NVIDIA unveils Vera Rubin work

NVIDIA’s Vera Rubin news is really about a change in what the company thinks it sells. Not just GPUs. Not even just servers. Basically, NVIDIA is pitching a whole AI factory — chips, racks, networking, orchestration software, and even reference designs for the building around them. That matters because the bottleneck in big AI systems has moved. Raw compute still matters, but power, memory movement, scheduling, and inference efficiency matter just. ### What is Vera Rubin, exactly? Vera Rubin is NVIDIA’s next rack-scale AI platform. It combines the Vera CPU, Rubin GPU, NVLink 6 switch, ConnectX-9 SuperNIC, BlueField-4 DPU, Spectrum-6 Ethernet, and a newly integrated Groq 3 LPU into one coordinated system. NVIDIA framed that as seven chips working like one supercomputer rather than a pile of separate parts. ### Why buy a few accelerators anymore? They are buying dense clusters that have to behave like one machine. Vera Rubin is built around NVL72 rack systems and broader POD-scale deployments, which means NVIDIA is optimizing the links between chips, the memory behavior, and the network fabric at the same time. The point is less “faster GPU” and more “less wasted infrastructure.” ### Why is inference so central here? Turns out the headline is not training. It is inference — especially agentic inference, where models reason, call tools, and keep serving lots of requests with ugly traffic patterns. NVIDIA also launched Dynamo 1.0, an open-source inference stack meant to orchestrate that mess across data-center scale clusters. NVIDIA says Dynamo can boost Blackwell inference performance by up to 7x, which tells you where the company sees the pain point: not just chips, but scheduling and serving. ### What is the biggest concrete claim? The eye-catching one is cost per token. NVIDIA says Vera Rubin NVL72 can deliver up to 10x lower inference cost per token than Blackwell-based systems. Separately, the company says its Vera Rubin DSX AI Factory reference design can raise usable AI infrastructure by up to 30% under fixed power constraints. Those are very NVIDIA-shaped numbers, of course, but they show the sales pitch clearly — cheaper inference and better power efficiency, not just more peak FLOPS. ### What is DSX doing in this story? DSX is NVIDIA trying to standardize the physical AI data center around its stack. The Vera Rubin DSX reference design covers how to build these facilities as codesigned systems, and the Omniverse DSX blueprint lets operators model them as digital twins before they build. That is a big deal because AI factories are now constrained by cooling, power delivery, and facility layout almost as much as by chip supply. ### Why does this widen NVIDIA’s moat? Because competitors can copy a chip faster than they can copy an ecosystem. NVIDIA now has the silicon, the interconnect, the inference software, the rack architecture, and the facility playbook. If a cloud provider or enterprise wants something turnkey, Vera Rubin makes NVIDIA harder to swap out of the stack one layer at a time. That is the strategic move here. ### Who is this really for? The obvious buyers are hyperscalers and model companies. But the design also fits governments, sovereign AI projects, and any organization trying to run very large training and inference workloads under tight power budgets. The language NVIDIA used — AI factories, mission-critical decisions, massive-scale agents — is enterprise and state-scale language, not startup tinkering. ### So what changed? The change is that NVIDIA stopped talking like a component supplier and started talking like the prime contractor for AI infrastructure. Vera Rubin is the clearest version of that yet. If Blackwell was about selling the next must-have GPU generation, Vera Rubin is about selling the operating model for the next generation of AI data centers.