You Can Buy GPUs but Not Electrons: The Power Bottleneck Is a Product Gap

The Problem

Ireland’s statistics office counted it: datacenters consumed 23% of the country’s metered electricity in 2025 (RTE). That beats every urban household combined, at 18%, and is up from just 5% a decade ago. Datacenter consumption grew 10% in a single year to 7,663 GWh while everyone else grew 2% (Data Center Dynamics). The Atlantic’s verdict on AI datacenters: this is no longer IT, it is heavy industry (The Atlantic). The software has not caught up with that shift. GPUs arrive if you pay; a grid interconnection in PJM or ERCOT territory can take years. Compute orchestration is a crowded shelf, Kubernetes to Slurm, but the layer that treats electricity as a first-class, metered, schedulable resource mostly does not exist. Few teams can say, per job, what burned how many kilowatts, when, and at what price.

Why Now

Goldman Sachs projects US datacenter power demand climbing from 31 GW in 2025 to 41 GW in 2026 and 66 GW the year after (Goldman Sachs). Cloud and AI infrastructure capex is headed toward roughly $1.5 trillion a year by 2027 (Yahoo Finance). The money and the chips are ready; the electrons are not. That moves the bottleneck. A large share of AI work, training runs and batch inference, can slip by hours without anyone noticing, yet shifting it into cheap or renewable-heavy windows is still a manual decision where it happens at all. Regulators are tightening too: Dublin has restricted new grid connections for years, and heavy consumers face growing disclosure and efficiency demands. As power gets scarcer, the operator who can prove throughput per megawatt wins the expansion permit and the supply contract. The measure-to-win dynamic just arrived, and no standard tool owns it.

flowchart LR
  A[Grid and tariff data] --> C[Power visibility layer]
  B[GPU workload queue] --> C
  C --> D[Power-aware scheduler]
  D --> E[Shift to cheap clean windows]
  E --> F[Throughput per megawatt report]

How to Build It

The wedge is metering. Pair rack- and cluster-level power telemetry with tariff and wholesale price feeds, then show, per job, what this training run cost in electricity and what it would have cost six hours later. Installation is light and the utility bill validates the numbers, so the sales motion is short. Scheduling comes second: let teams tag deferrable jobs, then move execution windows against price, carbon, and grid congestion signals as a layer on top of Kubernetes or Slurm. That unlocks the third tier, procurement. Enroll clusters in demand response programs, collect curtailment revenue, and bring measured data into power purchase negotiations. Charge a subscription per managed megawatt plus a share of verified savings. Skip the hyperscalers; they build this in-house. Start where the power bill hits the P&L directly but no dedicated energy team exists: mid-size GPU clouds, colocation operators, and enterprise AI platform teams that just started running their own clusters.

See the Structure as a Map

Power is one input in a bigger cost equation. The DeepThought brief Inference Economics maps, node by node, how inference costs now shape AI margins.