NVIDIA presents new AI infrastructure efficiency results at AI Infra Summit

NVIDIA presented new AI infrastructure results at the AI Infra Summit in Santa Clara. Ian Buck, the company’s vice president for hyperscale and high-performance computing, discussed Vera Rubin, DSX power management and new partner integrations.

In preliminary MLPerf Inference v6.1 results, the Vera Rubin NVL72 system delivered up to 3.7x higher throughput than NVIDIA GB300 NVL72. MLPerf is operated by the MLCommons consortium, which reviews results before publication.

The event also included results from Lambda, which reported a 23% improvement in performance per watt using NVIDIA DSX MaxLPS. Lambda ran 19 nodes within the power budget normally assigned to 16 full-power nodes and raised cluster throughput from about 4 million to 5 million tokens per second.

Vera Rubin and MLPerf results

A 288-GPU submission using four GB300 NVL72 racks reached 99% scaling efficiency. NVIDIA separately said software optimizations in MLPerf Inference v6.1 produced up to 1.6x higher performance than version 6.0.

NVIDIA describes its infrastructure as a full-stack platform covering Vera Rubin systems, Dynamo and NeMo software, NVLink, Spectrum-X Ethernet and ConnectX SuperNIC networking, plus BlueField technologies for context-memory storage and infrastructure security.

DSX manages AI factory power

Emerald AI and NVIDIA, working with Silicon Valley Power, demonstrated a flexible-load program for AI factories. The system responded to hundreds of grid signals while maintaining the performance of priority AI workloads.

DSX Flex can automatically pause lower-priority jobs during load-reduction requests and resume them afterward. It can respond to load-shedding, demand-response and electricity-pricing signals.

Partnerships and additional figures

Amazon’s Annapurna Labs is working with NVIDIA on custom NVHBM high-bandwidth memory technology. d-Matrix is integrating NVLink Fusion with Vera CPUs and Raptor XPU accelerators for low-latency inference, while Pinterest is using Blackwell and Dynamo for conversational visual discovery.

NVIDIA said Vera Rubin NVL72 can support up to 40% more GPUs within the same power budget in suitable deployment environments. Combined with Groq 3 LPX, the company also reported up to 35 times higher token throughput per megawatt than GB200 NVL72 for models exceeding 2 trillion parameters at long context.

On a 100,000-token-context Qwen 3.8 27B workload, Groq 3 LPX reached 2,529 output tokens per second per user. SemiAnalysis AgentX data showed Vera Rubin NVL72 delivering up to 30 times higher throughput per megawatt than GB300 NVL72 on the DeepSeek V4 Pro model.