The constraint on data centre construction used to be capital, then land, then chips. It is now increasingly electricity — not the price of it, but the physical ability to deliver it to a specific location within a useful timeframe.
This is a slower-moving problem than a chip shortage and considerably harder to solve, because the binding constraint sits in infrastructure with build times measured in years and regulatory approval processes measured in the same units.
The density change
A traditional server rack drew something in the range of five to ten kilowatts. Racks built for dense accelerator workloads draw several times that, and the trajectory is upward rather than flat.
This is not simply a larger version of the same problem. Two things change qualitatively:
- Cooling stops being incidental. Air cooling has a practical ceiling. Above it, liquid cooling becomes necessary — which means plumbing, pumps, heat exchangers and a facility designed around them. Retrofitting an air-cooled building is often uneconomic compared with building new.
- Power delivery becomes a site selection problem. A facility drawing tens or hundreds of megawatts is a substantial industrial load. It cannot simply connect to whatever capacity happens to be nearby.
The consequence is that the useful question is no longer “where is land cheap” but “where can a large, continuous load be connected, and when.”
The interconnection queue
Connecting a large load or a new generator to the transmission network requires a study process: what the connection does to system stability, what network reinforcement it requires, who pays for it. These processes are queued, and the queues in many markets are long.
The waits are frequently multi-year. Critically, the delay is not proportional to project size — a small project can wait as long as a large one, because the constraint is administrative and engineering throughput rather than physical capacity.
This produces a market dynamic that is easy to miss. Sites with existing grid connections of adequate size — decommissioned industrial facilities, retired generating stations, older data centres — carry a premium far above their physical worth. What is being bought is not the building. It is a position in a queue.
Why generation capacity is the wrong number
A common response is to observe that the grid has ample generating capacity in aggregate. It usually does. That is not the constraint.
Three distinctions matter:
- Location. Electricity must be delivered where it is consumed. Surplus generation several hundred miles from the site is useful only if transmission exists to move it, and transmission lines are among the slowest infrastructure to permit and build.
- Firmness. A data centre runs continuously. Intermittent generation supplies energy but not the guaranteed availability the load requires, which means either firm capacity, storage, or on-site backup at meaningful cost.
- Peak coincidence. Networks are sized for peak demand. A new continuous load adds to peak by definition, so it consumes headroom even in a system with substantial average surplus.
Where the cost actually lands
For the operator, the visible effect is a longer and less certain development timeline, and a shift in what a site is worth. Power availability has moved from a due diligence item to the first screen.
For the customer, it appears as capacity pricing. Compute in constrained regions costs more, availability of specific instance types becomes patchy, and contract terms lengthen as providers seek to underwrite the fixed investment. Committed-use discounts get deeper because the provider needs the commitment more than it needs the margin.
For anyone assessing the economics of running their own infrastructure, this is a genuine change in the comparison. Part of what a cloud provider sells is having secured power and connection years in advance — a thing that is now materially harder to replicate. That does not settle the question either way, but it is a factor that did not exist a few years ago, and it belongs in the analysis alongside the ones we set out in the economics of cloud repatriation.
What to watch
Announced capacity is a poor indicator, because announcements are cheap and timelines slip. More informative:
- Signed power purchase agreements, which represent a financial commitment rather than an intention.
- Interconnection queue positions and study completions, published by system operators in most markets.
- Capital expenditure actually spent versus announced, in operator filings.
- Behind-the-meter generation — operators contracting for dedicated on-site or adjacent generation are routing around the queue entirely, which is the clearest signal that the queue is binding.
The general lesson is familiar from other industries. When a fast-moving sector runs into slow-moving physical infrastructure, the infrastructure sets the pace, and the adjustment happens through price and timeline rather than through anyone announcing that a limit has been reached.
