PJM wholesale power costs up 76%, pushing AI data centers toward rack-level energy monitoring

Wholesale electricity costs in the PJM Interconnection region jumped 76 percent year over year in the first quarter of 2026, according to recent reporting highlighted in an essay published on dev.to. The piece, originally posted on the Sensaka blog, draws an operational conclusion for infrastructure teams: whatever caused the spike, power has shifted from a background utility expense to a scarce resource that now shapes capacity planning, operating cost and business risk.

What caused the spike is disputed

The essay is careful about attribution. How much of the PJM increase belongs to data center demand is contested, it notes, because gas prices, weather, generation retirements, transmission constraints and market design all feed into wholesale prices. Collapsing the number into a claim that AI workloads caused the entire rise would overstate the case. The lesson for operators survives the debate anyway: AI deployments are raising power density, grid access is getting harder in many markets, and electricity has become a genuine constraint rather than a fixed overhead.

Utility bills arrive too late to act on

Conventional energy management starts with utility invoices or facility-level meters, which serve accounting but not operations. When a monthly bill climbs, the essay points out, a chain of questions is still open: which rooms drew more power, which racks run densest, which GPU clusters added load, which teams owned that work, how much electricity idle gear burns, whether cooling overhead is tracking IT load, and how much headroom is genuinely safe to deploy against. Without that detail, a facility knows what energy cost but not how it was used — and large GPU deployments can move power and cooling demand faster than monthly reporting can capture.

Following power down the rack

The proposed fix is an energy model that mirrors the physical hierarchy — site, data hall, row, rack, power circuit, server, accelerator node — so operators can see where energy concentrates and where constraints are forming. Rack-level data gets particular weight. A rack can have empty U slots yet almost no power headroom left; another can show plenty of electrical headroom while its cooling zone or network links block further deployment. The useful question is therefore not the count of free racks but how much capacity remains deployable once power, cooling, physical space and a reserve margin are weighed together.

Idle GPUs and who pays the bill

Idle hardware costs more in AI facilities than elsewhere: an idle GPU server burns electricity while monopolizing accelerator capacity, rack power, cooling and capital at the same time. Linking energy data to utilization lets operators spot accelerators that have sat underused, see for how long, and identify the owning project — the basis for reclaiming resources, adjusting schedules and deferring new construction. Reclaiming stranded accelerators effectively adds usable supply inside the existing building, which the essay argues can be worth more than a marginal efficiency gain when power is scarce.

The same data enables cost attribution. When electricity is a shared overhead, individual projects have little reason to economize; when GPU hours, energy consumption and ownership are connected, model services can be compared on the energy they actually consumed rather than the accelerators allocated to them, and idle resources can carry a visible cost.

Scheduling as an energy lever

Where electricity prices vary by time, workload scheduling can become part of the energy strategy: flexible training jobs shift toward cheaper hours while latency-sensitive inference keeps priority, and multi-site operators can factor per-site capacity and energy cost into placement decisions. The essay tempers this with caveats — business priority, deadlines and reliability still govern — but argues that once energy and workload data are connected, operators gain one more scheduling input.

A vendor blog with a broader point

One caveat of its own: the essay originates from Sensaka, whose DCOS product monitors power, racks, environment and hardware, offering real-time consumption data, efficiency analysis, PUE tracking and forecasting, and links those physical signals to compute, GPU utilization, projects and services. The argument is commercially motivated, but it does not stand or fall on the product behind it.

Why it matters

The fight over whether data centers drove PJM's price spike will stay political and unresolved. The measurement requirement will not wait for it. If electricity is now a binding constraint on AI expansion, every serious operator needs to answer basic questions: how much power goes to facilities versus IT equipment, which racks are approaching their limits, which GPU capacity is productive and which is idle, and how much deployable headroom actually remains. On that reading, the cheapest megawatt available to an operator may not be the next one bought from the grid — it may be the one already inside the building, currently going to waste.