Artificial Analysis published the first AA-AgentPerf results for NVIDIA’s Vera Rubin NVL72 on October 15, 2026. In its Vera Rubin NVL72 benchmark, the rack serves up to about 5.7 times the total throughput per megawatt of the GB300 NVL72 it replaces on DeepSeek V4 Pro, about 5.4 times on DeepSeek V4.1 Flash, about 3.0 times on GLM-5.2 and about 2.8 times on Kimi K3. Each is the largest gap at one per-user speed; at peak throughput the lead is 1.14 to 1.45 times.
Vera Rubin NVL72 is NVIDIA’s rack-scale system of the Rubin generation, with 72 Rubin GPUs and 36 Vera CPUs in one NVLink domain. What NVIDIA has announced about the platform is collected in our Nvidia Rubin explainer, and the chip sits alongside its predecessors in our AI chips table.
On this page
What AA-AgentPerf measures
AA-AgentPerf is a hardware benchmark built around coding agents rather than chat. According to Artificial Analysis’s methodology, it replays real coding-agent trajectories, produced by agents resolving issues in public code repositories, at a range of concurrency levels. Each simulated agent works through three trajectories in sequence and pauses after each reply for a tool-call time taken from real tool-call durations.
That traffic looks different from fixed-length prompts. Requests carry up to 132K input tokens, replies are often a short tool call, and on average 97% of a request’s input tokens repeat the previous request. Artificial Analysis says this makes KV cache capacity and reuse, uneven prefill and time to first token decisive. Our KV cache explainer covers why a repeated prefix can be served from cache instead of being processed again.
The headline metric is total throughput (input and output tokens, about 96% of them cached input) per megawatt of measured GPU power, plotted against per-user output speed at the 25th percentile. Measured power covers each GPU and its HBM as reported by NVIDIA’s telemetry, and excludes CPUs, networking and cooling.
The results
| Model | Largest lead per MW vs GB300 NVL72 | At per-user speed | Lead at peak throughput (tokens/s per MW) |
|---|---|---|---|
| DeepSeek V4 Pro | ~5.7x | ~190 tok/s | 1.14x (272M vs 239M) |
| DeepSeek V4.1 Flash | ~5.4x | ~580 tok/s | 1.45x (571M vs 395M) |
| GLM-5.2 | ~3.0x | ~165 tok/s | 1.35x (272M vs 201M) |
| Kimi K3 | ~2.8x | ~67 tok/s | 1.26x (70M vs 56M) |
Source: Artificial Analysis, results as of October 15, 2026.
The pattern is that Vera Rubin’s advantage grows where the older rack’s throughput falls away as each user is served faster. Artificial Analysis gives one example: going from about 120 to about 165 tokens per second per user on DeepSeek V4 Pro costs GB300 NVL72 81% of its throughput per megawatt, against 16% for Vera Rubin NVL72.
Per GPU rather than per megawatt, the gap is wider: 2.0 to 2.5 times at peak throughput, and up to about 12 times on DeepSeek V4 Pro at roughly 165 tokens per second per user. Vera Rubin’s fastest configurations also set the highest per-user speeds on the leaderboard for DeepSeek V4.1 Flash (906 tokens per second, the fastest result for any model), GLM-5.2 (408) and Kimi K3 (255).
Who tuned the configurations
The serving configurations behind these numbers came from NVIDIA. Artificial Analysis says NVIDIA shared how it built both its Vera Rubin and GB300 configurations, and that the two sets are very similar, so most of the difference comes from the newer hardware alongside early Rubin-specific kernels and tuning. The techniques it lists include offloading KV cache to CPU memory over NVLink-C2C, session-aware routing with NVIDIA Dynamo, disaggregated prefill and decode with wide expert parallelism, speculative decoding tuned to batch size, and NVFP4 kernels tuned for Rubin. Every configuration is published in Artificial Analysis’s configurations browser.
The hardware differences Artificial Analysis lists per GPU, from NVIDIA’s own pages: the same 288 GB of memory as Blackwell Ultra but HBM4 at 19.2 TB/s instead of 8 TB/s, 3.0 TB/s of NVLink bandwidth instead of 1.8 TB/s, and 35 instead of 15 PFLOPS of dense NVFP4 compute. Measured GPU power in these results ranged from 0.5 to 1.9 kW on Vera Rubin and 0.3 to 1.0 kW on GB300.
How to read the numbers
These are first results, measured on four models and on configurations prepared by the chipmaker, and the largest multiples are taken at single points on each speed curve. Artificial Analysis’s argument for the per-megawatt view is that AI data centers are increasingly limited by the power they can draw rather than the number of accelerators they can install, so efficiency at a given user speed decides how many agents a site can serve. Our explainer on how much power AI data centers use gives the background, and the first production Vera Rubin NVL72 deployment we reported is CoreWeave’s with Cognition.
Scores published by labs and benchmark maintainers are collected on our LLM leaderboard; AA-AgentPerf measures serving hardware, not model quality, so its results are not comparable with model benchmark scores.
