CoreWeave Puts Vera Rubin NVL72 Into Production With Cognition — 4.8× SWE-2 Tokens, Not Another Shipping BOM

This morning’s Claude Sonnet 5.5 note was a routing ticket — everyday coding cost next to Opus 5.5. The live Supermicro Vera Rubin NVL72 shipping post answered the orderable BOM purchase. The live Vera Rubin MLPerf note answered vendor harness rows. On September 30, 2026, CoreWeave said NVIDIA Vera Rubin NVL72 is in limited availability on CoreWeave Cloud, with Cognition (Devin) as the first customer anywhere running production workloads — and Cognition’s own engineers publishing 4.8× and 3.8× token-throughput deltas vs a GB200 NVL72 baseline. Tonight is that neocloud production + customer agentic-throughput board.

Primary sources: CoreWeave news release, CoreWeave customer blog. Optional secondary: NVIDIA Fully Connected blog. Specs and benches below are vendor / customer claims. Treat 4.8× / 3.8×, tokens-per-megawatt, and rack bandwidth rows as claims until you reproduce them on your own harness.

Spec sheet (named numbers)

That is a different purchase sentence from Supermicro’s can I order the cold-plate-to-tower BOM post and from the MLPerf what does the vendor harness print post. This one is: is someone else’s agentic production loop already compounding tokens on Vera Rubin in a neocloud, and by how much vs GB200.

What Cognition actually measured

Cognition runs training, reinforcement learning, and production inference for Devin on CoreWeave. The company says it scaled from bridge capacity to thousands of GPUs in under nine months. For this generation, Cognition’s engineers ran the first customer-executed Vera Rubin inference benchmark against a GB200 NVL72 baseline on CoreWeave Cloud — not a slide from a silicon launch deck.

Two numbers matter for anyone who operates agentic coding fleets:

1. SWE-2 inference — up to 4.8× total token throughput at matched interactivity. Agentic coding is a serial compounder: read repo → reason → edit → run → evaluate → repeat. Throughput that lands on every step changes wall-clock work finished per GPU-hour, not just a single chat completion.
2. RL post-training — 3.8× output token throughput at matched interactivity. Cognition frames RL as generating large volumes of trial-and-error trajectories. Faster output tokens shorten the research loop that feeds the next production weights.

CoreWeave’s customer blog is explicit that both comparisons are per GPU at matched interactivity vs GB200 NVL72. That is the right unit for capacity planning; it is still their SWE-2 / RL harness, not yours. Reproduce before you reprice a reservation.

Silas Alberti’s quote is marketing-adjacent, but the operational claim underneath it is the one I care about: for a workload where every step waits on the last one, a 4.8× total-token jump compounds into finished engineering work, not a prettier tokens/s chart.

Where it sits next to the other Vera Rubin boards

Do not collapse this into the Supermicro shipping note. Shipping answers CapEx and facility integration. Cognition answers whether the rack is already earning tokens in someone else’s agent loop under the same CKS/SUNK/Mission Control path as their Blackwell fleet.

Caveats

1. Customer harness ≠ your fleet. Re-run SWE-style agent loops and your RL rollout generator on Vera Rubin before you move default Devin-class traffic off GB200.
2. Matched interactivity is load-defining. Ask for the exact concurrency, context length, and interactivity target behind 4.8× / 3.8×.
3. Limited availability. “Hundreds of Rubin GPUs across multiple regions” is not a blank PO. Get a briefing for capacity and onboarding.
4. 45× vs Blackwell is a CoreWeave platform claim on the customer blog — separate from Cognition’s SWE-2 / RL numbers. Do not cite it as Cognition’s measurement.
5. 10× tokens/MW is CoreWeave’s earlier DeepSeek R1 silicon claim — useful context, not tonight’s new Cognition delta.
6. NVLink 216 TB/s vs Supermicro datasheet 260 TB/s — same class of disagreement I flagged on the shipping post. Ask which figure is current on the SKU you rent or buy.
7. Not a Sonnet rematch, not a Helios deal rematch, not Everspin’s 1GB CXL demo. This is the production NVL72 customer-throughput board.

How I would evaluate

1. Same-harness bake-off — pin one SWE-style agent suite and one RL rollout generator; run GB200 NVL72 and Vera Rubin NVL72 at the same interactivity target; log total and output tokens/s, wall time to task done, and dollars per successful session.
2. Compounding check — measure end-to-end task completion rate, not only tokens/s. Agentic coding only pays if steps chain.
3. Ops path — confirm CKS / SUNK / Mission Control / Inference parity with your existing GB200/GB300 fleet so bring-up is days, not a second control plane.
4. Facility alternative — if you need owned CapEx, keep the Supermicro shipping BOM board; this CoreWeave ticket is rent-and-run production evidence.
5. Keep boards separate — MLPerf for vendor rows, Supermicro for orderable liquid racks, Helios for open Ethernet scale-up, Inferra / Leo X for KV tiers. Cognition is the customer agentic Δ on Vera Rubin in a neocloud.

Bottom line

On September 30, 2026, CoreWeave put NVIDIA Vera Rubin NVL72 into limited availability, with Cognition as the first production customer and Cognition-run benches claiming up to 4.8× total token throughput on SWE-2 inference and 3.8× output token throughput on RL versus GB200 NVL72 at matched interactivity. Durable facts: 72 Rubin + 36 Vera per rack, NVLink 6, CoreWeave’s 1,400 TB/s HBM4 / 216 TB/s NVLink rack numbers, Spectrum-X 102.4T, same fleet tooling as GB200/GB300, Dell PowerRack path, and early-September bring-up into production within days of rack handover.

That closes the gap between “NVL72 is shipping” and “someone’s agent loop is already compounding on it.” Not a Sonnet night, not a Helios procurement night, and not permission to retire GB200 without a same-harness bake-off. Measure tokens and finished tasks, then decide.

Create-only. Specs and benches from CoreWeave’s Sep 30 news/blog and NVIDIA’s same-day Fully Connected post. Update when CoreWeave publishes general availability, when a second named customer reproduces GB200 deltas under a common harness, or when NVLink rack bandwidth is reconciled across vendor datasheets.

Sources: CoreWeave news · CoreWeave × Cognition blog · NVIDIA blog.

Anpoo Sivanadi, Staff Software Engineer