NVIDIA Adds Cost Estimates to Cosmos Data Factory
NVIDIA's latest walkthrough for its robot and autonomous-vehicle data toolkit now includes performance and cost figures.
Updated on Nov. 19, 2026 in Robotics

NVIDIA added performance estimates to a two-prompt walkthrough for Cosmos Data Factory, a toolkit for preparing training data for robots and autonomous vehicles. The company said a benchmark server with eight NVIDIA RTX PRO 6000 Blackwell Series GPUs generated about 130 autonomous-driving edge-case clips per day.
Why it matters
Robotics and autonomous-vehicle developers often face high costs in collecting and labeling training data, and published cost figures let teams compare synthetic generation against physical data collection. The walkthrough builds on five reference blueprints NVIDIA outlined in an October Cosmos update.
The benchmark server used eight NVIDIA RTX PRO 6000 Blackwell Series GPUs in the cloud, with NVIDIA estimating a cost of about $40 per server-hour. Its Cosmos Evaluator checks generated video for physical plausibility, and some workflows automatically label their outputs.
The players
NVIDIA
NVIDIA is a American semiconductor and computing company best known for its graphics processing units, which are widely used to train and run artificial intelligence models.
The details
The first prompt directs a coding agent to clone and configure workflow orchestration tools, while the second selects a reference blueprint and runs it against the developer's data. The five blueprints cover autonomous-vehicle dataset generation, image attribute augmentation, video augmentation, event-video generation and defect-image generation, and they draw on six composable skills.
Timeline
Oct. 21: NVIDIA outlined five reference blueprints in a Cosmos update.
Nov. 18: NVIDIA published the walkthrough with performance and cost estimates.
The Tech Race
The update extends NVIDIA's Cosmos platform, which it expanded in October, by attaching published cost figures to synthetic data workflows. It positions NVIDIA's simulation tools against rival approaches that rely on collecting real-world road and sensor data.
Developers building robots or autonomous vehicles can use the published walkthrough to estimate what synthetic data generation may cost compared with physical road collection. The approach may reduce the time needed to assemble specialized datasets, such as defect images.
The takeaway
Teams weighing synthetic data against physical collection should compare the per-server-hour and per-scenario figures against their own workloads before committing to either approach.
Further reading
For more on how synthetic data is reshaping machine training, see the Robotics section.
Source note: This article includes information reported by TokenPost.



