Nvidia confirmed in an investor update that first commercial shipments of its Rubin generation accelerators have reached hyperscale customers, with Microsoft, Google, Meta and Oracle named as initial recipients. The company said volume ramp will continue through the second half of 2026, but acknowledged that enterprise and sovereign cloud orders placed today are quoted with delivery windows extending into the first and second quarter of 2027.
Rubin replaces the Blackwell Ultra generation as Nvidia's flagship training platform, with the company citing roughly 2.5 times the training throughput per rack and lower energy per token on inference workloads at FP4 precision. The accompanying NVLink fabric scales to 576 GPUs in a single coherent domain.
For enterprise IT planners the message is the same as the last two cycles. Co-location power, liquid cooling readiness and capital approvals need to be in motion well before silicon is allocated, and most workloads do not require the newest generation. Many vendors are reporting strong availability of prior-generation Hopper and Blackwell capacity at materially lower per-token economics.
The BroadVision view
Extended lead times for latest generation hardware require mid-market IT teams to focus on maximizing existing compute efficiency. Organizations may need to re-evaluate infrastructure lifecycles and consider hybrid cloud configurations to maintain performance while waiting for hardware availability. Strategic planning remains necessary to align procurement timelines with long-term digital objectives. BroadVision strategic IT services