GTC 2026: Vera Rubin in production, and the case for Blackwell now

NVIDIA’s GTC 2026 keynote, given by Jensen Huang at the SAP Center in San Jose on 16 March 2026, was built around Vera Rubin, the rack-scale platform that pairs the Rubin GPU with NVIDIA’s own Vera CPU. Coverage from Jon Peddie Research and GIGAZINE reported that mass production had begun and that AWS, Google Cloud and Microsoft Azure planned to offer Rubin services in the second half of 2026.

NVIDIA’s product page for Vera Rubin NVL72 lists 72 Rubin GPUs and 36 Vera CPUs per rack, 20.7 TB of HBM4 with 1,400 TB/s of bandwidth, sixth-generation NVLink at 216 TB/s, and 3,600 PFLOPS of NVFP4 inference (sparse). GIGAZINE reported 288 GB of HBM4 per Rubin GPU and NVIDIA’s claim of five times the cost-effectiveness of Blackwell-generation systems. The rack also brings ConnectX-9 SuperNICs and BlueField-4 DPUs, and NVIDIA showed a companion Groq 3 LPX inference rack.

The roadmap beyond Rubin

Jon Peddie’s report on the keynote set out the next steps. Vera Rubin Ultra, targeted for 2027, moves to the Kyber rack architecture with 144 GPUs in vertical compute trays. Feynman follows in 2028 with 3D die stacking, custom HBM and fabrication on TSMC’s A16 process, paired with a new Rosa CPU and an LP40 LPU developed with the Groq team. Huang also doubled NVIDIA’s target for combined Blackwell and Vera Rubin orders through 2027 to $1 trillion.

By late August 2026 NVIDIA said Vera Rubin was in full production and already running at CoreWeave, Google Cloud, Microsoft Azure, Oracle Cloud Infrastructure and Nebius. TrendForce expects Vera Rubin racks to ramp in 2027 and to cost roughly twice as much per system as GB300, while GB300 remains the main shipment driver through 2026.

What it means for buyers

For enterprises and regional cloud providers in the Gulf, Central Asia and Africa, the early Rubin deployments named so far are all at hyperscalers and large AI clouds, and those buyers are first in line for allocation. A project that needs hardware in the next two or three quarters should plan on Blackwell (HGX B300, HGX B200 or GB300 NVL72) and treat Rubin as a later expansion. Vera Rubin NVL72 is a liquid-cooled rack, so check facility readiness before assuming a migration path. When specifying Blackwell now, ask OEMs how long they will support the platform with spares and firmware, which software licences are included, and whether quoted lead times assume existing stock or new allocation. Export licensing applies to both generations equally, so end-user documentation should be prepared before the order rather than after.