All research

AI · Flagship research · 2026-08-28

China Can Reshape Memory Without Winning HBM

China can pressure conventional DRAM prices long before it reaches HBM parity. The result could be a split memory market, faster incumbent migration into custom HBM, and a more violent cycle when supply turns.

, Founder and publisher, AI Bottlenecks

China can redraw the economics of memory without matching SK hynix, Micron or Samsung at the frontier of high-bandwidth memory.

CXMT only needs enough qualified conventional DRAM, sold at enough scale, to influence the price of the marginal bit. That would reach incumbent revenue and margins well before Chinese HBM reaches technical parity.

Conventional DRAM and HBM share wafers, cleanrooms and engineering talent, but their competitive thresholds are moving apart. Conventional DRAM rewards yield, cost and scale. HBM adds thin-die handling, vertical interconnects, bonding, thermal control, custom logic and platform qualification.

> The first Chinese impact will probably appear in the price of ordinary memory, not in the HBM socket beside the world's fastest accelerator.

From aibottlenecks.app.

The thesis in four points

  1. CXMT has crossed the relevance threshold. Its IPO prospectus put its 2025 global DRAM share at roughly 7.7%, according to Reuters. A supplier at that scale can influence capacity decisions even before it becomes a global technology leader.
  2. The next credible Chinese milestone is mobile DRAM. CXMT has been linked to LPDDR6 for Xiaomi's Xring O3 processor. The supplier relationship remains unconfirmed, but a volume shipment would show that Chinese DRAM is moving into current-generation premium devices.
  3. The HBM barrier is rising while China catches conventional DRAM. Hot Chips 2026 and NVIDIA's new NVHBM architecture show HBM shifting toward taller stacks, harder thermal constraints and custom base-die logic.
  4. Price pressure should arrive before HBM share loss. Qualified Chinese DDR5 and LPDDR6 can weaken the price umbrella in conventional memory while frontier HBM remains concentrated among three incumbents.

One capacity pool, two competitions

Conventional DRAM

Servers, PCs, phones and industrial systems use conventional DRAM as working memory. Suppliers compete through cost per bit, power, speed, yield and the ability to qualify across a wide set of devices.

A domestic supplier with acceptable performance can win business through availability, policy support and supply security. Every qualified socket adds volume, factory learning and a larger installed base for the next product generation.

High-bandwidth memory

HBM feeds accelerators through a wide interface inside an advanced package. The DRAM die is only one component in a stack that must survive wafer thinning, vertical connections, bonding, heat, packaging and customer-specific validation.

The commercial unit is therefore closer to a qualified subsystem than a commodity chip. Memory performance, base-die logic, the surrounding package and the accelerator platform increasingly develop together.

> Conventional DRAM is a scale-and-cost competition; HBM multiplies that competition by packaging yield and system integration.

Four signals since August 24

1. CXMT has a possible LPDDR6 opening

Xiaomi says its Xring O3 mobile processor supports LPDDR6. Chinese supply-chain reporting links CXMT to the platform, while The Korea Herald describes the company as a prospective supplier rather than a confirmed volume vendor.

The evidence now sits on three rungs: interface support shows that Xiaomi built for LPDDR6, samples would show that CXMT has functional silicon, and a named device shipping in volume would add evidence on qualification, yield and repeatability.

If that final step arrives, CXMT will have moved beyond import replacement in older PC memory and into a current mobile standard. It would also give the company a demanding domestic customer whose product cycles can accelerate process learning.

2. HBM has reached a packaging wall

SK hynix used Hot Chips 2026 to put 12-layer HBM4 in production and 16-layer HBM4 under qualification. Its presentation showed a 775-micrometre package height and positioned hybrid bonding as the route to 20-layer stacks and beyond.

The slides concentrated on thin-die handling, warpage, thermal resistance, micro-bump pitch and stress from the surrounding package. Those are factory constraints with direct effects on yield, cost and shipment timing.

3. HBM is becoming custom silicon

NVIDIA announced NVHBM on August 26. The architecture moves NVIDIA's custom memory controller from the XPU into the HBM base die. NVIDIA claims up to 30% higher memory bandwidth, 15% lower HBM power and 25% more compute-die area compared with standard HBM4E.

Shipping systems will determine the realized gains. The design choice already changes the competitive requirement: a memory supplier must now integrate logic that belongs to the accelerator architecture, then qualify the complete stack across several providers.

NVIDIA also intends to standardize NVHBM across multiple memory suppliers. That can reduce switching friction among incumbents while raising the engineering threshold for any new entrant. Customization and multi-sourcing are advancing at the same time.

4. The shortage still distorts every comparison

TrendForce reported that conventional DRAM contract prices rose about 93% to 98% quarter on quarter in the first quarter of 2026. Its July outlook still called for another 13% to 18% increase in the third quarter, despite weaker consumer demand and a much higher base.

Reuters also found some CXMT server modules priced above comparable Samsung products in China. Scarcity and domestic availability can create pricing power before a supplier achieves cost leadership.

Current revenue and margins therefore reveal little about CXMT's position at a normalized clearing price. The first downturn will provide the cleaner test.

How Chinese supply reaches the rest of the market

The effect travels through four channels.

1. The marginal conventional bit gets cheaper

Memory pricing turns on small changes in supply and demand. Once CXMT can deliver qualified DDR5 or LPDDR6 beyond protected pilot volumes, incumbents face a difficult choice: defend share with price or protect margins by conceding sockets.

Neither response requires CXMT to become the industry's lowest-cost producer. It only requires a buyer to view the Chinese alternative as usable and available.

2. Incumbents move further upmarket

Samsung, SK hynix and Micron already direct more wafers, cleanroom space and engineering toward HBM and high-capacity server DRAM. Chinese competition in conventional products strengthens that incentive.

The resulting feedback loop leaves less incumbent capacity for ordinary DRAM, supporting near-term prices and generating the cash used to fund HBM. The same allocation raises future downside because any HBM slowdown can release capacity back into conventional products while CXMT continues to expand.

3. Domestic customers become part of CXMT's process development

Reuters reported in June that CXMT had reached a multiyear server-DRAM agreement with Tencent worth more than 20 billion yuan. Neither company confirmed the deal, and Reuters could not establish whether it included HBM. The report also named Tencent, Alibaba Cloud, ByteDance, Lenovo and Xiaomi as major customers and described a planned expansion from roughly 300,000 to 600,000 wafers per month. Several details rely on unnamed sources.

The strategic value sits in the feedback loop rather than the headline contract value. Domestic customers can provide volume, workload data and repeat qualification. Each cycle improves the supplier's understanding of yield, firmware, thermals, system behavior and failure modes.

4. The global and Chinese supply chains begin to diverge

Export controls, domestic-content policy and platform-specific qualification are pushing memory toward two overlapping ecosystems. Chinese accelerators and devices can adopt domestic memory earlier, accepting a different balance of performance, cost and supply security. Global platforms can continue around NVIDIA, TSMC, Samsung, SK hynix and Micron.

China could therefore win selected domestic sockets while remaining one or two product generations behind at the global frontier. Generation labels alone will understate the commercial progress inside China and overstate its transferability outside China.

Why HBM yield compounds

HBM assembly places several manufacturing processes in series:

  1. Fabricate DRAM dies with the required speed, power and reliability.
  2. Etch and fill microscopic vertical connections.
  3. Thin the wafers without bending or breaking them.
  4. Stack and bond known-good dies with precise alignment.
  5. Connect the stack through a logic base die and advanced package.
  6. Pass electrical, thermal, reliability and platform tests.

If five independent stages each yielded 90%, only 59% of units would survive all five:

0.9 × 0.9 × 0.9 × 0.9 × 0.9 = 0.59

The calculation illustrates compound yield rather than estimating a real factory. Manufacturers screen bad dies early, process stages are correlated and actual yields differ by step. Even so, a defect found late can destroy a stack containing several tested dies and a logic base die.

Lam Research describes millions of vertical connections that require precise alignment and copper fill. Applied Materials says HBM dies can be thinned to roughly one twenty-fifth the thickness of a standard wafer, increasing warpage and deformation risk.

Large DRAM capacity provides the starting material without guaranteeing acceptable end-to-end stack yield.

Qualification creates the commercial moat

The incumbents' disclosures show the sequence:

Sampling removes basic functional risk; qualification adds platform reliability; commercial shipment adds manufacturing repeatability; stable volume adds confidence in yield, supply and field performance.

Each program also teaches the supplier how its memory behaves with a specific accelerator, base die, package, power profile and workload. That knowledge transfers into the next product generation and shortens the next qualification cycle.

> An HBM sample is an engineering milestone. Qualified volume is a durable commercial position.

The read-through by company

CXMT

CXMT gains leverage as soon as customers trust its conventional products for repeat deployments. DDR5 server modules and LPDDR6 mobile shipments offer more useful evidence than total wafer capacity because they expose the company to demanding systems and current standards.

The largest uncertainty is cost after the shortage ends. Strong margins during extreme pricing can conceal weak yield, higher tool costs or a product mix concentrated in less demanding applications.

An ongoing Seoul case adds legal and political risk. A former Samsung engineer who has already been sentenced testified that CXMT's founding team sought Samsung process data rather than developing the process from scratch. The testimony concerns a roughly 600-step DRAM process plan, while the court has not reached a final finding against CXMT's leadership. The case could restrict Western adoption even if domestic qualification continues.

Micron

Micron has the clearest geopolitical protection as the only US-based DRAM producer, alongside direct exposure to Chinese restrictions and conventional price pressure. Its strategic defense is a faster mix shift toward HBM and high-capacity server memory.

If HBM growth stays ahead of conventional price erosion, Micron can improve its earnings quality. If HBM demand slows while Chinese conventional output rises, the company faces both weaker commodity pricing and underused premium capacity.

SK hynix

SK hynix owns the strongest current HBM position through yield, customer history and packaging execution. Custom HBM can deepen those relationships because more engineering moves into the base die and surrounding package, although the tighter link to a few accelerator roadmaps increases qualification expense and raises the cost of a customer delay. SK hynix also retains exposure to the wider DRAM cycle through its conventional bits.

Samsung

Samsung combines DRAM, logic fabrication and advanced packaging inside one group. That structure fits a market in which HBM base dies carry more custom logic and thermal design begins earlier.

Samsung also has the largest conventional DRAM pool exposed to Chinese price pressure. Its integrated model can become a stronger advantage in custom HBM while creating a larger earnings offset when ordinary memory prices fall.

Equipment suppliers

HBM increases the process intensity of every delivered bit through more etch, deposition, planarization, bonding, inspection and metrology. Chinese fab expansion creates another source of equipment demand, but export rules limit which tools can reach which facilities.

The durable position belongs to the supplier that controls a yield-critical step, can legally ship and remains difficult to replace. Broad exposure to wafer starts is less valuable than ownership of the process where scrap is accumulating.

Memory buyers

PC, phone and server manufacturers gain a second source in conventional DRAM, especially for products sold inside China. Once qualification broadens, that can improve bargaining power and supply resilience at the cost of more fragmentation: separate memory bills of materials for Chinese and global products, duplicate qualification work and changing export or procurement rules.

Three paths from here

1. Conventional catch-up with a persistent HBM gap

CXMT improves DDR5 and LPDDR6 yield, wins repeat domestic orders and adds share while qualified HBM remains concentrated among the three incumbents.

This is the most direct route to lower conventional prices, stronger incumbent investment in premium memory and a larger valuation gap between HBM leadership and commodity exposure.

2. A domestic HBM beachhead

Chinese accelerator designers, packaging firms and cloud operators co-develop memory with CXMT. Initial products trail the global frontier but ship on selected domestic platforms where supply security outweighs peak performance.

The commercial gap would narrow inside China before the technology gap closes globally. Equipment controls would become more targeted around bonding, metrology, base-die logic and the process steps constraining yield.

3. Tool access and yield stall the frontier

CXMT adds conventional capacity but fails to achieve stable yield on advanced products. Customers keep Chinese memory in policy-driven or lower-value sockets while HBM remains outside commercial reach.

This path can still hurt ordinary DRAM pricing because new supply enters the least differentiated part of the market. It would leave the HBM profit pool concentrated and make the conventional downturn more severe.

The dashboard that survives the cycle

> Track qualified output and cost per good bit. Announced capacity and peak-cycle margins can hide the economics that appear when supply loosens.

Signals that would force a revision

Evidence of a faster Chinese HBM advance would include repeat customer-qualified shipments, competitive yields on 12-layer or taller stacks, several accelerator customers at volume and credible power, heat and reliability data from complete systems.

Evidence of a weaker conventional threat would include poor DDR5 or LPDDR6 yield, inability to secure essential tools, customers limiting CXMT to low-value deployments or demand absorbing new capacity without lower clearing prices.

Conclusion

The next phase of the memory market is likely to split rather than converge. China can gain scale in conventional and mobile DRAM while the HBM leaders extend their advantage through packaging, custom base dies and platform-specific qualification.

The split changes every participant's strategy: CXMT can pressure the marginal conventional bit, incumbents can move further into premium memory, buyers can gain a new source while accepting a more fragmented supply chain, and equipment value can concentrate around the few process steps that determine yield.

The cycle may also become less forgiving. HBM absorbs capacity during the expansion, Chinese conventional output keeps rising, and any later rotation of incumbent wafers back from HBM can place both sources of supply into the same downturn.

China does not need the industry's best HBM to alter the economics below it. Qualified conventional output at scale is enough to start.

Selected sources

Research date: 27 August 2026. Educational analysis only. Not investment advice.