Memory now decides the AI race
The price of plain DRAM rose 90% in one quarter as capacity moved to HBM. China advanced where nobody was looking — 10% of global DRAM, 14% of NAND and the most valuable company on the Shanghai exchange — while American export control aims at yesterday’s bottleneck.
The shortage deciding the artificial intelligence race is a shortage of memory, and it has inverted who holds the bargaining power.
The mechanism is simple and brutal. An AI accelerator does not work without high-bandwidth memory stacked beside the processor, and that memory yields several times more per wafer than the common kind. Samsung, SK hynix and Micron moved capacity there. The result showed up in the price of what was left: plain DRAM contracts rose more than 90% in the first quarter of 2026, in what Goldman Sachs described as the worst memory shortage in fifteen years, according to TrendForce records.
With the big three allocated to HBM, the company holding spare common capacity was precisely the one Washington spent two years trying to contain.
What China does where nobody was looking
Counterpoint’s second-quarter figures show the size of the move. CXMT, the Chinese DRAM maker, reached 10% of the global market, double the 4% of a year earlier, in fourth place behind Samsung (38%), SK hynix (25%) and Micron (24%). In NAND, YMTC rose from 9% to 14%. South Korea’s semiconductor industry association publicly acknowledged that the gap with China has fallen to three years in HBM, two in DRAM and one in NAND, where the previous estimate exceeded five years.
The first name on that list deserves separate reading. In HBM, CXMT sells nothing: its disclosed portfolio runs from DDR4 to LPDDR5X, and its May 2026 listing document allocates 29.5 billion yuan to projects with not a single HBM line among them. When the Shanghai stock exchange asked directly about HBM, in December 2025, the company listed the product among its competitors’ rather than among its own.
The opening lies in NAND. Three people the market has consulted say CXMT will set up a NAND flash research and development line at the new Beijing fab, with its own research institute, and has already offered the chips to buyers — among them a young storage company for AI systems and supercomputers. No start date was disclosed, and the company has not said whether it is moving from experimental production to commercial volume. Samsung led NAND with 29.3% of revenue in the second quarter, according to TrendForce.
A research line does not threaten 2027 supply. It exists so that a path exists on the day access to equipment closes for good.
Why the encirclement did not close
The American instrument was designed to block the top of the chain: YMTC entered the entity list in 2022 and Chinese access to the HBM chips used with AI processors was restricted later. Without extreme ultraviolet lithography, which ASML controls and the American regime bars, the most advanced HBM cannot be made.
The Chinese answer was to go around rather than scale up. CXMT folded into its HBM design the X-Stacking technique, developed by YMTC for three-dimensional NAND stacking, according to the Taiwanese publication DigiTimes: two wafers bonded directly, which raises bandwidth and cuts heat without needing EUV. In Hefei the company is testing bonded DRAM, which separates the cell layer from the control layer into two wafers, with deep ultraviolet equipment and multiple patterning. Reuters further reported that China has built a prototype EUV machine, which changes the nature of the deadline, not the deadline.
And there is the capital. CXMT’s Shanghai debut opened up 471% and reached about 500 billion dollars in market value by midday, overtaking ICBC as the most valuable company on the continent’s exchanges, the first semiconductor firm to lead the A-share market in 35 years. In the first quarter of 2026 the company earned 24.76 billion yuan on revenue of 50.80 billion, with a gross margin of 79.2%.
The other end: capacity committed through 2029
On the side that still controls the technology, the answer is volume. Samsung is expected to more than double production of the HBM4 family in 2027, according to the Seoul Economic Daily: outsourced cleaning of glass substrates rises from 20,000 to 50,000 pieces a month, and monthly HBM wafer intake goes from about 180,000 to 250,000. The HBM4 family moves from about 40% of shipments to roughly 80%. Yield, which was below 60% at the start, has come close to 80%. The company has already presented the zHBM architecture, which stacks memory directly on top of the accelerator, and a target of a thousand tokens per second.
That is not slack. Nvidia, Alphabet and AMD may consume up to 85% of global HBM output in 2027, according to the analyses cited, and the three suppliers have sold practically all of that year’s capacity. New lines only arrive in volume in late 2027 or in 2028.
The instrument’s target
Put the two ends together and the American design appears with a targeting problem. Export control was built to deny what China does not have: EUV for the most advanced HBM. The restriction that actually moved the market, however, was on common DRAM and NAND capacity, precisely where Chinese share rose from 4% to 10% in a year and where the local exchange has just delivered capital on a scale no development budget can reach.
The coming months will test something other than whether CXMT makes HBM: whether the Beijing research line leaves the laboratory while common memory prices are still high enough to finance the build. A country that learned to do more with less equipment, and that now has market money to buy the rest, changes level without having to win the race its adversary chose.