Why cloud GPUs sit at 5%, and what companies pay for it

Two numbers from 2026 belong in the same sentence, and they rarely appear together.

The first comes from Gartner. Data centres worldwide will use 565 terawatt-hours of electricity this year, 26 percent more than in 2025. AI-optimised servers alone account for 175 TWh of that. By 2030 Gartner expects consumption to pass 1,200 TWh.

The second comes from Cast AI, which looked at roughly 23,000 Kubernetes clusters running on AWS, Google Cloud and Azure. On average, the GPUs in those clusters did useful work 5 percent of the time. CPUs managed 8 percent, memory 20 percent.

Put side by side, they describe an odd situation. The industry is planning power plants for new data centres, while most of the hardware companies already rent sits idle.

Why companies buy capacity they don't use

Nobody plans for 5 percent. It happens for reasons that each make sense on their own.

GPUs are scarce, so teams reserve them when they can and hold on to them. Cloud contracts and reserved instances reward commitment, not flexibility. Capacity gets sized for the busiest hour of the busiest day and then runs around the clock. And resources are usually set once, at deployment, while the workloads on them keep changing.

The result, in Cast AI's words, is that companies pay for roughly twenty times the GPU capacity their workloads actually use. Cast AI co-founder Laurent Gil put the difference in one line: "A GPU sitting idle costs dollars per hour. A CPU sitting idle costs cents."

Prices are moving the wrong way too. AWS raised prices for H200 Capacity Blocks by about 15 percent in January 2026, breaking a long run of falling cloud prices.

What idle capacity really costs

The bill is the visible part. Reserved GPUs are paid for whether a job runs or not.

The less visible part is energy. A server that is switched on and waiting still draws power and still needs cooling. Every reserved machine that waits for work is part of the 565 TWh. When a company buys new hardware because its existing capacity is "full" on paper, it adds to that number again, even though most of what it already has is doing nothing.

So the cheapest and cleanest compute is usually the compute a company has already paid for.

What helps

Three things move the number, roughly in this order.

Measure it continuously. Most teams know what they pay for, not what they use. Utilisation over a full day tells a different story from a dashboard checked at noon.

Use what is already paid for, first. Before a new instance or a new server is ordered, the question should be whether the work fits on capacity the company already rents or owns: idle cloud servers, on-premise machines, workstations that sit unused at night.

Decide per job, not per project. Allocation set once at deployment goes stale. A render, an inference batch or a large data transfer each need something different, and what is free changes by the minute.

How Symbola approaches it

This is the problem we are building Symbola for. Our orchestrator puts already-paid cloud servers to work first, then a company's own servers and computers, across CPU, GPU and bandwidth. For every job it checks what the job needs and which resources are free at that moment, then picks the fastest and most efficient way to run it: one well-suited resource, or several at once when splitting the work finishes sooner.

Our MVP already does this. In internal tests a 120-frame Full HD Blender render, split across four GPU nodes, finished 38 times faster than on a single CPU machine: 7.76 minutes instead of about 296. A 10 GB download through Bandwidth Pool, using several network paths at once, averaged 1,301 Mbps against 430 Mbps over one connection. We estimate the distributed render used 60 to 95 percent less energy than the single machine; metered figures will follow. Setup, method and limits are published on our benchmark page.

We are preparing pilots with industry partners now. If your company runs AI inference, rendering, video or large transfers and you suspect a good part of your cloud bill pays for waiting, we would like to hear from you: connect@symbola.io.

Suvom Baroi is founder and CEO of Symbola in Munich.