(September 18, 2026, Semiconductor Industry Review)
Since the start of September, China's domestic semiconductor industry has sent two converging signals. On one side, seven Chinese AI chip designers — Cambricon, Hygon, Moore Threads, Muxi, Biren, Iluvatar CoreX, and Enflame — have released their 2026 half-year reports, combining for approximately RMB 21.46 billion in revenue, nearly double the same period last year. On the other, CXMT (ChangXin Memory Technologies), China's only mass-volume DRAM manufacturer, announced the global debut mass production of LPDDR6 memory and launched equipment tenders for a new Shanghai fab, targeting a long-term monthly capacity of 600,000 wafers. Against a backdrop of surging global AI compute demand and persistently tightening export controls, China's chip industry is entering a phase of accelerated delivery across design, manufacturing, and memory.
The earnings performance of China's AI chip designers has accelerated markedly over the past half year. Cambricon posted H1 revenue of RMB 5.996 billion, up 108.13% year-on-year, with net profit attributable to shareholders of RMB 2.311 billion. Hygon Information recorded revenue of RMB 9.099 billion, up 66.52%. Moore Threads generated RMB 1.736 billion in H1 revenue, exceeding its full-year 2025 total of RMB 1.506 billion. Biren Technology posted RMB 1.236 billion for the half year, also above its 2025 full-year figure, while Enflame Technology's RMB 1.120 billion exceeded its RMB 990 million for all of 2025. Iluvatar CoreX saw revenue grow 191.6%, with its inference product line reaching RMB 654 million, a 651.8% surge that overtook its training business for the first time. On profitability, Cambricon's gross margin was 55.25%, rising to 56.10% in Q2; Muxi's was roughly 57.22%; Moore Threads reached 56.95%, with its net loss narrowing to just RMB 11.56 million, on the verge of breakeven on a main-business basis. This indicates that domestic AI chips have not been forced into a price war; in a tight supply environment they retain meaningful pricing power.
Demand-side tightness has outperformed expectations. According to CCTV Finance, domestic compute chips are being "swamped by orders," with some products seeing delivery schedules pushed out to a year, and domestic compute demand running at more than ten times current supply. Multiple cloud service providers report two consecutive years of explosive order growth, with customer counts up roughly tenfold over the past year. In late August, Zhipu disclosed deployment details for GLM-5.3-Flash, its first native multimodal model in the GLM-5 series, noting that all online traffic is carried by 100,000 domestic chips. Ahead of launch, Zhipu anonymously tested the model on overseas platforms OpenRouter and OpenCode, accumulating 62T token calls entirely on domestic silicon. After underlying architecture optimization and inference service tuning, cluster hardware efficiency and per-token cost have reached parity with mainstream NVIDIA GPU solutions. These large-scale production cases signal that domestic chips have moved beyond the "can light up" phase into "can reliably serve," and customer evaluation has shifted from single-card performance to system-level usability, concurrent scheduling, chip-to-chip communication, fault recovery, and total cost.
Commercialization is simultaneously shifting from selling cards to delivering systems. Moore Threads' Ku'e intelligent computing clusters are deployed in Beijing, Wuxi, and Hangzhou; on September 9, JD Cloud announced a domestic 10,000-card cluster built with Moore Threads and partners, with plans to scale to a 100,000-card fully functional GPU cluster. Muxi's domestically-processed Xiyun C600 entered mass production in May, with the Xijing S600 super-node in parallel development, and its GPUs are already in commercial use at thousand-card cluster scale. Biren, together with Shanghai Yidian, Xizhi Technology, and ZTE, launched a commercial "Guangyue" optical-interconnect GPU super-node already deployed at thousands of cards. Hygon's DCU has been adapted to over 400 mainstream large models. Training clusters remain the source of large-ticket purchases, while inference — driven by search, recommendation, customer service, content generation, and enterprise agents — is emerging as a steadier, higher-frequency compute load, opening a second market for domestic chip companies beyond training.
Breakthroughs in memory are equally significant. On September 3, CXMT officially announced mass production of its self-developed LPDDR6 chips, debuting on the Xiaomi 18 Fold flagship foldable — the world's first LPDDR6 deployment on a flagship handset, filling a domestic gap in high-end mobile memory mass production. Xiaomi's chip team was the first to complete adaptation. On the capacity side, CXMT currently operates three 12-inch DRAM fabs in Hefei and Beijing, with monthly output above 200,000 wafers, expected to surpass 300,000 by end-2026, including 50,000 wafers per month dedicated to HBM. On September 11, CXMT officially launched equipment tenders for a new Shanghai fab in two phases, with the Shanghai site alone poised to exceed 100,000 wafers per month. Per its roadmap, a new Hefei fab ramps in H1 2027, monthly capacity reaches roughly 420,000 by 2027 and targets 500,000 by end-2028, with the long-term goal of two additional fabs bringing total output to 600,000 wafers per month and challenging Micron by shipment volume by 2030. SemiAnalysis projects CXMT will reach roughly 350,000 wafers per month by end-2026, near Micron's estimated 385,000, making it the world's third-largest DRAM maker by wafer capacity.
On the global market share front, CXMT's Q2 2026 DRAM revenue share touched 10%, up from less than 1% in 2023. As AI servers and HBM absorb capacity at the three Korean and Japanese majors, tight supply and rising prices have pushed domestic phone and PC makers to diversify away from Samsung, SK hynix, and Micron, making CXMT a critical alternative. Notably, CXMT is ramping DDR5 faster than the incumbents because it does not need to divert capacity to HBM, giving it a differentiated position in mainstream PC and mobile memory. Meanwhile, CXMT's STAR Market IPO is advancing, with public listing expected to fund continued capacity expansion and HBM technology development.
Upstream manufacturing and equipment are benefiting in parallel. SMIC posted H1 revenue of RMB 38.635 billion, up 19.44% year-on-year, and explicitly noted strong demand for AI-related companion chips, with shortages in data board logic, BCD, and optical-module driver ICs, leading the company to negotiate price increases on constrained products. AMEC (Zhongwei) reported H1 revenue of RMB 6.691 billion, up 34.89%; Naura reached RMB 20.161 billion, up 24.90%. ACM Research (Shengmei) saw new orders grow 105% year-on-year in H1 with shipments up 40%, across cleaning, furnace, PECVD, coater-developer, and panel-level advanced packaging lines. CITIC Construction Investment notes that pricing power in the semiconductor supply chain is structurally shifting from downstream chip devices toward equipment and components.
Risks warrant attention. First, while revenue is surging across domestic AI chip designers, divergence has begun — Biren and Enflame remain loss-making by hundreds of millions, with R&D expense ratios of 57–65%, and as capacity comes online and pricing rebalances, companies without cluster delivery and software ecosystems face shakeout. Second, the current tenfold demand-supply gap is unsustainable; as new capacity lands over the next two years, the balance may swing around 2028. Third, CXMT still lags Samsung, SK hynix, and Micron in high-end HBM. Fourth, tightening export controls — from Japanese photoresist restrictions to U.S. equipment and EDA limits — keep supply-chain uncertainty elevated. Fifth, some AI chip names have rallied sharply and face pullback risk if earnings disappoint.
Looking ahead, the MIIT's 15th Five-Year Plan for the ICT sector calls for China's intelligent compute capacity to grow more than fivefold from 2025 to 2030, with ordered deployment of 10,000- and 100,000-card intelligent clusters and increased adaptation of domestic compute chips. Driven by policy, demand, and capital, China's domestic AI chips and memory are at a critical window of moving from "usable" to "combat-proven." The contest is no longer about single-chip design alone but about system-wide capability across design, manufacturing, memory, packaging, equipment, and materials. The companies that can convert one-off projects into repeatable clusters and settle those clusters into stable industry workloads will be the ones that ride through this AI-driven semiconductor cycle.