ChangXin Memory Technologies has won a roughly $3 billion memory supply deal with Tencent, locking in one of China's largest AI customers for a domestic DRAM maker still working to close the gap with Samsung, SK Hynix, and Micron. The contract, reported by Reuters citing sources familiar with the agreement, is the largest known commercial deal for CXMT to date and marks a clear shift in how China's largest cloud operators are sourcing the memory that AI workloads consume in bulk.
Tencent runs one of China's largest cloud and AI infrastructure footprints, and DRAM is a critical input for the GPU servers that train and serve large models. A $3 billion commitment to a domestic supplier signals that Tencent is prepared to absorb the engineering trade-offs of buying Chinese memory in exchange for supply security as US export controls continue to tighten around high-bandwidth memory and advanced chips.
CXMT, based in Hefei, is the leading mainland Chinese DRAM producer and a central piece of Beijing's industrial push to localize semiconductor supply. The company has been ramping production of standard DDR memory and is working toward higher-margin products including high-bandwidth memory, the variant stacked directly alongside AI accelerators. Until now, CXMT's customer base has been dominated by smartphone and PC OEMs rather than hyperscale AI buyers.
Key facts
- 01CXMT signed a roughly $3 billion memory supply deal with Tencent, the largest known commercial contract for the Chinese DRAM maker.
- 02The agreement gives Tencent a domestic alternative to Samsung, SK Hynix, and Micron for AI server memory.
- 03CXMT is China's leading DRAM producer and a central piece of Beijing's drive to localize memory supply.
- 04Tencent operates one of China's largest cloud and AI infrastructure footprints, where DRAM is a critical input.
The Tencent deal changes that profile. A multibillion-dollar AI infrastructure customer gives CXMT both the revenue scale to fund its next process nodes and a reference design partner with real workload pressure on memory bandwidth and latency. That is the same flywheel SK Hynix used to dominate HBM supply for Nvidia: a single anchor customer whose volume justified aggressive capex.
For Tencent, the calculation is harder to read from the outside but easy to infer. The company has been expanding its Hunyuan model family and its cloud AI services, and like every other large Chinese AI operator it is navigating a market where the most advanced foreign memory and accelerators are either restricted, allocated to other buyers first, or both. Buying $3 billion of DRAM from a domestic source removes one variable from a planning problem that has gotten more complicated each quarter.
The deal also fits a broader pattern across China's AI industry. Baidu's chip unit Kunlunxin is preparing a Hong Kong listing reportedly targeting a $50 billion valuation, Z.ai and Sakana-backed efforts have been shipping models pitched as alternatives to restricted US systems, and domestic GPU makers continue to win allocation from cloud buyers that would have defaulted to Nvidia two years ago. Memory is the missing layer in that stack, and Tencent's commitment to CXMT begins to fill it.
Samsung, SK Hynix, and Micron still hold the technology lead in DRAM density, yield, and HBM stacking, and CXMT is not yet a one-for-one substitute on the highest-end AI accelerators. The bulk of the Tencent contract is likely standard server DRAM rather than HBM, which would mean CXMT is supplying the memory around the AI accelerator rather than the memory bonded onto it. That is still a meaningful share of a data center's bill of materials.
Neither CXMT nor Tencent has publicly confirmed the deal, and the financial terms, delivery schedule, and product mix have not been disclosed. The reported figure should be read as a contracted supply commitment over multiple years rather than a single year of revenue, though even amortized it represents a step-change in CXMT's commercial profile.
The memory supply chain has become the quietest but most consequential battleground in AI infrastructure. Accelerator headlines get the attention, but every large training run is constrained by how much memory sits next to the compute and how fast it moves data. A Chinese hyperscaler paying $3 billion to a Chinese memory maker is the kind of contract that reshapes which suppliers get to fund the next generation of fabs — and which ones get squeezed out of the customers they assumed they would keep.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
