China’s CXMT is preparing limited production of HBM3E memory for AI processors, even as reported yields remain far below those of established South Korean rivals. The move reflects a serious domestic shortage, tighter US export controls, and Beijing’s push to build AI hardware at home, but it also shows how expensive catching up can be.
China’s AI memory shortage is forcing a difficult bet
The striking detail is the waste. Sources told Reuters that only about one-fifth to one-third of CXMT’s HBM3E output may currently meet specifications, leaving a large share to be discarded or reworked. For a product built through intricate layers of stacked memory, that kind of loss can quickly turn every usable chip into an expensive one.
Even so, CXMT is moving ahead because Chinese AI-chip companies need memory now. HBM, or high-bandwidth memory, sits alongside AI processors and moves data at extremely high speeds, making it essential for training and running large language models.
The company has reportedly sent samples to several domestic chip designers and is preparing a limited production run. That suggests the first goal may not be commercial efficiency, but simply securing a local supply line. Technology reporting from XenoSpectrum and TechTimes has also pointed to growing interest in CXMT’s HBM efforts among Chinese AI hardware companies.
Why HBM3E matters to China’s chip industry
HBM is made by stacking memory dies vertically, a design that packs speed and capacity into a small footprint. It’s technically demanding, and tiny manufacturing flaws can spoil an otherwise valuable package.
That matters because China is trying to develop alternatives to Nvidia-powered AI systems while access to advanced foreign technology is restricted. Industry analysis from iTechGuides notes that export controls have complicated access to equipment and components across China’s semiconductor sector, making domestic memory production more important than ever.
CXMT, formally ChangXin Memory Technologies, is already China’s largest DRAM maker and a significant supplier of conventional memory. But moving from standard DRAM to advanced HBM is a much steeper climb. The machinery, packaging expertise, testing, and reliability requirements all raise the stakes.
Low yields make local chips a costly solution
A yield of 20% to 30% doesn't mean the technology is useless. It means the manufacturing process isn't yet efficient enough to compete comfortably with the industry's leaders. Each successful chip must effectively help pay for the failed ones around it.
That cost pressure could flow through the AI hardware market. Reuters previously reported that several Chinese chip designers had raised prices for current and next-generation AI products by 20% to 50% as memory became harder and more expensive to obtain.
Unofficial supply channels have reportedly offered foreign HBM at several times the normal price. CXMT’s local option could therefore look attractive even if it costs more than comparable overseas products. For Chinese buyers, a pricier domestic chip may still be preferable to an uncertain imported one.
CXMT is chasing leaders that keep moving
The performance gap is just as important as the manufacturing gap. Sources said CXMT’s HBM3E testing was closer to SK Hynix’s previous-generation HBM3 than to the company’s newer HBM3E product.
That leaves CXMT competing against firms with enormous scale and years of specialised experience. Counterpoint Research data cited in the report put SK Hynix at 50% of global HBM revenue in the second quarter, followed by Samsung Electronics at 33% and Micron Technology at 18%. Together, those figures show how concentrated the market remains.
Meanwhile, the leaders aren't standing still. Tom's Hardware has reported that Chinese memory is beginning to appear in limited quantities in some non-US notebooks, while Samsung has been promoting newer HBM4 products. CXMT may be entering the market just as the technology frontier moves again.
What buyers and chip designers should watch next
The next useful signal will be whether CXMT can improve yields without sacrificing performance. Its reported two-version approach, including an updated design intended to improve manufacturing results, suggests the company is already treating production efficiency as a priority.
Customers will also be watching consistency. AI-chip designers need memory that works reliably across large batches, not merely impressive samples. A small number of successful packages can prove a concept, but high-volume deployment requires predictable supply, stable performance, and manageable costs.
For now, CXMT’s move looks less like a finished breakthrough than a determined opening move. It could give Chinese AI companies a badly needed source of memory, but the soft spot in the process remains clear: making enough good chips, at a reasonable price, before rivals move another generation ahead.
The road to AI independence may begin with imperfect chips, but production discipline will decide how far they go.
Disclaimer: This article may have been created with AI assistance and reviewed by our editorial team. It is provided for general informational purposes only. Readers should verify information independently before relying on this content.
