China’s CXMT will make AI memory despite yields as low as 20%, Reuters says

Reuters reported, citing two people familiar with the matter, that China’s CXMT plans to begin making HBM3E memory for AI chips, despite only 20%–30% of its chips being usable. The Next Web said it had not independently verified the report; no production timeline or market reaction was provided.

Analysis

The key variable is not whether CXMT can produce HBM3E, but whether it can convert output into qualified, economically usable stacks. If the reported 20–30% usable rate is accurate and measured at the relevant finished-product stage, the implied cost per usable unit could be several times that at full yield before accounting for packaging, qualification, and other costs. That makes near-term volume claims a weak proxy for competitive supply.

For established suppliers such as SK hynix, Micron, and Samsung, this is more a medium-term competitive signal than an immediate earnings threat. A credible Chinese alternative could eventually pressure pricing or customer allocation, particularly if domestic procurement supports demand despite weaker economics. Conversely, sustained low yields could consume capacity and investment without materially expanding marketable supply. HBM customers, including AI-chip designers, benefit only if CXMT clears qualification, reliability, and performance requirements—not merely if wafers or test units exist.

Near term, treat the report cautiously: it is attributed to unnamed sources and is not independently verified by the publisher. Over 1–3 months, watch for confirmed shipments, customer qualification, and evidence of usable output. Over 6–18 months, yield improvement, packaging capacity, and access to required equipment matter more than the initial launch. A meaningful deterioration in incumbent HBM pricing or guidance would falsify the view that this is currently a limited competitive threat.

Key Decisions for Investors

• No immediate trade on this report alone. Avoid treating a production-start headline as evidence of commercially meaningful HBM supply.
• Watch SK hynix, Micron, and Samsung for HBM pricing, customer allocation, and guidance changes; consider a relative-value position only if confirmed CXMT qualification coincides with evidence of incumbent pricing pressure.
• Set an alert for independently confirmed CXMT customer qualification, shipment scale, and yield progression. Without those data, treat the reported yield as a supply-chain risk indicator, not an earnings forecast.
• Reassess the cautious stance if CXMT demonstrates sustained usable yields and customer acceptance, or if established suppliers disclose lower pricing or lost allocations; reassess in the other direction if qualification or equipment constraints prevent scale.