NNebius published a benchmark on October 26 that runs NVIDIA's Cosmos 3 Super model four different ways on one server with eight graphics processing units, or GPUs. NVIDIA Cosmos is a family of world models, which are artificial intelligence models that generate realistic video of how things move in the physical world. Robotics and self-driving teams use that video as training data, so they generate clips by the thousands.

NVIDIA is promoting the results this week, with a joint livestream with Nebius on October 28. The findings matter to anyone renting GPU capacity for this kind of work, because they show how much the server layout changes the price of each video.

One Server Can Run the Model as One Large Copy or Eight Small Ones

Nebius loaded the model in four layouts. The first spread one copy of the model across all eight GPUs. The others split the server into two copies on four GPUs each, four copies on two GPUs each, and eight copies on one GPU each. Every copy worked on one video request at a time, using an open-source serving tool called vLLM-Omni.

A large copy divides each video among its eight GPUs, so that clip finishes sooner. Eight small copies each build their own clip at the same time, so the server finishes more clips per hour even though each one takes longer.

Think of eight cooks who can either work together on one order or each cook a separate order.

Eight Small Copies Made 42.5% More Video, and Each Clip Took Over Five Times as Long

Every test clip was 1280 by 720 pixels and 189 frames long, which plays for about eight seconds. On an NVIDIA HGX B200 server, one large copy returned each clip in 68.5 seconds on average. Eight small copies took 378.1 seconds per clip, but the server produced 42.5% more video per hour (Nebius, 2026).

At the same hourly server price, more video per hour means a lower cost per video. The eight-copy layout cut the server cost per second of video by 30% (Nebius, 2026). The pattern held on the older NVIDIA HGX H200 server, where eight copies made 25.7% more video and each clip took 6.3 times as long (Nebius, 2026).

The hourly price of the server stays fixed, so the layout you choose sets your cost per second of video.

Your Workload Decides Which Layout Pays Off

Nebius gives plain advice based on who is waiting for the video:

• If you are generating thousands of training clips overnight, use eight small copies, because total output matters more than the wait for any one clip.
• If a person is watching the screen for each clip, use one large copy, because speed per clip is what that person notices.
• If your wait limit falls in between, run as many copies as you can while each clip still finishes inside that limit.

Nebius also tested sending each copy a second request while it was still working on the first. Output rose by 3.2% at most, and the average wait grew by 33% to 49%, so the company recommends keeping one request per copy (Nebius, 2026).

Your Own Prompts Belong in the Next Test

Nebius ran every layout with one fixed prompt and the model's safety filters turned off. The company counted a clip as valid when the file opened without errors, had the right size, and showed motion. Teams that plan to train robots on these clips will want to add their own checks for realism and accuracy before they scale up.

Nebius says the same thing in its own advice: test your own model and workload before you commit to a fleet. The company published its full test method on GitHub so you can repeat it, and the October 28 livestream with NVIDIA walks through the four layouts.

A layout choice that looks like a technical setting ends up as a line item in your budget.