NVIDIA Vera Rubin NVL72 Sets New AI Agent Efficiency Benchmark

NVIDIA has announced itsVera Rubin NVL72platform, claiming a significant leap in efficiency for agentic AI workloads. Measured performance data indicates up to 30x higher throughput per megawatt and 35x lower token costs compared to the previous GB300 NVL72 system. This announcement positions NVIDIA (NASDAQ:NVDA) at the forefront of AI inference efficiency, a critical factor as agentic AI applications become more sophisticated and resource-intensive. Agentic workloads, which involve complex reasoning chains and extensive context handling, can consume significantly more computational resources than simpler tasks like chat queries. The Vera Rubin NVL72 performance was measured using the SemiAnalysis AgentX workload, designed to replicate real-world agentic coding sessions. This includes preserving context growth, tool calls, and sub-agent spawning, offering a more accurate benchmark for agentic AI performance. For power-constrained AI factories, this efficiency gain translates directly to performing more agentic computations within the same energy budget. The reduction in token costs, up to 35x lower per million tokens, further impacts the profitability of deploying AI agents at scale. StartupHub.ai data shows NVIDIA with a score of 82/100, placing it ahead of many competitors in the AI hardware space. For context, Broadcom (NASDAQ:AVGO) scores 83/100, while Celestial AI scores 72/100, indicating a competitive but high-performing segment. NVIDIA attributes these gains to extreme codesign across the platform's hardware and software stack. Techniques such as disaggregated serving, rate matching, large-scale expert parallelism, and distributed KV-caching are optimized for the demands of agentic AI. The platform's architecture, including the NVL72 scale-up domain, leverages next-generation NVLink interconnect technology for high-bandwidth, low-latency communication. The Vera Rubin platform is a seven-chip architecture that also incorporates the NVIDIA Vera CPU and integrates with other specialized components. This comprehensive approach to system design is aimed at maximizing inference efficiency for complex AI agents. This advancement directly challenges the cost and power efficiency metrics for companies developing or deploying large-scale AI agents. Competitors like Cerebras Systems (score 70/100) are also focused on AI compute efficiency, but NVIDIA's integrated hardware and software approach, as demonstrated with Vera Rubin, sets a high bar for performance per watt in agentic AI processing.