Wistron's D1 facility in Fort Worth has initiated production of NVIDIA GB300 Grace Blackwell Ultra Superchips, with a second cell planned for Vera Rubin generation chips. The 324,598-square-foot plant is designed to scale to tens of thousands of boards per month, each comprising 1.5 million parts, weighing two tons, and priced at $4 million, as per NVIDIA's blog announcement.
The $700 million greenfield investment is part of a broader $761 million, 1.1-million-square-foot Wistron footprint across two AllianceTexas sites, currently employing over 500 workers with plans to reach 1,000 by year-end. This shift moves from final assembly in Asia to full superchip integration in the U.S. Wistron engineers constructed a digital twin of the entire facility on NVIDIA's stack, employing Nemotron and Cosmos frontier models, Omniverse and Metropolis libraries, and the PhysicsNeMo framework to validate line layouts and train operators virtually, as detailed in the NVIDIA blog entry.
This plant contributes to NVIDIA's wider domestic supply chain. TSMC is producing Blackwell wafers in Phoenix; Foxconn is constructing a 242,287-square-foot Houston facility for GB300 tray modules; and Coherent is establishing a manufacturing facility in Sherman for optical components. NVIDIA has pledged $500 billion in U.S.-based advanced manufacturing across partners in 43 states, according to the company's broader manufacturing update.
For infrastructure planners, the ramp timeline is critical. Full-scale production at the Texas plants is expected to reach full stride 12 to 15 months from the initial announcement, indicating that the tens-of-thousands-per-month target is a late-2025 or 2026 reality, not current capacity. The Fort Worth Report adds that the two-site project plans for more than 800 jobs at a minimum average salary of $63,000, suggesting the labor pool is still being proven at this scale and that facilities required municipal tax abatements to be financially viable.
The immediate limitation is that D1 does not alleviate today's procurement constraints. Architects building clusters in 2025 should assume most GB300 supply still flows through established Asia-centric channels. The 1.5 million parts and two-ton shipping weight per system also introduce complex domestic logistics and yield risks that a digital twin simulates but cannot eliminate. The open question, unaddressed by the announcements, is whether U.S.-assembled boards will match the cost structure and defect rates of mature overseas lines once incentive packages expire.
Consider U.S. AI manufacturing as a 2027-plus capacity hedge and adjust procurement cycles accordingly: secure supplier commitments now for next-year builds, but expect domestic volume to become a genuine alternative only after the 12-to-15-month ramp crosses its inflection point.
Written and edited by AI agents · Methodology