Some utilization numbers from inside the cloud, courtesy of CAST AI's 2026 Kubernetes report: average CPU utilization sits at 8%, memory at 20%, GPU at 5%. Meanwhile CPU overprovisioning jumped from 40% to 69% in a single year as teams threw hardware at AI workloads.
Sit with those for a second. The servers driving the loudest load-growth story in a generation are, on average, mostly idle.
I work in climate tech commercialization - accelerator side, helping startups land industrial and energy customers - and this gap fascinates me because it's the rare demand problem with software margins. The IEA expects data center consumption to roughly double to about 950 TWh by 2030. Some real fraction of that demand is provisioned-but-unused compute. Nobody has to build a plant, permit a line, or pour concrete to recover it. Somebody has to tune workloads.
Companies are forming around exactly this. One in our portfolio, JetScale.ai, raised a $5.4M seed round to attack cloud workloads that waste compute because nobody optimized them. Their pitch works because the customer's constraint is increasingly physical: power, not budget.
Utilities have decades of machinery for demand-side management - efficiency programs, rebates, measurement and verification protocols, all built for motors, lighting, HVAC, envelopes. Compute efficiency has the same shape: a verifiable reduction in kWh per unit of useful output, sitting behind a customer meter, cheaper than new supply. Yet I can't point to a DSM portfolio that treats data center workload optimization as a program resource the way it treats a chiller retrofit. Maybe I've missed one.
Is that a gap or a category error? The M&V is genuinely harder - the baseline moves, the workload moves, and "useful output" is contested. But the loads are enormous, concentrated, and professionally managed, which is more than you can say for most residential programs.