A Korean newspaper reported on November 3, 2026 that Samsung Electronics will use its second foundry plant in Taylor, Texas to make base dies for high-bandwidth memory (HBM) for AMD and Nvidia, and that the plant is already fully booked. Reuters summarized the Maeil Business Newspaper report and added one fact of its own: Samsung declined to comment. That is the whole confirmed state of the story. The booking, the customer split and the 2030 timeline all come from anonymous industry sources in one outlet.
This post lays out what was reported, what can be checked, and what to take from it if you plan around AI hardware supply.
TL;DR: claimed versus verified
| Claim | Source | Status |
|---|---|---|
| Samsung will use Taylor fab 2 for HBM base dies for AMD and Nvidia | Maeil Business, via Reuters (Nov 3) | Reported, unconfirmed |
| Plant has secured orders for its entire output | Maeil | Reported, unconfirmed |
| Most of the booked volume is for AMD | Maeil | Reported, unconfirmed |
| Formal construction starts this month (November) | Maeil | Reported, unconfirmed |
| Equipment installed from February 2029, mass production 2030 | Maeil | Reported, unconfirmed |
| Samsung declined to comment | Reuters | Confirmed by Reuters |
| Nvidia will use Samsung Taylor for its custom NVHBM base dies | Not in the sources we found | Unsupported; another report points to TSMC |
What exactly did the report say?
Reuters, in a piece republished by The Star, gave these points from Maeil, which cited industry sources:
- The plant would use 2nm technology and cover about 2.7 million square feet, roughly 1.5 times the first Taylor facility.
- Construction of the second Taylor plant is due to start formally this month.
- The plant has secured orders for its entire output, and most of that secured volume is for AMD.
- Equipment is expected to be installed from February 2029, with mass production in 2030.
- Samsung Electronics declined to comment.
French outlet Pause Hardware, relaying Wccftech, repeats the size and 2nm details (about 251,000 square meters) and says Wccftech itself rates the story as only about 50 percent plausible, with weak corroboration. We treat the Wccftech rating as opinion.
What is an HBM base die, and why would a foundry matter?
High-bandwidth memory stacks several DRAM core dies on top of a logic die, called the base die. The base die connects the stack to the GPU or accelerator. For years it was made by the memory maker itself on a DRAM-style process. With HBM4, the base die moved to logic processes, which is why foundries now compete for it. The Maeil plan, as relayed, splits the work: base dies made in Texas, DRAM core dies made in South Korea, then combined into finished HBM stacks.
For Samsung, that would be a way to sell both halves of the stack. It is also a way to fill a US foundry that has had a hard time winning customers. We covered the yield side in Samsung's HBM4 yield report and the earnings side in Samsung's record Q3 and HBM4 long-term contracts.
Where does Nvidia fit?
This is the weakest part of the Maeil story. Reuters names Nvidia next to AMD in its headline. But Pause Hardware, citing an earlier report, says Nvidia has strongly encouraged Samsung, SK hynix and Micron to let TSMC make all of its custom NVHBM base dies. The reason given is to keep one common memory-controller architecture and reduce differences in latency, thermal behavior and validation across suppliers.
We looked at that claim in a separate post: Samsung and TSMC on NVHBM base dies: claimed versus verified. The short version: nobody has confirmed it either. So on Nvidia we have two unconfirmed reports pointing in different directions. Read "AMD and Nvidia" in the Reuters line as what Maeil's sources said, not as signed contracts.
What can we actually verify?
Three things hold up.
- The report exists and Samsung did not deny it. Reuters attributes it to Maeil and records "declined to comment." A decline is not a confirmation.
- Samsung's first Taylor fab is not yet in production. TrendForce quotes a Samsung North American foundry executive saying Taylor will begin production in 2027, per search results we reviewed, and Tom's Hardware reports that Tesla's AI6 chip has reportedly slipped by about six months over a delayed engineering run on Samsung's 2nm line. Those are reports, but they show the first fab is still ramping.
- HBM demand is real and tight. Our own coverage of SK hynix's record Q3 and the memory shortage and the Stanford memory price history show why every big supplier is hunting for capacity.
What we could not verify: any AMD statement, any Nvidia statement, any Samsung filing or press release on Taylor fab 2 orders, and the "entire output" claim. A fab that does not start equipment installation until February 2029 would be booking customers more than three years ahead. That is not impossible for a strategic anchor customer, but no document we found shows it.
Is there a conflicting report on the timeline?
Yes. Another item in the search results says Samsung will move the HBM4E base die to 2nm and that Taylor ramp supports 2nm before year-end. That conflicts with a 2030 start for the second plant. It most likely refers to the first Taylor fab, which is a separate site, but we could not confirm that from the source text. If you see "Taylor 2nm in 2026" and "Taylor 2nm in 2030" in the same week, check which fab each line is about.
What came before: why Taylor needs customers
Taylor is the Texas site Samsung has been building for US foundry work. Tom's Hardware's examination of Samsung's fab roadmaps ties the site to the yield problems behind a $16.5 billion Tesla deal, and notes that Samsung's current 4nm HBM4 base die is made in-house rather than bought from TSMC, as some rivals must. That context explains why a base-die order would matter to Samsung: it would give a logic fab a steady, high-value product that is not a phone chip.
Read the Maeil claim against that backdrop. A fab with a history of slow ramps and a customer base that is still forming is exactly the kind of site that would want to announce an anchor customer early. That does not make the report wrong. It does mean the incentive to talk up bookings exists on every side, including in the sources that brief Korean newspapers. This is one reason we label the report as unconfirmed rather than dismissing it.
There is also a practical gap between "booked" and "contracted." In chip supply, a reservation can mean a signed take-or-pay agreement, a letter of intent, or a capacity-planning conversation. Maeil's wording, as relayed, does not say which. If AMD or Samsung later describes the arrangement, the contract type will tell you far more than the headline does. Watch for words like prepayment, minimum volume or multi-year supply agreement.
What does this mean for what you build or pay?
Short answer: nothing changes in 2027 or 2028 from this report.
- GPU and accelerator supply through 2028 depends on HBM3E and HBM4 output at existing fabs, plus advanced packaging. A plant with first equipment in 2029 does not touch that.
- Pricing pressure on memory-heavy accelerators stays tied to the shortage described in our AI chip supply chain roundup.
- Long-range planning is where it matters. If AMD really does anchor a US-made base-die line, that is one more sign that HBM supply is moving toward logic-foundry manufacturing and toward US sites. That is a trend to watch, not a contract to bank on.
- Risk disclosure language is worth reading. See how Anthropic's S-1 treats TSMC and HBM suppliers for what AI labs themselves say about this dependency.
How can readers check this themselves?
- Watch for a Samsung Electronics statement or filing, or a Texas state or county announcement tied to groundbreaking in November. Groundbreaking is a public, checkable event.
- Watch for AMD commentary on its next earnings call or at a supply-chain event. AMD has an incentive to say something if it reserved a fab.
- Compare the Maeil original (Korean) with the Reuters summary and other relays.
- Treat any outlet's "plausibility" score, including Wccftech's, as an opinion, not evidence.
What should I take away?
Do not repeat "Samsung Taylor 2 is fully booked by AMD" as fact. Say "Maeil reported, citing industry sources, that..." and note that Samsung declined comment. If the November groundbreaking happens and AMD or Samsung speaks on the record, we will update this post.
Related reading
Accurate as of November 6, 2026. All claims about Taylor fab 2 come from one Korean newspaper via Reuters; no company has confirmed them.