Key Takeaways
- • NVIDIA Vera Rubin NVL72 delivers 30x higher inference throughput per megawatt compared to GB300 NVL72 on agentic AI workloads, with measured performance data from CoreWeave demonstrating 10x tokens-per-second per megawatt versus Blackwell NVL72 systems.
- • Token cost efficiency improves 35x on Vera Rubin NVL72 relative to GB300 NVL72, directly addressing the infrastructure challenge posed by agentic AI workloads that consume 15x more tokens than standard chat applications due to multi-step reasoning, tool invocation, and sub-agent spawning.
- • Vera Rubin platform represents a rack-scale, pod-scale AI supercomputing system engineered for pretraining, post-training, reinforcement learning, and agentic AI with advanced liquid cooling, GPU/CPU integration, and RoCE fabric architecture designed for long-context handling and distributed agent workflows.
- • Performance validation used SemiAnalysis AgentX workload consisting of recorded real-world agentic coding sessions, preserving actual context growth, tool calls, and sub-agent spawning patterns to benchmark production-grade agentic AI infrastructure requirements.
- • For power-constrained AI factories, the 30x efficiency gain translates to 30x more agentic workload capacity per megawatt of energy consumption, positioning Vera Rubin as critical infrastructure for enterprise-scale agent deployment across financial research, software development, and customer service applications.
Summary
NVIDIA introduced the Vera Rubin NVL72 platform as a next-generation AI supercomputing architecture purpose-built for agentic AI workloads. The system achieves breakthrough efficiency metrics that substantially exceed the current-generation Blackwell GB300 NVL72, with measured performance data indicating 30x higher inference throughput per megawatt and 35x lower token costs on agentic AI tasks. These gains directly address a fundamental infrastructure challenge: agentic AI workloads consume approximately 15x more tokens than traditional chat or document summarization applications, according to OpenRouter data cited by NVIDIA.
The performance differential stems from the nature of agentic AI execution patterns. When an AI agent tackles complex tasks—such as investment research, software development, or customer service—it employs multi-step reasoning chains where each step generates output that becomes the input for subsequent operations. Sub-agents are spawned for specialized subtasks, context windows expand as the agent accumulates reasoning history, and tool invocations multiply token demand across the workflow. A financial research agent, for example, might query databases, search news and regulatory filings, invoke peer comparison sub-agents, and synthesize findings before delivering a final recommendation—each step compounding token consumption. This architectural reality necessitates infrastructure optimized for sustained high-throughput token processing under power constraints.
NVIDIA validated Vera Rubin's efficiency claims using the SemiAnalysis AgentX workload, a benchmark consisting of recorded real-world agentic coding sessions that preserve actual context growth, tool invocation patterns, and sub-agent spawning behavior. Testing on CoreWeave infrastructure demonstrated that Vera Rubin NVL72 delivers 10x tokens-per-second per megawatt compared to Blackwell NVL72 on identical DeepSeek R1 workloads, establishing measured performance parity across independent cloud infrastructure platforms.
The Vera Rubin platform itself represents a significant architectural redesign. Built as a multi-rack pod-scale system, it integrates high-performance GPU and CPU components, implements advanced liquid cooling to manage thermal density, and employs RoCE (RDMA over Converged Ethernet) fabric for low-latency distributed communication—essential for coordinating multi-agent workflows and maintaining efficiency across scaled deployments. NVIDIA's engineering approach reflects what the company describes as "extreme co-design," uniting hardware and software optimization throughout the stack. The platform supports the full spectrum of AI infrastructure requirements: pretraining, post-training, reinforcement learning for agent improvement, and test-time scaling—the repeated computational passes agentic systems require to refine reasoning quality before task completion.
For enterprise deployment, the efficiency implications are material. Data centers operating under power budgets can effectively deploy 30x more agentic AI workload capacity per megawatt using Vera Rubin compared to prior-generation systems. As agentic AI transitions from research prototypes into production systems across financial services, enterprise software, and customer operations, infrastructure capable of handling the token economics and latency requirements becomes strategically critical.
The performance differential stems from the nature of agentic AI execution patterns. When an AI agent tackles complex tasks—such as investment research, software development, or customer service—it employs multi-step reasoning chains where each step generates output that becomes the input for subsequent operations. Sub-agents are spawned for specialized subtasks, context windows expand as the agent accumulates reasoning history, and tool invocations multiply token demand across the workflow. A financial research agent, for example, might query databases, search news and regulatory filings, invoke peer comparison sub-agents, and synthesize findings before delivering a final recommendation—each step compounding token consumption. This architectural reality necessitates infrastructure optimized for sustained high-throughput token processing under power constraints.
NVIDIA validated Vera Rubin's efficiency claims using the SemiAnalysis AgentX workload, a benchmark consisting of recorded real-world agentic coding sessions that preserve actual context growth, tool invocation patterns, and sub-agent spawning behavior. Testing on CoreWeave infrastructure demonstrated that Vera Rubin NVL72 delivers 10x tokens-per-second per megawatt compared to Blackwell NVL72 on identical DeepSeek R1 workloads, establishing measured performance parity across independent cloud infrastructure platforms.
The Vera Rubin platform itself represents a significant architectural redesign. Built as a multi-rack pod-scale system, it integrates high-performance GPU and CPU components, implements advanced liquid cooling to manage thermal density, and employs RoCE (RDMA over Converged Ethernet) fabric for low-latency distributed communication—essential for coordinating multi-agent workflows and maintaining efficiency across scaled deployments. NVIDIA's engineering approach reflects what the company describes as "extreme co-design," uniting hardware and software optimization throughout the stack. The platform supports the full spectrum of AI infrastructure requirements: pretraining, post-training, reinforcement learning for agent improvement, and test-time scaling—the repeated computational passes agentic systems require to refine reasoning quality before task completion.
For enterprise deployment, the efficiency implications are material. Data centers operating under power budgets can effectively deploy 30x more agentic AI workload capacity per megawatt using Vera Rubin compared to prior-generation systems. As agentic AI transitions from research prototypes into production systems across financial services, enterprise software, and customer operations, infrastructure capable of handling the token economics and latency requirements becomes strategically critical.
Go deeper than the headline
Research every firm in this story, track the ones that matter and get the daily briefing — free for 7 days.