Two years ago, the bottleneck on AI capability was getting hold of GPUs. A year ago, it was finding hyperscaler capacity to host them. This year, every story about AI infrastructure ends in the same place: a substation, a transformer, a queue at an interconnection.
Data center electricity demand is on track to reach roughly a thousand terawatt-hours in 2026 — somewhere between Japan and Russia in scale, if you wanted to stack it against a country. Interconnection queues stretch to multi-year lead times. Heavy electrical equipment is back-ordered. The geography of new compute is no longer being chosen by latency or talent or proximity to a customer base. It is being chosen by where the megawatts already are, or where they can be built quickly.
Hyperscalers have noticed. Conditional offtake agreements between data center operators and small modular reactor projects grew from twenty-five gigawatts at the end of 2024 to forty-five gigawatts now. Some operators are skipping the public grid entirely and building their own generation on-site — energy islands that answer to nobody but the load they were built to serve.
The operator's question, watching this, is not whether to admire the engineering. It is which side of the grid your work has to live on.
If your operation depends on a workload that requires frontier-scale compute, you are now in queue alongside governments and hyperscalers for the same megawatts. Your contract, your timeline, your cost curve are all subject to the same physics. If your operation can run its serious work on close compute — a desk, a rack, a modest cabinet — you are not in that queue. You are on a different grid.
This is not a clean break. There are workloads that genuinely need the frontier. There are operators who genuinely have to wait for the megawatts. But the assumption that the grid will quietly accommodate every AI ambition the industry generates is now broken, and pretending otherwise is how procurement teams find themselves on a six-year wait list for a capability they thought they could buy this quarter.
Every AI capability story ends in a substation. Plan accordingly.