Morgan Stanley analysts believe that despite lower production due to a significant but easily repairable design issue, Nvidia will still produce approximately 450000 artificial intelligence GPUs based on the Blackwell architecture. If this information is accurate and the company is able to sell these products this year, it could bring in over $10 billion in revenue.
According to reports, analysts from investment bank Morgan Stanley wrote in a statement to clients, "Blackwell chips are expected to produce 450000 units in the fourth quarter of 2024, which could bring potential revenue of over $10 billion to Nvidia
Although the figures of $10 billion and $450000 may seem impressive, Nvidia will sell the scarce Blackwell GPUs at around $22000 per piece, much lower than the rumored price of $70000 per piece. Although the actual pricing of data center hardware depends on factors such as quantity and demand, it is a bit strange to sell the first batch of ultra high end GPUs at a lower price compared to Nvidia's current generation H100.
Of course, Nvidia is more inclined to sell "reference" artificial intelligence server cabinets driven by its Blackwell GPU: the NVL36 is equipped with 36 B200 GPUs and is expected to sell for $1.8 million to $2 million; NVL72 is equipped with 72 B200 GPUs internally, with an expected starting price of $3 million. The profit from selling cabinets instead of GPUs, GPU modules, and even DGX and HGX servers is much higher, and it is expected to earn $10 billion by selling such machines. However, in this case, Nvidia does not need to provide 450000 GPUs to generate $10 billion in revenue.
In late August, in order to increase production, Nvidia announced that it had to change the GPU photomask design codenamed B100/B200. The company also stated that GPUs will enter mass production in the fourth quarter and continue until the 2026 fiscal year. In the fourth quarter of fiscal year 2025 (starting from the end of October), Nvidia is expected to achieve "billions of dollars in Blackwell revenue," far below the figures mentioned by Morgan Stanley analysts.
According to SemiAnalysis's report, the mismatch in coefficient of thermal expansion (CTE) between the Blackwell chip and packaging material resulted in the malfunction. Nvidia's B100 and B200 GPUs adopt TSMC's CoWoS-L package for the first time, which uses a passive local silicon interconnect (LSI) bridge integrated into the redistribution layer (RDL) intermediate layer to connect the chips. The precise placement of these bridges is crucial, but CTE mismatch between chips, bridges, organic intercalation layers, and substrates can lead to warping and system failures. According to reports, Nvidia had to redesign the top metal layer of the GPU to increase production, but did not disclose details, only mentioning new masks. The company clarified that it has not made any functional changes to the Blackwell silicon wafers and is only focused on increasing production.
Due to the need to change the design of B100 and B200 GPUs, and TSMC's CoWoS-L (a method that is vastly different from mature CoWoS-S technology) packaging capability, analysts predict that Nvidia will not ship too many Blackwell based GPUs in the fourth quarter of fiscal year 2024 or 2025.
