Nvidia is experiencing a big supply bottleneck for their RTX Pro 5500 AI accelerators in China, causing a huge unfulfilled demand with major cloud service providers. This results in a big delay for hyperscalers who have already contracted billions of dollars to secure the availability. Essentially this means that the plans for their infrastructure growth will be seriously delayed. For those who need these chips to immediately support AI training, be prepared for big delays (possibly until 2027).
Major Chinese cloud providers face severe delays with only 1% of their $70 billion RTX Pro 5500 order arriving in the next four months
Specifically, the focus of this huge bottleneck issue is around the Nvidia RTX Pro 5500 – a premium GPU for enterprise AI workloads. This chip model (RTX Pro 5500) is the one that Chinese tech companies have ordered ~6 million units for. This large volume of orders really showcases the fierce competition to get the latest and greatest Nvidia silicon out of China.
- Unit Price: >$11,000
- Total Order Volume: ~6,000,000 units
- Total Order Value: ~$70 billion
- Initial Delivery Rate: ~1% of total order
When you start looking into the actual financial numbers for all these orders, you'll really see the magnitude of the issue. The per unit price for the chip is well above $11,000. So the total orders amount to ~$70B.
Yet, the initial availability is still extremely low. In the next four months, customers will be getting around ~1% of their total order amount (~tens of thousands of units).
Due to manufacturing limitations, the delivery schedule is stretched out over a few years. The initial thousands of chips are set to arrive before the end of 2024. And then another wave of tens of thousands of units should arrive in February 2025. Nvidia says that the final shipping numbers could fluctuate depending on how things go with availability in the future.
However, looking into the future, given that they're now shifting to production of newer models, availability is murky at best. Suppliers estimate that they could deliver ~700,000 units to Chinese hyperscalers by the end of 2027. Essentially, this reflects the fact that Nvidia is really moving towards newer models (like Rubin). So they're probably going to focus more on the newer generations rather than the current demand.
one percent? that is brutal for a seventy billion dollar order 💀
The unit price exceeding eleven thousand dollars makes the financial impact severe. I assume the supply chain issues are due to advanced packaging constraints rather than raw wafer fab capacity. This delay will likely shift their training timelines to 2027.
i am just glad i do not need to rent one of these for my home server. the rack space alone would be terrifying.
does the feb 2025 delivery wave include the full contract value or just the initial batch?
The report specifies only tens of thousands of units arriving in February. It seems the seventy thousand total by 2027 is the realistic cap, not the six million ordered.
We need independent benchmark data before assuming the performance justifies the wait. Theoretical throughput does not equal real world AI training efficiency.
Driver stack stability will be tested when they finally scale this cluster. Kernel module conflicts usually appear only under heavy load.
is there a secondary market for these or do they ship directly to the hyperscalers?
The article implies the six million units are all RTX Pro 5500. The total order value suggests a broader mix of enterprise hardware might be included in that figure.