weekly · 2026-10-26 — 2026-11-01
Memory constraints and capacity investment frame domestic compute expansion
Reported memory sampling and MetaX training results broaden the evidence of domestic capability. Packaging investment and optical revenue show commercial scale, while HBM yields, material shortages and working capital constrain what can be inferred about sustainable output. No dated forecast changed.
Published 2026-11-02 · Evidence through 2026-11-01 · Completed period
Memory sampling advances while production economics remain uncertain
Reuters reported on October 28, citing unnamed sources, that CXMT had supplied HBM3E samples to domestic accelerator designers and expected limited production within two months. Reported yields were only 20–30%; CXMT did not respond. These are reported plans and estimates, not verified commercial output. Separately, CXMT’s October 30 investor call, published October 31, targeted first 4F2 DDR5 RDIMM samples by year-end, with production dependent on customer qualification and demand. Server DRAM and HBM are distinct products. Neither disclosure establishes qualified HBM volume, acceptable cost or a fully domestic manufacturing flow; earlier dated production claims remain preserved.
Reuters via Devdiscourse · Reuters relay: CXMT HBM3E production plans and yields (source)Reuters · CXMT plans limited HBM3E production despite low yields (source)CXMT (investor record via Sina) · CXMT October 30 investor call: 4F2 RDIMM sampling target (source)MetaX reports a larger training run, with narrowly scoped scaling evidence
MetaX’s October 26 announcement reports Pi0.5 supervised fine-tuning on RoboTwin with 1,024 C-series GPUs across 128 eight-GPU servers, completing 20,000 steps in 32 hours. Reported throughput was 2.75 samples per second per GPU, versus 2.8 at 512 GPUs. Its separate above-97% scaling claim concerns 8–64 GPUs running asynchronous PPO on ManiSkill, not the thousand-GPU SFT run. These are vendor measurements across different workloads. They strengthen evidence of usable domestic training infrastructure without establishing independent replication, frontier pretraining time-to-quality, matched cross-hardware efficiency or production reliability.
MetaX · MetaX announces RLinf upstream integration and thousand-GPU validation (source)Cambricon separates commercial growth from an unconfirmed shipment forecast
At its October 26 results discussion, Cambricon said it had never communicated the market’s million-chip shipment figure and provided no specific delivery schedule for Siyuan 690. Management said procurement reflects both customer orders and forecasts. Its reported nine-month inventory provision of RMB 501.6 million reinforces the need to distinguish stockpiling from accepted shipments. This clarification narrows what can be inferred from earlier growth disclosures: revenue, inventory and procurement commitments do not establish accelerator unit deliveries or customer utilization.
National Business Daily · Cambricon says it never communicated million-chip shipment figure (source)Packaging investment rises as equipment supply still constrains delivery
JCET’s October 27 results discussion, reported October 28, put nine-month fixed capital expenditure near RMB 8.9 billion. Its revised 2026 investment plan rises from RMB 11.01 billion to RMB 18.2 billion, including RMB 17.31 billion of fixed-asset investment; management says the increment principally targets advanced-packaging lines. The plan is consequential capacity investment, but is not itself customer-qualified output. NAURA’s October 30 unaudited filing separately reports growing orders and delivery pressure from shortages in some material categories. It does not identify those inputs’ provenance. Together these disclosures show demand and expansion alongside unresolved execution constraints, rather than proof of a self-sufficient production chain.
Securities Daily via Eastmoney · JCET adds advanced-packaging investment (source)NAURA / CNInfo · NAURA 2026 third-quarter report (source)Optical growth carries working-capital demands
October 30 unaudited filings put nine-month revenue at RMB 69.547 billion for InnoLight and RMB 31.850 billion for Eoptolink. InnoLight’s September-end inventory reached RMB 23.878 billion, attributed to materials prepared for orders. Eoptolink’s nine-month operating cash flow was RMB 1.464 billion, down 68.42% year on year. These are commercial and financing signals, not counts of shipped 800G/1.6T modules, production yields or evidence that foreign optical components have been replaced. The relevant next test is conversion of orders and inventory into accepted deliveries and cash, alongside component-level provenance.
Eoptolink / Sina filing mirror · Eoptolink 2026 third-quarter report (unaudited) (source)InnoLight / Sina filing mirror · InnoLight 2026 third-quarter report (unaudited) (source)Forecasts unchanged; publication-day checks remain separate
No dated Observatory forecast changes. The 2027–2030 HBM/advanced-packaging and late-2020s frontier-pretraining judgments remain intact. Prioritize qualified HBM deliveries and yields, actual RDIMM sampling, accepted packaging capacity, resolution of equipment-material shortages and reproducible training or inference comparisons with matched quality and operating conditions. Post-period context: the November 2 research check found no new matched contemporary Chinese frontier end-to-end comparison in the inspected MLPerf repositories and InferenceX coverage. That check was not an exhaustive artifact audit or a reconstruction of what each leaderboard displayed on November 1. Existing independent kernel evidence remains distinct from end-to-end system parity, and standardized submitter-run benchmarks are not automatically independent replications.
Reuters · CXMT plans limited HBM3E production despite low yields (source)SemiAnalysis · Continuous inference benchmark dashboard (source)Securities Daily via Eastmoney · JCET adds advanced-packaging investment (source)MLCommons · MLPerf Inference v6.1 closed submissions (source)MLCommons · MLPerf Training v6.0 submissions (source)