Vera Rubin NVL72 Delivers 5x More Tokens per GPU Than GB300 on DeepSeek V4.1 Flash

Everyone is used to comparing accelerators with a single number: "X times faster." The latest Artificial Analysis data on Vera Rubin NVL72 shows well why one number is not enough. NVIDIA has submitted updated Rubin configurations for the DeepSeek V4.1 Flash model, and the outcome depends on what speed you promise to each user or agent.

What exactly was measured

The metric here is throughput per GPU (tokens per second) at a given speed per agent. This is an important nuance. You can load a card so that it produces the maximum total tokens, but then each individual request will run slowly. Or you can maintain a high speed per agent, and then total output per GPU drops. Every platform lives on a curve between these extremes, and what you need to compare are points at the same speed.

The comparison is against improved GB300 results, that is, against a generation that has already been optimized in software. This is fairer than comparing against its launch figures.

The numbers

At 260 tokens per second per agent, Vera Rubin delivers 5.0x more tokens per GPU than GB300. At 80 tok/s the gap narrows to 1.7x. For the single-agent case the result is: 1,249 tok/s for Rubin versus 848 for GB300, or 1.47x.

The picture is clear. The higher the speed that must be guaranteed to each agent, the more the new architecture wins. At low speeds the advantage is more modest; at high speeds it is a multiple. The post does not explain why, and it is not worth guessing at the causes without configuration details. But the practical takeaway does not change.

Why this matters for agents

Agentic workloads are sensitive precisely to per-session speed. An agent makes dozens of sequential calls: it plans, calls tools, checks the result, rewrites. If each step is generated slowly, the delays add up, and the user waits minutes instead of seconds. That is why the "fast and for everyone" mode is closer to reality for agents than the "maximum total tokens" mode, which suits batch processing.

In this mode, judging by the numbers, Rubin comes into its own: 5x per GPU means the same volume of fast agent sessions can be served by a noticeably smaller fleet of cards. For an inference provider this directly affects the cost per token and which speed-guaranteed tiers it can afford to offer.

Limitations

A few things are worth keeping in mind. First, these are results submitted by NVIDIA itself, and the configurations are tuned for a specific model, DeepSeek V4.1 Flash. With a different model or context size the ratios may differ. Second, the post reports ratios, not price: how much an NVL72 rack on Rubin costs and when it will be available in clouds does not follow from the data. Third, 1.47x for a single agent shows that for low-concurrency workloads the upgrade has a much smaller effect.

What to do in practice

If you are building a product on agents and rent inference yourself, ask your provider not "how many tokens per second," but "how many per GPU at my target per-agent speed." Determine what speed you need: 80 tok/s and 260 tok/s give completely different economics. If you are planning purchases or long-term contracts, assume that the gain from the new generation is nonlinear and shows up most strongly in fast interactive scenarios. And if your workload is background batch processing, waiting for Rubin to get a multiple-fold effect is most likely not worth it: there the difference will be closer to moderate.