Microsoft Azure Leads DeepSeek-R1 Interactive Inference in MLPerf v6.1

Russ Fellows
November 9, 2026

Independent Signal65 analysis of audited DeepSeek-R1 results shows Azure’s ND GB300 v6 submission, built on NVIDIA GB300 NVL72, leading Interactive throughput by 10.3% over the next cloud submission.

The MLCommons MLPerf Inference v6.1 results provide an important reference point for organizations evaluating inference options for large AI systems. As enterprises move from standard dense models to reasoning-centric models based on mixture-of-experts (MoE) architectures and longer generation cycles, datacenter infrastructure must support greater compute, memory and networking demands. MLPerf Inference v6.1 offers an audited, objective basis for comparing these capabilities under consistent operating scenarios. Signal65, a leading independent third-party AI benchmarking and validation firm, analyzed the audited published results for Microsoft Azure’s DeepSeek-R1 submissions and compared them with other Available systems at matched GPU generation and count. This is Azure’s first DeepSeek-R1 submission to MLPerf, and it includes a measured four-rack, 288-GPU ND GB200 v6 system, built on NVIDIA GB200 NVL72, alongside single-rack ND GB200 v6 and ND GB300 v6 results. Azure is the only cloud provider in this round with a multi-rack NVL72 result for DeepSeek-R1, which gives customers a direct measurement of multi-rack capacity.

Microsoft Azure’s MLPerf Inference v6.1 results support Azure as an inference platform for both single-rack deployments and larger installations, combining Interactive leadership, competitive single-rack scale-up throughput and nearly linear scale-out across four racks. Together, these results give infrastructure teams two measured Azure capacity points, one for a competitive single-rack platform and one for a four-rack, 288-GPU deployment. The clearest scale-out result comes from Azure ND GB200 v6, where expanding from 72 to 288 GPUs delivers 3.75x the Server throughput and 3.86x the Offline throughput, equal to 93.6% and 96.4% of ideal fourfold scaling efficiency, respectively.

Azure also leads the MLPerf results for DeepSeek-R1 Interactive throughput. Its 72-GPU ND GB300 v6 submission reaches 260,098 output tokens per second, which is 10.3% above the next-highest cloud submission and 2.6% above NVIDIA’s own GB300 reference. These comparisons use the same model, scenario, GPU generation and GPU count. MLPerf measures three scenarios. Interactive applies the tightest latency limits and is most relevant to applications where users wait on generated output, such as conversational tools and agentic coding assistants. Server measures throughput as requests arrive over time, and Offline measures batch throughput.

Four implications stand out for organizations evaluating inference platforms:

Near-linear scaling beyond one rack. The measured 288-GPU GB200 system produces 1.60 million tokens per second in Server and 2.09 million in Offline. Azure retains most of its single-rack throughput per GPU as the tested system grows from one to four racks.
Leadership in strict interactive tests. Azure’s GB300 Interactive score is the highest among the five Available Interactive submissions analyzed. Interactive applies substantially tighter response-time limits than the Server category.
Strong results at same GB200 scale. Azure leads all four Available 72-GPU GB200 systems in both Server and Offline. Its margins over the closest cloud submission, from a large neocloud provider, are 1.7% and 2.3%, respectively.
Higher throughput with GB300. At 72 GPUs, Azure GB300 increases Server throughput by 38.1% and Offline throughput by 22.0% over Azure GB200 in the same benchmark round.

These findings are specific to the MLPerf scenarios analyzed in this report, each of which uses different measurement conditions and, for Server and Interactive, different latency service level objectives (SLOs). Azure’s 72-GPU GB300 Server and Offline results sit within 2.3% and 4.2% of the highest scores in a four-system peer group whose results span less than 5%. All results in this report come from MLPerf’s Available category, which excludes preview and research systems.

All throughput figures are aggregate output tokens per second. Selected result IDs, calculations and five public references appear in the appendix.