Huang unveils Vera Rubin
Jensen Huang used GTC to introduce the Vera Rubin superchip system, another step up in NVIDIA’s hardware roadmap that’s meant to accelerate large AI workloads. Market commentary ties Rubin and NVIDIA’s Blackwell systems to massive revenue expectations — one outlet quoted NVIDIA forecasting up to $1 trillion in related revenue across 2026–2027 — underscoring how investors see these platforms as multi‑year profit drivers. (x.com) (investing.com)
NVIDIA is no longer selling just chips. On March 16, 2026, Jensen Huang walked onstage at the GPU Technology Conference in San Jose and pitched Vera Rubin as a full artificial intelligence supercomputer made from seven chips and five rack types that work as one system. (nvidianews.nvidia.com) That shift matters because the old unit of sale was a server, and the new unit is a whole “artificial intelligence factory” that can train models, tune them after training, and run them for users in the same stack. NVIDIA said Vera Rubin combines its Vera central processor, Rubin graphics processor, NVLink switch, ConnectX network card, BlueField data processing unit, Spectrum Ethernet switch, and a Groq inference chip. (nvidianews.nvidia.com) A graphics processor is the part that does the heavy math for artificial intelligence, like adding more lanes to a highway so more cars can move at once. NVIDIA built its rise by selling those graphics processors, and now it is bundling the roads, the traffic lights, and the power grid around them. (blogs.nvidia.com, nvidianews.nvidia.com) The immediate reason is that artificial intelligence workloads changed. Huang said the market is moving from simple chatbot replies to “agentic” systems that generate far more tokens, which are the small chunks of text a model reads and writes, and that pushes customers to buy faster inference machines. (blogs.nvidia.com, cnbc.com) Blackwell is the current generation doing that work today, and Rubin is the next one meant to extend it. In NVIDIA’s February 25, 2026 earnings release, Huang said Grace Blackwell with NVLink is “the king of inference today” and said Vera Rubin would push that lead further. (nvidianews.nvidia.com) The money behind that roadmap is already enormous. NVIDIA reported $68.1 billion in revenue for the quarter ended January 25, 2026, including $62.3 billion from data center sales, and full-year revenue of $215.9 billion. (nvidianews.nvidia.com) Huang used GTC to put an even bigger number on the table. CNBC reported that he told the audience he sees $1 trillion in purchase orders for Blackwell and Vera Rubin systems through 2027, up from a $500 billion opportunity estimate the company discussed a year earlier. (cnbc.com) That is why investors treat Rubin less like a science project and more like the next assembly line. If customers are buying entire racks instead of individual chips, NVIDIA captures more of the spending on compute, networking, storage, and system design in one sale. (nvidianews.nvidia.com, cnbc.com) NVIDIA also tied Rubin to the industry’s biggest bottleneck: electricity. CNBC reported that the Vera Rubin system contains 1.3 million components and that NVIDIA claims it can deliver 10 times the performance per watt of Grace Blackwell, which is a direct pitch to cloud companies trying to fit more artificial intelligence output inside fixed power budgets. (cnbc.com) So the launch was not just a product reveal at a developer conference. It was NVIDIA saying the next fight is over who supplies the entire artificial intelligence factory, and Vera Rubin is the machine it wants customers to build that factory around. (nvidianews.nvidia.com, blogs.nvidia.com)
