TapTechNews August 2nd news, Tianfeng Securities analyst Ming-Chi Kuo released a market investment brief today, analyzing the potential factors that affect Nvidia's GB200NVL36 and NVL72 cabinet products in the short term.
Ming-Chi Kuo said that currently, the stock prices of most supply chain companies involved in Nvidia's GB200NVL36 and NVL72 products have risen sharply, but short-term investment decisions cannot rely solely on supply chain investigations and need to consider more aspects.
TapTechNews attaches the relevant analysis content of Ming-Chi Kuo as follows:
According to TrendForce's prediction, the proportion of AI servers in the overall server shipments in 2024 and 2025 is about 8.8% and 12.2% respectively.
The power consumption of each GB200NVL36 cabinet is about 80 kilowatts, and according to AMAX's survey in April this year, currently less than 5% of the data centers in the world can support 50-kilowatt servers per cabinet. Therefore, before purchasing GB200NVL36, it is necessary to ensure that there is enough space for installation.
The single cabinet version of GB200NVL72, with a power consumption of 130 kilowatts per cabinet, cannot be mass-produced in the short term. Although there is a NVL72 plan that provides a dual NVL36 cabinet, when NVL72 is proposed, the key selling point for customers is to provide the largest AI computing power with the smallest space, so it is necessary to observe whether the dual NVL36 cabinet plan that takes up more space can be favored by customers.
BlackwellUltra and VeraRubin will be taped out in 3Q24 and 1H25 respectively, will customers change to buy BlackwellUltra or VeraRubin AI servers with better performance or more cost-effective in the foreseeable future, and affect the short-term demand for GB200?
If customers change to buy BlackwellUltra and VeraRubin AI servers, Nvidia is undoubtedly still the biggest winner, but are the supply chain beneficiaries the same manufacturers as the current GB200?
There are many models of AI servers using Blackwell chips, not just GB200NVL36/72. If customers first purchase other Blackwell's AI server models that ship first due to the demand for AI computing power (such as the GB version with fewer GPUs than NVL36, or the x86 scheme), will it affect the demand for GB200NVL36 in the short term?
The computing power advantage of GB200NVL36 is undeniable, but it also faces many unprecedented design and production challenges, so can it ensure a large number of shipments on schedule (the market consensus is September-October)?
In terms of assembly, currently the EVTs of L10 and L11 are both about 8 weeks (the longest time-consuming ones are the reliability verification and safety testing processes respectively), both shorter than the assembly EVT of the iPhone.
The system development requirements of AI servers should be significantly higher than that of consumer electronics. Can such a compact development schedule solve unprecedented design and production challenges and ensure on-time mass production?
Related reading:
TrendForce Consulting: Nvidia's Blackwell high energy consumption promotes cooling demand, and it is estimated that the penetration rate of the water-cooling scheme for AI servers will reach 10% by the end of the year.