Nvidia hopes to sell AI supercomputers made in the United States worth up to half a billion dollars over the next four-year period, with the assistance of Taiwan Semiconductor Manufacturing Co., aka TSMC and its partners.
According to the pledge, Nvidia plans to begin production in the United States of its latest Blackwell accelerators and system within the next 12-15 month. It also assumes there will be sufficient buyers, domestic or otherwise, of this half-a-trillion-bucks of made-in-America gear over the coming years.
This promise comes at a time when American companies are grappling against the Trump regime’s tariffs on foreign products.
“The engines of the world’s AI infrastructure are being built in the United States for the first time,” Jensen Huang, CEO of Nvidia, said in a canned statement on Monday. “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.”
Nvidia has contracted more than a milion square feet of manufacturing space in Arizona to assemble chips and build servers. According to Huang, these sites are expected to support the deployment of datacenters with tens of gigawatts in the United States.
Production of Nvidia’s Blackwell chips – which relies on TSMC’s chip-on-wafer-on-substrate (CoWoS) packaging technology – is said to have begun earlier this year at TSMC’s Phoenix, Arizona, chip fabs.
Nvidia used TSMC CoWoS technology to integrate high bandwidth memory modules into its highest-end accelerations, but only outside the US. TSMC announced late last year that it would bring its CoWoS technology and integrated fan-out packaging to America. This was part of a partnership with Arizona-based Amkor, a chip packaging and testing provider, who will work with Taiwan-based Siliconware Precision Industries Co. to assemble and to test its US-made Blackwell Accelerators. Nvidia has partnered with Foxconn and Wistron, two contract manufacturing outfits, to set up manufacturing facilities in Houston, Texas, and Dallas, Texas, respectively, in order to produce its increasingly powerful servers and rack systems. Wistron appears to have a $50 million budget to purchase land or property in the US. These plants will be using Nvidia’s Omniverse Digital Twin and Isaac GR00T Robotics Platforms to automate production.
The final assembly is performed in the US to benefit Nvidia, its American customers and Nvidia itself. AI systems can be unique and it would be difficult to fix a fault or misconfiguration of an imported system. Final checks in the US can save a lot back and forth.
We actually used to have chassis standards. We had power standards. We had keep-out zones. [Nvidia’s server design spec] HGX is one of those standards. But they didn’t comply with wattage, or even rack height.” Moor Insights and Strategy Chief Analyst Patrick Moorhead told El Reg. “These days, almost all of them are bespoke.”
According to Moorhead, Nvidia has made its system design more complex over the last few years. The tolerances have also tightened as the GPU giant has moved from semi-standard air cooled systems to rack scale machines that contain more than a 100 liquid-cooled GPUs and CPUs connected by miles and miles of copper cable.
What exactly is $500 billion?
It’s unclear if Nvidia will sell 500 billion dollars of its own chips, DGX-branded servers, or if it will also include machines made in the US by Dell, HPE and Supermicro. The announcement from the GPU giant is not very clear.
Moorhead argued that if the $500 billion was just homegrown DGX sales, then it would be ambitious. If it included HPE et. toit would be conservative in order to avoid being interpreted by some as a long-term guide. He said that this would also be a change in business model, as the ecosystem currently offers a mix of Nvidia servers, equipment for hyperscalers and traditional boxes such as those from HPE and Dell. Nvidia declined comment.
- China raises tariffs on US products to 125% and calls Trump’s war a joking joke
- Nvidia spent $1M at Mar-a-Lago, US scrapped AI chip export crackdown later
- Nvidia may be forced to increase the price of its AI suite due to Jensen GPU math error
- AI Datacenters want to go Nuclear. It’s a shame they didn’t need it yesterday.
The GPU giant’s commitment to US Manufacturing comes a week after it was reported that the Trump regime would pause controls that prohibited the export of Nv’s H20 GPUs into China, following a appearance by Huang at an $1 million-per-head dinner in the US President’s Mar-a-Lago home. Nvidia’s move to manufacture in the USA is not unique
as the White House encourages corporations to “reshore”. TSMC announced last month plans to spend up to $100 billion in the States to build three new fabrication facilities, two advanced manufacturing centers, and an R&D centre.
In February, Apple announced that it would spend $500 billion on the US. This included a commitment to buy a large number of chips from TSMC’s Arizona fab. It also planned to build a server manufacturing plant in Houston – the same city in which Foxconn will be setting up facilities for Nvidia. It sounds suspicious that Foxconn’s factory could serve both tech giants. (r)
Boot note
Apple, Nvidia and others are also interested in TSMC’s US production capacity. AMD announced that it had validated its 5th generation Epyc datacenter chip, codenamed Turin, for production at TSMC Fab 21 Arizona. AMD also announced that its 6th-gen Venice Epycs will be built using TSMC’s 2nm technology when they arrive in 2020.
