Cognition, the applied AI lab behind coding assistant Devin, just became the first company anywhere running production workloads on Nvidia’s new Vera Rubin NVL72 system, deployed through CoreWeave. The number that actually matters here isn’t the announcement itself. It’s what Cognition’s own engineers measured once they got their hands on it: a 4.8 times increase in total token throughput for SWE-2 inference workloads, compared to the previous GB200 NVL72 generation.
For a tool like Devin, that speed increase isn’t just a nice benchmark to cite. Agentic coding work is inherently sequential, each step waits on the output of the last one, so throughput gains compound rather than simply add up. Silas Alberti, Cognition’s SVP of research, described it directly: that kind of speedup translates into real work Devin actually completes, not just a faster number on a chart somewhere.
The timeline behind this is worth sitting with too. Cognition worked with CoreWeave to stand up the Vera Rubin cluster in early September, and the company’s own engineers ran the first customer-executed benchmark on the platform almost immediately after. Chen Goldberg, CoreWeave’s EVP of product and engineering, framed the speed of that rollout as the direct payoff of years spent building infrastructure that works consistently across GPU generations, rather than something the company pulled together specifically for this launch.
That consistency claim has real history behind it. CoreWeave’s relationship with Nvidia goes back to 2017, starting with the Volta generation, hardware that’s reportedly still running in commercial service on CoreWeave’s cloud today. The company’s software stack, including CoreWeave Kubernetes Service and its Mission Control platform, gives customers a single consistent way to deploy workloads whether they’re running older Nvidia hardware or adopting the newest architecture the day it becomes available.
Beyond Devin specifically, Cognition also recorded a 3.8 times boost in output token throughput for reinforcement learning workloads running on the same Vera Rubin setup. For an applied AI lab racing against competitors building similar coding assistants, that kind of infrastructure advantage can matter just as much as whatever improvements happen inside the model itself, since faster infrastructure means faster iteration cycles across the board.
