watching the same pattern emerge in real time.
Some operators have stopped waiting for the public grid to catch up. The recently proposed DATA Act would allow AI data center operators to build dedicated off-grid power infrastructure and bypass federal electricity regulations. Whether you read that as visionary pragmatism or a dangerous workaround that could compromise reliability for everyone else on the grid depends on your perspective. But its very existence tells you the industry has given up waiting. The 14 largest publicly listed data center operators are collectively spending close to $750 billion in 2026 — an extraordinary bet that power access will follow.
Source: ConstructConnect, 2026
The Engineering Escape Routes
Three technical responses are emerging simultaneously, each operating on a very different timeline. Liquid cooling is the most immediately deployable. Direct-to-chip and full-immersion cooling systems can reduce data center energy overhead by up to 30% compared to traditional air cooling, while enabling far greater compute density per rack. NVIDIA’s GB200 NVL72 systems are already too thermally dense for conventional air cooling to handle at full utilisation. The industry is mid-transition, and the pace is accelerating. This is the lever that matters most for 2026.
Small modular reactors (SMRs) are the longer game. Google has signed agreements with Kairos Power for 500 MW across six reactors. Amazon has invested directly in X-energy targeting over 5 GW of nuclear capacity by 2039. These are serious commitments from serious operators. But SMRs carry a realistic deployment horizon of 5 to 10 years from contract signing to first kilowatt-hour. They are the right answer for 2035, not 2026.
Demand flexibility may be the most underappreciated lever in the near term. EPRI research suggests that a 1 to 2% reduction in data center peak load could reduce electricity rates by up to 2.8% across the entire grid — meaningful relief for millions of households. The 96 MW Aurora AI Factory in Manassas, Virginia, built specifically to demonstrate AI workload flexibility at scale, is an early proof point. If hyperscalers embrace this at scale, it could buy grid planners critical breathing room.
The $1 Trillion Gamble
Here is my honest read. The data center power demand trajectory is real, and the people most worried about it are not issuing press releases. They are engineers quietly running numbers inside grid operators and utilities, and they are losing sleep. What I find remarkable — and I do not use the word lightly — is how rapidly the physical consequences of digital decisions have become visible. In 2020, the idea that a single data center development could meaningfully affect electricity rates and grid reliability for millions of households would have seemed hyperbolic. In 2026, it is simply true.
The AI infrastructure boom is also beginning to export its power problem. European and Middle Eastern data center markets are seeing surging investment precisely because operators found power access too constrained in Virginia and Texas. The grid bottleneck is reshaping the global geography of compute — a consequence that was not in any hyperscaler’s five-year plan three years ago.
By 2027, the US data center power demand picture will be clearer. Either the combination of liquid cooling, demand flexibility, on-site generation, and accelerated grid investment will have created a functional bridge to the nuclear era — or the grid will have become the binding constraint on AI expansion, with real consequences for who can build where and at what cost. I suspect the industry will find a way through. Humans have a remarkable track record of building the infrastructure that transformative technology demands. But the next 24 months are going to be genuinely uncomfortable for anyone responsible for keeping the lights on.
