Rubin platform unveiled

NVIDIA just drew a new map for the artificial intelligence data center. At its GTC conference on March 16, 2026, Chief Executive Officer Jensen Huang unveiled the Vera Rubin platform, a rack-scale system built around seven chips that NVIDIA says are now in full production for what it calls “agentic AI.” (nvidianews.nvidia.com) To understand why that launch landed so loudly, start with how most artificial intelligence systems have worked so far. A traditional inference system takes a prompt, runs a model once, and returns an answer, like a calculator that gives you a number after one button press. Agentic artificial intelligence is closer to a worker with a checklist: it breaks a task into steps, calls tools, stores context, checks results, and keeps going until it reaches a goal. (nvidianews.nvidia.com) That change sounds subtle, but it puts pressure on hardware in a different place. A chatbot that answers one question mostly needs raw compute. An agent that plans a trip, writes code, searches documents, and verifies its own work also needs fast memory movement, low-latency networking, storage access, and a way for many processors to behave like one machine. NVIDIA’s Rubin pitch is that the bottleneck is no longer just the chip, but the whole system around it. (nvidianews.nvidia.com) That is why Vera Rubin is not being presented as a single processor. NVIDIA says the platform combines the Vera central processing unit, the Rubin graphics processing unit, the sixth-generation NVLink switch, the ConnectX-9 SuperNIC, the BlueField-4 data processing unit, the Spectrum-6 Ethernet switch, and a newly integrated Groq 3 low-latency processor unit. The company packages those parts into rack-scale and pod-scale systems meant to cover pretraining, post-training, test-time scaling, and agentic inference. (nvidianews.nvidia.com) In plain English, NVIDIA is selling a factory floor, not just a faster engine. Its Vera Rubin NVL72 system ties together 72 Rubin graphics processors and 36 Vera central processors in one rack-scale computer, then links those racks with networking and storage gear so customers can build what NVIDIA calls artificial intelligence factories. On the product page, NVIDIA says the design is aimed at “reasoning” workloads and long-context jobs, with more tokens per watt and lower cost per token than the Blackwell generation. (nvidia.com) The company has been building toward this for more than a year. In March 2024, NVIDIA introduced the Blackwell platform as the next major step after Hopper. In January 2026, it launched Rubin as a six-chip platform and said it could cut inference token cost by up to 10 times versus Blackwell for some workloads. By March 2026, NVIDIA had added a seventh chip, the Groq 3 low-latency processor unit, and reintroduced the system as Vera Rubin for the agentic artificial intelligence era. (nvidianews.nvidia.com) That sequence helps explain the naming, too. “Rubin” refers to Vera Rubin, the astronomer whose work helped confirm the case for dark matter. NVIDIA says the full platform now pairs a Vera central processing unit with Rubin accelerators, turning the codename into a two-part product identity. (nvidianews.nvidia.com) Huang paired the product launch with a forecast designed to get Wall Street’s attention. During the March 16 keynote, he said cumulative orders or revenue opportunity for Blackwell and Vera Rubin systems would reach at least $1 trillion through the end of 2027, a sharp step up from the company’s earlier $500 billion figure tied to Blackwell and Rubin demand through 2026. CNBC reported the updated projection the same day. (cnbc.com) That number matters because NVIDIA’s biggest customers are no longer buying chips one server at a time. The buyers watching Vera Rubin most closely are hyperscalers such as Microsoft, Amazon Web Services, Google Cloud, Oracle, and CoreWeave, which build giant data centers and rent computing power to everyone else. NVIDIA highlighted work with several of those companies during the GTC keynote and had already said in January that Microsoft’s next-generation Fairwater facilities would scale to hundreds of thousands of Vera Rubin superchips. (blogs.nvidia.com) There is also a competitive subtext inside the launch. NVIDIA’s March announcement says Vera Rubin now includes the Groq 3 low-latency processor unit, a product tied to technology from Groq, the startup founded by former Google engineers. That lets NVIDIA address a part of the market where customers care less about training giant models and more about serving answers quickly and cheaply at scale. (nvidianews.nvidia.com) The deeper message is that artificial intelligence infrastructure is moving up a level. For years, the industry compared graphics processing units the way car buyers compare horsepower. Vera Rubin is NVIDIA arguing that the new unit of competition is the rack, then the pod, then the entire data center, because autonomous models spend as much time moving information between components as they do doing math. (nvidianews.nvidia.com) That is why investors care about a product that most people will never touch. If agentic artificial intelligence becomes the standard way software gets built, then the winners may be the companies that supply the plumbing for millions of those decisions every second. NVIDIA is betting that Blackwell started that buildout, and Vera Rubin is the system that keeps it going into 2027. (cnbc.com)