Nvidia's Blackwell

TapTechNews July 26th news, Morgan Stanley institution newly released an analysis report, estimating that only the revenue from the Blackwell GB200 AI server order alone will bring Nvidia an annual income of 210 billion US dollars (TapTechNews note: currently about 1.52 trillion Chinese yuan).

The GB200 chip was released on March 19th, consisting of two B200 Blackwell GPUs and one Arm-based Grace CPU. The inference performance of the large language model is 30 times higher than that of the H100, and the cost and energy consumption are reduced to one twenty-fifth.

Morgan Stanley said that the latest GB200 AI server order is in a state of short supply, and it is estimated that it will continue beyond 2025.

The annual shipment volume of Nvidia's Blackwell GB200 AI server is 60,000 to 70,000 units. Calculated according to 2 to 3 million US dollars per server, only the revenue from the Blackwell server alone can bring Nvidia an annual income of 210 billion US dollars.

Nvidia's GB200 AI chip server has two specifications of NVL72 and NVL36. Due to the increasing number of AI start-ups choosing the more economically viable NVL36 server, NVL36 is more favored.

TapTechNews reported in June this year that Nvidia's brand-new GB200 series AI chips are in short supply. After Nvidia increased the advanced process wafer input to TSMC, it also placed follow-up orders with the back-end packaging and testing plants. The relevant order volume of ASE and King Yuan Electronics in the fourth quarter will double compared to the previous quarter.

The test process of the GB200 and B series AI chips is significantly longer than that of the previous-generation H series. It must go through four consecutive procedures, including the final test, the Burn-in aging test, and then return to the final test, and finally conduct the SLT system-level test.