Samsung Texas Fab 2 Booked by AMD, Nvidia for 2nm HBM Base Dies

A factory sold out before it exists

Samsung Electronics has reportedly found buyers for every wafer its second fab in Taylor, Texas, is expected to produce, even though formal construction is only starting this month. South Korea's Maeil Business Newspaper reported on November 3, citing industry sources, that the plant will make base dies for high-bandwidth memory (HBM) for AMD and Nvidia. AMD reportedly accounts for most of the volume booked so far.3 Reuters carried the report, and Samsung declined to comment to it.3 A Samsung official told Maeil that matters involving customers were difficult to confirm.34

The reported specifications are large. Fab 2 is expected to run a 2-nanometer process across roughly 2.7 million square feet, which makes it about 1.5 times the size of the first Taylor plant.36 Equipment is due to start arriving in February 2029, and mass production is targeted for 2030.18 Site work such as levelling has already begun, according to Maeil.3

Part of the schedule is on the record. On October 27, Samsung spokeswoman Michele Glaze said construction would begin in November and that the facility should be operating by 2030. Samsung first set that 2030 target in July.4 The customer names are what remain unconfirmed.

What a base die is, and why the details matter

Much of the coverage has been loose about what this plant will actually make. Several outlets described the output as "advanced memory chips" or "memory-chip parts."110 The more careful reports describe something narrower: a logic die built on a foundry process. It sits at the bottom of an HBM stack, beneath the stacked DRAM, and handles data traffic between the memory and the processor.34 TokenPost noted that a base die is one component of an HBM package. It is not a finished AMD or Nvidia processor.4

The distinction affects how to read Samsung's strategy. According to Maeil, the Texas base dies will be paired with DRAM made in Korea and delivered to customers as finished HBM.35 In practice, the foundry order and the memory sale come as one package. That is something Samsung can offer because it runs both a leading-edge foundry and a memory business. SK Hynix and Micron do not have the first.

The technology timeline explains why 2nm matters. Samsung's HBM4, now in mass production, and its sampled HBM4E both use 4nm base dies. Maeil expects Fab 2 to make base dies on 2nm or something more advanced for later generations.3 Customers have started asking for custom logic inside the base die, and more companies are ordering base dies separately, according to Maeil.3 The base die is turning from a commodity layer into a custom chip, and that is the trend Samsung is building this fab for.

Where the reporting diverges

The coverage largely agrees on the core facts: the site, the size, the node and the dates. The differences are in emphasis and in caveats.

The first concerns how full the order book is. Most outlets say the plant has booked its entire output.178 Sammy Fans says almost the entire planned capacity.5 That is a small gap, but no outlet has reported a contract value, wafer volume or exact split between AMD and Nvidia.39

The second, and more important, concerns Nvidia. Most of the short-form coverage presents Nvidia as a straightforward customer. FourWeekMBA's longer account of Maeil's reporting shows the Nvidia business is much less settled. Maeil reports that Nvidia has asked all three memory makers, Samsung included, to have the base dies for NVHBM, its next custom HBM, made by TSMC. The reported aim is to reduce performance differences between suppliers and cut the validation work.3 Samsung is reportedly checking whether its own foundry could make NVHBM base dies as a backup if TSMC runs short of capacity.35 Today Samsung supplies HBM4 for Nvidia's Vera Rubin with a base die made on its own 4nm process. Maeil warns that if NVHBM spreads, Samsung could lose that edge, because SK Hynix and Micron already use TSMC base dies.3

My reading is that AMD is the solid part of this order book and Nvidia is the uncertain part. The headline pairing of AMD and Nvidia makes the deal look more balanced than the details support.

Taylor's long road to anchor customers

The reaction makes more sense against the history of the first Taylor fab. Samsung broke ground there in November 2022 with an initial $17 billion investment and planned to start operating in October 2024. That date slipped because Samsung struggled to win orders.18 The schedule also moved because Samsung dropped its original 4nm plan and went straight to 2nm Gate-All-Around transistors, which are harder to manufacture.13

Tesla changed the picture. In July 2025, Samsung signed a $16.5 billion contract to make Tesla's AI5 and AI6 chips at Taylor through 2033.1318 Samsung held an equipment move-in ceremony in April 2026.20 By September, Munhwa Ilbo was reporting that Fab 1 was fully booked for 2nm before operations began, with orders tied to Tesla, Broadcom and Arm, and an initial capacity of 50,000 wafer starts per month.12 Trial production of Tesla's AI5 reportedly started that month.14

The pattern is consistent. Fab 1 sat partly built for years without customers, then filled up once Tesla signed. Fab 2 is following the opposite order, with customers secured first and construction second. Maeil says that is how foundry investment works: committed customer volume decides when the money is actually spent.3 It also matters that Samsung asked equipment suppliers in September to get SEMI safety certifications early for Fab 2 tools, before the fab's detailed specifications were final.12 That suggests the company expected demand like this.

The capital behind the bet

The Fab 2 report came five days after Samsung's earnings, where it said it would raise capital spending by more than 60% this year to 84.4 trillion won, or about $63 billion. Samsung also said its high-end HBM is sold out through 2027.6 The company tripled HBM4 sales in the third quarter and signed 10 long-term supply deals with major clients.6 In its third-quarter results, Samsung credited growth in HBM base-die sales and advanced-node sales for the improvement in its foundry business, though it did not name customers or link those sales to Taylor.4

Investors are not uniformly enthusiastic. Samsung shares fell about 4% after the earnings, when chip-division profit missed estimates and the company postponed new shareholder payout plans until January.6 Rival SK Hynix says its HBM supply is sold out through 2029.6 Several outlets note that the bookings do not immediately raise Samsung's share of the foundry or HBM markets.12 Whalesbook points to the gap of more than three years between construction and revenue. If AI demand weakens before 2030, margins and cash flow could come under pressure.10

The market's reaction favored the customer over the supplier. AMD shares rose more than 4% in morning trading on the report.9 That fits the idea that the deal gives AMD a second source of advanced manufacturing capacity, at a time when TSMC capacity is the main constraint for the whole AI hardware industry.9

Why this matters for U.S. chip manufacturing

From a construction and investment point of view, three things stand out.

First, long-term capacity is being reserved further ahead than before. AI chip companies are now committing to specialized capacity years before any wafers exist.1 By FourWeekMBA's calculation from Maeil's dates, there are 27 months between the formal start of construction and the first tool move-in, and the customers are already named for that whole period.3

Second, more of the HBM supply chain is moving to the U.S. Maeil's companion analysis expects the finished HBM to be packaged with accelerators from AMD, Broadcom and others made at TSMC's U.S. plants. Broadcom is also seen as a possible future customer for Taylor base dies.3 If that happens, Texas would supply logic layers for chips assembled elsewhere in America, which links two of the largest foreign chip investments in the country.

Third, Samsung's foundry can win business through its memory unit. Samsung has trailed TSMC on advanced-node yields for years, and Tom's Hardware notes that its foundry is still roughly ten times smaller by that measure.20 Selling base dies bundled with its own HBM gives Samsung a way into advanced customers that does not depend on beating TSMC on a standalone logic chip.

The bottom line

This looks like the strongest sign yet that Samsung's U.S. foundry has moved past its period of empty buildings. Fab 1 is booked, and Fab 2 is reportedly booked before construction. The 2030 date is still a plan, and 2nm production at this scale is something Samsung still has to demonstrate. None of the three companies has confirmed the orders.3 The key question is Nvidia's NVHBM decision. If Nvidia keeps its next base dies at TSMC, the order book for Fab 2 rests mainly on AMD. If Samsung can win NVHBM work as a backup supplier, the fab gains a second major customer. The milestones to watch are confirmed customer contracts and steady construction progress through 2029.1