Samsung’s Second Texas Fab Reportedly Books AMD-Led HBM Orders

The area comparison does not establish a corresponding increase in chip output. The report provides no monthly wafer capacity, order price or contract duration. It supports a case for committed demand, without enough information to calculate Fab 2’s revenue or investment return.

Memory expertise is an advantage, but customers have alternatives

A base die is the logic chip beneath HBM’s vertically stacked DRAM memory layers. It manages communication between the memory and an AI processor. An accompanying report describes how customers increasingly want computing and control functions built into that layer for their own accelerator architectures, making base-die design and manufacturing more consequential.

Samsung’s advantage is that it has memory and foundry operations within the same company. In its February 12 announcement, Samsung said it had begun HBM4 mass production and commercial shipments using a 4nm logic base die. The reported Texas plan would take that work to a newer process generation by 2030. Process labels alone do not establish the eventual product’s performance or manufacturing economics.

AMD’s reported allocation builds on an existing relationship. Samsung said in its March 18 collaboration announcement that the companies had worked together for nearly two decades across graphics, mobile and computing technologies. Samsung was already AMD’s primary HBM3E partner for its Instinct MI350X and MI355X AI accelerators.

Their March memorandum covered alignment on primary HBM4 supply for the next-generation Instinct MI455X accelerator, a building block for AMD’s Helios rack-scale computing platform. It also provided for discussions about Samsung manufacturing next-generation AMD products. That agreement explains the route toward closer cooperation; it does not confirm the reported Taylor allocation.

Samsung’s integrated approach faces an established alternative. Rival memory manufacturer SK hynix announced a partnership in April 2024 to develop HBM4 with TSMC, the Taiwanese contract chipmaker. SK hynix said it would move from its own base-die technology through HBM3E to TSMC’s advanced logic process for HBM4, and work on integrating its memory with TSMC’s CoWoS packaging. Customers can obtain closely coordinated memory, logic and packaging through partnerships as well as within Samsung.

Samsung’s foundry improvement also predates the proposed factory. In its October 29 results announcement, Samsung attributed better foundry revenue and earnings to growing HBM base-die sales and advanced-node sales, alongside further orders centered on 2nm high-performance-computing customers. It expected operations at its new Taylor fab to begin in 2027—a separate timeline from the reported 2030 start for Fab 2. Those results do not establish the second factory’s profitability.

Texas production would still rely on Korean memory

The reported supply plan combines Texas-made base dies with DRAM manufactured in Korea to deliver completed HBM. The accompanying report envisages that memory being packaged with processors made at TSMC’s U.S. operation. It describes Samsung’s longer-term ambition to manufacture customers’ processors and offer packaging as well, but that is broader than the reported base-die orders.

Nvidia already has a U.S. manufacturing relationship with TSMC. In its April 2025 announcement, Nvidia said production of its Blackwell AI chips had begun at TSMC’s Phoenix plants, with Amkor and SPIL as Arizona packaging and testing partners. The Taylor project would add a U.S. component-making step while retaining reliance on Korean DRAM.

Samsung also has a substantial U.S. manufacturing history. The U.S. Department of Commerce said in its December 2024 CHIPS award announcement that Samsung had manufactured chips in the country since 1996. The department awarded up to $4.745 billion in direct funding through the federal semiconductor-manufacturing incentive program to support more than $37 billion in planned investment.

That award covered two logic fabs and a research fab in Taylor, plus expansion in Austin, with disbursements tied to construction, production and commercial milestones. Neither the investment total nor the subsidy is a figure for Fab 2 alone.

Commerce’s December 2024 award provides up to $4.745 billion for Samsung’s broader Taylor and Austin program, not Fab 2 alone. The ceiling does not represent money disbursed. Graphic: Magica. Source: U.S. Department of Commerce.

The next test is who manufactures Nvidia’s customized base dies

The original Korean report says Nvidia has required the three HBM suppliers—Samsung, SK hynix and Micron—to have base dies for its customized NVHBM approach manufactured by TSMC. It attributes that requirement to Nvidia’s effort to reduce performance differences between suppliers and the burden of validation.

The accompanying report describes NVHBM moving the memory controller from the graphics processing unit (GPU) into the base die, with the processor designer responsible for the die’s design. Under that arrangement, memory suppliers could sell the DRAM layers while TSMC manufactures the logic beneath them. That would narrow Samsung’s opportunity to combine a memory sale with its own foundry services. The accompanying report says Micron, like SK hynix, already uses TSMC base dies for HBM4.

The reporting allows for standard HBM and NVHBM to coexist because Nvidia supplies several GPU generations at once. It also says Samsung is testing whether its foundry process could manufacture NVHBM base dies if TSMC cannot meet demand. That does not establish that Nvidia has qualified Samsung as an alternative manufacturer.

The immediate evidence to watch is confirmation of the bookings, their financial terms and progress toward equipment installation in 2029. Beyond the AMD-led allocation, Samsung’s opportunity depends on whether customers accept its manufacturing process for customized base dies or direct that work to TSMC. Filling the reported plant and winning control of the next generation of memory logic are separate tests.

Google’s school AI controls offer no AI-free Search

Google plans to switch off Gemini and Search for school-managed K–12 student accounts unless administrators enable them. Schools will gain control over access, but Google says keeping Search without its AI features will not be an option.