Nvidia Corp. today announced plans to mass-produce artificial intelligence supercomputers in the U.S.

The company disclosed in conjunction that Taiwan Semiconductor Manufacturing Co. has started producing Blackwell chips at its Arizona fab complex. The hub currently includes one operational processor plant that came online last year. In the coming years, TSMC plans to build five more fabs on the site along with two advanced packaging facilities.

Nvidia is also partnering with several other suppliers on its new stateside manufacturing initiative. The group includes Amkor Technology Inc., Wistron Corporation, SPIL Co. and Foxconn.

Nvidia’s Blackwell chips include 288 gigabits of onboard memory that can be used to store artificial intelligence models’ data. This memory is integrated with the chips’ logic modules, the components that can carry out calculations, using an interconnect technology called packaging. Two of the partners that are involved in Nvidia’s new manufacturing initiative, Amkor and SPIL, are among the world’s top providers of chip packaging.

Amkor is the larger of the two companies with more than $6 billion in annual revenue. It’s currently building a $2 billion packaging plant not far from TSMC’s chip complex in Arizona. Amkor says that the factory will supply packaging components for, among other products, AI systems.

TSMC’s fabs presumably produce only the logic modules of Nvidia’s Blackwell chips. The latter historically relied on SK Hynix Inc. to supply the high-speed memory it integrates with the logic modules. Earlier this year, Samsung Electronics Co. Ltd. also received approval to make memory for Nvidia processors.

The chipmaker’s flagship Blackwell graphics cards can provide 20 petaflops of performance when processing data stored in the FP4 format, which is widely used by AI applications. One petaflop corresponds to 1,000 trillion computations per second. Nvidia recently previewed an improved version of the chip, the Blackwell Ultra, that promises to provide 50% higher performance for inference workloads than the original.

Nvidia detailed today that though its Blackwell chips will be produced in Arizona, its AI supercomputers are set to be made in Texas. The company sells servers under the DGX brand that can be linked together into supercomputers. The most advanced DGX servers combine eight Blackwell chips with nearly 2 terabytes of flash storage and other auxiliary components.

According to Nvidia, Foxconn and Wistron will make supercomputers in Houston and Dallas, respectively. Foxconn, officially Hon Hai Precision Industry Co., recently disclosed that it manufactures most of the world’s Nvidia-powered AI servers. Wistron is another major electronics manufacturer that is reportedly investing $50 million in the U.S. to build an AI server plant.

Nvidia stated that the suppliers involved in the initiative will together dedicate more than a million square feet worth of manufacturing space to making its products. According to the chipmaker, the Foxconn and Wistron facilities will expand production over the next 12 to 18 months.

Nvidia will help design and operate the new plants. The company plans to do so using its Omniverse and Isaac GR00T product suites. 

Omniverse is a set of software tools that can be used to create digital twins, virtual replicas of physical systems. Manufacturers use digital twins to simulate their production lines and find areas for improvement. Isaac GR00T, the other product suite that Nvidia plans to use in the project, makes it easier to develop AI models for factory robots. 

“The engines of the world’s AI infrastructure are being built in the United States for the first time,” said Nvidia Chief Executive Officer Jensen Huang. “Adding American manufacturing helps us better meet the incredible and growing demand for AI chips and supercomputers, strengthens our supply chain and boosts our resiliency.”

Image: Nvidia