U.S. AI stocks had just come through a sharp pullback, and one popular explanation in the market was that Kimi K3 had revived doubts about the high-cost model strategy pursued by U.S. AI companies and the return on heavy chip capital spending.
Semiconductor shares were under visible pressure last week. The PHLX Semiconductor Index briefly fell into technical bear-market territory, while Nvidia, Micron, equipment names, and memory stocks were all caught in the adjustment.
Then The Information published a fresh set of figures that pushed the conversation in a different direction. Nvidia’s next-generation Vera Rubin server system has entered customer testing, according to the report, and dozens of customers including CoreWeave, Microsoft, OpenAI, Anthropic, and SpaceXAI have received a small number of test racks.
Rubin racks are priced above the current generation
Each Rubin rack includes 72 GPUs and carries a price of about $7 million to $8 million, above the roughly $5 million price tag of the current Grace Blackwell 300 rack.
The bigger trigger for the market was capacity. Andrew Bell, Nvidia’s senior vice president of hardware engineering, said more than 10 manufacturing partners will ultimately have the ability to produce as many as 1,000 Rubin racks per day.
On a rough calculation, if that capacity were fully utilized, it would imply at least $630 billion in potential rack revenue over one quarter. The figure is simply the product of theoretical output and rack pricing, not something that can be treated as Nvidia revenue guidance, but it gives a sense of how quickly the scale of the AI infrastructure race is still rising.
Kimi K3 and Rubin are pulling the AI trade in opposite directions
That tension sits at the center of the current AI trade. High-performance, lower-cost, open-weight models such as Kimi K3 have weakened confidence in the premium once assigned to U.S. closed models. They have also pushed investors to question the idea that ever-higher spending automatically preserves a lead.
At the same time, Rubin’s rollout shows that key customers including OpenAI, Anthropic, and Microsoft are still competing for next-generation compute. The spread of cheaper models could lower inference costs, but it could also broaden AI use cases and, in turn, keep demand rising for GPUs, memory, networking, and cooling systems.
Tuesday’s rebound in U.S. equities reflected that push and pull. Technology stocks snapped a losing streak, the Nasdaq rose about 1.3%, the PHLX Semiconductor Index posted its biggest one-day gain in a month, and memory names including Micron and SanDisk rebounded sharply.
The market has not abandoned AI hardware. What it is doing is recalculating whether the efficiency shock from Kimi will reduce chip demand or encourage more developers and enterprises to deploy AI.
Nvidia is pushing beyond GPUs into full AI infrastructure
Nvidia is also moving to position itself less as a standalone GPU supplier and more as a full AI server infrastructure provider. The company is not only selling GPUs; it is also building out CPUs, network switches, cables, storage, and cooling technologies.
As companies such as Google develop their own AI inference chips, Nvidia’s strategy is to keep selling networking, server components, and supporting systems even when customers adopt alternative chips.
That changes the variables investors need to watch. The market used to focus mainly on how many GPUs Nvidia could sell. Now it also has to track whether Rubin racks can be mass-produced on schedule, whether customer data centers are ready to install them, whether memory and power bottlenecks can be eased, and whether cloud companies can turn expensive racks into real revenue.
In the short term, Kimi K3 is still likely to weigh on the “high-cost moat” narrative embedded in AI valuations. Over the medium term, Rubin capacity planning is a reminder that AI infrastructure spending has not stopped. In the coming earnings season, cloud capital spending guidance and progress across Nvidia’s supply chain will shape whether this rebound is just a technical recovery or the start of a broader repricing in the AI hardware trade.