What was actually broken ground on
Samsung broke ground on November 12, 2026 at its Onyang campus in Asan, South Korea. The headline figure is 46 trillion won, about $32.9 billion. It splits in a way that matters. 6 trillion won, roughly $4.3 billion, is for the Onyang Line 5 building. 40 trillion won, roughly $30 billion, is for phased equipment procurement. About 87 percent of the money is machines, not concrete.
The building is 31,000 square meters of cleanroom floor across three stories. That expands the site by 56 percent, to 420,000 square meters of total floor area, according to municipal planning records. Asan municipal authorities put the daily construction headcount at an average of 3,400 workers and count 700 permanent manufacturing jobs at the end of it.
Onyang has handled memory assembly and final testing since late 1991. It is being converted into a primary high-bandwidth memory hub, which Samsung Executive Chairman Lee Jae-yong stated directly. The stated purpose is backend lines for 16-high HBM4 and HBM4E units. Samsung moved its completion timetable forward by 12 months, citing tightening server memory backlogs. Initial line operations are scheduled for June 2028, commercial mass production for May 2029.
The stated requirements for those generations are sub-micron interconnects and hybrid bonding, to keep the stacks from overheating during intensive model training. That is the whole event in one sentence. This is not a DRAM fab. It is a bonding line, and the equipment budget is the tell.
The bottleneck is the interface, not the memory
Advanced packaging connects separate silicon dies on an intermediate silicon interposer. A high-bandwidth stack is DRAM dies stacked vertically over a logic base die. The bandwidth a stack delivers is a property of the interface between those dies and the interface between the package and the substrate. Stack more dies on the same interface and the number does not move.
Taiwan Semiconductor Manufacturing Co. controls over 80 percent of global capacity in 2.5D advanced packaging, and packages flagship graphics chips for Nvidia through its Chip-on-Wafer-on-Substrate process. Industry procurement filings describe a two-quarter bottleneck on global accelerator deliveries caused by that capacity. That is why chip designers have been seeking alternative packaging sources, and it is why a $30 billion equipment budget is going to a bonding line rather than to wafer fabs.
The supply chain already shows the split. SK hynix supplies more than 50 percent of the world’s high-bandwidth memory market, and it still outsources custom logic base wafer production to TSMC for its upcoming HBM4 units. A memory company that owns the DRAM does not own the base die. That is the structural fact behind the Asan build, and it is why the pull-forward is on the packaging line.
The base die is booked, and it is a foundry problem
Samsung’s second plant in Taylor, Texas is reported to produce HBM base dies for AMD and NVIDIA. Orders reportedly cover the entire planned production capacity of Taylor Fab 2, with AMD accounting for the majority, and the capacity could be stretched by AMD’s orders alone. The facility is planned for a 2nm process across 2.7 million square feet, about 250,000 square meters, roughly 1.5 times the size of Taylor Fab 1.
Site preparation is underway. Full-scale construction was scheduled to begin in early November 2026, equipment installation is expected to start in February 2029, and mass production in 2030. Those two dates are the point. A packaging line that starts running in June 2028 and a base-die fab that starts in 2030 are not the same timeline, and a finished stack needs both. The 12-month pull-forward applies to the bonding line. It does not move the logic die.
Samsung’s mass-produced HBM4 and sampled HBM4E use 4nm base dies. Taylor Fab 2 is intended for 2nm or more advanced. Meanwhile NVIDIA has reportedly asked all three major memory manufacturers, including Samsung, to outsource NVHBM base die production to TSMC. The DRAM maker, the foundry and the customer are pointing at three different fabs for one part of one package. The spec side of that split is what a 13 percent bandwidth cut on Rubin already showed: the interface number is the one that moves a decode loop.
Samsung’s third-quarter foundry revenue benefited from HBM logic die sales and advanced 2nm orders from hyperscalers, and the business remains loss-making. A loss-making segment that is fully booked is not a pricing story yet. It is a capacity story, and capacity stories become prices later.
What the interface looks like on a local build
The same physics exists on a desktop board at a much smaller scale, and it is measurable. I ran a copy pass on the serving box, a Ryzen 9 5900X with 2 x 32 GB of DDR4 running at 2667 MT/s, executed on that machine over SSH rather than on the box I sent the request from. A 128 MiB buffer, 20 copies per pass, five passes. A read loop through Python is CPU-bound, so this is a copy pass, not a scan.
| Pass | Bytes moved (GB/s) | Memory traffic (GB/s) |
|---|---|---|
| 1 | 16.75 | 33.50 |
| 2 | 17.31 | 34.62 |
| 3 | 17.09 | 34.18 |
| 4 | 16.99 | 33.98 |
| 5 | 17.19 | 34.38 |
| Median | 17.09 | 34.18 |
A copy touches memory twice, so the two columns are the same run counted two ways. The spread across the five passes is 3.3 percent of the median, tight enough to treat the median as the number.
The channel arithmetic is mine, not a published board figure. 2667 MT/s across two 64-bit channels is 42.67 GB/s. The measured pass reaches 80.1 percent of that channel counted as memory traffic, or 40.1 percent counted as data moved. Both readings describe one measurement, and quoting one as the other is the usual way this gets wrong.
The efficiency figure is the part that carries over to a purchase. It is not a property of the memory kit. It is a property of the path from the controller to the memory. The earlier sweep on this same box put the memory controller ceiling at 68.7 percent on one thread and 84.4 percent on two, and the measured cache cliff showed the same shape one level up, where a buffer that never reaches DRAM cannot be expressed as a share of the channel at all.
A kit rated 6000 MT/s at CL30 changes the theoretical channel. It does not change the fraction of the channel a copy pass actually reaches. Applying the measured 40.1 percent to a dual-channel DDR5-6000 kit gives about 38.5 GB/s of bytes moved. That is arithmetic carried from a platform I did not measure, so read it as an estimate, not as a result.
What the money is actually for
A memory kit is priced for an interface, not for capacity. That is the same reason a bonding line gets 87 percent of a 46 trillion won budget. The reader’s ceiling is the controller path, and the kit only sets the theoretical channel above it. The best hardware for running local LLMs in 2026 covers the ladder, so this section only covers the two parts of a build that the interface argument actually touches.
The 64 GB AMD EXPO U-DIMM kit at G.SKILL Trident Z5 Neo, 2 x 32 GB, 6000 MT/s CL30 reads $1,279.99 as of November 15, 2026, checked on the listing through a real browser. The Intel XMP variant of the same physical spec, Trident Z5 64 GB CL30, reads $1,349.99 on the same day. Both are 288-pin U-DIMM desktop modules. Neither is a laptop kit, and a 64 GB SODIMM kit is the wrong advice for a desktop board even when the capacity matches.
The storage tier is where a local build can actually change a number, because it is the tier below the interface. A WD_Black SN8100 2 TB, PCIe 5.0 x4, M.2 2280, reads $479.99 on the listing today. That is a rated sequential figure, not a per-request figure. At 4 KB the effective per-request cost sits between 19 and 31 microseconds depending on which loop produced it, and a freshly written test file sits inside the drive’s pSLC cache, so large-block figures can exceed the drive’s rating. The block-size and queue-depth measurements on this site are the ones that describe what a KV cache tier actually does.
What the Asan and Taylor timelines mean for a buyer is a date, not a price. Sub-micron bonding capacity for 16-high stacks arrives in mid-2028 on the packaging side and in 2030 on the logic-die side. A build bought now is priced against the interface available now, and the DRAM price pressure that a 3x DRAM price already put on local builds is the same shortage showing up at the consumer end of the same line.
Decision rules
- Treat a stack’s bandwidth as an interface number. More dies on the same bonding line does not move it.
- Measure the fraction of the channel, not the kit’s rating. 40.1 percent of a 42.67 GB/s channel is what a copy pass on this box actually reached.
- Report a copy pass as two numbers. Bytes moved and memory traffic differ by a factor of two, and the percentage depends on which one you divide by.
- Run at least five passes and quote the median. The five passes on the same buffer ranged from 16.99 to 17.31 GB/s.
- Do not plan a build around a 2030 base-die timeline. The 2028 date is the packaging line, and the logic die is not on it.
Two things in this article are measurements and one is arithmetic. The five-pass copy run is a measurement, taken on the serving box over SSH. The 42.67 GB/s channel is my own arithmetic from the DIMM configuration read on that machine, and no vendor publishes a bandwidth figure for it. The 38.4 GB/s extrapolation to a DDR5-6000 kit is arithmetic carried across a platform I did not measure, and it is labelled as such.

