Thought Leadership Summary
Power moved from a facilities line item to the first question in every siting decision. Grid operators cannot match the pace of AI demand, so capacity now shapes architecture instead of following it. That pulls IT into rooms it used to skip.
Interconnection is the gate. PJM’s queue holds terawatts of proposed projects and the backlog runs into multiple years. Price follows scarcity. In a recent PJM capacity auction the clearing price hit 329 dollars per MW-day, and analysts attributed roughly 40 percent of the cost, about 6.5 billion dollars, to data center load. That bill lands on ratepayers and on the facilities budget. Energy and capacity are separate purchases. Capacity is the right to draw at peak, so a flat load pays for it around the clock and gives the operator nothing to curtail. A power purchase agreement locks an energy price and builds no wire. Utility-scale capacity factors sit near 24 percent for solar and 35 percent for wind, so a constant load needs nameplate capacity built several times over. Watch the contract term against the hardware term.
Site reviews used to start with workloads. Now they start with watts. Grid operators cannot hand over capacity on the schedule data center teams need, so power became the constraint that shapes every choice downstream. IT leaders sit in those conversations now, because the answer determines what they deploy and when.
Interconnection is the bottleneck. A large load or a new plant has to request a connection and wait for a series of studies before anyone breaks ground. PJM interconnection queues hold terawatts of proposed projects, and the backlog runs into multiple years. Reform efforts sped up the study process and cut the withdrawal rate, but the queue is still not a formality. It is a gating item that decides whether a site exists at all.
Capacity Is Priced, and Data Centers Move the Price
Grid capacity is auctioned. In a recent PJM capacity auction the clearing price reached 329 dollars per MW-day, and data center demand drove a large share of the increase. Analysts put data center load at roughly 40 percent of capacity cost in that auction, about 6.5 billion dollars. That money lands on ratepayers and on the facilities budget. The PJM capacity market is where a load forecast turns into a bill.
Energy and capacity are different purchases, and the distinction matters more every year. Energy is what you consume. Capacity is the right to consume at peak, and it is priced on your contribution to system stress. A load that runs flat and never drops is the worst possible customer on both axes. It pays for capacity around the clock and it gives the operator nothing to curtail.
That is why IT gets pulled in. A capacity model needs a load forecast, and a load forecast comes from hardware plans. Nobody in facilities knows how many accelerators you plan to rack in 2028. Procurement asks. Someone has to answer with a number, and the number is an architecture decision wearing a finance costume.
A PPA Buys Energy, Not a Wire
A power purchase agreement locks a price for energy. It does not build transmission to your site, and it does not make the resource firm. Teams that read a PPA announcement as a solved problem learn this the hard way. Contracting clean energy at a distance and delivering steady power to a specific building are two different contracts, often with two counterparties and two schedules.
Renewables have a second limit, the shape of their output. Utility-scale capacity factors sit near 24 percent for solar and near 35 percent for wind. A data center draws the same power at 3 AM in January as at noon in July. Meeting a constant load from wind, solar, and batteries in 99 percent of hours means overbuilding nameplate capacity many times over the load. Firm power means a resource that runs when you ask, and you have to contract for that separately.
Siting Decides the Architecture
Power availability now precedes architecture. The question is no longer which racks to buy. It is where you can secure hundreds of megawatts on a schedule. NERC reliability assessments keep flagging the gap between load growth and new supply. Some developers skip the queue entirely and build behind the meter, adding on-site gas turbines to serve the load themselves. Gas turbine lead times are measured in years, so even that shortcut has a queue of its own.
There is a timing mismatch nobody budgets for. A server refresh is a three to five year decision. A transmission upgrade, an interconnection agreement, and a power contract run for fifteen to twenty five years. The infrastructure you commit to for power will outlive three generations of the hardware it feeds. Signing that contract to keep a 2028 roadmap on track is a long bet on a short plan.
A hyperscaler can afford to wait in the queue. A mid-size enterprise that just signed a ten year colocation deal cannot. Delay has a price, and it is paid in stranded assets. An operator that commits to land and shell before it has a firm power date ends up holding a building it cannot fill. The reverse happens too. A site with power and no workloads burns the reservation while the hardware queue slips. Both failures come from treating power as a downstream detail.
The practical shift is ownership. If power gates your roadmap, then data center power procurement belongs in the same room as capacity planning. IT teams that treat it as a facilities problem end up waiting on someone else’s timeline. Bring a load forecast, a load shape, and a date. The grid decides the rest.
Every announced commitment in this space is tracked with its source in our AI data centre power commitments record.
Related reading. why data centers run out of power before demand, and how data centers are rewriting how they buy power.
Get the next one before it is old news
Independent analysis of cloud-native infrastructure, Kubernetes and data centre economics. No vendor spin.