Skip to content
State of GPU Compute 2026

Report · 2026 edition

Preview

State of GPU Compute 2026

Where GPU-hours went over twelve quarters, how big clusters really are, and how busy rented hardware is when nobody is watching.

Get the reportSee what is inside

PDF / 38 pages / 7 chapters / no sales call

Nexpod research2026

State of
GPU Compute

Rental patterns, utilisation and
the B300 effect on training

2026 editionnexpod.ai/reports

What is inside

Three years of GPU‑hours. Six charts.

One chart from each chapter, Q1 2024 to Q4 2026. The figures on this page are sample data and will be replaced by the published numbers.

61%

of GPU-hours on inference

  • inference 61%
  • training 31%
  • batch 5%
  • agents 3%
Chapter 01Sample data

Most hours now serve users

Split of all GPU-hours in Q4 2026, by workload.

B300 8%
Q1 '26
B300 21%
Q2 '26
B300 35%
Q3 '26
B300 46%
Q4 '26
  • B300
  • B200
  • H200
  • H100
Chapter 02Sample data

The B300 year

Share of new rentals by GPU family, each quarter of 2026.

inferencetraining
Chapter 03Sample data

Inference took the lead

Share of GPU-hours spent serving versus training, by quarter.

  • 1 node64%
  • 2–422%
  • 89%
  • 16+5%
Chapter 04Sample data

Most jobs fit on one node

Share of jobs by node count.

  • Batch92%
  • Training87%
  • Serving58%
  • Agents34%
Chapter 05Sample data

Batch keeps GPUs busiest

Median utilisation by workload type.

EU-AMS-1
EU-FRA-1
EU-HEL-1soon
Chapter 06Sample data

Pinned to a region

Where EU teams keep training data.

Contents

Seven chapters, 38 pages.

Read it end to end in half an hour, or jump to the chapter that answers the question you have this week.
  1. 01

    Where the GPU-hours went

    Training, inference, batch and agents, quarter by quarter.

    p. 4
  2. 02

    The B300 year

    How fast teams moved to the newest family, and why some did not.

    p. 9
  3. 03

    Inference takes the lead

    The quarter serving passed training, and what changed in the bill.

    p. 14
  4. 04

    How big clusters really are

    Node counts per job, from single servers to 32-node runs.

    p. 19
  5. 05

    Busy or idle

    Measured utilisation by workload, and where the idle hours hide.

    p. 24
  6. 06

    Data residency in practice

    Which teams pin regions, and what they pin them for.

    p. 29
  7. 07

    Methodology

    Sample, time window, anonymisation and the limits of the data.

    p. 35

Get the full report

PDF, 38 pages. We email you the link.

Placeholder

We use your details to send the report and, if you tick the box, the newsletter. See the privacy policy.

Method

How the numbers are made.

Metered usage, not a survey. The methodology chapter on page 35 has the full detail.
From meter to chartPlaceholder
  1. 01

    Sample

    Aggregated GPU-hours from the meter in every Nexpod region, grouped by workload type and GPU family.

    source
    billing meter
    unit
    GPU-hours
    regions
    all live
  2. 02

    Window

    Twelve quarters, each weighted the same, so a busy quarter does not tilt the trend lines.

    from
    Q1 2024
    to
    Q4 2026
    quarters
    12
  3. 03

    Anonymise

    Figures are rounded, and any group too small to hide a single customer is folded into other.

    rounding
    1%
    small groups
    folded into other
    per customer
    never shown

FAQ

About the report

Who is the report for?
Engineering leads, platform teams and finance people who plan GPU capacity and want numbers from real usage rather than list prices.
Is any customer data identifiable?
No. At launch every figure is aggregated across projects and rounded, and no single customer can be picked out. The methodology chapter explains how.
Can I quote the figures?
Yes, with a link to this page. The figures shown here today are sample data and should not be quoted.
Will there be a webinar?
A walkthrough with the authors is planned after launch. Leave your email in the form and tick the newsletter box to hear about it.

Put it to work

Test the numbers yourself. One node, by the second.

Pick a server, pick a region, and connect over SSH. Stop it whenever you are done; the meter stops with it.

Get startedTalk to sales
On-demand ratesEUR · excl. VAT
  • B300€6.00/GPU·hr
  • B200€5.00/GPU·hr
  • H200€4.00/GPU·hr
  • H100€3.00/GPU·hr