CXMT's share of global DRAM revenue fell to 9% in the third quarter (Q3) of 2026. According to research published by Counterpoint Research on November 5, that is down 1 percentage point from 10% in the previous quarter. However, a smaller share of total market revenue does not necessarily mean CXMT's own revenue or shipment volume declined. With the DRAM market expanding rapidly on rising prices, the key question is how to interpret the 9% figure.
Compared with a year earlier, CXMT's share is still much higher. Counterpoint previously reported a 5% share for Q3 2025, so the current 9% is 4 percentage points higher. This is a reference comparison using data published at different times, but it shows that year-over-year share growth and a recent quarterly decline can coexist. Note also that this is a revenue share, not a measure of manufacturing capacity or the number of chips actually sold.
A Lower Share Doesn't Necessarily Mean Lower Revenue
According to Counterpoint, global DRAM market revenue in Q3 2026 reached about $200 billion, up 369% from a year earlier. That is 4.69 times the figure for the same period last year. The firm attributes the surge more to rising prices than to growth in shipment volume. When total market revenue swings this much, market share and actual revenue need to be considered separately.
Revenue share is a company's revenue divided by total market revenue. If the whole market is growing, a company's revenue can rise even as its share falls. The same applies to CXMT: using the previous quarter's 10% and the current 9%, we can calculate the condition under which its revenue grew.
Let R be CXMT's revenue and M be global DRAM market revenue, with subscripts 2 and 3 denoting Q2 and Q3. The quarter-on-quarter change in CXMT's revenue can be expressed as follows.
R₃/R₂ = (9/10) × (M₃/M₂)
From this, CXMT's revenue grows quarter on quarter when M₃/M₂ > 10/9. In other words, if total market revenue grew by more than about 11.1% quarter on quarter, CXMT's revenue would have increased even with its share slipping from 10% to 9%. This is a rough estimate based on the published shares, however, and does not mean CXMT's revenue actually grew. Counterpoint's public article does not give the quarter-on-quarter change in Q3 market revenue, so it cannot be confirmed whether this condition was met.
The 369% global figure is a year-over-year number and is not evidence that the quarter-on-quarter condition of more than 11.1% was met. Nor can the decline in share alone be taken to mean that CXMT's sales volume, total memory capacity sold, or profit decreased.
In the same Q3 research, Samsung's revenue share was 38%. That figure alone, though, does not establish that Samsung's production increase directly caused CXMT's share to fall. For companies buying memory, it is also difficult to read actual available supply from revenue share. Assessing supply capability requires checking which types of memory each company can ship, and how much.
HBM Is Growing, but Demand for Conventional DRAM Remains
Counterpoint notes that from Q4 2026, expanding shipments of high-priced HBM4 could put further pressure on CXMT's revenue share. This is an outlook on product mix, however, and does not identify the cause of the Q3 decline. When revenue from high-priced memory rises, it also lifts total market revenue, which can reduce the relative share of manufacturers focused on other products.
On the other hand, looking at AI memory demand through HBM alone overlooks the importance of the DRAM used on the CPU side. HBM is mainly used as high-bandwidth memory for GPUs, but AI systems also need large amounts of memory for CPUs. A concrete example is NVIDIA's Vera Rubin NVL72.
According to the Vera Rubin NVL72 specifications published by NVIDIA, one rack contains 72 Rubin GPUs and 36 Vera CPUs, each using a different type of memory.
| Component | Quantity | Memory type | Total rack capacity |
|---|---|---|---|
| Rubin GPU | 72 | HBM4 | 20.7TB |
| Vera CPU | 36 | LPDDR5X | Up to 54TB |
Note: Capacities are totals for each component within the same rack. The CPU-side capacity is a maximum.
In other words, within the same AI rack, HBM4 on the GPU side and LPDDR5X on the CPU side are used together. The two play different roles, and a larger capacity for one does not mean it can replace the other. The difference in installed capacity also does not allow a simple comparison of memory price, bandwidth, or business value.
Counterpoint's market research likewise found that HBM revenue nearly doubled, while DDR and LPDDR revenue grew even faster, lowering HBM's share of total DRAM revenue. This does not mean HBM demand has fallen. When the whole market is expanding rapidly, even a product with strongly growing revenue can see its relative share decline.
LPDDR5X, which CXMT mass-produces, is the same type of memory used in AI racks. But NVIDIA's specification table does not identify memory suppliers or qualification status, and does not mean CXMT products have been adopted. To judge actual demand wins, it is necessary to confirm not only that products meet the required specifications, but also customer qualification and adoption track record.
What It Takes for G5 Mass Production to Translate into Real Supply
On September 20, 2026, CXMT announced mass production of its new-generation G5 and two LPDDR5X products using 24Gb dies. The main targets are smartphones and portable consumer devices, so the announcement cannot be directly linked to expanded supply capacity for AI racks. The 24Gb figure also refers to the storage capacity of an individual die, not of an entire memory module.
Regarding G5 manufacturing technology, CXMT says the number of dies per wafer (DPW) has increased by more than 50% compared with the earlier G4. This figure needs to be interpreted carefully, however. According to a footnote to the company's announcement, the comparison is based on the gross number of dies before sorting good from bad, normalized to 8Gb equivalents.
Therefore, the announcement that G5's DPW increased by more than 50% does not mean the shipment volume of good dies rose by more than 50%. A higher gross die count per wafer leaves room for better production efficiency and larger supply. But the quantity that can actually be sold depends on post-manufacturing yield. Moreover, unless products are qualified and adopted by customers, higher manufacturing capability does not necessarily translate directly into revenue. The improvement of more than 50% is also a figure published by CXMT itself, not a measurement by an independent third party.
The start of mass production for a new product and a decline in company-wide revenue share are not necessarily contradictory. It can take time for mass-produced products to contribute to revenue, and if total market revenue expands further in the meantime, share can fall. The G5 announcement is evidence of progress in manufacturing technology, but it does not directly show the volume of good-die supply across all of CXMT's fabs or the extent of customer adoption.
If expanding HBM4 shipments change the revenue composition of the DRAM market going forward, it will become even harder to evaluate a manufacturer's competitiveness from revenue share alone. For CXMT as well, it is important to check shipment volume by product, good-die supply, and continued adoption by customers. Whether G5's manufacturing efficiency gains lead to stable LPDDR supply and wider adoption is a clue to the company's real strength that the 9% revenue share alone cannot reveal.