Huang noted that he had been asked to bring manufacturing back to the U.S. to contribute to national security and jobs. The first Blackwell wafers have been produced at a factory in Phoenix, Arizona, as a result of NVIDIA’s collaboration with TSMC. NVIDIA also said that systems based on Blackwell will be assembled in the United States.
The Blackwell GPU is manufactured on TSMC’s 4nm process and is one of NVIDIA’s most advanced AI chips. It contains 208 billion transistors, more than 2.5 times that of the previous-generation Hopper chip. Blackwell systems also include the Grace CPU and NVLink switches, are equipped with 192GB of HBM3E memory, and support fifth-generation NVLink technology.
Over the past four quarters, shipments of NVIDIA’s Blackwell GPUs have reached six million units. Huang expects Blackwell and next year’s Rubin chips to generate a combined $500 billion in GPU sales for the company.
NVIDIA also plans to collaborate with the U.S. Department of Energy to build seven supercomputers, some of which will be used for research and stewardship of the U.S. nuclear stockpile. These collaborations will further strengthen NVIDIA’s position as a “national AI hub.”