The bottleneck in building a datacenter stopped being the building about three years ago. It is the interconnection queue, and the queue is now longer than the useful life of the accelerator the building is meant to hold.
In several major markets the wait for a new high-voltage connection has stretched to five or six years, with some requests quoted beyond 2030. A building shell takes eighteen to thirty months. So the shell can be finished, powered, and sitting empty for years while the utility works through a queue of requests submitted by people who were guessing.
That mismatch is the single most important fact in infrastructure right now, and it explains a set of decisions that otherwise look irrational.
Why the queue got long
Interconnection queues were designed for a world where new generation was built by utilities and large independent producers, one project at a time, with a planning cycle to match. They were not designed for hundreds of speculative requests arriving at once from developers who have not yet decided whether to build.
The result is a queue full of projects that will never be built, each of which must still be studied because the study determines who pays for network upgrades. The studies are serial, the upgrades are shared, and the whole process is sensitive to any earlier request changing its parameters.
The gap between the two lines is where the entire behind-the-meter industry lives. On-site generation is more expensive per megawatt-hour and arrives years earlier, and for a workload with a two-year payback horizon, arriving earlier is worth more than being cheaper.
The cost of speed
The economics look bad on a spreadsheet and good on a timeline. On-site gas generation typically lands somewhere between 1.4 and 1.7 times the levelised cost of a grid contract, depending on the market and the utilisation. That premium is the price of skipping the queue.
But levelised cost is the wrong metric here, because the two options do not deliver the same product. A grid contract delivers firm capacity with a long lead time. On-site generation delivers firm capacity in two years with a fuel supply chain you now own and a permit you now have to defend.
The third bar is the interesting one. Grid plus storage firming is a middle path: take the interconnection you can get, add batteries to shape the load so that the connection you were granted can carry more useful work. It costs less than on-site generation and arrives later than both, but it does not require you to become a power producer.
What this does to the price of a token
Here is the part that reaches people who will never read a grid study.
Compute prices are, to a first approximation, a function of the cost of a megawatt-hour and the cost of capital. When the industry shifts from grid contracts to on-site generation, the marginal megawatt-hour becomes more expensive. That cost does not disappear; it is passed through, and it shows up as a floor under the price of inference.
This is why the price of AI features is unlikely to fall as fast as the price of the underlying models. The model gets cheaper every quarter. The electricity, the land, the interconnection and the turbine do not.
There is a second, subtler effect on ordinary households. Behind-the-meter generation removes a large load from the shared network without removing the network's fixed costs. Those costs are recovered from the remaining customers, which means residential bills rise as the largest customers leave. That is a political problem before it is an economic one, and it is already visible in several markets.
The market arithmetic
My estimate is that behind-the-meter capacity for datacenters grows from a small share of new builds today to somewhere between a fifth and a third of new capacity by 2030, concentrated in markets with long queues and liberal self-generation rules.
The more interesting forecast is about the shape of the contract. As on-site generation becomes normal, the buyer's problem shifts from "how do I get a connection" to "how do I manage fuel price risk", which is a problem the industry has no experience with. Expect a wave of long-dated gas supply agreements and, eventually, a wave of hedging products that look a lot like the ones airlines use for jet fuel.
What to do if you are the one building
Three things, in order.
Measure your actual load shape before you negotiate anything. A datacenter that can flex its load by ten percent is a much easier interconnection customer than one that cannot, and flexibility is now a tradeable asset in the queue.
Get the interconnection request in before the design is final. The queue position is worth more than the design optimisation, and it is the only part of the project you cannot buy your way out of later.
Price the schedule risk explicitly. If the connection arrives two years late, what does that cost you in stranded capital and lost revenue? That number, not the levelised cost, is the one that should drive the decision between waiting and building.
The counter-argument: demand response instead of steel
The strongest objection to the build-your-own-plant conclusion is that it treats load as fixed. It is not. A datacenter that can curtail or shift a meaningful fraction of its draw is a different kind of customer, and in several markets that flexibility is now worth more than the generation it replaces.
The mechanism is straightforward. Grid operators pay for capacity that can respond within minutes. A training cluster that can pause for twenty minutes, or an inference service that can defer batch work to a cheaper hour, is exactly that product. Done at scale, it turns a liability into a revenue line and removes the need to build anything.
The catch is that the flexibility has to be real. A promise to curtail that depends on a customer's willingness to accept a slower response is not capacity; it is an option that will not be exercised when the grid actually needs it. Regulators have started to notice the difference, and several markets now require proof rather than a contract.
My estimate is that demand response covers somewhere between ten and twenty percent of the gap, concentrated in batch training rather than in latency-sensitive inference. That is meaningful and it is not enough, which is why the turbines are still being ordered.
The realistic outcome is a mix: some on-site generation, some firming storage, some contracted flexibility, and a much more adversarial relationship with the utility than the industry is used to.
What is actually transferable
The lesson is that in any infrastructure market, the binding constraint migrates to whatever has the longest lead time, and the price of everything else adjusts around it. For two decades that constraint was capital. Then it was silicon. Right now it is a permission to connect.
When the constraint moves, the businesses that win are the ones that reorganised around the new one before it was obvious. The rest spend years optimising a cost that is no longer the problem.
If this saved you time, a tip is appreciated.
HCqPmPJM3NVtS6YrmVaTZYnuP8XZ2PFgQZKT94VK1LkS
Top comments (0)