Blackwell shipments are likely to ramp up very quickly
Nvidia's H100 GPU was so popular in 2023 that the company had a whopping 98% share in the market for AI data center chips. It's still a top seller today, but data center operators are lining up to buy Nvidia's latest chips, which are built on its new Blackwell architecture. They offer a significant leap in performance, meaning they can process high volumes of data more quickly in AI training and AI inference workloads.
The Blackwell-based GB200 NVL72 system can perform AI inference at 30 times the speed of the equivalent H100 system. A single GB200 GPU within the NVL72 system sells for around $83,000, and while that's roughly double the price of an H100, the 30-fold increase in inference performance translates into significant cost savings for any company deploying AI.
Microsoft is reportedly the largest buyer of Blackwell GPUs so far. It will use them to develop AI for its own purposes, but it will also rent the computing capacity to other AI developers for a fee through its Azure cloud platform. AmazonWeb Services, Alphabet's Google Cloud, and Oracle are likely to be large Blackwell customers for the same reasons.
Nvidia shipped 13,000 Blackwell samples to customers during its recent fiscal 2025 third quarter (ended Oct. 27), but Jensen Huang says demand is "staggering," so that number will probably grow rapidly. Morgan Stanley predicts that Nvidia will ship up to 300,000 units in the last three months of calendar 2024, followed by 800,000 units in the first three months of 2025