Production of base dies for high-bandwidth memory for AMD and Nvidia will move to Samsung's second chip plant in Taylor, Texas, South Korea's Maeil Business Newspaper reported on Tuesday, citing industry sources. Most of the booked volume is assigned to AMD, and the plant has already secured orders for its entire output. Formal construction is due to start this month, with mass production scheduled for 2030. For buyers of AI infrastructure, this points to more US-based capacity for the logic layer that feeds AI processors.
Second Taylor fab terms and timeline
The Taylor facility will produce base dies, the logic layer at the bottom of stacked memory chips that feed AI processors. The report describes the second fab as running on Samsung's advanced 2-nanometre process and covering about 2.7 million square feet, roughly 1.5 times the size of the first plant. Equipment is expected to arrive from February 2029, followed by mass production in 2030. Samsung declined to comment, Reuters said it could not confirm the report, and TNW has not independently verified it either.
The project builds on a large existing campus in Taylor. The first Taylor plant grew from a $17 billion project to a $37 billion total investment and is due to start mass production in 2027. Its anchor customer is Tesla, which signed a $16.5 billion contract for its AI5 and AI6 chips. The 1,268-acre site has room for up to 10 fabs. The first plant's financing includes a $4.75 billion subsidy under the US CHIPS Act.
Signs of the expansion have been visible for months. Taylor's city council approved engineering support for the second fab's permits back in March, and Samsung has been hiring for the campus. A second fab filled before ground is broken would mark a turnaround for Samsung's foundry arm, which has lost money for years. Kiwoom Securities expects the unit to return to profit from the fourth quarter. Samsung has projected that its 2nm orders will grow more than 130% this year, helped by AI demand and strained capacity at rival TSMC.
What this means for AI hardware buyers
For companies buying AI servers and accelerators, additional US output of HBM components addresses a bottleneck where demand has outrun supply. Memory rivals from SK Hynix to China's CXMT are racing to catch up, while AMD has been locking in capacity as its AI chip business scales up. A plant with output already allocated suggests tighter near-term availability but a broader supplier base by 2030. Small buyers will feel this through lead times and vendor allocations, while large buyers gain another negotiation point for long-term supply.
Procurement teams should treat the timeline and booking status as constraints on planning. Equipment arrival in February 2029 and mass production in 2030 leave several years in which HBM supply remains tight. The report rests on unnamed industry sources, and key commercial terms such as pricing, volumes by customer, and delivery schedules were not disclosed. It also does not by itself guarantee yields or qualification of 2-nanometre base dies for specific AMD and Nvidia products. Questions to vendors include allocation guarantees, qualification milestones, and fallback supply if ramps slip.
The marker to watch is progress at Taylor against the stated dates, starting with construction activity this month and equipment installation from February 2029. Confirmation of customer qualifications and sustained foundry profitability from the fourth quarter would show the turnaround is real. If those steps hold, US-based HBM supply becomes a firmer part of capacity planning for AI deployments.
