CoreWeave announced NVIDIA Vera Rubin NVL72 availability and CoreWeave Forge during its Fully Connected event, according to NVIDIA’s September 30, 2026 account. Cognition, the company behind Devin, is identified as an early production customer. The announcement describes a new infrastructure offering rather than a performance guarantee for every cloud workload.
Agentic applications create demand for both model inference and environments where tools can run. That makes the connection between serving, evaluation and training operationally significant: an efficient model endpoint can still wait on a slow or unreliable execution environment.
Read the throughput claim with its baseline
NVIDIA reports that Cognition’s early software-engineering tests achieved up to 4.8 times total token throughput on Vera Rubin NVL72 compared with a GB200 NVL72 baseline. The account specifies SWE-2 inference workloads. It does not establish an equal gain in completed coding tasks, correctness or end-user response time.
Token throughput measures one part of a system. A coding agent also spends time reading files, running checks and interpreting failures. Compare complete tasks under the same tool access and acceptance criteria. Otherwise, a faster stream of tokens can be mistaken for faster delivery of a correct change.
Related sources: NVIDIA on CoreWeave’s Vera Rubin and Forge announcements ↗
Training and evaluation need repeatable environments
The announcement describes Forge as a connected environment for training, evaluating and improving models and agents, and says CoreWeave will offer NVIDIA Vera CPUs. These are related parts of the platform, but they should not be collapsed into a claim that every component is already available under identical access terms.
For an agent evaluation, preserve the environment image, dependency versions and starting state. Test isolation and cleanup between runs. If one attempt leaves files or credentials behind for the next, the score may reflect the environment’s history rather than the agent’s capability.
Related sources: NVIDIA on CoreWeave’s Vera Rubin and Forge announcements ↗
Questions for Canadian AI teams
Ask about the exact capacity, access status, regions and commercial terms for the proposed service. A named early customer does not establish immediate capacity for every new buyer. Data residency and retention questions should cover tool environments and stored evaluation artifacts as well as inference requests.
The useful comparison is cost per verified completed task at the concurrency the product needs. Include failed attempts, sandbox time, storage and operational support. The announcement expands infrastructure options; a representative trial establishes whether the platform improves a particular application.
Sources and editorial method
We use the linked original statements and reports to distinguish documented facts from our analysis. Company announcements describe their authors’ plans and claims; they are not independent verification of future outcomes. This article is dated and will be revised if material evidence changes. Read our editorial standards and corrections process.
- NVIDIA on CoreWeave’s Vera Rubin and Forge announcements ↗ — September 30, 2026; vendor account of availability, early customer and benchmark setup.