From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI

On September 30, 2026, NVIDIA published a blog post announcing that CoreWeave, at its Fully Connected conference in San Francisco, has made NVIDIA Vera Rubin NVL72 systems with Spectrum-X 102.4T Ethernet networking available on its cloud. Cognition, the applied AI lab behind the Devin AI software engineer, is the first customer running production workloads on Vera Rubin. CoreWeave will also offer NVIDIA Vera, which NVIDIA describes as the first CPU built for AI agents, and it launched CoreWeave Forge, a connected environment for training, evaluating and improving models and agents on NVIDIA accelerated computing. Taken together, the three announcements sketch one continuous chain: training, inference, sandboxed execution, and evaluation feeding back into training. A note on sourcing before going further. This analysis rests mainly on NVIDIA's own blog. Every performance figure below is an early-test result disclosed by a vendor or its customer, and none has been independently reproduced. Read them as directional signals, not settled facts.

Start with the headline number. Cognition runs training, reinforcement learning and production inference for Devin on CoreWeave, and it scaled to thousands of GPUs on the platform in nine months to power its inference workloads. Earlier this month CoreWeave received its first Vera Rubin NVL72 production racks. Shortly afterward, Cognition benchmarked Vera Rubin inference against a GB200 NVL72 baseline using a real-world software engineering workload: it sampled a subset of tasks from FrontierCode and deployed AI agents to solve them. In those early tests, Vera Rubin NVL72 delivered up to a 4.8x increase in total token throughput for SWE-2 inference workloads over GB200 NVL72. Three details matter. First, the words “up to” mean 4.8x is a ceiling, not an average. Second, the baseline is the previous generation, GB200 NVL72, not a much older Hopper part. Third, the workload is agentic coding, not chat or single-turn question answering. Silas Alberti, of Cognition's founding team, put the logic plainly: agentic coding means long contexts, high concurrency and token volumes where cost per token decides what the company can ship. For Devin, more throughput means faster real-time code generation and more responsive multistep reasoning. One more fact deserves weight. CoreWeave stood up a production Vera Rubin cluster for Cognition in days, which NVIDIA credits to codesign across the stack, from infrastructure to tokens served. For a customer, time from delivery to production is often as valuable as peak compute.

The second thread is the CPU. NVIDIA's argument is that agentic AI squeezes infrastructure from two directions. Serving agents demands low-latency compute at scale, while improving them through post-training requires thousands of isolated environments running at once. The second demand is largely a CPU problem, not a GPU one. As the blog frames it, the key measure is how many isolated agent environments can run at once, and how consistent and performant each stays as the count grows. CoreWeave's Vera deployment places 128 CPUs and 11,264 cores in a single rack, enough for more than 11,000 concurrent environments at one core each. Dividing the two totals gives roughly 88 cores per CPU; that is our arithmetic, not a figure the blog states. The environments run through CoreWeave Sandboxes, hardware-isolated and alongside the training jobs they support, with Spectrum-X Ethernet switches and BlueField-4 DPUs securing agent communication at low latency. In testing, CoreWeave measured more than 3x faster agent sandbox startup on Vera CPUs, and a 1.7x performance gain across all passing tasks on Terminal-Bench. Sandboxes are the execution layer for reinforcement learning, agent tool use and model evaluation, letting teams run code in isolated environments. Startup time matters because every reinforcement learning rollout may need a fresh environment, so slow starts leave expensive accelerators waiting.

The third thread is Forge. The blog says models and agents improve by running a loop: production behavior informs the next training run, and each evaluation sharpens the next version. That loop has historically been split across tools from different vendors, with signal lost at every handoff. CoreWeave Forge is positioned to connect those stages in one environment so that production feedback flows back into training. We should be candid here. The blog text available to us is cut off in the middle of the Forge section, so we cannot confirm its components, interfaces or pricing, and readers should consult CoreWeave's original announcement. What we can confirm sits on the software side: Vera Rubin capacity can be operated through CoreWeave Kubernetes Service, SUNK, CoreWeave Mission Control, CoreWeave Sandboxes and CoreWeave Inference. From scheduling and cluster management to sandboxes and serving, CoreWeave is trying to let one pool of capacity handle training, evaluation and live inference. That matches how Cognition itself uses the platform, running training, reinforcement learning and production inference in one place.

Finally, consider the industry meaning. The point NVIDIA's Ian Buck stresses is value across generations. CoreWeave's NVIDIA V100 GPUs are still running customer workloads nearly a decade after Volta launched, even as CoreWeave brings Vera Rubin NVL72 into production. When AI infrastructure is financed in the tens of billions of dollars, whether an asset keeps earning across several hardware cycles shapes returns and the cost of capital. That is what the blog's subtitle means by infrastructure that stays productive, durable and fungible. Buck also speaks of putting the right GPU on the right workload, which implies old and new generations coexisting and tiered by job. Our reading is that the announcement carries three messages. Inference economics is shifting from peak compute toward cost per token and concurrency. Agents make the CPU a bottleneck candidate again, so sandbox density becomes a new purchasing metric. And a closed loop across training, evaluation and production is becoming a point of product differentiation among clouds. Open questions remain: whether independent benchmarks reproduce figures such as 4.8x and 3x, what Vera Rubin supply and power costs look like, and how widely customers adopt Forge. Until those are answered, the safest reading is that the direction is clear and the numbers still need checking.