CoreWeave Brings Nvidia Vera Rubin to Its Cloud, With Cognition First in Production

At its Fully Connected conference on September 30, CoreWeave announced that Nvidia’s Vera Rubin NVL72 was available on its cloud, with Cognition identified as the first customer running production workloads on the system. The announcements also covered a planned standalone Vera CPU offering and CoreWeave Forge, a platform for developing and improving models and agents. 21733

The performance headline needs a precise caveat: Nvidia reported that a CoreWeave benchmark on DeepSeek-R1 achieved 10 times the throughput per megawatt of GB200 NVL72. That is a result for a particular workload and metric—not evidence that every application runs 10 times faster. 10

Vera Rubin NVL72: available on CoreWeave, with Cognition first

CoreWeave said Cognition was the first customer anywhere running production workloads on Vera Rubin NVL72. It also said customers could use the system with the same operating model and tooling as their existing GB200 NVL72 and GB300 NVL72 fleets, supported by CoreWeave performance engineering. 2

The benchmark figure is separate from that availability announcement. Nvidia described CoreWeave’s DeepSeek-R1 result as a 10× improvement in tokens per second per megawatt compared with Grace Blackwell NVL72. Because the result is tied to DeepSeek-R1 and power-normalized throughput, it should not be read as a universal performance comparison across models or workloads. 10

Vera CPU and sandbox results: what the available information says

CoreWeave said it would offer Nvidia Vera, a CPU designed for AI-agent workloads. The announcement establishes the product’s intended role, but the provided sources do not report a numerical figure for how many agents it can support or a benchmark of its standalone performance. 17

CoreWeave has described its Sandboxes as secure, isolated environments for reinforcement learning, agent tool use and model evaluation. The source provided about that capability does not give a sandbox performance result, so there is no substantiated figure to compare here. 35

Forge connects model and agent development

CoreWeave launched Forge as a development layer for teams building and improving models and agents. The company describes it as bringing training, inference, evaluation and agent development into one environment, open across models, frameworks and other clouds. 33

The available Forge announcement details the platform’s purpose and scope but does not identify early adopters. That leaves an important distinction: the company named Cognition as a Vera Rubin production customer, but the provided sources do not establish that Cognition or another named customer adopted Forge at launch. 233

Intrator’s view: orders, data-center limits and business outlook

CEO Mike Intrator said CoreWeave had seen “no reduction in orders at all.” He argued that restrictions on data-center development could affect where infrastructure is built without eliminating demand. These are his statements about the company’s orders and his view of the market, not an independent measure of industry-wide demand. 446

On the broader outlook, Intrator expressed confidence in the economics of additional AI infrastructure through 2027 and into early 2028. He also argued that infrastructure economics could outpace interest-rate concerns. Those comments convey management’s expectations; they are not a guarantee of future demand, margins or returns. 447