we_are_coded.by CODE · The world, decoded
БГ

NVIDIA · Chips

Vera Rubin is now running at CoreWeave, and first customer Cognition reports up to 4.8 times more tokens

On 30 September NVIDIA said CoreWeave had made Vera Rubin NVL72 systems available and that Cognition, the company behind Devin, is the first customer running real work on them. In early tests Cognition saw up to 4.8 times higher total token throughput for SWE-2 compared with GB200 NVL72. The figures are NVIDIA's and the customer's, not independent.

Source
NVIDIA Blog
Published
The event
Checked on
Responsible editor
Cvetelin Ivanov
How we work
MethodCorrections

Four point eight times. The number is a customer's, even though NVIDIA publishes it, and that makes a difference.

Cognition, who make Devin, have not been selling Vera Rubin to anyone. They took tasks from FrontierCode, set agents to solve them and compared against the previous GB200 generation. One customer's test on one workload, but at least it is their workload.

Fact

The facts: on 30 September 2026 NVIDIA said CoreWeave had announced availability of NVIDIA Vera Rubin NVL72 systems with Spectrum-X 102.4T Ethernet networking, and that Cognition, the company behind Devin, is the first customer running production workloads on Vera Rubin. CoreWeave received its first production racks earlier in the month. In early tests Cognition saw up to 4.8 times higher total token throughput for SWE-2 inference compared with GB200 NVL72, on a sample of FrontierCode tasks. CoreWeave will also offer the Vera CPU: 128 CPUs and 11,264 cores in a single rack, enough for over 11,000 concurrent isolated environments at one core each; in testing, agent sandbox startup was more than 3 times faster. CoreWeave also launched Forge, which brings together Weights & Biases, OpenPipe and marimo for continuous improvement of models and agents. Per NVIDIA's Ian Buck, CoreWeave's V100 GPUs are still running customer workloads nearly a decade after Volta launched.

Part of the weight is in the CPU. An agent needs a sandbox, and reinforcement learning needs thousands of sandboxes at once. A rack with more than eleven thousand isolated environments answers exactly that need.

Agent sandboxes are now counted per rack.

The line about the V100s is nice too: the old cards still work nearly a decade on. It is advertising, of course, buy today and it will run for years. But it is also a rare reminder that old iron does not die on the day the new one comes out.

The next number has to come from someone the seller is not quoting.

The visual is generated code-art, no third-party images.