← Briefing archive

weekly · 2026-10-19 — 2026-10-25

Software and operating evidence deepen; frontier efficiency remains unproven

Ascend software and independent kernel evidence became more inspectable, while a specialist model developer reported completed domestic-hardware training. Cambricon’s filing and optical suppliers’ disclosures strengthened commercial evidence, with working-capital, delivery and dependency limits intact. No dated forecast changed.

Published 2026-10-26 · Evidence through 2026-10-25 · Completed period

Inspectable software and independent kernels strengthen the Ascend evidence

DeepSeek’s October 20 DeepInfer release supports V4.1-Flash offline inference on Ascend 950 and NVIDIA SM100, with CANN 9.2 required for Ascend. Complete serving infrastructure is excluded, and the published performance plots use NVIDIA hardware: four SM100 GPUs for prefill and 64 for decode. Separately, SemiAnalysis’s October 21 study measures 1,055 BF16 GEMM shapes on Ascend 950PR and memory behavior. Its independent kernel measurements must be distinguished from its analysis of a Huawei-supplied model trace. High kernel utilization does not establish full-model throughput, latency, training time-to-quality or operating cost. Together these sources make software support and kernel behavior more inspectable without establishing a matched frontier end-to-end result.

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)

Completed specialist training and prospective frontier adoption remain distinct

iFlytek’s October 22 Spark-Audio paper reports using Ascend 910B and Hygon BW1000 throughout its training pipeline. It combines a 0.65-billion-parameter audio encoder with a 30B-A3B language model. This is developer evidence of completed specialist training; matched hardware efficiency, reliability and supply-chain provenance are not established. Separately, October 19 reporting relayed by Investing.com from The Information describes planned DeepSeek training on Ascend 950DT from late 2026 and preferential 960DT supply. The accessible relay is not independent corroboration; the original article was paywalled. Plans cannot be counted as completed frontier training.

The Information · DeepSeek–Huawei chip-design reporting (source)Investing.com, relaying The Information · DeepSeek teams up with Huawei to challenge Nvidia (source)iFlytek Spark Audio Team · Spark-Audio 1.0 Technical Report (v1) (source)

Cambricon’s growth also requires substantial working capital

Cambricon’s October 24 unaudited filing reports Q3 revenue of RMB 3.531 billion, up 104.47%, and attributable profit of RMB 1.404 billion, up 147.74%. Q3 operating cash flow was negative RMB 83.68 million. September-end inventory reached RMB 11.665 billion, compared with RMB 4.944 billion at December 2025; prepayments increased to RMB 3.044 billion from RMB 744.54 million. The commercial expansion is consequential, but these accounting values do not reveal shipped accelerator units, product mix, customer utilization, qualified supply or HBM provenance. The report covers an earlier operating period but became public during this week.

Cambricon · Cambricon 2026 third-quarter report (board dated October 23; disclosure dated October 24) (source)

Optical disclosures extend visibility while leaving quantities unknown

Following October 22 operating assurances, October 23 primary filings provide more specific company-attributed evidence. InnoLight says major overseas customers have placed orders through 2027, while some key customers have provided separate 2028 demand guidance. It reports co-development and sampling on unspecified next-generation products. Eoptolink says 800G and 1.6T are its principal delivery products and that 1.6T shipments are continuing to ramp. Orders, guidance and samples are different stages; neither filing establishes future deliveries, shipment quantities, production yields or freedom from foreign-component dependencies.

Eoptolink · Eoptolink: voluntary production and operations disclosure (2026-061) (source)InnoLight · InnoLight: voluntary production and operations disclosure (2026-100) (source)CNR / reporter Niu Guyue · InnoLight and Eoptolink respond on operating conditions (source)

Cleaning-tool validation advances a process step, not an entire domestic flow

ACM Research, Inc. announced on October 21 that its 170°C single-wafer hot-SPM system had been validated across more than 50 process layers at an unnamed foundry and achieved 70% sulfuric-acid recycling. Roughly 400 tons of annual acid savings per tool remains an expectation. This US-parent disclosure is relevant to the Shanghai equipment profile, but does not separately verify a Shanghai shipment, named-customer production acceptance, tool count, node, yield or independent operating economics. It supports a narrower process-capability claim rather than proof of a replenishable domestic fabrication flow.

ACM Research, Inc. · ACM Research enhances single-wafer hot SPM system with sulfuric acid recycle capability (source)

Forecasts remain unchanged; the next tests concern operating results

No dated Observatory forecast changes. The week strengthens the case for useful domestic compute and optical demand, while leaving frontier training efficiency and supply-chain independence unresolved. The archive’s period-end benchmark review and the publication-day check did not verify a matched current Chinese-versus-frontier end-to-end comparison; existing independent Ascend kernel results remain explicitly recognized. MLPerf is standardized submitter-run evidence, not automatically an independent replication. Prioritize completed DeepSeek training with quality, elapsed time, recovery and energy data; reproducible cross-hardware inference under matched service conditions; accepted deliveries and HBM provenance; optical shipment quantities and yields; and foundry-confirmed tool economics. The late-2020s frontier-pretraining and 2027–2030 HBM/advanced-packaging judgments are preserved.

DeepSeek-AI · DeepInfer initial release: supported platforms, service exclusions and benchmark hardware (source)SemiAnalysis · Continuous inference benchmark dashboard (source)MLCommons · MLPerf Inference v6.1 closed submissions (source)MLCommons · MLPerf Training v6.0 submissions (source)SemiAnalysis — Bryan Shan and Dylan Patel · The Huawei-DeepSeek Symbiosis: Ascend 950PR microbenchmarks (source)