← Briefing archive

weekly · 2026-10-12 — 2026-10-18

Adoption and earnings scale while sovereignty proof stays incomplete

Third-party shipment data and preliminary company forecasts strengthened the evidence for domestic accelerator adoption and commercial scale. Investment and acquisition activity also targeted hybrid-bonding equipment, but closing, production qualification and domestic-component evidence remain incomplete. YMTC's reported DRAM work stayed at research and testing scale, and no matched current frontier accelerator comparison or dated forecast change was established.

Published 2026-10-19 · Evidence through 2026-10-18 · Completed period

Domestic accelerators gained adoption, not demonstrated performance parity

Reuters reported IDC data showing 2.008 million AI accelerators shipped in China in the first half of 2026. Huawei accounted for 572,000 units and 28% share, while T-Head held 7%, Cambricon 5% and Kunlunxin 4%. Non-GPU accelerator shipments rose nearly 80% to 1.016 million units and exceeded GPU shipments by volume; the four domestic vendors represented 95% of that segment. These figures establish meaningful market adoption, but not the product mix, buyers, accepted deliveries, utilization, workload performance, prices, memory or fabrication provenance. Export controls also constrain Nvidia's accessible China product mix, so shipment share is not a latest-versus-latest performance comparison.

Reuters · Nvidia's China AI-chip share falls as Huawei closes gap, report finds (source)

CXMT and MetaX forecasts strengthened commercial-scale evidence

CXMT forecast January-September revenue of RMB 265.0-270.0 billion and attributable net profit of RMB 131.0-132.5 billion, versus RMB 32.084 billion of revenue and a RMB 5.280 billion loss a year earlier. It cited higher DRAM prices, high utilization, higher output per wafer, scale and product mix. MetaX forecast revenue of RMB 2.541-3.106 billion and attributable net profit of RMB 586.8-717.2 million, versus a RMB 345.5 million loss, while saying supply-chain investments also contributed to the change. Both forecasts were preliminary and unaudited. Neither filing supplied the unit, customer, yield, workload, HBM-volume or component-provenance evidence needed to translate financial scale into current production or performance conclusions.

CXMT Corporation · CXMT 2026 first-three-quarter earnings forecast (source)MetaX Integrated Circuits (Shanghai) Co., Ltd. · MetaX 2026 first-three-quarter earnings forecast (source)

Hybrid-bonding transactions target capability before proving production use

NAURA's subsidiary signed an agreement to acquire hybrid-bonding equipment developer Nuohe Jianwei for RMB 404.6 million, subject to closing conditions. AMEC's board separately approved up to RMB 35 million of additional investment in Qinghe Jingyuan within a RMB 250 million round intended partly for hybrid-bonding and atomic-beam equipment research, market development and customer delivery; that investment agreement had not been signed. The moves could broaden domestic advanced-packaging capability, but reported validation, shipments and repeat orders do not establish named-customer acceptance, production use, throughput, yield, uptime, reliability or domestic component provenance.

AMEC · Announcement on external investment and related-party transaction (source)NAURA Technology Group Co., Ltd. · NAURA announcement on acquisition of Nuohe Jianwei (source)

YMTC's DRAM effort remained research rather than commercial output

Counterpoint reported that YMTC was researching and testing a second-generation 16 Gb LPDDR5 device at roughly 5,000 wafers per month and had no plan to mass-produce that version. It described a faster 9,600 MT/s version as a development target for Fab 3. This is process-learning evidence, not qualified output, yield, customer acceptance, HBM capability, equipment provenance or a production date. Counterpoint also cited a current 16.5% NAND shipment share without publicly reconciling its basis with its earlier 14% second-quarter estimate, so the Observatory retains both figures with the discrepancy disclosed.

Counterpoint Research · YMTC files for IPO after climbing to No. 3 in global NAND race, with DRAM next (source)

Benchmark evidence and dated forecasts did not move

A period-end review found no matched current Chinese accelerator result against Rubin, GB300, MI355X, TPU v7 or OpenAI Jalapeño in MLPerf Training v6.0, MLPerf Inference v6.1 or InferenceX's published coverage. MLPerf results are standardized but submitter-run, and individual result artifacts were not exhaustively audited. Market share, earnings forecasts and pre-production memory research do not change the existing late-2020s frontier-pretraining view or the 2027-2030 assessment for HBM and advanced packaging. No dated forecast changed.

Reuters · Nvidia's China AI-chip share falls as Huawei closes gap, report finds (source)Counterpoint Research · YMTC files for IPO after climbing to No. 3 in global NAND race, with DRAM next (source)CXMT Corporation · CXMT 2026 first-three-quarter earnings forecast (source)SemiAnalysis · Continuous inference benchmark dashboard (source)MLCommons · MLPerf Inference v6.1 closed submissions (source)MLCommons · MLPerf Training v6.0 submissions (source)

Next evidence to watch

Prioritize product-level and accepted-delivery breakdowns behind IDC's shipment totals; audited CXMT and MetaX results with unit, customer, yield and product-mix detail; completion of NAURA's acquisition and AMEC's investment; production qualification and operating metrics for hybrid-bonding tools; YMTC DRAM qualification, yield and volume-production decisions; HBM and fabrication provenance; and matched training time-to-quality or inference throughput, latency, energy and cost measurements on current Chinese and frontier foreign systems.