Is the AI Loop, Not Just the GPU, CoreWeave’s Real Product Now?

Key Highlights

• CoreWeave launched Forge at Fully Connected 2026, a single development layer that runs the full improvement loop for models and agents, open across models, frameworks, and other clouds.
• Caterpillar's Brandon Hootman spoke on the company's relationship with CoreWeave on physical AI, and CEO Mike Intrator has described Caterpillar as building its own AI clusters with CoreWeave.
• Capital One presented an evaluation-first architecture for multi-agent systems, combining deterministic checks and LLM judges under human oversight, a close match to the evaluate stage at the center of Forge.
• NVIDIA Vera Rubin NVL72 is in limited availability on CoreWeave, with Cognition the first customer in production, reporting up to 4.8 times the total token throughput per GPU at matched interactivity versus a GB200 NVL72 baseline on its own SWE-2 inference benchmark.
• With Caterpillar and Capital One sharing the stage with Cognition, CoreWeave appears to be pitching Forge to enterprises that want the improvement loop run alongside their capacity.

The News

CoreWeave used its Fully Connected 2026 conference in San Francisco (September 29 to October 1) to launch CoreWeave Forge, a development layer that runs the full AI improvement loop in one environment, open across models, frameworks, and clouds. Enterprise customers carried much of the keynote program, with Caterpillar on physical AI and Capital One on evaluation-first agent architecture. They were joined by Cognition, the first customer in production on NVIDIA Vera Rubin NVL72, now in limited availability on CoreWeave. The announcements, which also included plans to offer the NVIDIA Vera CPU and a new CoreWeave Partner Network, appear designed to extend CoreWeave's customer relationships from capacity into the software layer where models and agents get improved, with enterprise buyers as a primary audience. Details are in CoreWeave's Forge announcement: https://www.coreweave.com/news/coreweave-forge-launches-turning-the-ai-loop-production-run-into-a-better-model-and-agent

Analyst Take

Our team, Steven Dickens and Stephen Sopko, attended Fully Connected in person at Moscone South, and the customers on the keynote stage set the tone: Caterpillar, Capital One, and Cognition, very different businesses making the same case that AI has to earn its place in production every day. The questions we heard most in the hallways were about iteration speed: how quickly a production failure becomes a fix, and who owns that path. CoreWeave, which built its name on getting NVIDIA systems into production before almost anyone else, used the stage to claim more of that question, and it chose enterprise names to make the argument.

The skeptic's case is worth stating plainly. Caterpillar and Capital One both appear to run deep in-house engineering organizations, so their presence may say more about AI-native habits inside a handful of enterprises than about the broad market. Forge, meanwhile, has no disclosed revenue figure. We think that reading is overly simplistic. Enterprise platforms tend to spread through their most technical buyers first; operators who lived through the early database and virtualization cycles will recognize the pattern. Pieces of Forge such as serverless post-training and a free tier appear aimed at the next wave of teams with thinner benches, and the first measure of success is likely to be how much of a customer's workflow lands on the platform.

What Was Announced

Forge is a wiring decision first. It brings together: ARIA (now generally available), an agent that reads experiment and observability data and proposes the next runs worth doing; Weights & Biases Models for experiment tracking; Notebooks built on marimo; Sandboxes (also generally available) for isolated agent, RL, and evaluation runs; Registry for versioned checkpoints and agent configurations; post-training services spanning serverless RL, serverless fine-tuning, and model distillation; and inference, now folded into Forge.

Agent Lens and Model Distillation arrive in preview. Free, Pro, and Enterprise editions let a team start without a procurement cycle, which matters more to enterprise adoption than any single feature.

Registry stores artifacts in open, portable formats with lineage, a deliberate choice on openness that we think fits the teams CoreWeave courts, many of whom have lived with proprietary MLOps stacks before. RL Rollouts, in preview within Dedicated Inference, hot-load new checkpoints into a live deployment so a reinforcement learning loop keeps turning without a redeploy. For agent builders, redeploy friction is where iteration speed quietly dies. Agent Lens classifies production failures, and CoreWeave's launch release reports it catches 20 percent more critical failures at one-tenth the cost of a general-purpose frontier LLM (the company's Forge blog cites the same detection gain at half the cost, without that comparator), so we expect customer deployments to settle the economics quickly. Anyone who has run a model team knows the real integration layer is usually a shared spreadsheet and one tired staff engineer. Forge is aimed squarely at that person.

Below the software, Cognition runs Vera Rubin NVL72 under the same operating model and tooling as its GB200 and GB300 NVL72 fleets, with the platform currently in limited availability. CoreWeave still runs NVIDIA Volta GPUs in commercial service, nearly a decade on, under the same platform that now hosts Vera Rubin. For a customer, that continuity is what makes adopting new silicon a matter of days. The planned Vera CPU offering targets the CPU-bound side of agentic work (sandboxes, tool calls, RL environments), running bare metal beside the training fleet under the same consumption model. The Partner Network rounds out the picture with integrations CoreWeave says are tested under production conditions, from partners including CrowdStrike, VAST Data, and ClickHouse. Separately, Exa, Parallel Web Systems, and You.com were announced at Fully Connected as search partners for agents, with integration work following in the weeks after.

Market Analysis

Two competitors in the space moved above the rack in the same week. One day after Forge launched, Nebius announced its acquisition of Inferize, folding cold-start and capacity-elasticity technology into its Token Factory inference platform. Different layer, same instinct. As next-generation racks reach multiple providers, the contest moves to the software around them, and CoreWeave's bet sits further up the stack than most, in the improvement loop, one step past serving. In the GPU business, capacity is the harvest; Forge seems to be CoreWeave's bid to own the irrigation.

The customer roster may be the more durable signal from the week. CoreWeave gave Caterpillar and Capital One prominent keynote time, and both brought workloads far from the frontier-lab profile that built its business. Caterpillar's Brandon Hootman, who leads physical AI platforms and construction autonomy, have a real world perspective on physical AI, a workload measured on jobsites, far from the chat window. Intrator has described Caterpillar as building its own AI clusters with CoreWeave. Capital One's Maulin Patel walked through an evaluation-first architecture for multi-agent systems, grading outputs with deterministic code and LLM judges under human oversight. In a regulated bank, an agent that cannot show how it was graded is unlikely to ship, and that discipline sits close to the evaluate stage Forge is built around. Ennoble Care's selection of CoreWeave for clinical inference adds a healthcare example.

For enterprise buyers, the open posture lowers the technical barrier, though not the organizational one. CoreWeave's own Forge material describes the split the product has to bridge: research teams that know what good output looks like, and application or SRE teams that own production and watch uptime dashboards. Forge runs across other clouds, so the estate can stay where it is; the harder question is which of those two teams signs for it. Capital One's emphasis suggests one answer. In regulated industries, the team that owns the grading may end up owning the loop.

On its August earnings call, CFO Nitin Agrawal disclosed an approximately 25 percent price increase across SKUs, implemented in July. Sell-side checks, led by UBS, point to a further roughly 10 percent in the two to three months since. A provider able to raise prices into demand does not need software to support margins, so Forge looks additive: a second layer of value on top of a capacity business that Intrator, in a Bloomberg interview recapped by TheStreet, described as largely sold out for years. He also said customers had effectively pulled CoreWeave up the stack, and that Forge should be accretive to margins.

Intrator argued that proposed data center moratoriums shift where capacity gets built without reducing demand for compute. In the week after Fully Connected, CoreWeave put that thesis into practice with AdaniConneX: a planned 240 MW deployment in Navi Mumbai running Vera Rubin, with the first phase expected in mid-2028. The timeline signals capacity reserved years ahead of need, and India becomes a second Asia-Pacific market expansion for CoreWeave after Indonesia.

Looking Ahead

The key trend we'll be monitoring is whether enterprises like Caterpillar and Capital One deepen from infrastructure into Forge, and whether less engineering-heavy enterprises follow them. The earliest evidence should surface in CoreWeave's upcoming third-quarter results, where any disclosure on Forge adoption or enterprise mix, even qualitative, would be telling. Agent Lens and Model Distillation exiting preview will offer a natural checkpoint for whether trial usage converts to paid usage, and a published Agent Lens methodology would resolve the cost figures. On the hardware side, we will be watching for Vera Rubin to move past limited availability. A second named production customer would suggest the Cognition result extends to other agentic workloads. Our read is that the strongest signal will come from customers and partners speaking for CoreWeave. If an enterprise outside the early-adopter tier, or partners such as CrowdStrike and VAST Data, begin citing production wins inside Forge workflows, the platform claim gains outside validation. If they stay quiet through the first half of 2027, that silence will be informative too.