Nvidia has claimed the top position in AgentPerf, the first industry benchmark specifically designed to evaluate AI hardware performance on agentic AI workloads — autonomous AI systems that must reason, plan, use tools, and take actions over extended periods. The result, achieved with the GB300 NVL72 system, demonstrates that Nvidia’s Blackwell Ultra architecture delivers decisive advantages not just in traditional AI training and inference, but in the emerging agentic AI workloads that are rapidly becoming the dominant use case for enterprise AI deployments.
AgentPerf was developed to address a measurement gap in the AI benchmarking ecosystem. While existing benchmarks like MLPerf evaluate static training and inference performance, agentic AI workloads have fundamentally different characteristics: they run for extended periods, make multiple sequential model calls, interact with external tools and databases, maintain state across interactions, and must deliver consistent low latency throughout. Standard benchmarks that measure throughput in isolation don’t capture these requirements adequately.
Data from SemiAnalysis’s InferenceX benchmark provides quantitative support for the AgentPerf results. The GB300 NVL72 delivers up to 50 times higher throughput per megawatt and 35 times lower cost per token compared with the previous Hopper generation, with the sharpest gains at low latency where agentic applications operate. These efficiency gains are particularly significant for agentic AI because agents typically make many sequential model calls per task, making per-call latency and cost critical operational metrics.
The AgentPerf benchmark results have practical implications for enterprises investing in AI infrastructure. As organizations move from deploying AI models for isolated tasks to building AI agent systems that autonomously handle complex workflows, the hardware infrastructure they choose will directly determine the responsiveness, cost, and scalability of their AI capabilities. The benchmark gives buyers a meaningful technical basis for evaluating infrastructure investments specifically for agentic use cases.
Nvidia’s performance leadership in agentic AI benchmarks extends its competitive moat beyond the training workloads where it has historically dominated. By demonstrating that the same GB300 hardware that wins training benchmarks also leads in agentic inference, Nvidia reinforces the value proposition of a unified AI infrastructure platform — organizations can use the same hardware for training their models and deploying them as agents, simplifying procurement and operational management.
The emergence of agentic AI benchmarking as a discipline marks a maturation of the AI industry’s approach to measuring what matters. As AI agents become embedded in enterprise workflows — managing customer interactions, automating business processes, conducting research, and making decisions — the performance characteristics captured by AgentPerf become directly relevant to business outcomes. Nvidia’s AgentPerf leadership suggests that its infrastructure investments are well aligned with where enterprise AI demand is heading.
