Nvidia Validated Nebius Blackwell Ultra Server Racks
The cloud provider has confirmed its B300 hardware achieves Nvidia’s official performance targets for AI training.
Updated on Oct. 10, 2026 in Data Centers
Live Poll
Do you trust that certified hardware benchmarks improve the quality of AI services you use?
Nvidia has granted Nebius Exemplar Cloud status for its HGX B300 racks after verifying the hardware against its reference architecture. This validation confirms that the systems, which utilize Blackwell Ultra technology, function according to design standards.
Why it matters
The validation process ensures that third-party cloud infrastructure consistently meets Nvidia's specific performance benchmarks for training massive AI models. Achieving this status signals that the hardware can execute high-compute workloads reliably, a critical requirement for enterprise AI adoption.
The HGX B300 platform integrates 8 Blackwell Ultra GPUs per node with approximately 2.1 TB of HBM3e (high-bandwidth memory, a type of RAM designed for high-speed AI processing) memory. The system utilizes fifth-generation NVLink, a high-speed interconnect technology that allows GPUs to share data at faster rates than standard PCIe.
The players
Nvidia
A leading designer of GPUs and AI infrastructure stacks that dictates the reference architectures for high-performance computing.
Nebius
A cloud infrastructure provider that operates specialized AI-focused data centers and trades on the NASDAQ exchange.
The details
Nvidia performs an engineering analysis by reviewing benchmarking results generated when cloud operators run standardized test recipes on production hardware. In this instance, Nebius tested the hardware using the MLPerf Training 6.0 suite, which measures performance across specific AI model architectures including Llama-3.1-8B and GPT-OSS 20B. These benchmarks demonstrate that the system performs as expected within the Nvidia stack, confirming the integrity of the server integration.
Timeline
October 8, 2026: Nvidia announced the formal validation of Nebius HGX B300 racks.
The Tech Race
This validation follows the performance standards established by the MLPerf Training 6.0 benchmarking suite. It marks the latest step in Nebius’s strategy to differentiate its stack by mirroring Nvidia's official reference architecture, following earlier certifications for the H200 and GB300 NVL72 configurations.
Enterprise users requiring high-compute capacity can now deploy Nebius infrastructure with the expectation that it aligns with Nvidia's performance baselines. The hardware is currently in production, providing a new option for organizations building on the Blackwell Ultra architecture.
The takeaway
Nebius has cleared the engineering hurdles necessary to integrate Blackwell Ultra systems into its high-performance cloud offering. The next indicator of progress will be the release of updated performance figures in subsequent MLPerf cycles as the B300 architecture matures in live production environments.
Further reading
For more on how infrastructure providers certify compute capacity, visit our coverage of Data Centers.
Live Poll
Do you trust that certified hardware benchmarks improve the quality of AI services you use?
