Cognition's $48 billion Devin runs on Nvidia's newest racks, rented from CoreWeave

First in line on Nvidia's newest racks

The first production workloads anywhere on Nvidia's▼−0.36% — Nvidia, down 0.36 percent today Vera Rubin NVL72 went live on September 30↗, and the tenant is not a frontier lab ¹. It is Devin, the coding agent from CognitionPVT~$10B — Cognition, private, latest valuation ~$10B, which runs training, reinforcement learning and inference on rented CoreWeave▲+0.21% — CoreWeave, up 0.21 percent today racks and scaled from bridge capacity to thousands of GPUs in under nine months ¹.

A run rate that doubled in four months

Cognition was on track in September to generate $1 billion in annualized revenue↗, roughly double the $492 million of May, on a figure the company declines to confirm; it had said that month its run rate topped $900 million ². Weeks earlier it closed $2 billion at a $48 billion valuation, three months after raising $1 billion at $26 billion ²³⁴. Nvidia is both a backer and a customer, alongside Citigroup and Mercedes-Benz ²³.

Free weights, one reinforcement-learning run

SWE-2, the model behind Devin since September 11, is post-trained from Moonshot's open-weight Kimi K3, 2.8 trillion parameters, with a reinforcement-learning run Cognition says adds five to six points ⁵. That puts it within a point of Fable 5.1 on FrontierCode 1.1 Main, at a cost Cognition puts 64% lower, about a quarter of GPT-6 Astra ⁵.

Provenance is the catch: Cognition wrote the benchmark, filled in missing competitor scores itself, and no independent test of SWE-2 exists ⁵. There is no printed per-task price and no API; it runs only inside Devin ⁵.

The landlord gets paid either way

CoreWeave, listed on Nasdaq, booked $2.58 billion of revenue in the June quarter↗, up 112% year on year, while spending $6.4 billion on capex in the same three months ⁶. Cognition's own engineers measured up to 4.8 times the total token throughput for SWE-2 inference on Rubin racks against a GB200 baseline ¹; CoreWeave separately claims 10 times the tokens per megawatt, the number that binds when power, not floor space, caps a data hall ¹⁷. Whoever wins the model layer, the rack owner bills them all.

The discount is racing the price cuts

Above Cognition, the frontier is getting cheaper. AnthropicPVT$350B — Anthropic, private, latest valuation $350B cut Haiku 5.5 prices 90% on October 7↗ to $0.10 per million input tokens, matching OpenAI'sPVT$1.18T — OpenAI, private, latest valuation $1.18T Luna, and its launch table scores that $0.10 model 46.4% on the same FrontierCode 1.1 main evaluation ⁸. OpenAI halved Pro 200's usage allowance and now lets subscribers spend plan usage inside Devin, one of 16 partner tools ⁹. Every cut narrows the discount that justifies Devin's price; on the hardest terminal benchmark, SWE-2 already trails GPT-6 Astra 27.3% to 57.9% ⁵.

What would settle it

Watch for one independent SWE-2 benchmark, one printed per-task price, and the first sign of frontier cuts in Cognition's renewal rates. The talent arbitrage continues regardless: Cognition's former chief of staff raised $60 million on October 8 for the agent startup Hone, with Cognition itself a customer ¹⁰. Free Chinese weights, one reinforcement-learning run, rented American racks: that is the whole cost structure, and the exposure sits with the secondary buyers now marking Cognition at $48 billion.