Datacenter GPUs are hitting a power wall, not a compute wall

For most of the last three years, the story of AI infrastructure was a chip story: who could get enough GPUs, and when. That story is over. The newest accelerators are available in volumes that would have seemed impossible in 2023, and the real constraint on bringing a new AI datacenter online in 2026 isn't silicon — it's electricity.

A fully built AI training cluster today can draw as much power as a mid-sized city. Getting that much power to one site, reliably, is now the longest item on the critical path for a new facility, routinely outlasting the time it takes to manufacture and ship the GPUs themselves.

The queue problem

Connecting a large new electricity load to the grid requires an interconnection study, and in most of the US and much of Europe, the queue for those studies now runs three to five years. That timeline hasn't moved much even as demand has exploded, because grid operators are running the same environmental and engineering review processes designed for a world where large new loads were rare. A hyperscaler that wants a gigawatt-class campus online in 18 months cannot wait for a conventional interconnection. It has to either buy into an existing grid connection, generate its own power on-site, or pick a location where capacity already exists for other reasons.

Why companies are going around the grid

This is the direct cause of the wave of on-site and dedicated generation deals that have defined 2026. Natural gas turbines, sited directly next to new datacenter campuses, have become the fastest way to get firm power without waiting in an interconnection queue — even though it runs against the public commitments several of these same companies have made on emissions. Nuclear has moved from a talking point to signed contracts: long-term power purchase agreements with existing reactors, plus a growing list of small modular reactor deals that won't deliver power for years but lock in future capacity now, while it's still available to lock in. Amazon's nuclear power agreements and similar deals from other hyperscalers aren't philanthropy or PR. They're a direct response to a grid that cannot deliver power on a timeline that matches chip deployment.

The secondary effect is site selection. Datacenter campuses are increasingly chosen based on proximity to existing power generation and transmission capacity, rather than proximity to fiber, customers, or cheap land — the priorities that used to dominate. A region with surplus hydro, nuclear, or stranded gas capacity is now more valuable to a hyperscaler than a region with good connectivity, because connectivity can be built in months and a new substation cannot.

Cooling is a power problem too, not a separate one

The latest generation of AI accelerators draws over 1,000 watts per chip, and rack densities have climbed to the point where air cooling can no longer dissipate the heat fast enough. Liquid cooling — direct-to-chip cold plates or full immersion — has gone from a niche option to a requirement for any new high-density deployment. That matters for the power story because liquid cooling systems themselves consume a meaningful share of a facility's total power draw, and retrofitting an air-cooled facility for liquid cooling is often more disruptive than building new. Several hyperscalers have simply written off existing air-cooled capacity for next-generation accelerators and are building new liquid-cooled halls instead, which adds another multi-year construction timeline on top of the power problem rather than solving it.

What this means for the rest of the industry

Companies without the balance sheet to sign a decade-long power purchase agreement or build a gas plant are increasingly priced out of frontier-scale training, and are shifting toward renting capacity from neo-clouds that have already solved the power problem, or focusing on inference workloads that can run on smaller, more distributed deployments where a single site's power ceiling matters less. The gap between companies that control their own power supply and those that rent compute by the hour is becoming one of the primary dividing lines in who can actually train frontier models versus who is stuck fine-tuning and serving them.

Takeaways

• When evaluating a cloud provider or colocation deal for AI workloads, ask about the power source and interconnection timeline before asking about GPU availability — power is now more likely to be the bottleneck.
• Expect on-site generation (gas, and increasingly nuclear PPAs) to keep expanding as a workaround for grid interconnection queues, not as a transitional stopgap.
• Factor liquid cooling retrofit costs into any plan to deploy next-generation accelerators in existing facilities — a new build is often cheaper than a retrofit.
• Treat proximity to surplus power generation as a genuine competitive advantage in site selection, on par with fiber connectivity.
• For workloads that don't require frontier-scale training, consider distributed inference deployments that avoid dependence on a single power-constrained megasite.