News · 2026-08-08

China's biggest memory maker is booked through 2027

China's largest memory maker has reportedly sold out its DRAM production through the end of 2027. DIGITIMES reported on 17 July 2026 that PC brands rushed to lock in supply from ChangXin Memory Technologies, with shipments "booked through the end of 2027." Nearly a month on, consumer memory prices have not come back down -- and the memory in question is not the exotic stuff bolted to AI accelerators, but the ordinary kind that goes in every laptop, desktop and phone.

Key facts

• Reported by Aaron Lee for DIGITIMES on 17 July 2026: CXMT shipments "booked through the end of 2027."
• The product is mainstream DRAM -- DDR5, DDR4, LPDDR5X, LPDDR4X -- not the high-bandwidth memory used on AI accelerators.
• A tracked DDR5-6400 CL32 32GB kit currently shows a $390 minimum, up 52.9% over 60 days and roughly flat over the last 30.
• Primary sources: the DIGITIMES report, CXMT's product pages, and Shanghai Stock Exchange coverage of its IPO.

The distinction between DRAM and HBM is the whole story, so it is worth being precise. High-bandwidth memory is stacked in towers next to an AI accelerator and sold almost entirely into data centres; when it is tight, the pain lands on cloud providers. DRAM is the flat sticks in the slots on a motherboard. CXMT builds the latter, and the Shanghai exchange's coverage of the company's IPO describes deep cooperation with Alibaba Cloud, ByteDance, Tencent, Lenovo, Xiaomi, Transsion, HONOR, OPPO and vivo -- a customer list spanning cloud, PCs and phones rather than accelerators.

So this is not a data-centre story that trickles down. It is a story about the memory in the machine on your desk, and the reason it hits AI users directly is a shift in how people run large models. As we covered when a 284-billion-parameter rig turned out to be 768GB of server RAM and when Kimi K3 was shown running in 8GB at 33 seconds per token, the community's answer to enormous open-weight models has been to stop trying to fit them on a graphics card. Mixture-of-experts models activate only a fraction of their weights per token, so the working set can be small even when the model is huge -- provided you have somewhere to keep the rest. That somewhere is system RAM, or an SSD. Which means the price of a capable local AI machine is now set by a DRAM allocation decision made in Hefei, not by a GPU launch.

The market data supports the squeeze without needing the booking claim to be exact. TrendForce's spot pricing showed mainstream DDR4 8Gb averages climbing through early July, and a live US retail tracker for a DDR5-6400 CL32 32GB kit shows the minimum price at $390 -- up nearly 53 percent over 60 days and essentially flat over the last 30. That flatness is the informative part: the spike did not spike and settle. It rose and stayed. Consumer memory has been pinned near its highs for weeks, which is what an allocation shortage looks like from the buyer's end.

What is not verified deserves equal billing. The accessible primary record does not disclose wafer counts, gigabyte commitments, or which OEMs booked what. Write-ups naming specific PC brands are downstream of the original report, not independent confirmations of it. And "booked through 2027" is a capacity-allocation statement from supply-chain sources, not an audited order book. The defensible version is narrower than the headline: CXMT's DRAM output was reported as committed well into 2027, and the customer-by-customer detail is not public.

The community's response is the same one it always reaches for when hardware gets expensive: extract more from what you already own. That means quantization to shrink the weights, expert-streaming setups that keep a 26-billion-parameter model in 2GB by paging experts off the SSD, and multi-GPU tuning to squeeze bandwidth out of consumer boards. The ceiling on all of it is real, though. llama.cpp's own multi-GPU documentation warns that peer-to-peer transfer support is generally restricted to workstation and datacentre cards and can produce crashes or corrupted output on consumer hardware. Software tricks buy time. They do not manufacture DRAM.

The honest caveat: DIGITIMES is a subscription trade publication and the full article is paywalled, so the booking claim rests on the accessible summary plus the surrounding evidence -- CXMT's own product positioning, its IPO disclosures, and a retail price curve that has refused to fall. Those corroborate the direction well. They do not independently confirm the end-2027 horizon. If you are budgeting a machine to run local models next year, the practical read is that memory, not compute, is the binding constraint, and that it is likely to stay that way for several quarters.

Primary source, verified: read the paper →

Ground Truth. (2026, August 8). China's biggest memory maker is booked through 2027. Ground Truth. https://groundtruth.day/news/chinas-biggest-memory-maker-is-booked-through-2027.html

@misc{groundtruth:chinas-biggest-memory-maker-is-booked-through-2027,
  title  = {China's biggest memory maker is booked through 2027},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {aug},
  url    = {https://groundtruth.day/news/chinas-biggest-memory-maker-is-booked-through-2027.html}
}

Topics: hardware · memory · supply-chain · china · local-ai