Nvidia says Vera Rubin delivers 30x more AI work per watt

Nvidia ↗ 2.78% on Monday released performance data showing its next-generation Vera Rubin NVL72 platform delivers up to 30 times higher inference throughput per megawatt than its current GB300 NVL72 system on agentic AI workloads, while cutting the cost per million tokens by up to 35 times. The company simultaneously presented the architecture of its new Vera CPU at the Hot Chips 2026 conference and introduced Scale-In as a fifth pillar of its AI networking infrastructure.

Agentic Efficiency at Scale

The throughput claims were measured using the SemiAnalysis AgentX benchmark, which captures real-world agentic coding sessions including context growth, tool calls, and sub-agent spawning. Running the DeepSeek V4 Pro model, Vera Rubin NVL72 achieved the 30x improvement over GB300 NVL72, which itself delivers up to 15 times better throughput per megawatt than the prior Hopper architecture. The results are pending SemiAnalysis review and do not yet reflect Vera CPU performance for tool calling.blogs.nvidia+1

The gains matter because agentic workloads consume roughly 15 times more tokens than a simple chat request, according to OpenRouter data cited by Nvidia. Combined with Nvidia's DSX MaxLPS power management technology — which can provision up to 40 percent more GPUs within the same megawatt budget — the platform targets a fundamental constraint facing AI data center operators: how much useful work each available megawatt can deliver.developer.nvidia+1

Vera CPU Architecture Unveiled

At Hot Chips 2026, Nvidia detailed the Vera CPU's 88-core design built on its custom Olympus core architecture. The chip prioritizes single-threaded performance and low latency over core count, a departure from the industry trend toward weaker but more numerous cores. Vera features a 10-wide decode front-end, spatial multithreading, eight LPDDR5X memory controllers delivering up to 1.2 TB/s of bandwidth, and a six-die construction with NVLink-C2C interconnect.nvidianews.nvidia+1

"Agentic AI requires a new kind of computing system — one built not only to generate answers, but to take action," said Ian Buck, vice president of hyperscale and high-performance computing at Nvidia.nvidianews.nvidia

Scale-In Networking and SpaceXAI Partnership

Nvidia also introduced Scale-In network infrastructure as the fifth pillar of its AI networking, powered by BlueField-4 DPUs and designed to secure, manage, and operate agentic AI factories at scale. The announcement complements Nvidia's existing Scale-Up, Scale-Out, Scale-Across, and storage networking pillars.developer.nvidia

In a separate announcement, SpaceXAI said it will deploy Vera CPUs across its AI infrastructure for Grok and plans to extend an optimized Vera Rubin NVL72 system into orbit aboard its first-generation Starmind AI satellite.nvidianews.nvidia