Nvidia Vera Rubin Systems Begin Shipping

• On July 21 NVIDIA confirmed the Vera Rubin platform had reached full production, with Ian Buck, vice president of accelerated computing, telling reporters at NVIDIA headquarters that systems are now shipping to customers including OpenAI, CoreWeave, Google Cloud, Microsoft Azure, Meta and Dell^[1]
• NVIDIA had said at GTC Taipei on May 31 that the platform was ramping into full production, with Taiwan's top server makers manufacturing Vera Rubin-based systems at scale^[1]
• The sources disagree on when shipments were due: the May 31 newsroom announcement, as^[1] cites it, said "production shipments of Vera Rubin are set to begin starting this fall", while the July 21 briefing has systems already shipping; the newer account is the shipping one^[1]
• They also disagree on the platform's production status at launch - the CES press release of January 5, 2026 already said "NVIDIA Rubin is in full production, and Rubin-based products will be available from partners the second half of 2026"^[2], while ServeTheHome's report of the same event called it "an 'everything remains on schedule' announcement" and noted "the hardware is not shipping yet, and the ramp-up for production wo not even start until the second half of the year"^[3]
• The number of chips in the platform is also stated inconsistently: both CES sources describe six new chips - Vera CPU, Rubin GPU, NVLink 6 Switch, ConnectX-9 SuperNIC, BlueField-4 DPU and Spectrum-6 Ethernet Switch^[2][3] - while^[1] says that on March 16, 2026 at GTC in San Jose NVIDIA announced the platform had expanded to seven chips in full production with the NVIDIA Groq 3 LPU, and elsewhere in the same article still calls it six^[1]
• Rubin-based products are to be available from partners in the second half of 2026, with AWS, Google Cloud, Microsoft and OCI among the first public clouds, alongside NVIDIA Cloud Partners CoreWeave, Lambda, Nebius and Nscale^[1][2][3]
• Microsoft will deploy Vera Rubin NVL72 rack-scale systems in next-generation AI data centres including future Fairwater AI superfactory sites, which "will scale to hundreds of thousands of NVIDIA Vera Rubin Superchips"^[1][2]
• CoreWeave, among the first cloud providers to receive the hardware, told Bloomberg its NVL72 racks are delivering ten times the token output of the previous generation^[1]
• Dell Technologies, HPE, Lenovo and Supermicro, with ASUS, Foxconn, GIGABYTE, Pegatron, Quanta Cloud Technology, Wistron and Wiwynn, are adopting NVIDIA DSX to accelerate the Vera Rubin ramp^[1]
• NVIDIA ran an aggressive 18-month development cycle from Blackwell's launch to Rubin production, against the 24 to 30 months such a step typically takes^[1]
• The CES release carries a forward-looking-statements notice and warns that "Many of the products and features described herein remain in various stages and will be offered on a when-and-if-available basis"^[2]

Notable features

• The flagship Vera Rubin NVL72 rack integrates 72 Rubin GPUs and 36 Vera CPUs connected by NVLink 6, with ConnectX-9 SuperNICs and BlueField-4 DPUs^[1][2][3]
• Each Rubin GPU carries 288 GB of HBM4 memory with up to 22 TB/s of bandwidth, against the Blackwell Ultra B300's 8 TB/s^[1]; ServeTheHome puts the same 288GB and 22 TB/second at 2.8x Blackwell's HBM3e bandwidth, on a dual-die 336-billion-transistor part built on TSMC 3nm^[3]
• NVIDIA claims Rubin reaches 5x Blackwell's inference performance and 3.5x its training performance, with 50 PFLOPS of NVFP4 inference and 35 PFLOPS of NVFP4 training per GPU^[1][2][3]
• The platform is pitched at up to a 10x reduction in inference token cost versus Blackwell; the NVL72 rack is quoted at up to 10x higher inference throughput per watt at one-tenth the cost per token^[1][2]
• A full NVL72 rack is rated at 3.6 EFLOPS of inference and 2.5 EFLOPS of training compute, 220 trillion transistors, 54TB of LPDDR5X and 20.7TB of HBM4^[3], with 3.6TB/s of NVLink bandwidth per GPU and 260TB/s across the rack^[2]
• The rack is 100 percent liquid cooled with a cable-free modular tray design; the sources give the assembly saving three ways - installation cut "from two hours to five minutes"^[1], rack assembly "from 100 minutes on Blackwell to just 6 minutes on Rubin"^[3], and "up to 18x faster assembly and servicing than Blackwell"^[2]