← Briefing archive

monthly · 2026-10-01 — 2026-10-31

Adoption expands while memory and production dependencies constrain scale

October strengthened evidence for domestic accelerator adoption, usable software, optical deliveries and capital formation. HBM yield uncertainty, material shortages and the gap between kernel measurements and full-system results still limit claims of a sovereign frontier stack.

Published 2026-11-01 · Evidence through 2026-10-31 · Completed period

Adoption is measurable, but shipment share is not performance parity

October reporting brought stronger evidence of domestic substitution. Reuters, citing IDC, reported Huawei at 28% of first-half Chinese AI-accelerator shipments versus Nvidia at 34%; these are January-June observations published in October, not October shipment growth. Product mix and export restrictions complicate comparisons. The result supports meaningful market adoption without measuring accepted deployments, utilization or frontier capability. Cambricon's unaudited nine-month revenue of RMB 9.526 billion, up 106.76%, adds commercial evidence, but revenue cannot be converted into accelerator counts or effective compute.

Cambricon · Cambricon 2026 third-quarter report (board dated October 23; disclosure dated October 24) (source)Reuters · Nvidia's China AI-chip share falls as Huawei closes gap, report finds (source)

Software adoption became inspectable; training and inference remain distinct

DeepSeek's pinned October DeepInfer release explicitly supports Huawei Ascend 950 alongside NVIDIA SM100. It supplies offline inference, while a complete service still needs KV caching, disaggregation and replica load balancing. Its published performance charts use NVIDIA hardware, so the release is evidence of an available Ascend software path rather than a matched hardware result. Separately, the Spark-Audio authors report a speech foundation model trained on domestic computing infrastructure. That is developer-reported completed training in a specific domain, not independently reproduced frontier pretraining efficiency. These developments strengthen the usefulness of domestic hardware while leaving software coverage, operational reliability and energy per completed workload unresolved.

DeepSeek-AI · DeepInfer initial release: supported platforms, service exclusions and benchmark hardware (source)iFlytek Spark Audio Team · Spark-Audio 1.0 Technical Report (v1) (source)

Memory progress does not yet remove the scale bottleneck

Reuters sources said CXMT had sent HBM3E samples to domestic chip designers and intended limited production within two months, with estimated yields of 20-30%; CXMT did not comment. This is a prospective, anonymously sourced ramp, not verified qualified volume. CXMT's October 31 quarterly filing separately reports sixth-generation DRAM engineering-sample validation. Conventional DRAM development and HBM stacking are different milestones: neither establishes sufficient accelerator memory, customer acceptance or competitive production economics. The memory constraint remains central to whether Chinese accelerator demand can translate into sustained output.

Reuters via Devdiscourse · Reuters relay: CXMT HBM3E production plans and yields (source)Reuters · CXMT plans limited HBM3E production despite low yields (source)CXMT / Sina filing mirror · CXMT 2026 third-quarter report (source)

Equipment scale increased while input constraints persisted

AMEC reports more than 9,300 cumulative reaction chambers in production lines by September 30. Chambers are not complete tools or wafer capacity, and the disclosure does not supply customer-level yield or uptime. NAURA's quarterly filing reports expanding orders but pressure on some deliveries from tight material supply; affected inputs and order values are unspecified. Together, these filings support growing domestic equipment participation while showing why tool sales cannot prove a replenishable domestic fab flow. Customer qualification, critical-component provenance and sustained process performance remain the decisive missing evidence.

AMEC / Sina filing mirror · AMEC reports more than 9,300 reaction chambers in production (source)NAURA / CNInfo · NAURA 2026 third-quarter report (source)

Optical demand and capital allocation extend the production opportunity

InnoLight's October 23 filing says major overseas customers placed orders through 2027, with separate demand guidance for 2028. Eoptolink says 800G and 1.6T are its principal delivery products and that 1.6T volume continues to expand. Orders, guidance and shipments must remain separate; neither statement quantifies qualified output or resolves component provenance. Capital also moved toward future capacity: Biren completed its October placement with approximately HK$4.0197 billion net proceeds, while JCET management described raising its annual investment plan from RMB 11.01 billion to RMB 18.2 billion, mainly adding advanced-packaging lines. Financing completion is observable; planned expenditure is not completed installation or accepted packaging capacity.

Shanghai Biren Technology / HKEX · Completion of placing of new H shares and amendments to articles (source)Eoptolink · Eoptolink: voluntary production and operations disclosure (2026-061) (source)InnoLight · InnoLight: voluntary production and operations disclosure (2026-100) (source)Securities Daily via Eastmoney · JCET adds advanced-packaging investment (source)

Benchmark evidence improved at kernel level, not yet at matched system level

SemiAnalysis published Ascend 950PR compute and memory microbenchmarks, including 1,055 BF16 GEMM shapes. Its own measurements must be distinguished from its analysis of a Huawei-supplied model trace. Kernel utilization and memory behavior are informative engineering evidence, but cannot establish matched full-model throughput, latency, training time-to-quality, energy or total cost. The reviewed October record establishes no new matched Chinese frontier training or inference comparison against contemporary foreign systems. This is a limit of verified coverage, not proof that no such experiment exists.

DeepSeek-AI · DeepInfer initial release: supported platforms, service exclusions and benchmark hardware (source)SemiAnalysis — Bryan Shan and Dylan Patel · The Huawei-DeepSeek Symbiosis: Ascend 950PR microbenchmarks (source)

Forecast implications and next milestones

No dated forecast changes. October supports the existing direction: useful domestic AI compute and optical networking advance faster than complete supply-chain independence. Confidence in further adoption is stronger than confidence in a timetable for frontier parity. Watch for customer-qualified HBM shipments and repeatable stack yields, accepted packaging capacity after investment, material-supply relief at equipment vendors, and independently reproducible large-system workload results with quality, device counts, software, latency, power and recovery disclosed. Financing announcements and engineering samples alone do not justify bringing manufacturing or HBM forecast dates forward.

Reuters · CXMT plans limited HBM3E production despite low yields (source)Securities Daily via Eastmoney · JCET adds advanced-packaging investment (source)NAURA / CNInfo · NAURA 2026 third-quarter report (source)SemiAnalysis — Bryan Shan and Dylan Patel · The Huawei-DeepSeek Symbiosis: Ascend 950PR microbenchmarks (source)