Nvidia Rubin testing shifts AI hardware trade back toward supply concerns
A sharp sell-off in U.S. AI stocks had recently been tied to concerns that Kimi K3, a high-performance and lower-cost open-weight model, could weaken confidence in the expensive closed-model path pursued by major U.S. AI companies. That view hit semiconductor names last week, pushing the PHLX Semiconductor Index briefly into technical bear-market territory and dragging Nvidia, Micron, equipment makers, and memory stocks lower. But a new report from The Information gave investors a different set of numbers to focus on. Nvidia’s next-generation Vera Rubin server systems have entered customer testing, with CoreWeave, Microsoft, OpenAI, Anthropic, and SpaceXAI among dozens of customers that have received a small number of test racks. Each Rubin rack includes 72 GPUs and is priced at roughly $7 million to $8 million, above the roughly $5 million price tag for the current Grace Blackwell 300 rack. Nvidia senior vice president of hardware engineering Andrew Bell said more than 10 manufacturing partners are ultimately expected to gain the capacity to produce as many as 1,000 Rubin racks per day. That theoretical output, multiplied by rack pricing, points to a massive infrastructure buildout still underway, even as the market debates whether lower-cost AI models will reduce chip demand or expand deployment enough to keep spending elevated.
