On 30 September 2026, CoreWeave Inc. said NVIDIA Vera Rubin NVL72 is available on CoreWeave Cloud, with Cognition — the applied AI lab behind Devin — as the first customer anywhere running production workloads on the system. The announcement landed during Fully Connected, CoreWeave’s AI cloud conference in San Francisco. For operators buying rack-scale inference and RL capacity, this is a generation handoff: same Kubernetes and Mission Control path as GB200/GB300 NVL72, new silicon under the hood, and a set of throughput claims that need cold reading.
Cowry Tech News is covering this as a US AI/cloud infrastructure piece. It is not a funding tip sheet. Capital desk owns raise frames. It is also not an independent lab benchmark of Vera Rubin. The numbers below come from CoreWeave’s investor release and from NVIDIA’s same-day blog on the CoreWeave–NVIDIA agentic stack. Where a figure is customer- or vendor-reported, that label stays in the open.
What CoreWeave said, in order
CoreWeave’s IR release states:
- Vera Rubin NVL72 is available on CoreWeave.
- Cognition is the first production customer on that system.
- Customers run Vera Rubin under the same operating model and tooling as existing NVIDIA GB200 NVL72 and GB300 NVL72 fleets, with performance engineering from CoreWeave’s team.
- Cognition scaled from bridge capacity to thousands of GPUs for training and inference on CoreWeave in less than nine months.
- Cognition and CoreWeave stood up a Vera Rubin NVL72 cluster in early September 2026.
- Cognition’s own engineers ran the first customer-executed Vera Rubin inference benchmark against a GB200 NVL72 cluster baseline.
- On those runs, Cognition measured up to a 4.8× increase in total token throughput for SWE-2 inference workloads versus GB200 NVL72, and a 3.8× boost in output token throughput for reinforcement-learning workloads.
- CoreWeave, with Dell Technologies, was among the first cloud providers to deploy Dell PowerRack systems with GB200 and GB300 NVL72, and is one of the first to deploy Vera Rubin.
- CoreWeave says it published measured silicon performance showing 10× token throughput per megawatt over GB200 NVL72 on the DeepSeek R1 reasoning model at matched interactivity — a CoreWeave platform claim, not an independent CTN measurement.
- The release cites MLPerf results and a SemiAnalysis ClusterMAX Platinum ranking “three times in a row” as company-stated competitive positioning.
Chen Goldberg, CoreWeave EVP of product and engineering, is quoted on bringing Vera Rubin up quickly and on agentic pressure — long contexts, repeated model calls, thousands of concurrent tasks — as a full-stack problem across compute, networking, and software.
Silas Alberti, SVP research and founding team at Cognition, is quoted on agentic coding needing long contexts, high concurrency, and rapid reasoning, and on the 4.8× SWE-2 total-token throughput figure compounding into more work Devin gets done when each step waits on the last.
Ian Buck, NVIDIA VP of Hyperscale and HPC, is quoted praising CoreWeave’s multi-generation infrastructure expertise and Cognition’s production use of Vera Rubin NVL72.
Those are attributed quotes from company materials. They are not CTN original interviews.
What NVIDIA’s blog adds (same day)
NVIDIA’s 30 September 2026 blog (“From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI”) aligns on Vera Rubin NVL72 availability on CoreWeave with Spectrum-X 102.4T Ethernet networking, Cognition as first production customer, and the 4.8× SWE-2 total-token throughput figure versus GB200 NVL72. It adds useful ops context:
- Early September: CoreWeave received its first Vera Rubin NVL72 production racks; Cognition then benchmarked inference against a GB200 NVL72 baseline using a real-world software-engineering workload sampled from FrontierCode tasks with agents solving them.
- Capacity can be operated through CoreWeave Kubernetes Service, SUNK, CoreWeave Mission Control, CoreWeave Sandboxes, and CoreWeave Inference.
- CoreWeave will also offer NVIDIA Vera CPU (described as the first CPU built for AI agents). NVIDIA/CoreWeave testing claims more than 3× faster agent sandbox startup times on Vera CPUs, and a 1.7× performance gain on Terminal-Bench across passing tasks — again, company-reported test results.
- CoreWeave Forge is positioned as a connected environment for training, evaluating, and improving models and agents (Weights & Biases, OpenPipe post-training expertise, marimo notebooks). New/expanded pieces named: ARIA (GA), Agent Lens, Sandboxes (GA), serverless SFT/RL, and RL Rollouts in private preview powered in part by NVIDIA Dynamo.
- Ennoble Care is cited as selecting CoreWeave for clinical AI inference on reserved NVIDIA RTX PRO 6000 GPU capacity via CoreWeave Kubernetes Service — a separate customer anecdote from the Cognition Vera Rubin storyline.
For Brief and story hygiene: Vera Rubin NVL72 production + Cognition benchmarks are the spine of this piece. Forge, Vera CPU, and Ennoble Care are adjacent Fully Connected news. Do not flatten them into one “everything shipped” claim without naming which product is GA, which is preview, and which is a customer selection story.
How to read the 4.8× and 3.8× numbers
Operator English, not slideware:
- Workload-specific. The 4.8× figure is total token throughput on Cognition’s SWE-2 inference workloads versus their GB200 NVL72 baseline on CoreWeave. It is not a promise that every LLM, every batch size, and every serving stack will see 4.8×.
- Customer-executed, on CoreWeave. Cognition’s engineers ran the benchmark. Stronger than a pure vendor slide. Still not a third-party lab with a public method CTN can re-run tomorrow.
- Agentic compounding. Alberti’s point — each step waits on the last — is why token throughput matters for Devin-class agents. Latency and scheduling still decide whether that throughput becomes user-visible speed.
- RL is a different number. The 3.8× boost is output token throughput for reinforcement-learning workloads. Do not merge SWE-2 inference and RL into one “almost 5× everything” headline.
- Power and silicon claims are separate. The 10× tokens-per-megawatt DeepSeek R1 figure is a CoreWeave-stated silicon/platform measurement at matched interactivity. File it next to MLPerf and ClusterMAX as competitive positioning, not as Cognition’s SWE-2 result.
If you are buying capacity, ask for the exact model family, precision, concurrency, context length, and serving stack used in the SWE-2 run — then ask what your own workload looks like on the same rack.
Same ops model as GB200/GB300 — what that buys you
The practical pitch is continuity. CoreWeave says Vera Rubin customers keep the operating model used on GB200 and GB300 NVL72 fleets: Kubernetes Service, SUNK, Mission Control, Sandboxes, serverless inference. Multi-generation continuity is the product claim NVIDIA’s Ian Buck and CoreWeave both lean on — Volta-era V100 capacity still in commercial service on CoreWeave while Vera Rubin enters production.
What that means for eng leads:
- Migration tax. If your charts, autoscaling, and runbooks already assume CoreWeave’s NVL72 path, a Vera Rubin node should look familiar. If you are greenfield on another cloud’s Vera Rubin offering, you are buying silicon plus a new ops dialect.
- Fungibility across generations. Matching workload to generation (keep older GPUs on tolerant jobs; put agentic inference on new racks) is the stated value. It only works if your scheduler and cost model actually place work that way.
- Days, not quarters — for them. Goldberg and NVIDIA both stress Cognition getting production workloads running within days after racks arrived. That is a CoreWeave–Cognition co-engineering outcome. Your timeline depends on quota, networking, model compile, and whether your security review treats a new GPU generation as a change control event.
Agentic load is a systems problem
Agentic coding and tool-using agents stress more than peak FLOPS. Long contexts, high concurrency, repeated model calls, and isolated sandbox environments push networking, CPU isolation, and inference scheduling as hard as the GPU. That is why the Fully Connected package pairs Vera Rubin NVL72 with Spectrum-X Ethernet, Vera CPU for dense sandbox concurrency, and Dynamo-backed inference/RL Rollouts.
Buyers should separate three questions:
- Do I need Vera Rubin NVL72 rack-scale GPUs for this model’s decode/prefill profile?
- Do I need Vera-class CPU density for sandboxes, evals, and tool execution?
- Do I need Forge-style post-training/observability loop — or do I already have W&B + internal RL infra?
Buying all three because Cognition did is how slideware wins and capacity sits idle.
Operator checklist
- Baseline. Ask CoreWeave (and your own bench harness) for GB200 or GB300 NVL72 numbers on your serving stack before celebrating any Vera Rubin multiplier.
- Workload card. Record model, precision, context, concurrency, batching, and whether the job is inference, RL, or mixed. Map claims to that card.
- Networking. Confirm Spectrum-X / fabric assumptions for your AZ and whether multi-node jobs need RDMA-class paths you already validated on GB200/GB300.
- Tooling continuity. Verify Kubernetes Service, Mission Control, and inference paths behave the same under Vera Rubin quotas and drivers.
- Sandbox/CPU path. If agent tool use and RL sandboxes dominate cost, evaluate Vera CPU and Sandboxes as a separate line item from NVL72 GPU spend.
- Preview vs GA. RL Rollouts is private preview in NVIDIA’s write-up. Forge components have mixed GA/new labels. Do not write contracts as if preview equals production SLA.
- Disclosure and independence. Throughput and tokens-per-MW figures are not CTN lab results. Prefer customer-run methodology docs if Cognition or CoreWeave publish them later.
- Health and compliance. Healthcare anecdote (Ennoble Care) is a separate RTX PRO 6000 reserved-capacity story. Do not imply Vera Rubin NVL72 is cleared for clinical workloads without the customer’s own compliance path.
Failure modes buyers actually hit
Rack-scale generation changes fail in boring ways. Driver skew between GB200 and Vera Rubin nodes in the same namespace. Quotas that look open until the gang of NVL72 peers cannot land in one fabric domain. Serving stacks tuned for GB200 memory footprint that OOMs or under-batches on a new interconnect profile. Agentic harnesses that measure “tokens per second” while the wall-clock budget is eaten by tool sandbox cold starts — which is why Vera CPU and Sandboxes show up in the same Fully Connected narrative even when your PO only names GPUs.
Another failure mode is methodology drift. If Cognition’s SWE-2 subset, concurrency, and serving path are not written down in a form your team can approximate, the 4.8× becomes a conference story. Ask for the harness. If you cannot get it, run your own golden prompts on GB200 NVL72 and Vera Rubin NVL72 under identical client code before you reprice a reserved fleet.
Conference optics matter too. Fully Connected is a customer and partner event. Availability announcements at events are real when capacity and APIs are reachable the next Monday. Confirm region, waitlist, and whether “available” means GA for your contract tier or early access for named accounts. Cognition’s early-September stand-up is evidence for one customer path. It is not a universal SLA.
Finally, keep generation fungibility honest. Older Volta capacity still earning revenue on CoreWeave is a durability argument for NVIDIA’s stack and for CoreWeave’s multi-gen software. It does not mean your Vera Rubin reservation will be cheap to exit, or that your security team will treat a new GPU SKU as a no-op change. Plan the change window like any other production infra upgrade: canary, observe tokens and error rates, then shift traffic.
What this piece is not
- Not a CoreWeave earnings or raise analysis (Ellis / Capital).
- Not an NDPA or AppSec review of Devin or Forge (Arjun).
- Not a claim that every AI cloud has Vera Rubin in production — CoreWeave says it is among the first.
- Not permission to reprint Cognition’s 4.8× as an industry-standard Vera Rubin scoreboard without the SWE-2 / CoreWeave baseline caveats.
Related: CoreWeave Vera Rubin NVL72 Brief (thin wire).
Related: Modal Clusters go GA: multi-node GPU with @modal.clustered (AI/cloud full).
Sources
- CoreWeave IR: “CoreWeave Delivers NVIDIA Vera Rubin NVL72 Performance at Production Scale, Starting With Cognition” (2026-09-30). https://investors.coreweave.com/news/news-details/2026/CoreWeave-Delivers-NVIDIA-Vera-Rubin-NVL72-Performance-at-Production-Scale-Starting-With-Cognition/default.aspx
- NVIDIA Blog: Stuart Pitts, “From Training to Production, NVIDIA and CoreWeave Close the Loop on Agentic AI” (2026-09-30). https://blogs.nvidia.com/blog/coreweave-agentic-ai-vera-rubin/