CoreWeave launched Forge Wednesday, a software platform that combines its tools for running, observing, retraining and testing AI models and agents, the company announced at its Fully Connected 2026 conference. Nvidia said the same day that CoreWeave now offers Vera Rubin NVL72 systems and is adding Nvidia's Vera CPU for agent workloads.

What Happened

Forge brings together software CoreWeave already offered with new tools. The pieces include Weights & Biases Models for experiment tracking, post-training services that grew out of OpenPipe, notebooks based on the open-source marimo project, and CoreWeave's own inference service. The company describes the result as a single environment for what it calls the AI loop: deploy a model or agent, watch how it performs in production, curate data, retrain and evaluate the change.

Several components are new or newly generally available. ARIA, a coding agent that analyzes experiment and observability data and proposes next experiments, is now GA, as are Sandboxes, isolated environments for agent tool use, reinforcement learning and evaluation. Forge also adds managed Notebooks, a Registry for versioned checkpoints and agent configurations, and Agent Lens, which analyzes traces from production agents. RL Rollouts, a reinforcement-learning inference capability, is in preview.

RelatedGoogle Cloud Unveils Gemini Agent That Gets Its Own Work Identity Inside Companies

CoreWeave claims Agent Lens finds 20% more critical failures than general frontier LLMs and fixes them at a tenth of the cost, and that its serverless reinforcement learning trains 1.4 times faster at 40% lower cost than self-managed setups. Forge comes in Free, Pro and Enterprise editions with a 30-day trial. MasterClass and Canva are named early users, and Data Center Knowledge reported launch integrations with VAST Data, CrowdStrike and ClickHouse.

Why It Matters

Forge is CoreWeave's clearest attempt to sell more than GPU hours. Corey Sanders, senior vice president of product management, told Data Center Knowledge the company has shifted from focusing on AI infrastructure to offering a broad set of AI services. IDC analyst Dave McCarthy told the outlet that reaching mainstream enterprises will require a larger software ecosystem and more turnkey options than a specialist GPU cloud has typically offered.

Chen Goldberg, executive vice president of product and engineering, framed the product around the gap between a model in testing and one in production, saying teams need to find the signals that matter, improve the next version and check whether the change worked.

The Hardware Side

According to Nvidia, CoreWeave's Vera Rubin NVL72 systems use Spectrum-X Ethernet at 102.4 terabits per second. AI coding startup Cognition, maker of Devin, is the first production customer. Testing on real software-engineering tasks, Cognition saw up to 4.8 times the total token throughput of GB200 NVL72 systems for inference, Nvidia said. Cognition has scaled to thousands of GPUs on CoreWeave within nine months, and Data Center Knowledge reported it had Vera Rubin workloads running within two days.

The Vera CPU targets a different bottleneck. Agents spend much of their time running code and tools in sandboxes that need CPU cores rather than GPUs. A Vera rack holds 128 CPUs with 11,264 cores, enough for more than 11,000 isolated environments at one core each, Nvidia said, and testing showed more than three times faster sandbox startup. CoreWeave is offering Vera on bare metal, with early access starting now and broader testing to follow. Nvidia also named Capital One as a Forge customer and said Ennoble Care chose CoreWeave for clinical AI inference on RTX PRO 6000 GPUs.

What to Watch

CoreWeave has not published detailed pricing for the paid Forge tiers. The next questions are how quickly Vera Rubin and Vera CPU capacity spread across its regions, and whether Forge draws enterprise customers that are not already heavy GPU renters.