Samsung Electronics plans to use its second Samsung Texas chip plant in Taylor to manufacture base dies for high-bandwidth memory destined for AMD and Nvidia processors, according to a report by South Korea’s Maeil Business Newspaper on Tuesday. Formal building work begins this month, with volume already fully booked.
Key points
- The second Taylor fabrication facility will build high-bandwidth memory base dies on Samsung’s advanced 2-nanometre manufacturing process.
- AMD has secured the majority of the planned manufacturing volume, with Nvidia taking the remainder of the fully allocated capacity.
- Equipment installation is scheduled to begin in February 2029, leading to commercial mass production in 2030.
- The site spans 2.7 million square feet, making it roughly 50 percent larger than Samsung’s initial Taylor chip factory.
What the second Taylor facility will build
The new facility represents a major expansion of Samsung’s semiconductor footprint in the United States. According to industry sources cited by Maeil, the plant has already locked in purchase commitments for its entire future output before breaking ground. Advanced Micro Devices will take the bulk of that allocation, with Nvidia securing the rest.
The plant will focus on 2-nanometre technology. Samsung declined to comment on the report, while international news agencies noted that they could not independently verify the commercial agreements. Even so, local preparations have been under way for months. The municipal council in Taylor approved engineering permits and technical support for the project in March.
Building work is scheduled to start formally in November 2026. Samsung expects cleanroom equipment to arrive from February 2029, pointing to commercial volume production during 2030. Covering roughly 2.7 million square feet, the second fab is approximately 1.5 times the size of the initial production facility on the same campus.
Why HBM base dies matter for AI accelerators
High-bandwidth memory, known as HBM, has become the chief physical bottleneck in building high-performance hardware for machine learning. Processors designed for training and running large models consume vast quantities of data every millisecond. Conventional memory architectures cannot move data quickly enough without overheating or stalling execution.
HBM solves this by stacking multiple dynamic random-access memory dies vertically. Those memory layers connect to each other through microscopic vertical channels called through-silicon vias. At the very bottom of that vertical stack sits the base die, sometimes called the buffer die or logic die.
The base die controls the entire stack. It acts as the traffic controller, routing electrical signals between the stacked memory cells and the primary computing processor, such as an Nvidia graphics processing unit or an AMD Instinct accelerator. Because the base die must handle high clock speeds and complex routing, manufacturing it on an advanced logic node helps lower heat and improve bandwidth.
Moving base die fabrication to a 2-nanometre node reflects how demanding next-generation memory standards have become. Older generations of HBM relied on legacy manufacturing lines for their base dies. Newer specifications require logic processes comparable to those used for flagship microprocessors to keep power consumption within manageable limits inside modern data centres.
A turnaround for Samsung’s foundry business
Booking the plant’s entire output provides welcome commercial relief for Samsung’s contract manufacturing arm. Known internally as Samsung Foundry, the business unit has absorbed financial losses for several years while trailing Taiwan Semiconductor Manufacturing Company in securing premier customers.
Analyst estimates from Kiwoom Securities suggest the division could return to profitability from the fourth quarter of 2026. Stronger demand for advanced packaging and packaging components has lifted financial forecasts across the group, supporting Samsung’s operating profit surge during recent quarterly earnings updates.
Samsung previously told investors that it expects orders for its 2-nanometre production nodes to grow by more than 130 percent during 2026. Much of that momentum stems from severe capacity constraints at TSMC, whose advanced packaging lines and leading-edge nodes remain heavily oversubscribed by Apple, Qualcomm and Nvidia.
By securing AMD as the anchor tenant for base dies in Texas, Samsung gains a marquee customer for its American logic operations. AMD has spent the past two years attempting to secure every available slice of packaging and memory capacity to expand its share against Nvidia in the enterprise market.
The expanding Taylor semiconductor campus
The broader Taylor campus is one of the largest foreign direct investments in American manufacturing history. Total capital expenditure for the complex grew from an initial budget of $17 billion to roughly $37 billion as plans expanded to support multi-stage packaging and additional cleanrooms.
Federal funding has helped back the development. Samsung secured a $4.75 billion grant under the US CHIPS and Science Act to support construction across its Texas properties, which also include older operational sites in Austin. The full Taylor property covers 1,268 acres, offering enough land to accommodate up to 10 separate fabrication units over the coming decades.
The first Taylor plant remains on track to begin mass production in 2027. Tesla serves as the anchor customer for that first facility, having signed a $16.5 billion agreement for automotive and autonomous computing hardware, including its AI5 and AI6 processors.
Adding dedicated HBM logic manufacturing gives the Texas hub a dual role. Rather than functioning purely as an automotive or general logic foundry, the site is shifting toward high-value artificial intelligence infrastructure components that US hyperscale cloud operators want produced domestically.
What happens next
Ground preparation and foundation work in Taylor will accelerate over the winter months as contractor teams move onto the site. Equipment vendors will monitor progress closely, with major tool procurement orders expected to begin formalising in 2028 ahead of the planned February 2029 machinery delivery date.
For AMD and Nvidia, the long lead time highlights how far ahead hardware makers must commit capital to secure components for the next decade of data centre expansion. Neither chip designer has commented publicly on the Taylor volume allocations.
