Vera Rubin GPU vs GB200 and Cognition’s 4.8x Token Means

The Vera Rubin GPU vs GB200 new AI chip usually arrives with a headline number that sounds almost too good to be true. This time the number is 4.8x. Cognition AI, the company behind the Devin coding agent, reports up to a 4.8x increase in total token throughput for its SWE-2 inference workloads on Nvidia’s Vera Rubin NVL72 compared with GB200 NVL72. That is a big claim for the Vera Rubin GPU, and it deserves a closer look. This guide explains what was measured, what the number does and doesn’t mean, and how it connects to AI inference cost, so you can read the headline with the right context.

Vera Rubin GPU vs GB200 Brings to the Rack

Vera Rubin GPU vs GB200 Nvidia Vera Rubin is the platform that follows the Blackwell-era GB200. An NVL72 rack packs 72 Rubin GPUs and 36 Vera CPUs together, so the whole rack works as one unit. That matters because modern AI hardware is sold as a whole system. In a rack-scale design, the GPUs, CPUs, and networking work together as one large machine, so performance gains come from the full stack and not just a faster chip.

The GB200 NVL72 is the baseline in this comparison. It is the previous-generation rack-scale system that many AI teams run today. CoreWeave announced on September 30 that Vera Rubin NVL72 systems were in limited availability on its cloud, with Cognition as the first customer running production workloads. So this isn’t a lab demo. It is an early look at Nvidia Vera Rubin handling a real agentic workload.

What Cognition AI Actually Measured

The details of the test matter more than the headline. The 4.8x figure applies to SWE-2 inference workloads drawn from FrontierCode tasks, measured per GPU at matched interactivity against a GB200 NVL72 baseline. Cognition also reported a second result. Output token throughput on reinforcement learning workloads was 3.8x higher per GPU than the baseline.

Three phrases in those results are worth understanding:

• “Up to” means 4.8x is a best-case ceiling, not a guaranteed average.
• “Per GPU” means the comparison is normalized by chip count, not by total rack cost.
• “Matched interactivity” means each user’s generation speed was held roughly constant between the two systems.

Why Token Throughput Is Not the Same as Speed

Throughput and speed are easy to confuse. Throughput is how much total work a system completes over a period of time. Latency is how long one request waits. A system can handle far more requests overall without making any single request finish sooner.

Because the test held interactivity constant, the 4.8x gain is best read as a capacity improvement. One GPU can serve far more tokens, and so more simultaneous coding sessions at the same per-user responsiveness. Higher token throughput does not equal a higher task success rate or greater intelligence. Devin won’t necessarily solve problems five times faster or better. It may be able to serve far more people at once.

What the Gain Means for AI Inference Cost

Agentic coding is token-hungry, with long contexts, many parallel sessions, and enormous token volumes. Cognition’s own framing centers on cost, since cost per token decides what the company can ship.

The logic is simple. If each GPU produces more tokens at the same responsiveness, the hardware cost is spread across more output, and AI inference cost per token can fall. Two things limit how far that logic goes:

• Pricing is unknown. The published results don’t include rack pricing or cloud rates, so a 4.8x throughput gain doesn’t automatically mean a 4.8x cost reduction.
• Energy matters too. CoreWeave also published results showing up to 10x more token throughput per megawatt than GB200 NVL72 on the DeepSeek R1 reasoning model at matched interactivity. Power is a major part of data center cost, so this efficiency angle may matter as much as the raw throughput figure.

For teams buying inference as a service, the lasting effect is likely to be on pricing pressure and capacity over time, not an overnight drop in bills.

Caveats Worth Keeping in Mind

Before treating 4.8x as a universal benchmark, remember the following:

• It’s company-run. The figure is an early, company-run throughput result, not proof of cheaper inference or more completed engineering work.
• It’s one workload. A coding agent with long contexts behaves differently from image generation, training, or short chatbot queries. Gains on other workloads will vary.
• It’s early. Until a second customer publishes numbers, 4.8x is one lab’s early reading on one coding workload.

The practical approach is to treat the result as a strong signal and not as settled fact.

Final Thought

Cognition’s result is a meaningful early data point. It moves the conversation about the Vera Rubin GPU from specification sheets to a named customer running a live agentic workload. The 4.8x gain is best understood as a jump in capacity at matched interactivity, measured on one workload against GB200 NVL72, with real potential to lower AI inference cost once pricing and availability settle.

If you’re planning AI infrastructure, an important move is to watch for independent benchmarks, ask providers for cost per token on your own workload, and avoid reading the Vera Rubin GPU’s headline number as a promise of a five-times-faster experience. The hardware clearly has momentum. The proof will come as more customers test it.