Dell Delivers Mass Production Rubin Racks to AI Cloud Provider CoreWeave
Dell has delivered the world's first mass-produced NVIDIA Vera Rubin NVL72 racks to CoreWeave, transitioning the Rubin architecture from prototype validation to scalable deployment. The racks consist of Dell's PowerRack combined with PowerEdge XE9812 liquid-cooled servers, integrating 72 Rubin GPUs and 36 Vera CPUs per cabinet, with a bandwidth of 260TB per second via sixth-generation NVLink.
This is not the same event as the "first power-on validation" that occurred mid-year. The Livingston data center in New Jersey previously completed end-to-end power, liquid cooling, network, and computing integration, passing NVIDIA's L11 diagnostics and approximately 147 hours of testing; Dell claimed at that time that the time from arrival to production readiness could be compressed to under 6.5 hours. This announcement clearly states that mass production racks are being shipped, indicating that the supply chain has shifted from trial cabinets to repeatable deliveries.
CoreWeave subsequently released silicon performance metrics: under the DeepSeek R1 inference load and matching equivalent interactive latency conditions, the Vera Rubin NVL72 generates approximately ten times the tokens per megawatt per second compared to the GB200 NVL72, with official comparisons showing about a quarter of the GPU count and roughly a tenth of the cost per million tokens. Each cabinet is rated for approximately 3.6 exaFLOPS of NVFP4 inference capability, with memory upgraded to HBM4 and native support for NVFP4 in the Transformer Engine.
To operate the entire cabinet as a programmable device, CoreWeave implemented two sets of self-developed components: Valvey creates a software-defined control interface for each cabinet's liquid cooling valves, monitoring flow, temperature, pressure, and leaks in real-time, allowing isolation of individual cabinets without compromising shared cooling sources; Racky aggregates power, cooling, and environmental sensors. The backend network is laid out with approximately 1.6Tb/s of multi-lane RoCE per GPU, introducing Spectrum-X and BlueField-4 for tenant isolation.
NVIDIA has designated Vera Rubin for full-scale production, with deployment plans for the second half of the year covering OpenAI, Google Cloud, Microsoft Azure, Oracle Cloud, as well as cloud providers like Lambda, Nebius, and Nscale. OpenAI is expected to scale up in the third quarter. Earlier this year, Microsoft claimed to have completed power-on validation first, while CoreWeave positions "the first production-level cabinet validation for AI cloud providers" as a differentiating narrative.
In market dynamics, buyers are model companies and hyperscale clouds looking to lock in next-generation inference unit prices, while sellers are NVIDIA's cabinet production capacity and Dell's system integration. Funding has shifted from "securing Blackwell quotas" to "who can first turn Rubin cabinets into sellable tokens." Beneficiaries are cloud providers and foundries that can maximize token output per megawatt; those under pressure are GPU leasing inventories still depreciating under old generations, as well as second-tier data centers unable to keep up with the density of entire cabinets in terms of power and liquid cooling. The event-driven nature is clear: once mass production delivery is established, computing futures will transition from PPT to signable cabinet schedules.
In supplementary data, CoreWeave previously secured a computing power contract with Meta worth approximately $21 billion, with its stock price experiencing about a 14% single-day fluctuation around the first validation announcement; the number of cabinets and monetary value for this mass production shipment has not yet been disclosed.
Source: Public Information
ABAB AI Insight
CoreWeave's path to the forefront of Rubin was not sudden. The company transitioned from mining power to AI cloud, consistently positioning itself with NVIDIA's every generation of cabinets: H100, H200, GB200 NVL72, and GB300 NVL72 all followed the same script of "Dell delivers the cabinet first, CoreWeave powers it on first." After going public under the ticker CRWV, it has used "being the first to run new cabinets" as a credit credential for customers and the debt market. Controversies have persisted: high-leverage electricity rentals, long-term GPU locks, and high customer concentration mean that if the next generation of cabinets is delayed, the rental and depreciation of old cabinets will backfire.
Capital mobilization is very specific. Dell is responsible for making the liquid-cooled cabinet with 72 GPUs and 36 CPUs into a transportable unit, CoreWeave is responsible for the cooling source, power, and orchestration software on the data center side, and NVIDIA is responsible for making NVLink 6 and HBM4 into an inseparable closed system. The motivation is not just to sell a few more cards, but to change the pricing unit from "renting one GPU for a month" to "how many tokens can one megawatt produce." Whoever completes this pricing change first will be able to write the next generation of inference costs into contracts with long-term clients like OpenAI, Anthropic, and Meta.
A comparable example is when Equinix acquired network cabinets, AWS developed Graviton to capture x86 profits, and TSMC turned advanced packaging from an added service into a capacity bottleneck. CoreWeave is in the "control interface" phase during its expansion: the cabinets still belong to NVIDIA, but the valves, cabinet controllers, and scheduling software are starting to become its moat. Microsoft, Google, and Oracle will deploy the same cabinets into their own supercomputing facilities, shifting competition from "whether there is Rubin" to "who can sell the same cabinet first."
Structural judgment involves the overlap of technological substitution and the transfer of pricing power. The mechanism is that when the output of tokens per megawatt falls by an order of magnitude, the rental anchor for old-generation assets breaks. Buyers no longer pay for the number of cards, but for the interactive inference traffic; thus, pricing power shifts from GPU spot sellers to a few system vendors who can simultaneously control cabinet delivery, liquid cooling, and power density. The mass production of Rubin is not just a generational launch event, but a moment when the electricity bill is rewritten into the gross margin of model companies.
ABAB News · Cognitive Laws
The first to power on will be the first to change the pricing unit.
Chip iterations sell performance, while cabinet mass production sells electricity costs.
Whoever can manage the cabinet as a single computer will take away the profits of the cloud.