NVIDIA explains Vera Rubin spec changes as platform enters volume production
NVIDIA has responded to questions surrounding its recently updated Vera Rubin specifications, explaining that the revised figures are not a downgrade in real-world performance but a more accurate reflection of what the platform sustains in qualified volume production.
The changes drew attention because Vera Rubin is one of NVIDIA’s most important next-generation AI computing platforms, designed for massive data center deployments and demanding agentic AI workloads. Earlier specifications listed higher peak figures for HBM4 memory bandwidth and NVLink 6 interconnect bandwidth. The latest official figures now show lower values in some areas, prompting concern that the final hardware might perform below the original targets.
According to NVIDIA, the earlier numbers were preliminary architectural goals shared during the platform’s development. The final specifications now published for Vera Rubin represent what the systems can consistently deliver at production scale.
The company emphasized that this is a normal transition from early target specifications to finalized production data. In other words, the updated numbers are meant to provide customers with a clearer and more realistic view of the hardware now being manufactured and shipped.
When NVIDIA first introduced Vera Rubin, each Rubin GPU was listed with 288 GB of HBM4 memory and 22 TB/s of memory bandwidth. It was also shown with 3.6 TB/s of sixth-generation NVLink bandwidth.
In the revised specifications, the HBM4 capacity remains unchanged at 288 GB per GPU, but memory bandwidth is now listed at 19.2 TB/s. NVLink 6 bandwidth has also moved from 3.6 TB/s to 3 TB/s. That represents a reduction of roughly 12.7% for HBM4 bandwidth and 16.7% for NVLink bandwidth compared with the earlier development targets.
The changes apply across the Vera Rubin lineup, including large-scale rack configurations such as Vera Rubin NVL72 and smaller optimized systems such as HGX Rubin NVL8.
For the Vera Rubin Superchip, HBM4 bandwidth has shifted from 44 TB/s to 38.5 TB/s, while NVLink bandwidth has moved from 7.2 TB/s to 6 TB/s. Networking bandwidth, however, has increased from 0.8 TB/s to 0.9 TB/s.
For Vera Rubin NVL72, total HBM4 capacity remains 20.7 TB. HBM4 bandwidth is now listed at 1,400 TB/s instead of 1,580 TB/s, and NVLink bandwidth has changed from 260 TB/s to 216 TB/s. Networking bandwidth has improved from 28.8 TB/s to 32.4 TB/s.
For HGX Rubin NVL8, HBM4 capacity is now listed at 2 TB instead of 2.3 TB, while HBM4 bandwidth has moved from 176 TB/s to 130 TB/s. NVLink Switch bandwidth has changed from 28.8 TB/s to 24 TB/s, and networking bandwidth has increased from 1.6 TB/s to 1.8 TB/s. Compute figures also show slight revisions, with NVFP4 training moving from 280 PFLOPS to 270 PFLOPS and FP8/FP6 training shifting from 140 PFLOPS to 130 PFLOPS.
While those changes may look significant on paper, NVIDIA argues that individual component specifications do not define application performance, especially in the type of AI workloads Vera Rubin is built to handle.
The company says Vera Rubin is designed for agentic AI, where models do far more than generate simple responses. These workloads can reason, call tools, retrieve information, evaluate results, and repeat those steps thousands of times for a single task. At that scale, performance depends on the entire data center system, not just one memory or interconnect figure.
This is why NVIDIA says it focuses on platform-level performance, including the combined effect of compute, memory, networking, software, model optimization, rack-scale design, and system architecture. The company describes Vera Rubin as the result of deep co-design across hardware and software rather than a collection of isolated components.
That point is especially important for AI factories, where overall throughput, power efficiency, model responsiveness, and cost per token matter more than peak bandwidth figures in isolation. NVIDIA says this is why it publishes performance data using named AI models and real agentic configurations instead of relying only on component-level specifications.
NVIDIA also stated that third-party benchmarks already available for Vera Rubin were measured using production specifications, not the older preliminary targets. That means existing performance evaluations should already reflect the final hardware configuration now entering full-scale production.
According to the company, Vera Rubin NVL72 is in full production and shipping now. NVIDIA says published agentic AI benchmark results show strong gains over the previous-generation GB300 NVL72, including up to 30 times more performance per watt and 30 times more tokens per dollar.
The broader message from NVIDIA is that customers should not view the revised spec sheet as a sign that Vera Rubin has lost performance. Instead, the company frames the update as a move from early engineering targets to verified production specifications.
Building a platform like Vera Rubin is far more complex than producing a single GPU. It involves GPUs, CPUs, high-bandwidth memory, NVLink switches, networking hardware, cooling infrastructure, servers, racks, software stacks, cables, and power delivery systems. Once all of those elements are validated for large-scale deployment, final sustained production numbers can differ from early architectural goals.
For data center buyers and AI infrastructure planners, the key takeaway is that Vera Rubin’s finalized specifications are intended to represent real-world sustained operation at volume, not theoretical peak targets. While some memory and interconnect numbers have been adjusted downward, networking bandwidth has improved in several configurations, and NVIDIA continues to emphasize platform-level gains for agentic AI.
The updated Vera Rubin specifications mark an important step in the platform’s move from development to commercial deployment. With volume production underway, NVIDIA is positioning Vera Rubin as a major engine for next-generation AI factories, where performance is measured not by one component alone but by the efficiency and scale of the complete system.
