Blackwell Ultra Tops AgentPerf, Runs 20x More Agents per MW

Artificial Analysis published the first public AgentPerf results and the NVIDIA Blackwell Ultra NVL72 platform led the initial round, with published figures highlighting an efficiency advantage: the NVL72 ran 20x more agents per megawatt in the tested scenarios. The benchmark is billed as the industry’s first focused on "agentic" AI - workloads characterized by autonomous, multi-step agent interactions - and these early results are being presented as an infrastructure comparison point for developers, enterprises and cloud providers.

AgentPerf’s arrival matters because agentic AI changes where infrastructure pressure appears: instead of single-model peak throughput, successful deployments require sustained concurrency, low-latency interactions across many agents, and predictable energy consumption under continual load. The benchmark’s reported agents-per-megawatt metric directly targets that shift, giving purchasers a metric that ties software behavior to electricity and rack-level capacity planning. In the published round, NVIDIA’s NVL72 produced the top numbers across the suite of agentic workloads the benchmark exercised, a result the company and benchmark sponsor highlight as evidence of operational efficiency.

What is new here is twofold: first, a benchmark explicitly designed for agentic workloads is now available publicly; second, the NVL72’s initial showing frames efficiency - agents per unit power - as a central competitive axis. Benchmarks reshape buyer conversations; introducing a benchmark focused on agentic patterns changes what vendors optimize for and what infrastructure buyers request in RFPs. NVL72’s performance signal will push providers to present agent-density and energy metrics alongside traditional latency and FLOPS figures when arguing TCO and capacity claims.

The operational implications are concrete for data center and cloud capacity planners. Higher agent density per megawatt implies fewer racks and lower incremental power and cooling needs for a given agent fleet size, which can compress capital expenditure and reduce ongoing energy spend. For enterprises projecting tens of thousands of agents or cloud providers provisioning multi-tenant agentic services, a platform that demonstrably increases agents per megawatt could materially change server counts, PDU sizing and chilled-water load projections. Those downstream effects - procurement cadence, facility upgrades and power contracts - are where the benchmark’s numbers translate into dollars.

Who is affected extends beyond chip buyers. Developers and engineering teams will need to validate that the benchmarked agent behaviors match their real-world traces: agent types, inter-agent communication patterns, and statefulness can all alter efficiency outcomes. Cloud operators and hosting providers will use early results to argue service differentiation and price point; infrastructure procurement teams will request benchmark-aligned test runs from vendors. Vendors themselves face a near-term incentive to tune drivers, libraries and orchestration stacks to improve AgentPerf figures, which could produce rapid performance gains but also complicate apples-to-apples comparisons.

The broader context is familiar: benchmarks often set agendas, but early rounds rarely settle them. Past industry-standard tests have driven both useful optimization and short-term gaming. AgentPerf’s first public outing provides a useful data point, yet it is only one round and limited in scope. Methodology details, workload diversity, and independent cross-vendor submissions will determine whether the benchmark becomes a stable procurement tool or a promotional vehicle for early leaders. Transparent disclosure of test parameters and rerunnable suites will be crucial if organizations intend to lean on AgentPerf for procurement and capacity planning.

Looking ahead, readers should watch three things: additional published rounds and cross-vendor comparisons; expanded methodology disclosure so teams can align internal testing to the benchmark; and vendor software updates that target agentic efficiency metrics. For infrastructure buyers, the immediate next step is to get hands-on: request benchmark runs that mirror production agent mixes and translate agents-per-megawatt into TCO models that include power, cooling and floor-space implications. The initial AgentPerf results put NVIDIA’s Blackwell Ultra NVL72 in a leading position for the tested scenarios, but the operational and procurement impact will depend on whether that lead persists across more tests and real-world workloads.