China’s CXMT plans HBM3E production despite low yields, Reuters reports

China’s memory maker CXMT plans to begin limited production of HBM3E, a key memory component for AI processors, within the next two months, Reuters reports. But reported yields of just 20% to 30% show how difficult it remains to challenge established suppliers.

Watch Desk analysis

What happened

CXMT has sent HBM3E samples to several domestic AI chip designers, according to two people familiar with the plans cited by Reuters. The company is preparing two versions, including one with changes intended to improve yields. Reuters says tests of the chips put their performance closer to SK Hynix’s previous-generation HBM3 than to its HBM3E.

CXMT did not respond to Reuters’ request for comment. SK Hynix did not immediately respond to a request for comment on the yield estimate. Reuters says it is reporting the production timeline and yield details for the first time. Read the report.

Key findings

• Production is planned, not yet a proven rival at scale
  Reuters reports limited production is expected to start within two months, with samples already sent to domestic AI chip designers.
• Reported yields are well behind the leader
  Sources told Reuters CXMT’s HBM3E yield is around 20% to 30%, compared with roughly 70% for SK Hynix.
• The performance gap remains
  One source told Reuters that CXMT’s HBM3E performed closer to SK Hynix’s older HBM3 than to its HBM3E.

Why it matters

High-bandwidth memory is essential to many AI processors used to train and run large models. More Chinese supply could help domestic chip makers facing shortages, but low yields make production costly: more chips have to be discarded or reworked. Reuters reports that CXMT’s HBM3E is expected to cost more than comparable products from overseas rivals, so a local source is not automatically a cheap one.

Our read

This is a meaningful attempt to build a domestic link in China’s AI-chip supply chain, not evidence that CXMT has caught the leaders. The reported yield and performance gaps are substantial, and production plans are not the same as dependable volume. For AI hardware buyers, the useful question is whether those gaps narrow as CXMT moves from samples to production.

What to watch

• Whether limited production begins on the reported timeline.
• Whether the yield-improved version reaches customers and improves usable output.
• How CXMT’s HBM3E performs in later tests against current products from established suppliers.

Discussion spark: If domestic AI-chip supply is strategically important, how much higher a cost and how large a performance gap should buyers tolerate to build it?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.