CXMT (ChangXin Memory Technologies) put its LPDDR6 into production and says its commercial debut took place in the foldable Xiaomi 18 Fold, with the SoC Xring O3. The news matters because Chinese DRAM reached the start of a new mobile memory standard in the same cycle as the leaders. This does not make CXMT the leader of the entire memory industry, but it changes the conversation about supply chains and on-device AI.
This text starts from the video CHINA JUST WON THE NEXT-GEN RAM RACE, from the channel Eastern Engine, but was revised using technical confirmation from CXMT, the JEDEC standard, Counterpoint data, and SK hynix’s position. The video title is provocative; the facts below are more precise.
What LPDDR6 changes
The standard JESD209-6 was published in July 2025. LPDDR6 is not “laptop DDR6”: it is a low-power DRAM family, usually soldered into the device’s design. The relevant change is how the bus is organized to serve short, parallel, energy-sensitive workloads — exactly the profile of on-device AI.
- An x24 channel in two x12 subchannels: JEDEC splits each die into two 12-line data paths. This provides more concurrency without forcing the system to activate the entire interface for a small transfer.
- Granularity: the standard provides for 32- and 64-byte accesses. For workloads alternating among inference, camera, modem, and interface, transferring less useless data is as important as increasing peak throughput.
- Speed: the standard’s range reaches 14.400 MT/s. CXMT’s component is specified at up to 12.800 Mbps per pin; the company says that, with the 18 Fold’s SoC, it operates at 10.667 Mbps.
- Power and reliability: CXMT cites dual-rail DVFS, dynamic efficiency mode, Link ECC, and On-Die ECC, plus PRAC against Rowhammer and MBIST for self-testing. These are product and ecosystem features; actual savings depend on the device and workload.
At 12,8 Gbps, a board trace stops being a “wire” and becomes a transmission line: impedance, crosstalk, length, and PHY training begin to matter. Swapping the chip is not enough: the SoC, firmware, packaging, and board must have been designed for LPDDR6.
From promise to commercial debut
On August 29, the public news was that CXMT had begun mass production and would supply the 18 Fold. On September 7, the manufacturer itself published the product’s technical page and stated its commercial debut in the device. It describes a package of 16 GB, made up of dies of 16 Gb, in PoP FBGA packaging with 1.295 balls.
The 12.800 Mbps figure should not be read as the guaranteed speed of any phone. CXMT separates the component’s maximum from the 10.667 Mbps rate used with the SoC, and the gains of 60% in bandwidth and 20% lower power consumption are internal measurements compared with LPDDR5X. The useful takeaway for a news reader is this: there is a commercial product; the percentages require independent tests on the final device.
The package also reveals why manufacturing is not trivial. CXMT says the ball area of the 1.295-pin FBGA is 50% larger than that of the 496-ball FBGA used as a reference for LPDDR5, and that it uses High-K EMC molding to improve heat dissipation. This is an integration and yield detail, not just a frequency detail.
CXMT + Xiaomi 18 Fold: what is confirmed
The relevant case is the pairing of a Chinese SoC + Chinese DRAM in the same high-end product: Xring O3 and CXMT LPDDR6. CXMT’s confirmation is stronger than a supply-chain leak because it includes capacity, data rate, and packaging. Reuters also reported the start of production and supply for the foldable.
There are clear limits to interpreting the news. The manufacturer has not published yield, wafer volume, distribution across all phone configurations, or independent battery life and sustained performance results. “Mass production” proves the product has left the sampling stage; by itself, it does not reveal scale equivalent to a global smartphone line with tens of millions of units.
The race in numbers
CXMT was founded in 2016 and entered the DRAM market later than Samsung, SK hynix, and Micron. Even so, Counterpoint ranked it fourth in Q2 of 2026. The table measures global DRAM revenue, not installed capacity, unit volume, or LPDDR6-specific share:
| Manufacturer | Share of global DRAM revenue — Q2/2026 |
|---|---|
| Samsung | 38% |
| SK hynix | 25% |
| Micron | 24% |
| CXMT | 10% |
| Nanya | 2% |
| Others | 1% |
CXMT jumped from 4% in Q2 of 2025 to 10% in Q2 of 2026, according to the same series. This shows commercial progress, but is not enough to establish technological leadership across all memory classes.
In the direct LPDDR6 comparison, SK hynix reported in March the development of a 16 Gb die on the 1c process, above 10,7 Gbps, and planned supply in the second half of the year. CXMT’s news is that it documented a commercial smartphone debut; rivals’ timelines and product maturity must be compared by product, capacity, rate, power consumption, and volume — not just the first announcement.
Did China win the race?
It is not accurate to summarize this as “China won RAM.”. What the confirmation demonstrates is an early commercial LPDDR6 debut by a Chinese company and a domestic alternative for a premium smartphone. It is a product execution and supply-chain victory — not a single measure of DRAM leadership.
What remains open:
- Scale and yield: CXMT has not published yield, wafer volume, or field return rates for LPDDR6.
- Fair comparison: peak Mbps, capacity per package, power consumption, and price must be compared on the same device and in the same configuration. “20% less power” is an internal CXMT figure.
- Different memories, different markets: mobile LPDDR6 does not measure, for example, the ability to produce stacked HBM for AI accelerators. A conclusion from one segment should not be transferred to another.
- Ecosystem: being the first commercial product does not mean a monopoly on the standard. SK hynix has already announced its LPDDR6, and other suppliers have their own roadmaps.
Beyond phones — and what this changes for Linux
CXMT cites smartphones, PCs, wearables, vehicles, edge servers, and robotics as potential applications. This breadth comes from the combination of bandwidth and power consumption, but adoption depends on compatible controllers, modules, and board designs.
- AI PCs and unified-memory SoCs: more bandwidth helps feed NPUs and integrated GPUs. But LPDDR is generally soldered: it is not a DIMM upgrade you buy later.
- Edge and robotics: the profile of local data, batteries, and thermals favors low-power memory, provided firmware and the platform support the standard.
- Automotive: the announced RAS features — ECC, PRAC, and self-testing — are relevant, but automotive product qualification requires industry-specific criteria.
- Local LLMs: with Ollama on Linux, bandwidth influences speed when the model uses shared RAM; however, capacity still determines whether it fits. LPDDR6 does not replace dedicated VRAM or make a current laptop compatible through software.
For PC users, the practical conclusion is simple: compare RAM capacity, platform bandwidth, and expandability before looking only at “LPDDR6” on the spec sheet.
References
- CXMT — LPDDR6 enters production and debuts in the Xiaomi 18 Fold (7 Sep. 2026)
- CXMT — LPDDR6 product page
- Reuters — CXMT supplies LPDDR6 for the Xiaomi 18 Fold
- Counterpoint — DRAM revenue share in Q2 of 2026
- JEDEC — LPDDR6 (JESD209-6)
- SK hynix — LPDDR6 1c, 16 Gb
- Video: Eastern Engine — CHINA JUST WON THE NEXT-GEN RAM RACE
The new information is not just “12,8 Gbps”: it is that CXMT now documents a commercial LPDDR6 product, with specifications and a first device. The next relevant test is not a technology-race slogan, but scale, yield, price, and independent measurements of power consumption and performance.