Nvidia grants Nebius Exemplar Cloud validation for HGX B300 training racks
In brief
- Nvidia granted Nebius Exemplar Cloud validation for HGX B300 training workloads, announced October 8, 2026.
- Nebius's customer-facing racks matched Nvidia's reference architecture, the validation confirmed.
- Nebius said its systems posted the fastest single-node B300 times in MLPerf Training 6.0's closed division.
- Each HGX B300 node packs 8 Blackwell Ultra GPUs and roughly 2.1 TB of HBM3e.
How the Exemplar Cloud check works
Under the program, cloud operators run a set of standardized benchmarking recipes on their live production hardware. Nvidia then measures those results against reference targets it defines. It isn't a test on a hand-picked lab machine (the racks in question are the ones serving paying customers).
Nebius qualified after an engineering review and an analysis of its benchmark results, per the report. The validation specifically targets training, even though Nvidia designed the HGX B300 system for two jobs: training large models and running high-throughput inference.
Nebius has been through this process before.
The company previously earned Exemplar status on other Nvidia platforms, including the H200 and GB300 NVL72 configurations.
The MLPerf result
Nebius, which trades on NASDAQ under the ticker NBIS, paired the validation news with a second result. According to Crypto Briefing, Nebius's systems posted the fastest single-node B300 times in the closed division of MLPerf Training 6.0. That's the company's result as reported; it's a single-node B300 figure in one category.
MLPerf's closed division is the stricter category, where participants have to train the same models under the same rules. Nebius's results covered two large language models: Llama-3.1-8B and GPT-OSS 20B.
What's inside an HGX B300 node
Each node packs 8 Blackwell Ultra GPUs. The platform carries approximately 2.1 TB of HBM3e memory, and it uses fifth-generation NVLink (Nvidia's high-speed interconnect) to tie the hardware together.
Nebius isn't the only name around the platform. The broader HGX B300 ecosystem includes server makers Supermicro, ASRock Rack and ASUS, along with cloud providers such as CoreWeave. Validation results for those other cloud providers weren't part of this announcement, so the report doesn't say where they stand.
Frequently asked questions
What is Nvidia's Exemplar Cloud validation?
Under the Exemplar Cloud program, cloud operators run standardized benchmarking recipes on their live production hardware. Nvidia then measures the results against reference targets it defines. Nebius qualified after an engineering review and an analysis of its benchmark results.
What is the MLPerf closed division?
MLPerf's closed division is the stricter category, where participants have to train the same models under the same rules. Nebius said its systems posted the fastest single-node B300 times in the closed division of MLPerf Training 6.0, covering Llama-3.1-8B and GPT-OSS 20B.
What hardware is in an Nvidia HGX B300 node?
Each HGX B300 node packs 8 Blackwell Ultra GPUs. The platform carries approximately 2.1 TB of HBM3e memory and uses fifth-generation NVLink, Nvidia's high-speed interconnect. Nvidia designed it for training large models and running high-throughput inference.
