BlockBeats News, July 22nd. The US stock market just experienced a sharp drop in AI trading, and a popular explanation in the market is that the appearance of Kimi K3 has made investors concerned again about the high-cost route of the US AI model and the return on chip capital expenditure. Last week, the semiconductor sector was under significant pressure, with the PHLX Semiconductor Index falling into a technical bear market territory. Nvidia, Micron, equipment stocks, and storage stocks were all caught in the adjustment.

However, The Information's latest report provided the market with another set of more direct numbers: Nvidia's next-generation Vera Rubin server system has begun customer testing, with dozens of customers such as CoreWeave, Microsoft, OpenAI, Anthropic, and SpaceXAI receiving a small number of test racks. Each Rubin rack contains 72 GPUs, with a unit price of approximately $7 to $8 million, higher than the current price of around $5 million for the Grace Blackwell 300 rack.

What is most stimulating the market's imagination is the production capacity. Nvidia's Senior Vice President of Hardware Engineering, Andrew Bell, stated that more than ten manufacturing partners will eventually have the ability to produce up to 1000 Rubin racks per day. Rough calculations suggest that if running at full capacity, this would equate to a potential rack revenue scale of at least $630 billion per quarter. Of course, this is only the product of theoretical capacity and rack prices, and cannot directly equate to Nvidia's financial guidance, but it is enough to illustrate that the scale of the AI infrastructure arms race is still rapidly rising.

This is the contradiction in the current AI market. Models like Kimi K3, with high performance, low cost, and open-weight characteristics, have weakened the market's confidence in the premium on closed-source models in the US, and also made investors question the capital expenditure logic of "spending more to stay ahead." At the same time, the progress of Rubin shows that core customers such as OpenAI, Anthropic, and Microsoft are still seizing next-generation computing power. The proliferation of cheap models may reduce inference costs, expand AI use cases, and ultimately continue to drive up demand for GPUs, memory, networking, and cooling systems.

The rebound in the US stock market on Tuesday has already reflected this tug-of-war. Tech stocks ended their continuous decline, with the Nasdaq rising by approximately 1.3%, and the PHLX Semiconductor Index recording its largest single-day gain in a month, with storage stocks such as Micron and SanDisk also rebounding significantly. The market has not completely abandoned AI hardware; it is just recalculating: Will the efficiency impact brought by Kimi weaken chip demand, or will it stimulate more developers and companies to deploy AI?

In addition, Nvidia is transforming itself from a GPU supplier to a full-stack AI server infrastructure provider. The company not only sells GPUs but is also expanding into CPUs, network switches, cables, storage, and cooling technology. Faced with companies like Google developing their own AI inference chips, Nvidia's strategy is: even if customers use alternative chips, it still hopes to sell networks, server components, and complete systems.

This means that a key variable in AI hardware transactions is changing. While the market used to focus on how many GPUs NVIDIA sold, it now also considers whether Rubin cabinets can be mass-produced smoothly, if customer data centers are ready for installation, if memory and power supply bottlenecks can be alleviated, and if cloud providers can turn expensive cabinets into actual revenue.

In the short term, Kimi K3 will still suppress the "high-cost moat" narrative in AI valuation; in the medium term, Rubin's production capacity planning reminds investors that AI infrastructure construction is still ongoing. In the upcoming earnings season, cloud provider capital expenditure guidance and NVIDIA's supply chain progress will determine whether this rebound is a technical recovery or the beginning of a repricing in AI hardware transactions.