Business case · Cost of slow data

What does waiting on data cost you?

Slow data movement bills you three ways: idle machines you pay for by the hour, people who sit waiting on data that hasn’t arrived, and the power those machines burn doing nothing. An AI factory is the sharpest case — idle GPUs billing by the hour while a multi-petabyte training set crawls in for days. Pick your industry for realistic starting numbers, then adjust to yours.

Industry
Machines waiting (GPUs)iMachines that sit idle waiting on data. The 2,048-GPU default is conservative — frontier AI clusters run 24,000–100,000+ GPUs. Pick a size, or choose Custom to enter your own.
Cost per machine-hour (USD)iRental rate, or amortized cost if you own. Pick a preset, or choose Custom to type your own.2026 H100, per GPU-hr: neocloud ~$2–3 · AWS P5 ~$7 · GCP / Azure ~$11–12. Non-GPU nodes ~$1–3.
People waiting (e.g. ML & data engineers)iAnyone whose work blocks while data moves — engineers, scientists, analysts, operators, ops teams. The role shown is just a typical example for the industry; count everyone idled on data. Often the bigger cost of the two. Pick a team size, or choose Custom to enter your own.
Loaded cost per person-hour (USD)iFully-loaded: salary + benefits + overhead (~1.3× pay). Pick a preset, or choose Custom to type your own.Loaded $/hr: researcher / artist ~$70 · engineer ~$90 · scientist ~$100 · ML eng / quant ~$120–130. (US BLS, Levels.fyi × ~1.3.)
Time spent waiting on data (%)iShare of time work stalls waiting on data movement — dataset staging, checkpoint copies, and inter-site or cloud transfers. It varies by environment, so the 15–20% defaults are illustrative estimates, not benchmarks. A demo measures yours.Illustrative default; replace it with your own figure.
Cost of slow data
—
Turn this into a one-page pitch →

These are starting estimates — tap the iEvery field has an info icon explaining where the starting number comes from and its source. The defaults are reference figures for a mid-to-large deployment, not your environment — type your own values into any field, including a custom percentage for time waiting. icon on any field for its basis and source, and type your own numbers into any field. Machine and people costs both scale with the share of time work stalls on data; Zettar’s job is to drive that toward zero by moving data at line rate. “Time waiting on data” here is the data-movement slice of GPU idle — the part Zettar can recover — not total GPU idle, which is often far higher, from scheduling and over-provisioning we don’t address. Where slow data adds up →  ·  Get a measured estimate →

Why it bills you three ways

Slow data costs machines, people, and power.

A GPU waiting on a checkpoint is one of the most expensive idle resources in your datacenter — but a scientist waiting on a sequencing run, an artist on the dailies, or an analyst on the overnight risk batch is a cost too, and often the bigger one. And every machine left on to wait still draws power and emits carbon for nothing. Zettar moves the data at line rate so far less of it is wasted.

ZettarZettar — Data-movement business case

What Zettar returns: line rate keeps machines fed and people working — recovering the data-movement share of this cost — plus the discoveries, decisions, and deadlines that no longer slip. Proven: 1 PB in 29 hours at 96% utilization with SLAC and the U.S. DOE.

Next step: zettar.com — schedule a demo for a measured estimate on your environment.

Prepared · Formula: machines × $/machine-hour × 8,760 hours/year × share of time waiting on data, plus people × $/person-hour × 2,080 hours/year × the same share. Estimates from your inputs — not a quote.

Where does GPU time actually go to data — and how do you get it back? How to keep GPUs fed →

FAQ

Common questions

How is the cost calculated?

Two costs: idle machines (their count times cost per machine-hour times the share of time they wait on data, across a year) plus people waiting (your team times loaded cost per hour times the same data-stall share). Pick your industry for realistic starting numbers, then adjust every field to yours.

Why include people, not just machines?

Because waiting on data is not only a machine cost. A scientist waiting on a sequencing run, an artist on the dailies, or an analyst on the overnight risk batch is a cost too, and often the bigger one. Zettar moves data at line rate so both wait far less.

Can I take these numbers to my team?

Yes — the calculator builds a one-page business case you can download or print, with your inputs, the cost breakdown, and what Zettar returns. Book a demo for a measured estimate on your environment.

Get started

Put a number on it — then let us prove it.

Run the figures, download the business case, and we’ll validate them against your real environment.