China's CXMT Hit $8 Billion in Memory Revenue — HBM3 Production Starts Before Year End

Abhishek GautamAbhishek Gautam9 min read
China's CXMT Hit $8 Billion in Memory Revenue — HBM3 Production Starts Before Year End

Quick summary

ChangXin Memory Technologies grew revenue 130% to $8 billion in 2025. Wafer capacity tripled. HBM3 — the chip inside every Nvidia AI GPU — starts mass production by end of 2026. For the first time, Samsung, SK Hynix, and Micron face a credible Chinese challenger across all memory tiers.

For three decades, three companies controlled the global memory chip market: Samsung (South Korea), SK Hynix (South Korea), and Micron (United States). No other company had successfully scaled DRAM production to compete at the high end. That is changing in 2026. ChangXin Memory Technologies, known as CXMT, posted $8 billion in revenue in 2025 — a 130% year-on-year jump — and is on course to begin mass production of HBM3, the high-bandwidth memory that powers every Nvidia AI accelerator, before the end of 2026.

The Revenue Numbers: $8 Billion, Up 130%

CXMT reported 2025 revenue of more than 55 billion yuan, equivalent to approximately $8 billion at current exchange rates. The 130% year-on-year growth rate is not a base effect from a small starting point. CXMT was already a significant producer before 2025. The jump reflects a combination of volume growth, product mix shift toward higher-value DDR5 and LPDDR5, and price tailwinds from the memory market tightening driven by AI demand.

For context: Samsung's memory division revenue in 2025 was approximately $70 billion. SK Hynix — the leader in HBM specifically — posted memory revenue of roughly $45 billion. Micron reported approximately $25 billion. CXMT at $8 billion is not yet in the same tier, but the trajectory is the story. The company projects H1 2026 revenue of 110-120 billion yuan — a 613-677% year-on-year increase — as HBM volume ramps and DDR5 pricing holds.

If that H1 2026 guidance holds, CXMT could exit 2026 with annualised revenue approaching $35 billion, putting it directly in Micron's range.

What CXMT Actually Makes

CXMT's current product portfolio spans the major DRAM categories that matter for AI infrastructure:

DDR5: The latest generation of standard server memory. DDR5 is now the default memory for new AI server configurations, offering double the bandwidth of DDR4 at lower power per bit. CXMT is in mass production on DDR5 and has passed qualification at several major Chinese server OEMs.

LPDDR5: Low-power DDR5, used in smartphones, laptops, and edge AI devices. CXMT's LPDDR5 production serves the Chinese consumer electronics supply chain — Huawei, OPPO, Xiaomi — reducing their dependence on Samsung and SK Hynix for mobile memory.

HBM3: High-bandwidth memory — the product that matters most for AI training. HBM3 is stacked memory bonded directly to AI accelerator chips. Nvidia's H100 and H200 use SK Hynix HBM3. The GB200 uses HBM3e. Samsung and SK Hynix supply virtually all HBM globally. CXMT began limited HBM production in early 2026 and is targeting mass production of HBM3 by end of 2026.

The HBM3 timeline is the critical data point. If CXMT achieves volume HBM3 production in H2 2026, China gains a domestic supply of the memory type most essential for AI training at scale.

Capacity: 100K to 290K Wafers Per Month

The production capacity numbers illustrate why CXMT's revenue trajectory is credible rather than aspirational.

At the start of 2024, CXMT was producing approximately 100,000 wafers per month — a meaningful but second-tier volume. By the end of 2025, capacity had reached approximately 290,000 wafers per month. The target for 2026 is 300,000 wafers per month.

This near-tripling of capacity in two years reflects both greenfield fab construction and equipment procurement at scale. CXMT has invested aggressively in both Dutch-origin ASML DUV lithography equipment (older generation, not restricted by export controls as of the most recent BIS rules) and Chinese domestic equipment from suppliers like NAURA and Advanced Micro-Fabrication Equipment (AMEC). The domestic equipment proportion in CXMT's fabs is higher than at any previous point — a deliberate strategy to reduce dependence on Western equipment suppliers.

At 300,000 wafers per month and current DDR5/LPDDR5 revenue per wafer, CXMT's capacity alone supports the H1 2026 revenue guidance.

HBM3: Why This Is the Critical Product

High-bandwidth memory is the pinch point in global AI chip supply. Every Nvidia H100, H200, and B200 GPU requires HBM. The AI training clusters that OpenAI, Google, Microsoft, and Meta are building consume HBM in volumes that have consistently outpaced supply.

SK Hynix held approximately 53% of the HBM market in 2025. Samsung held approximately 35%. Micron held the remainder. No other company produced HBM in meaningful volume. This near-duopoly gave SK Hynix significant pricing power and put every AI lab in a position of dependency on Korean memory supply.

CXMT's entry into HBM3 production changes the supply dynamics in two ways. First, it adds global supply that can relieve the shortage that has constrained AI hardware production. Second, it provides Chinese AI labs — Baidu, Alibaba, ByteDance, DeepSeek — with a domestic HBM source that is not subject to US export controls or South Korean diplomatic pressure.

The US Commerce Department has not yet placed HBM-specific export controls targeting CXMT. The combination of Chinese domestic memory supply and the energy advantage discussed separately in our China AI energy analysis means Chinese AI training clusters have a credible path to full domestic supply chains for AI compute by 2027-2028.

YMTC Plus CXMT: China Covers Both Memory Types

Memory chips divide into two categories: DRAM (dynamic random-access memory, the working memory of a computing system — this is what CXMT makes) and NAND flash (the storage memory that retains data without power — solid-state drives, phone storage). Historically, South Korean and US companies controlled DRAM while NAND was more contested.

China's Yangtze Memory Technologies Corporation (YMTC) reached 232-layer NAND production in 2024 and has been pouring capacity into the 2026 memory price spike. Together, CXMT in DRAM and YMTC in NAND mean China now has credible domestic production across both major memory types. The remaining gaps — HBM yield rates, advanced packaging for stacked die, leading-edge EUV lithography — are closing.

Our Analysis: Why Micron Should Be More Concerned Than Samsung

Samsung and SK Hynix are at risk from CXMT, but they have the R&D budget and process technology lead to stay ahead on the highest-performance products for several years. Micron is in a different position.

Micron is the only US DRAM producer and has benefited from both SK Hynix's HBM focus (which diverted some of SK Hynix's commodity DRAM capacity) and China's exclusion from the market by US export controls. CXMT's rise threatens Micron's commodity DRAM and DDR5 market share directly, without the HBM premium that protects SK Hynix. If CXMT achieves DDR5 cost parity with Micron through its $0.05/kWh electricity advantage and lower labour costs, Micron loses its commodity DRAM pricing power.

For the global AI supply chain: CXMT's growth is net positive for AI infrastructure availability. More memory supply means less HBM scarcity and lower GPU costs over time. For developers building on AI infrastructure in India, Singapore, and Southeast Asia — where Chinese components are not subject to the same procurement restrictions as in the US — CXMT memory appearing in servers and edge devices is already a reality.

See our global semiconductor market $1.5 trillion analysis for where CXMT fits in the broader chip market picture.

Key Takeaways

  • CXMT revenue hit $8 billion in 2025, up 130% year-on-year: the growth reflects volume scale, DDR5 and LPDDR5 product mix, and AI-driven memory price tailwinds
  • H1 2026 revenue guidance is 110-120 billion yuan (+613-677% YoY): if it holds, CXMT exits 2026 approaching Micron-tier annual revenue
  • Wafer capacity tripled from 100K to 290K per month (2024-2025): 300K targeted in 2026, reflecting both greenfield fabs and domestic equipment substitution
  • HBM3 mass production targeted end of 2026: this is the critical product — it powers every Nvidia AI accelerator and is currently monopolised by SK Hynix and Samsung
  • YMTC (NAND) plus CXMT (DRAM) = China with domestic memory across both major types: the remaining gap is HBM yield and advanced packaging, both closing in 2026-2027
  • Micron is the most exposed Western company: CXMT's DDR5 directly threatens Micron's commodity DRAM market share without the HBM premium that partially protects SK Hynix
  • IPO on mainland China exchange expected imminently: CXMT will become a publicly traded company, giving Chinese investors direct exposure to the domestic memory buildout

Sources

FAQ

Frequently Asked Questions

What is CXMT and why does it matter for AI?

CXMT (ChangXin Memory Technologies) is China's leading DRAM memory chip manufacturer. It matters for AI because every AI accelerator — Nvidia H100, H200, B200 — requires High Bandwidth Memory (HBM), a type of DRAM. Samsung and SK Hynix supply virtually all HBM globally. CXMT is entering HBM3 production in 2026, which means China will gain a domestic supply of the memory type most essential for AI training. CXMT's 2025 revenue was $8 billion (up 130% year-on-year), with H1 2026 guidance of 110-120 billion yuan — a potential 613-677% year-on-year increase.

When will CXMT start producing HBM3?

CXMT began limited HBM production in early 2026 and is targeting mass production of HBM3 by the end of 2026. HBM3 is the high-bandwidth memory used in Nvidia H100 and H200 GPUs. Currently, SK Hynix holds approximately 53% of the HBM market and Samsung holds 35%. If CXMT achieves volume HBM3 production in H2 2026, it will be the first non-Korean, non-American company to supply HBM at commercial scale, reducing China's dependence on South Korean memory supply for AI training infrastructure.

How fast has CXMT grown its production capacity?

CXMT's wafer capacity tripled from approximately 100,000 wafers per month at the start of 2024 to approximately 290,000 wafers per month by end of 2025, with 300,000 per month targeted for 2026. This growth was achieved through a combination of greenfield fab construction, procurement of Dutch ASML DUV lithography equipment (not covered by current export controls), and increasing use of Chinese domestic equipment from suppliers like NAURA and AMEC. The domestic equipment proportion is higher than at any previous point in CXMT's history.

Does CXMT threaten Samsung, SK Hynix, or Micron more?

Micron is the most directly exposed Western company. CXMT's DDR5 and LPDDR5 production targets the commodity DRAM market where Micron competes, without the HBM premium products that partially protect SK Hynix. Samsung has the R&D scale to maintain a process technology lead for several years. SK Hynix's dominance in HBM is not immediately threatened by CXMT. But Micron, as the sole US DRAM producer, faces commodity market share pressure from CXMT's cost structure — particularly CXMT's energy cost advantage from $0.05/kWh industrial electricity in China's western provinces.

What is the difference between CXMT and YMTC in China's memory chip strategy?

CXMT produces DRAM — the working memory used in computing systems, AI accelerators, smartphones, and servers. YMTC (Yangtze Memory Technologies) produces NAND flash — the storage memory used in SSDs, smartphones, and data centers. Together they give China domestic production across both major memory types for the first time. YMTC reached 232-layer NAND in 2024 and has been adding capacity into the 2026 memory price cycle. CXMT covers DRAM including DDR5, LPDDR5, and now HBM3. The remaining gaps are HBM yield quality, advanced packaging for stacked die, and EUV lithography access — all of which are closing.

Free Weekly Briefing

The AI & Dev Briefing

One honest email a week — what actually matters in AI and software engineering. No noise, no sponsored content. Read by developers across 30+ countries.

No spam. Unsubscribe anytime.

Written by

Software Engineer based in Delhi, India. Writes about AI models, semiconductor supply chains, and tech geopolitics — covering the intersection of infrastructure and global events. 917+ posts cited by ChatGPT, Perplexity, and Gemini. Read in 167 countries.