TL;DR

  • Global data center electricity use is projected to grow 26% in 2026 to 565 terawatt-hours, driven by AI servers that encounter memory supply and power grid limits.
  • AI inference servers average roughly 20% utilization, leaving four-fifths of paid hardware capacity and floor space idle while continuing to draw power.
  • Decentralized compute networks pool idle GPUs across facilities into an open inference market, directing contributed cycles straight to active AI workloads with minimal overhead.
  • Selling unused accelerator hours maximizes existing infrastructure efficiency, serving more AI tokens without needing to build, cool, or plug in additional hardware.

# # #

Global data center electricity use will grow 26% in 2026, to 565 terawatt-hours, according to Gartner. Most of that growth will come from AI-optimized servers. New clusters also run into limits on memory supply and grid access.

Hyperscalers want more control over hardware costs, and several are building their own chips to get it. Microsoft plans to begin deploying its Maia 300 accelerator in 2027 and claims it will deliver 40% more AI tokens per megawatt than leading GPU-based systems.

Having built a decentralized AI compute network, I think the industry underuses a cheaper lever, which is getting more productive work out of the GPUs it has already racked, powered, and paid for.

Installed Capacity and Delivered Compute Are Drifting Apart

Training clusters run close to capacity, and Lawrence Berkeley National Laboratory’s latest U.S. modeling assumes they average about 80% utilization. Inference fleets are sized for peak traffic and spend long stretches of the day underused.

The same report assumes AI inference servers will reach about 20% average utilization by 2030, half the figure its previous edition used, and that idle machines still draw a fifth of their rated power. It also estimates that grid interconnections run at around half their rated capacity because of redundancy requirements and maintenance.

At 20% average utilization, four-fifths of what an inference server costs in depreciation and floor space pays for capacity that does no work.

Hardware Design Raises the Price of Every Idle Hour

In large-model inference, the amount and speed of high-bandwidth memory (HBM) on each card often caps throughput before raw compute does; the IEA expects the current HBM shortage to last through at least the end of 2027. Models split across many GPUs slow down sharply once the interconnects between them saturate.

Rack-scale systems respond with denser packaging. The IEA estimates that AI server power density rose elevenfold between 2020 and 2025 and will quadruple again by 2027.

At these densities, liquid cooling has become a requirement, because each AI rack has to shed as much heat as 30 natural gas boilers put out. Liquid-cooled AI halls helped bring the U.S. average power usage effectiveness (PUE) down from 1.55 to 1.45 between 2018 and 2024, according to the Berkeley Lab report. All of this adds to the capital cost a GPU has to earn back during the hours it actually works.

Spare GPU Cycles Can Find Buyers Across a Wider Network

A decentralized compute network pools GPUs from data centers, cloud providers, and independent operators into one inference market. Requests from anywhere can go to whichever accelerators have spare capacity, so a GPU that one facility can’t keep busy still earns.

When we designed our network’s protocol, the goal was to send nearly all contributed compute straight to AI workloads and keep consensus overhead to a minimum. Hosts prove their capacity in a short, transformer-based compute sprint, and outside that window their hardware serves inference. Other participants re-check only a small random sample of the results.

Data center operators can earn on capacity they built for peak demand. Owners of older A100 and H100 fleets can keep those cards earning next to newer ones, because a node can mix NVIDIA GPUs anywhere from the A100 to the B300 as long as it pools at least 320 GB of memory.

Inference is the workload that suits open networks today. A shared pool can absorb demand spikes and narrow the gap between peak provisioning and average load. Frontier-scale training is another matter, since it still needs ultra-fast fabrics inside a single facility.

Useful Work Belongs in Every Capacity Plan

Power is going to stay tight for the rest of the decade, and custom chips will keep squeezing a bit more work out of every watt. Operators don’t have to wait on either one, though, because idle accelerator hours are inventory they can start selling today. Every token served on a GPU that would otherwise have sat dark is compute nobody had to build, cool, or plug into an already stretched grid. My bet is that utilization ends up right next to megawatts and PUE on every AI infrastructure scorecard, and open networks will be one of the things nudging that number up.

# # #

About the Author

David Liberman is a serial founder and operator who had spent two decades building across AI, media, fintech, and augmented reality. He is a co-creator of Gonka, a decentralized network for high-efficiency AI compute that directs nearly 100% of GPU cycles toward productive AI inference and training rather than the wasted computation of conventional blockchains.