NVIDIA Supercomputers Reach Texas While Chips Still Leave Arizona

NVIDIA said on April 14, 2025, that it would produce NVIDIA AI supercomputers entirely in the U.S. The White House called the move the Trump Effect in action.

A Taiwanese contractor has since opened a Fort Worth plant that builds GB300 boards Huang priced at $4 million each. The wafers that become those chips still leave Arizona for packaging in Taiwan.

NVIDIA Promised Entire Supercomputers on U.S. Soil

The company’s newsroom said it was working with manufacturing partners to design and build factories that would, for the first time, make NVIDIA AI supercomputers wholly in the United States. Together they had commissioned more than a million square feet of space to build and test Blackwell chips in Arizona and AI supercomputers in Texas.

Blackwell production, NVIDIA said that day, had already started at TSMC’s plants in Phoenix. Supercomputer plants would rise in Texas with Foxconn in Houston and Wistron in Dallas. Mass production at both was expected to ramp in the next 12 to 15 months, a window that ran from April 14, 2026, to July 14, 2026.

Amkor and SPIL were named for packaging and test work in Arizona. Within four years, NVIDIA planned to produce up to half a trillion dollars of AI infrastructure in the United States through those partners, plus TSMC, Foxconn and Wistron. Founder and chief executive Jensen Huang tied the shift to demand, supply-chain strength and resiliency.

The engines of the world’s AI infrastructure are being built in the United States for the first time

Jensen Huang, founder and CEO, NVIDIA newsroom, April 14, 2025

The earlier U.S. supercomputer pledge landed as Washington was threatening new duties on chips and electronics. NVIDIA designs the GPUs. Contract factories cut the wafers, attach the memory and screw the racks together. That split is the whole story that followed.

THE 15-MONTH CLOCK

1. April 14, 2025: NVIDIA says entire AI supercomputers will be produced in the United States and names Texas plants with Foxconn and Wistron.
2. April 14, 2026: The 12-month mark of the promised mass-production ramp arrives.
3. July 14, 2026: The 15-month mark of that ramp arrives.
4. July 21, 2026: Wistron opens its Fort Worth D1 plant, seven days after the 15-month mark, and starts GB300 board work.
5. Early 2028: Amkor slates production at its Peoria packaging campus, the U.S. step that still is not running.

The April blog also said NVIDIA would use Omniverse digital twins and Isaac GR00T robots to design and staff the new lines. That software promise showed up later on the Houston floor, where dual-arm cells are being taught Blackwell assembly tasks that used to be done by hand.

Fort Worth Opened Just Past the 15-Month Mark

On July 21, 2026, Wistron opened its first U.S. manufacturing facility in Fort Worth, the Dallas plant in all but the city name. NVIDIA’s April note had said Dallas. The working factory sits in Fort Worth, a 324,000-square-foot greenfield site NVIDIA described as a $700 million commitment.

The D1 plant runs two cells. One produces the NVIDIA GB300 Grace Blackwell Ultra Superchip. The other is reserved for the Vera Rubin Superchip. NVIDIA said D1 is scaling this year toward tens of thousands of boards per month.

More than 500 jobs have been added, with plans to reach 1,000 by the end of 2026. Huang stood with Wistron chairman Simon Lin at the opening and said manufacturing is an essential pillar for every economy and every country. He also said chip plants, packaging plants, computer-system plants and AI factories were letting the United States reindustrialize.

Engineers had already built the factory once in software. Wistron designed and fully simulated the site in a digital twin on NVIDIA’s open tools, including Omniverse, before construction, then used that model to set line layouts and train workers. Huang signed the first GB300 Superchip produced on site.

WHERE THE WORK RUNS

Foxconn’s Houston line is the quieter half of the Texas plan. NVIDIA named it in April 2025 as a supercomputer plant, and the site is building NVIDIA AI server systems. Robot cells on that floor are now being used for high-precision Blackwell assembly, which is a labor story as much as a geography story. The public ribbon-cutting that NVIDIA documented in detail was Wistron’s.

Huang Signed the First $4 Million Board

Huang called the system built around the GB300 Grace Blackwell Ultra Superchip the most powerful AI supercomputer in the world. He said it has one-and-a-half million parts inside, weighs two tons, and costs $4 million, and that Wistron is producing them like phones because the world needs the machines.

That is the unit Texas can actually make. A GB300 board is a finished computing tray, not a wafer and not a packaged GPU. Fort Worth’s job is to put a huge number of parts onto a board, test it, and send it on as the building block of a rack. The ceremony made that board the patriot object. The silicon inside it has a longer commute.

NVIDIA’s own newsroom posted from the floor the day D1 opened, with Huang and Lin on site as the first U.S. GB300 boards came off the line.

NVIDIA CEO Jensen Huang joined Wistron Chairman Simon Lin as Wistron opened its first U.S. facility in Texas, producing the GB300 Grace Blackwell Ultra Superchip.

We’re proud to partner with Wistron as its global expertise strengthens U.S. manufacturing and the supply chain… pic.twitter.com/iwi4agzSxs

— NVIDIA Newsroom (@nvidianewsroom) July 21, 2026

Lin called the project a vision that ran from the whole AI world down to community and country. Huang talked about jobs in plumbing, construction, manufacturing and electrical work. Those jobs are real. They attach to board assembly, not to the step that turns a Phoenix wafer into a working GPU.

Packaging Is the Step Arizona Cannot Do

TSMC can now cut advanced NVIDIA logic in Phoenix. Paul Rousseau, TSMC’s North America packaging solutions head, said on April 8, 2026, that the company still sends 100 percent of chips to Taiwan to be packaged, including chips made at the Phoenix fab. He did not give a completion date for TSMC’s Arizona packaging sites.

The missing step has a name. CoWoS, chip on wafer on substrate, is how TSMC sits a GPU die next to stacks of high-bandwidth memory on a silicon interposer so the two can talk at the speed the model needs. Rousseau said that method is growing at an 80 percent compound annual rate. NVIDIA has reserved a majority of TSMC’s most advanced CoWoS capacity.

Without that attach, an Arizona wafer is not yet a Blackwell accelerator and cannot go into an NVL72-class system. Analysts who follow the line have been blunt about the same handoff: Phoenix can print the die, and Taiwan still finishes it. Jan Vardaman of TechSearch International has said putting packaging next to the Arizona fab would cut the round trip. That building is not the one that opened in July.

THE PATH FROM WAFER TO RACK

• Wafer step: TSMC’s Phoenix plants fabricate Blackwell silicon in the United States.
• Packaging step: Those wafers go to Taiwan for CoWoS, where the GPU die is joined to high-bandwidth memory.
• Board step: The finished GPU comes back for Superchip work at Wistron in Fort Worth.
• Rack step: Foxconn in Houston builds the server systems around those boards.
• U.S. packaging step: Amkor’s Peoria campus is meant to do advanced packaging and test, with production slated for early 2028.

A Texas-built rack can therefore carry a wafer that left Arizona, became a GPU in Taiwan, and returned for screws, cooling and burn-in. Procurement teams that treat origin as a single country field will misread that chain. The useful map is the handoff, not the press release.

The Contractors Who Own the Factory Floor

NVIDIA does not run these plants as a chip foundry. TSMC fabs. Amkor and SPIL package and test. Foxconn and Wistron build boards and racks. The April 2025 list was the point: the U.S. footprint is a partner map, and four of the five names are Taiwanese groups expanding American sites.

Wistron put $700 million into Fort Worth and staffed it as its first U.S. factory. Foxconn already knew how to build NVIDIA AI racks at volume and is now teaching robots to do Blackwell assembly in Houston. TSMC’s Phoenix wafers are the patriotic photograph. The company’s packaging muscle, and most of NVIDIA’s CoWoS allocation, still sits in Taiwan, with more packaging halls ramping there as well.

Amkor is the American-listed packaging house in the plan. Apple will be its first and largest customer in Peoria. NVIDIA is on the customer list too. Huang said the companies were rebuilding the supply chain and moving the AI stack onshore. The sentence is a goal. The campus is still a construction timetable.

That is why the loudest Made in America framing from April 2025 aged so fast. Texas can point to boards coming off a line. Arizona can point to wafers. The contractor who closes the GPU is still, for now, in Taiwan, and the contractors who close the rack are Taiwanese firms on Texas deeds.

What the $500 Billion Figure Counts

NVIDIA’s number is easy to misread as a check the company is writing. The April blog said that within the next four years it planned to produce up to $500 billion of AI infrastructure in the United States through partnerships with TSMC, Foxconn, Wistron, Amkor and SPIL. That is a production-value target through 2029, not a disclosed NVIDIA capital budget.

The Fort Worth $700 million sits inside that umbrella as Wistron’s plant cost. Amkor’s Arizona campus is a $7 billion two-phase bet, lifted from an earlier $2 billion plan by another $5 billion, with 750,000 square feet of cleanroom space and about 3,000 jobs attached. NVIDIA said making the chips and supercomputers would create hundreds of thousands of jobs over coming decades. That is a different clock from Wistron’s plan to go from more than 500 jobs to 1,000 by the end of 2026.

Huang’s other claim from April was narrower and easier to check. American manufacturing, he said, would help meet demand for AI chips and supercomputers, strengthen the supply chain and boost resiliency. Demand is not the debate. Foxconn’s server work has stayed hot, and Wistron’s Texas line exists because the order book needed a U.S. node. Resiliency is the debate, because a Phoenix-to-Taiwan-to-Texas loop still depends on the same packaging bottleneck that was tight before the ribbon-cutting.

Amkor’s Arizona Campus Will Not Run Until 2028

Amkor broke ground on a new semiconductor advanced packaging campus in Peoria, Arizona, and has called it the first U.S.-based high-volume advanced packaging facility. Construction is expected to finish in mid-2027. Production is slated to begin in early 2028. The project has $407 million in CHIPS Act funding, plus tax-credit and state support.

TSMC is also building packaging capacity in Arizona. Rousseau’s April 2026 account was that those sites were going up and that, until they run, every Phoenix die still goes to Taiwan. NVIDIA has looked at Intel packaging as another U.S. option after taking a stake in the company. That is not a substitute for CoWoS volume in 2026.

THE PACKAGING GAP

• What runs now: TSMC Phoenix wafers, Wistron GB300 boards in Fort Worth, Foxconn rack work in Houston.
• What still leaves: 100 percent of TSMC’s chips, including Arizona output, packaged in Taiwan as of April 8, 2026.
• What is booked: NVIDIA holds a majority of TSMC’s most advanced CoWoS capacity, which Rousseau said is growing at 80 percent a year.
• What is dated: Amkor production in early 2028; TSMC’s Arizona packaging halls without a public finish date in that April interview.

Until Peoria is actually attaching memory to logic, a Blackwell wafer made in Phoenix still has to leave the country to become a GPU. Fort Worth can then mount that GPU on a two-ton, $4 million board and ship an American supercomputer that began, and paused, somewhere else.