CoreWeave Launches Forge, Brings Vera Rubin NVL72 Online and Says It Will Offer Vera CPU Racks for Agent Sandboxes
At its Fully Connected event, CoreWeave announced Vera Rubin NVL72 availability, plans to offer NVIDIA Vera CPUs, and the Forge platform for training and agent improvement.
Overview
CoreWeave used its Fully Connected event to announce three infrastructure moves. According to NVIDIA’s blog, published September 30, 2026, the cloud provider announced availability of NVIDIA Vera Rubin NVL72 systems with Spectrum-X 102.4T Ethernet networking, said it will also offer the NVIDIA Vera CPU, and launched CoreWeave Forge, “a connected environment for training, evaluating and improving models and agents on NVIDIA accelerated computing.” The event was held in San Francisco, per the same post.
What We Know
Vera Rubin NVL72 and the Cognition benchmark
NVIDIA’s blog says Cognition, the applied AI lab behind the Devin AI software engineer, is the first customer running production workloads on Vera Rubin. In early tests, according to the post, Cognition saw Vera Rubin NVL72 deliver up to a 4.8x increase in total token throughput for SWE-2 inference workloads over GB200 NVL72. The post says Cognition generated the workload by sampling a subset of tasks from FrontierCode and deploying AI agents to solve them.
The post also says CoreWeave stood up a production Vera Rubin cluster for Cognition in days, and that Cognition scaled to thousands of GPUs on CoreWeave in nine months. Vera Rubin capacity can be operated through CoreWeave Kubernetes Service, SUNK, CoreWeave Mission Control, CoreWeave Sandboxes and CoreWeave Inference, NVIDIA wrote.
Vera CPU for agent environments
CoreWeave’s planned Vera deployment puts 128 CPUs and 11,264 cores in a single rack, which the post describes as enough for more than 11,000 concurrent environments at one core each. In CoreWeave’s own testing, according to NVIDIA’s blog, agent sandbox startup was more than 3x faster on Vera CPUs, and CoreWeave saw a 1.7x performance gain on Terminal-Bench across all passing tasks. The Vera chip was the subject of earlier Machine Herald coverage when NVIDIA hand-delivered the first Vera CPUs to AI labs.
CoreWeave Forge
NVIDIA’s blog says Forge unifies Weights & Biases, post-training expertise from OpenPipe and the open source marimo notebook project, and that it stays open across models, frameworks and clouds. Components named in the post:
- CoreWeave ARIA, now generally available, which analyzes runs and proposes experiments.
- CoreWeave Agent Lens, a new service that the post says improves failure detection by 20% and fixes issues at half of the cost.
- CoreWeave Sandboxes, now generally available, for running agents, tool calls, reinforcement learning (RL) and evaluations in isolated CPU or GPU environments.
- Serverless RL, which the post says trains 1.4x faster at 40% lower cost than a self-managed setup.
The post adds that NVIDIA Dynamo, an open source inference framework, powers CoreWeave’s managed inference service as well as RL Rollouts, now in private preview. According to NVIDIA, RL Rollouts loads new checkpoints into a live deployment while it is running, so reinforcement learning continues without redeploys. Canva, Capital One and MasterClass are among the first companies building on Forge, per the same post.
Strategic reading
SiliconANGLE’s keynote analysis quoted Vellante, chief analyst at theCUBE Research, saying: “What you’re seeing CoreWeave do strategically is they’re expanding out beyond compute.” He pointed to networking, storage and a software layer, and said CoreWeave announced Forge that day. SiliconANGLE disclosed that theCUBE is a paid media partner for the Fully Connected event.
What We Don’t Know
- The performance figures (4.8x, 3x, 1.7x, 1.4x, 20%) come from NVIDIA’s blog post, which attributes them to Cognition or to CoreWeave’s own testing. The sources reviewed do not describe independent verification.
- The NVIDIA post says CoreWeave “will also offer” Vera; the sources reviewed do not give a general availability date or pricing for Vera or for Vera Rubin capacity.
- The sources reviewed do not describe how Forge is priced or how its components are packaged commercially.
Analysis
Taken together, the announcements show a cloud provider extending from raw accelerators into the tooling around model training and agent operation: sandboxes for RL and evaluation, observability for agent traces, and serverless post-training. NVIDIA’s post frames this as closing a loop that “has historically been split across tools from different vendors.” Whether customers adopt a single vendor’s loop, or keep it open as the post says Forge does, will depend on details not yet published in the sources reviewed.