Nvidia's daily production capacity of 1,000 cabinets is only the upper limit; Rubin's real bottleneck is in the computer room
Author: Groove BlockBeats
Original title: Nvidia Rubin's latest developments: Test cabinets have been delivered, what does the daily production capacity of 1000 cabinets mean?
TL; DR
· According to The Information, Vera Rubin test cabinets have been delivered to dozens of customers, about $7 million to $8 million per cabinet.
· Each cabinet contains 72 Rubin GPUs and 36 Vera CPUs, and manufacturing partners target production capacity of up to 1000 units per day.
· 1000 cabinets are only the upper limit of production capacity, not equal to the order. The customer data center's ability to connect electricity, liquid cool, and go online is still limited.
According to The Information, Nvidia's next-generation Vera Rubin server system has delivered a small number of test cabinets to dozens of customers. The price of a single cabinet is about 7 million to 8 million US dollars, and the manufacturing partner's ultimate target production capacity can reach up to 1,000 units per day.
NVIDIA Vera Rubin NVL144 CPX rack with tray
This set of numbers puts the focus of Nvidia's next round of AI hardware upgrades quite bluntly: it's not just selling more powerful GPUs, but more expensive, more complex full-cabinet server systems. According to official Nvidia data, the Vera Rubin NVL72 is a rack-scale system containing 72 Rubin GPUs and 36 Vera CPUs. Compared to the current flagship Grace Blackwell 300 rack's price of approximately $5 million, the Rubin single cabinet price has further increased.
CoreWeave announced in June that bring-up and verification of the Vera Rubin NVL72 has been completed. Nvidia also recently claimed that Vera Rubin has entered the full capacity climb phase and is already running machines with partners such as CoreWeave, Google Cloud, Microsoft Azure, and OCI. In other words, Rubin is no longer just a paper product, but the $7 million to $8 million single cabinet price revealed by The Information, dozens of test customers, and a target production capacity of 1,000 containers per day are still not Nvidia's official revenue guidelines.
Up to $8 million for a single cabinet, I bought a complete set of computing power units
Rubin's price is not simply a chip price increase, but a continuation of Nvidia's full-cabinet system route.
After Blackwell, the core products Nvidia sold to major customers were increasingly not isolated GPUs, but systems that packaged GPUs, CPUs, networking, cooling, power, software, and rack-level connectivity. The customer purchased a computing power unit that can be used for data center planning, rather than components that can be assembled from scratch by themselves.
This is why the price of a single cabinet can reach 7 million to 8 million dollars. According to Nvidia's technical data, the Vera Rubin NVL72 is about 4,000 pounds, which is close to the weight of a pickup truck. For cloud vendors and AI companies, purchasing Rubin is not only about placing an order chip, but also simultaneously preparing the computer room load, power supply, liquid cooling, network connection, and online commissioning.
The Information paraphrased Ian Buck, Nvidia's vice president of high-performance computing, as saying that the company wants to sell to all customers, but the actual distribution will be linked to whether the customer has the ability to physically install and bring servers online. Who can actually connect these cabinets to the data center is more likely to get more goods.
1000 cabinets a day is amazing but not an order
The easiest thing to trigger the market's imagination is “1000 cabinets a day.”
The Information paraphrased Andrew Bell, Nvidia's senior vice president of hardware engineering, as saying that the more than ten manufacturing partners working with Nvidia to produce Rubin racks will eventually be able to produce up to 1,000 racks per day. At $7 million to $8 million per cabinet, this represents a huge potential revenue potential.
Based on this production capacity scale, the report estimates that if 1,000 containers continue to be produced every day, the previous quarter could theoretically correspond to at least 630 billion US dollars in revenue. By comparison, Nvidia's revenue for the quarter ended April 26, 2026 was $81.6 billion.
However, this figure can only be understood as a theoretical estimate of “production capacity multiplied by unit price.” It is not a revenue guide given by Nvidia management, nor is it an order confirmation, let alone indicate that manufacturing partners will be fully productive for a long time. Actual shipping also depends on customer orders, supply chain, data center construction, inspection pace, and revenue recognition rules.
For investors, 1000 cabinets/day is more like Nvidia showing a manufacturing limit for next-generation systems rather than a short-term revenue figure that can be put directly into a profit model. What can really change the pace of revenue is whether these expensive racks can be delivered stably, installed in the customer's computer room, and up and running.
Blackwell's lessons let Rubin solve manufacturing challenges first
The first thing Rubin wants to prove is not just performance, but whether it can build, install, and run more smoothly than Blackwell.
“We almost messed everything up,” Bell said, recalling the problems Blackwell racks had last year. There were problems with hardware, software, diagnostic systems, and manufacturing. After entering a rack size that had never been reached before, Nvidia almost tore the system down and rebuilt.
This experience explains why Rubin's manufacturing details are emphasized by executives. Next-generation racks aren't “cable-free.” According to the Nvidia technology blog, the NVLink spine has pre-integrated cable cartridges at the rear, containing approximately 5,000 copper cables. The change is that more cables and components are made into modular, pre-integrated structures, reducing manual wiring work on site and production lines, and improving consistency and assembly speed.
This is critical for Nvidia. AI customers buy Rubin not only for book computing power, but also for stable delivery and quick launch. If Blackwell-like hardware, software, diagnostic, or manufacturing issues reappear in next-generation racks, the pace of customer capital expenditure will slow down, and Nvidia's shipping and revenue recognition will also be affected.
Nvidia is also locking customers into a complete data center system
Another implication of Rubin is that Nvidia is moving its position from a GPU vendor to an AI data center systems vendor.
In a conversation at headquarters, Nvidia executives emphasized that the company not only sells GPUs, but also sells non-GPU components such as CPUs, network switches, cables, storage, and server cooling technology. This is not simply expanding the category, but continuing to maintain key aspects of data centers as cloud vendors and big model companies look for alternative AI chips.
The competition is getting more complicated. According to public information from Google, the eighth-generation TPU is divided into training TPU 8t and inference TPU 8i, and the inference chip and training chip routes are diverging. Cloud vendors and big model companies are also looking for cheaper and more controllable computing power solutions. If customers use non-Nvidia AI chips in the future, Nvidia still wants to sell them networking, connectivity, cooling, and other server components.
This is also the reason why Rubin's single cabinet price has attracted attention. It is not an offer for a single chip, but a new price scale formed by Nvidia after packaging key aspects of AI data centers. The more customers purchase the full cabinet system, the more Nvidia can earn additional revenue from networking, CPU, cooling, and system integration.
If the expensive rack actually runs, it still gets stuck in the data center site
Rubin's production capacity is already climbing, but large-scale commercial deployment is likely to continue until 2027. Customers need to complete electricity, liquid cooling, rack installation, network coordination, and software adaptation. Any delay will affect the actual pace of use of Rubin.
Supply chain constraints aren't limited to advanced chip manufacturing either. Nvidia executives mentioned that the company has a team that continuously tracks the supply chain and establishes connections with companies involved in natural resources such as aluminum and indium phosphide. These materials relate to servers and network equipment, indicating that after the entire cabinet system is released, bottlenecks may occur in more traditional industrial processes, not just wafers and packages.
The goal of 1000 cabinets per day gave the market a huge room for imagination, but what Rubin really needed to overcome was a series of realistic limitations: can manufacturing partners assemble stably, can customers receive and install it, can the data center provide sufficient power and cooling, and whether non-GPU components can be released simultaneously.
Nvidia has pushed the next-generation AI hardware business towards higher unit prices and greater system complexity. What determines Rubin's success is not just the computing power of 72 GPUs per cabinet, but whether these multi-million dollar racks can get into the data center as planned and actually run.
Twitter:https://twitter.com/BitpushNewsCN
Compare the TG exchange group:https://t.me/BitPushCommunity
Compare TG subscriptions:https://t.me/bitpush
