ChangXin Memory 4F2 impacts memory stocks: Goldman Sachs believes the capacity gap is narrowing, but the gap in advanced DRAM and HBM technology remains large

After the ChangXin 4F2 news, the market capitalization of storage stocks evaporated by about $738 billion.Goldman Sachs' assessment: the market underestimated ChangXin’s capacity expansion, but overestimated the technological breakthrough; Samsung and SK hynix have fallen below their 2028 net asset value, which already reflects a downward cycle.

A single tech news, $700+ billion market cap swings

On October 10th, there were reports that ChangXin Memory had made progress on a new DRAM memory cell architecture called 4F2. The company’s president stated the plan is to launch new server memory products based on the 4F2 architecture by the end of 2026.In the following trading days after the news, global storage company stock prices fell across the board: Samsung Electronics dropped 20%, SK hynix by 25%, Micron by 11%, SanDisk by 11%, Kioxia by 26%, and ChangXin itself fell 15%. According to Goldman Sachs’ statistics, the total storage sector lost about $738 billion in market value.

There are two inconsistencies in these declines with the explanation of a “ChangXin technology threat.”First, ChangXin itself also fell 15%. If the market truly saw 4F2 as a weapon for ChangXin to take market share, its stock price should not have fallen in line with its competitors. Second, Kioxia and SanDisk primarily produce NAND flash, which is not strongly related to DRAM cell architecture, yet their declines matched SK hynix, with Kioxia’s even larger. Therefore, Goldman believes other factors are at play, including concerns about credit, AI financing, and higher-than-expected CPI data.

Even more noteworthy is the divergence between price and stock price.DRAM spot prices are usually regarded as a market sentiment indicator because they reflect what distributors and small-to-medium customers are willing to pay now. In the same period when Samsung and SK hynix’s market caps dropped, the DDR5 16Gb spot price was trending up from mid-September to mid-October, with spot prices peaking even as stock prices fell. In other words, stock trading reflects concerns about future supply, not current supply and demand.

This is precisely the question Goldman Sachs aims to answer: which concerns about ChangXin's 4F2 news are well-founded, and which have already been excessively priced in.Their conclusion is two-fold. The market may simultaneously underestimate the speed of Chinese storage capacity ramp-up and the potential for technological breakthroughs; Goldman agrees with the first, as ChangXin’s capacity is indeed catching up rapidly; but on the latter, Goldman thinks the market has overestimated, because customer qualification, mass production with high yield, and narrowing the gap in advanced DRAM such as HBM will all take significant time.

What is 4F2: Increasing density through cell structure

DRAM chips are composed of hundreds of millions of memory cells, each storing data as electrical charge.In the industry, F represents the smallest feature size achievable by the process, and cell area is measured in F squared. Mainstream DRAM currently adopts 6F2 layouts, with each cell occupying an area of 3F×2F; the 4F2 layout compresses this to 2F×2F per cell.

With the same F, shrinking cell area from 6F2 to 4F2 allows for about 50% more cells in the same area, thus raising bit density.For a wafer fab, this means more capacity can be cut from the same wafer, and the unit bit cost has a chance to go down.

The tradeoff is just as direct.The smaller the cell, the more prone transistors become to leakage, and it’s easier for stored charge to dissipate. ChangXin's solution has two parts: first, VSG (vertical surround gate) transistors, putting the current channel vertically and letting the gate fully surround it to better control current and reduce leakage; second, W2W (wafer-to-wafer) hybrid bonding, fabricating the memory cell array and control circuitry on separate wafers using their optimal processes, then precisely aligning and bonding them together.

Goldman also listed the manufacturing challenges for 4F2: vertical transistors must be produced with highly consistent height; capacitors must retain enough charge in smaller area; denser wiring must be precisely designed to avoid interference; hybrid bonding adds steps and requires precise alignment between the memory and control layers.All these challenges could drive up manufacturing costs and reduce yields, possibly offsetting the benefits of “packing more cells in the same area.”

Why ChangXin chose 4F2: A detour under EUV limitations

There are usually two ways to increase DRAM density: shrink F, or reduce the number of F2 occupied by each cell.For over a decade, the industry mainly followed the first path, increasingly relying on finer lithography equipment for advanced nodes, especially ASML’s EUV (extreme ultraviolet lithography).

ChangXin can’t get EUV equipment, so there’s a limitation on shrinking F.Goldman views ChangXin’s shift to 4F2 as a response to this constraint: Moving from 6F2 to 4F2 increases density without fully relying on finer lithography. This explains why ChangXin is willing to take on the manufacturing challenge of 4F2 first and highlights that their trade-offs differ from Korean manufacturers.

A deep-dive report in August discussed the same constraints: Chinese storage remains weak in immersion DUV lithography, high-end metrology, some process control software, and HBM packaging qualification. There is a gap between “available” equipment and “production-worthy” equipment due to concerns about stability, throughput, and yield.4F2 is a structural path carving around the lithography bottleneck, but it shifts the core difficulties to vertical transistors, capacitors, and bonding, rather than making the process inherently easier.

Samsung and SK hynix are also moving toward 4F2, just not in a hurry

4F2 is not a route unique to ChangXin.Korean media report that Samsung has already produced the world’s first below-10nm DRAM working chip using a 4F2 architecture and VCT (vertical channel transistor, similar to ChangXin’s VSG), plans to complete development of a 4F2-based 10a node DRAM by the end of 2026, initiate customer verification in 2027, and ramp to mass production in 2028. Samsung reportedly plans to retain 4F2 and VCT for consecutive nodes—10a, 10b, 10c—and shift to 3D DRAM from 10d node onward.

SK hynix has also included 4F2 in its next-gen technological roadmap: roll out a 4F2 vertical gate platform for 10nm and below nodes, adopting wafer bonding to place control circuitry under the memory cell array.In July 2026, SK hynix verified read/write operations on a 4F2 test array.

The difference among the three companies isn’t whether they “can make” 4F2, but why and when they must do so.Samsung and SK hynix have EUV, so they can continue to improve density and performance by shrinking F without taking on early yield loss and higher cost from 4F2; their plan places 4F2 at sub-10nm nodes for around 2028, as a relay once the current roadmap nears its limit. For ChangXin, which lacks EUV, 4F2 is a necessity rather than a choice.

Therefore, Goldman argues that the essence of the 4F2 debate isn’t whether it can be made in the lab, but whether this architecture can deliver true manufacturing efficiency and economic payoff.Samsung and SK hynix are advancing at their own pace; even if ChangXin samples first, it does not mean the Korean players’ technical roadmaps are disrupted.

From design to mass production: What steps does ChangXin still need?

Goldman Sachs believes it’s too early to evaluate 4F2’s fundamental impact on Samsung and SK hynix.A true assessment requires at least five types of information, which are currently lacking publicly: whether the development stage is merely R&D samples or customer samples; cost competitiveness versus products already in the market; expected yield; customer qualification schedule; commercial rollout schedule.

According to present reports, ChangXin is now in the product design phase with a goal to enter the sampling phase by the end of 2026.If that timeline is met, it is a technological achievement in its own right. But Goldman stresses that going from sampling to customer validation and then to mass production is a completely different level of challenge.

This is especially true for server memory.Server customers require far higher standards for stability, compatibility, and long-term reliability than consumer electronics. A new product must pass multiple rounds of testing from CPU platforms, motherboards, and cloud vendors before it makes the procurement list; after qualification, it must also achieve yield rates on mass production lines that are cost competitive. Samsung’s own 4F2 timeline is end of 2026 for development completion, 2027 for validation, and 2028 for mass production—about two years between development and volume production. If we extrapolate with Samsung’s pace, after ChangXin's 2026 year-end sampling, the real impact on market supply will likely be around 2028, and only if both yield and cost pass muster.

This is also the main reason for Goldman’s view that “technological breakthroughs are being overestimated”: the market has already priced in results that will take at least two years to materialize, dependent on yield and cost, based merely on a piece of product design-stage news.

Node gap: 1z vs. 1b, 1c

In addition to the new 4F2 route, there’s still a marked gap between ChangXin and Korean players in the current 6F2 technology.DRAM nodes have evolved from 1x, 1y, 1z to 1a, 1b, 1c, with later numbers/letters indicating greater advancement. According to structure compiled by TF and company data, Samsung and SK hynix’s capacity is currently centered around the 1b node, transitioning toward 1c; ChangXin remains mainly on 1z, most likely still stuck at this node in 2027.

Goldman believes that without EUV, it will be very difficult for ChangXin to expand to more advanced nodes like 1b, 1c.Global DRAM makers have begun using EUV in some process layers several years ago, with SK hynix implementing it from 1a node and Samsung introducing it even before 1a. ChangXin might continue scaling without EUV via other means, but will face technical limitations and incur higher costs.

The generational product gap is similarly clear.Take low-power memory LPDDR used in mobile phones and laptops, for instance: by 2026, about 85% of Samsung’s and 76% of SK hynix’s LPDDR output will be LPDDR5 or 5X, while only about 29% for ChangXin, with the rest still in the previous generation LPDDR4 or 4X. ChangXin’s ratio goes from nearly zero in 2024 to about 14% in 2025—a decent catch-up—but at 2026 will still lag behind the 2024 levels at both Korean firms.

Capacity gap: The only metric closing rapidly

If the node and product generation gaps are hard to bridge in the short term, capacity is where ChangXin is catching up fastest.Goldman forecasts that by the end of 2026, Samsung’s and SK hynix’s DRAM monthly capacity will reach 770,000 and 660,000 wafers, respectively, with ChangXin at 290,000 wafers. By the end of 2028, ChangXin will rapidly rise to 513,000 wafers, a 77% two-year increase; during the same period, Samsung and SK hynix will expand to 930,000 and 770,000, up 21% and 17%, respectively.

According to Goldman’s estimates, from year-end 2026 to end-2028, ChangXin will add about 223,000 wafers per month in new capacity, more than Samsung’s 160,000 and SK hynix’s 110,000. The three combined will add 493,000, with ChangXin accounting for about 45%.ChangXin’s monthly output, as a percentage of Samsung’s, will rise from about 38% to about 55%; as a percentage of SK hynix’s, from about 44% to around 67%. Of total output among the three, ChangXin’s share goes from about 17% to 23%.

This forecast is largely consistent with Morgan Stanley’s numbers cited in the August deep-dive report: they estimate ChangXin’s monthly output will reach about 300,000 in 2026 and about 500,000 in 2028, noting that ChangXin will account for about 13% of global DRAM wafer capacity in 2026, but only about 11% of bit shipments.This gap between wafer share and bit share is because ChangXin’s nodes lag behind, so each wafer yields less capacity. The two firms’ capacity views are close; their main divergence is over when these capacity increases will impact prices.

Goldman still expects that by 2028, Korean manufacturers will hold a substantial lead in capacity, but acknowledges ChangXin’s speed and resolve in expanding.Memory is a price-transparent commodity; even capacity additions from less advanced nodes, as long as they enter mainstream applications, will change the marginal pricing. Capacity catch-up affects DRAM pricing more directly than technology catch-up—hence Goldman’s agreement that “the market had previously underestimated the capacity ramp-up.”

HBM: The biggest gap lies in the base die

In HBM (High Bandwidth Memory), Goldman thinks that barriers for ChangXin to catch up are even larger.HBM vertically stacks multiple DRAM chips, connects them with TSVs, and packages this stack together with GPUs—making it the most critical supporting component for AI accelerators. Its competitiveness derives not just from DRAM itself, but also from packaging technology and the base die at the bottom of the stack.

Early generation base dies like HBM3 could still be fabricated using mature process technologies.For next-generation specs, though, to support higher bandwidth and efficiency, the base die has migrated to leading-edge foundry nodes. For example, Samsung’s HBM4 base die uses 4nm foundry technology. For ChangXin to make a competitive next-gen HBM, domestic foundries need to mass-produce advanced node base dies; yet export controls limit domestic access to EUV and other state-of-the-art equipment, making the base die an ongoing bottleneck for catching up with HBM leaders.

This extends ChangXin’s technology gap from “memory process” to “foundry process.”Even if 4F2 allows ChangXin to take a step forward in mainstream DRAM density, it doesn’t solve HBM’s dependence on advanced logic process for the base die. Thus, ChangXin’s 4F2 news has limited direct impact on the HBM business of the two Korean companies.