Chips & Export Controls

Commerce Ministry Announcement No. 48 makes a filing mandatory, effective immediately, before a Chinese company licenses restricted model weights or an AI chip design abroad. The draft export catalogue, open for comment until 15 November, restricts weights of models trained with at least 10^26 FLOPs or with at least 2 trillion parameters. The ministry's Q&A puts API, app, web and SaaS access outside its scope, so the rule governs who holds the weights, not who calls the model.

Beijing ESWIN Computing started trading in Hong Kong on 9 October 2026 under the code 01256 and closed at HK$1.535, below the HK$1.55 offer price at which it had sold the shares the night before. It opened at HK$1.540, reached HK$1.640 at the top of the session and HK$1.365 at the bottom, 11.9 percent under the offer, on turnover of HK$260.0 million across 172,184,400 shares. TechNode has the listing. The exchange filing is the more useful reading: ESWIN sold 1,617,310,000 H shares including a partial upsize of 47,106,000 shares, 3.00 percent of what was originally on offer, for gross proceeds of HK$2,506.8 million and HK$2,385.7 million net of HK$121.2 million of listing expenses. The Hong Kong retail tranche drew 18,776 valid applications of which 7,649 got stock, was 9.09 times covered and was 9.71 percent of the deal; the international offering took the other 90.29 percent and was 1.95 times covered across 45 placees. Eleven cornerstone investors had committed to 749,354,000 shares before the book opened, 46.33 percent of the entire global offering, 7.34 percent of the H shares in issue and 3.40 percent of the company, with E-Town Shenghai Fund taking 224,382,000 shares on its own, 13.87 percent of the offer. Two of the eleven are close associates of existing shareholders and the exchange waived strict compliance with Listing Rule 10.04 to let them participate. ESWIN was founded in September 2019 by Wang Dongsheng, 69, who led the founding of BOE in 1993 and ran it as chairman and chief executive until 2016 and as chairman until 2019. It qualified under Rule 8.05(3), the market capitalisation and revenue test, having no profit record: the prospectus puts revenue at 1,752.2 million yuan in 2023, 2,025.4 million in 2024 and 2,431.2 million in 2025, against losses of 1,837.5 million, 1,547.4 million and 1,515.8 million yuan in those same three years. Market capitalisation at the close was HK$33.83 billion, about 28.85 billion yuan or 4.31 billion dollars. Until this year ESWIN was mostly a screen company by its own account, and human-machine interaction chips for screens were 1,855.7 million yuan of 2025 revenue, 76.3 percent, against 319.8 million for computing chips, 13.2 percent, and 41.7 million for IP module licensing, 1.7 percent. In the first quarter of 2026 the mix inverted: screen chips fell to 227.6 million yuan, 46.0 percent of revenue, from 331.9 million in the same quarter of 2025, while computing chips went to 223.0 million, 45.1 percent, from 44.7 million. The prospectus is plain about the method, saying it is leveraging a price-for-volume market strategy to promote the adoption of RISC-V within the industry, a strategy it had already adopted in 2023 to cope with weak consumer electronics demand. Gross margin on computing chip products was 26.2 percent for 2025 and 13.4 percent in the first quarter of 2026 as that line grew five-fold, and group gross margin went from 18.6 percent to 14.8 percent over the same step. R&D spending was 1,444.7 million yuan in 2023, 1,337.0 million in 2024 and 1,041.6 million in 2025, down 27.9 percent while revenue rose 38.8 percent, so the entire 321.6 million yuan improvement in the annual loss came out of the research line rather than out of gross profit. The timing was not a choice: ESWIN closed 2025 with 188.3 million yuan of cash in hand against 1,171.0 million yuan burned in operations that year, about two months of runway, and the first quarter of 2026 was rescued by 631.5 million yuan of financing inflows. RISC-V here answers the architecture question and not the export-licensing one, and the two get run together often. At HK$1.535 the public market is paying 11.9 times last year's revenue for a chip designer whose revenue is still roughly half screens and whose research spending is lower than it was three years ago.

Metastone (是石科技), a Chinese compute operator that describes itself as an independent full-stack operator unbound to any model vendor, published three numbers for the same eight-card PCIe machine with no hardware changed: DeepSeek-V4.1-Flash input throughput of 1,932 tokens per second on the community default deployment, 5,850 after a first adaptation pass, and 13,274 after a second, a 6.87-fold lift with a million-token context still supported. The mechanism is unglamorous and that is the interesting part. A card outside a framework’s validation matrix does not throw an error; it quietly takes a slow generic path, so the first pass was repair work, aligning metadata definitions, correcting page size, restoring a sparse-MLA prefill fast path, replacing a wrong FP8 dense GEMM kernel and widening the threshold for fast-path PCIe inter-process communication. The second pass fused attention and projection operators, overlapped computation with collective communication, and gave prefill and decode separate parallel configurations instead of one global setting inherited from hardware with NVLink. The same treatment took DeepSeek-V4-Flash from 14,546 to 22,584 tokens per second and GLM5.3 from 3,236.78 to 6,222.72, where P95 first-token latency fell from 141.6 seconds to 46.6 and the context ceiling went from 270,000 tokens to roughly 1.05 million. The headline claim that 1.5 units of the RTX 6000D beat one B300 holds only for video generation with cache reuse and a Turbo LoRA on top. On text inference the same document reports B300 input throughput at about 4.9 to 5.6 times that entire eight-card machine on DeepSeek-V4-Flash and 3.5 to 4.7 times on GLM-5.3. This is also a vendor benchmark, published with authorisation and with all four charts credited to Metastone, which the issue says plainly rather than presenting it as independent.

SJSemi (盛合晶微, 688820.SH) is the only company in China doing 2.5D advanced packaging at mass-production scale and holds about 85 percent of that domestic market, and it is running its 2.5D lines at roughly 60 percent utilisation with an order book it describes as full. Its own prospectus puts 2.5D capacity utilisation at 63.42 percent in the first half of 2025, and revenue has been flat enough since to imply about the same now. The reason is that a 2.5D package is a logic die sitting beside stacks of HBM on a silicon interposer, and SJSemi makes neither input: HBM comes from SK Hynix, Samsung and Micron, and the large low-warpage through-silicon-via interposers come from TSMC, Samsung and Intel. LatePost puts industry expectations for HBM tightness at 2028 and for high-end interposers later still. The prospectus line that carries the argument is that customer-supplied interposer material went from 3 percent in 2023 to about 25 percent in the first half of 2025, which is what happens when customers stop trusting a supplier to source a component. The filed numbers underneath the story are worse than the headline: half-year net profit of 449.4 million yuan minus a first quarter of 191.3 million leaves 258.0 million for the second quarter against 308.6 million a year earlier, a fall of 16.4 percent on revenue up 1.9 percent, while the half-year reads plus 7.2 and plus 3.3 percent. The market is paying about 245.25 billion yuan, roughly 36.5 billion dollars at Friday’s 6.7114 onshore rate, or 262 times the 935.0 million yuan earned in the twelve months to June. Customer A is 74.40 percent of sales, up from 40.56 percent in 2022; the filing never names it and LatePost reads it as China’s leading chip designer, adding that SJSemi is the core 2.5D supplier for Huawei’s Ascend accelerators. In the Briefing: Nikkei, citing internal company data, put CXMT’s second-quarter EBIT margin at 82 percent against SK Hynix at 76 and Samsung’s semiconductor arm at 70, while CXMT’s filed half-year report shows 150.3 billion yuan of revenue and 77.6 billion of net profit against a 2.3 billion yuan loss a year earlier; China Mobile Cloud showed a hybrid inference system splitting DeepSeek V4 across a domestic GPU and a Lynxi neuromorphic chip for a claimed doubling of price-performance; Shanghai launched the Pearl Constellation orbital-compute programme; and Zhipu closed about 5 billion dollars in placement and convertible bonds.

Biren’s first interim report as a listed company showed 1.2355 billion yuan of half-year revenue, up 1,997.6 percent, with a net loss 76.4 percent narrower than a year earlier. The number that decides the story is the float. Moore Threads carries 250.94 billion yuan of market value on 6.43 percent of its shares actually trading, and MetaX 270.03 billion yuan on 4.63 percent. Two Shanghai listings hold 520.97 billion yuan of combined market value, about 77.4 billion dollars, priced each day by 28.6 billion yuan of tradable stock. Biren, listed in Hong Kong with just over half its shares in public hands, trades at about 71 times half-year revenue against Moore Threads at about 144.

Profits at China’s large industrial enterprises rose 17.6 percent across January to July, and AI-related electronics accounted for 9.3 percentage points of that, more than half of all the profit growth Chinese industry produced this year. July on its own grew 11.2 percent, down 3.9 points from June, so the headline is decelerating while the electronics contribution holds. The same fact reads from the bottom up in the interim reports filed that morning: SMIC’s first-half net profit rose 94.2 percent on revenue up 19.4 percent, profit growing about five times faster than sales, with the extra coming from price rather than volume.

Loongson chairman Hu Weiwu told the company’s 25th-anniversary meeting in Beijing on 19 August that its taped-out 3B6600 processor runs on a 1Xnm domestic process and performs at the level of mainstream x86 CPUs built on 7nm or better. The claim is Loongson’s own, made at Loongson’s own event, with no independent benchmark published and no workload named. Its audited filings show FY2025 revenue of 635.3 million yuan, about 89 million dollars, against a 455.1 million yuan net loss attributable to shareholders, a third straight nine-figure annual loss. Hu said the main contradiction has moved from research to sales, and that 2025 to 2027 is a transition from the policy market to the open market.

China switched on its first fully domestic 100,000-chip AI supercluster, and the National Development and Reform Commission confirmed it is running at the Zhengzhou node of the national compute grid. Sugon’s 8000, nicknamed Dengfeng, is Chinese at every layer, the accelerators, the interconnect and the immersion cooling, with no restricted imports; the honest caveats are that its workload today leans on double-precision science rather than model training and that per chip it still trails Nvidia on raw performance and the CUDA stack.

Cambricon, the designer of China’s most advanced homegrown AI chips, reported it earned more in the first half of 2026 than in all of 2025: revenue of about 5.99 billion yuan (up 108 percent) and net profit of about 2.31 billion yuan (up 123 percent), against a full-year 2025 net profit near 2.06 billion yuan. Chinese buyers will take every accelerator it can design; how many China can actually manufacture, at what node and with whose memory, is the fabless ceiling that remains.

DeepSeek finished shipping V4, and its day-0 domestic-chip support runs on Huawei Ascend (via Huawei’s CANN, its answer to Nvidia’s CUDA), Cambricon and Hygon. After the launch Alibaba, ByteDance and Tencent ordered hundreds of thousands of Ascend 950 processors.

On Nvidia’s earnings call, Jensen Huang said the quiet part out loud: China revenue is essentially zero, and the company has "largely conceded" the market to Huawei.

The full Nvidia-to-China chip sequence, from the A800 and H800 to the H20 halt and a 4.5 billion dollar write-off, one of the largest write-downs in semiconductor history.

Washington cleared the H200 for ten Chinese firms with a 25%-of-revenue condition. Beijing told them not to buy, keeping the money pointed at domestic chips. Deliveries stayed at zero.

China’s national computing hub in Zhengzhou runs 60,000 accelerators made by Sugon, a firm tied to the Chinese Academy of Sciences, with zero chips from Nvidia, AMD or any Western supplier.

A Chinese system topped the global TOP500 at 2.198 exaflops on 47,000 domestic CPUs and zero GPUs. On the AI-relevant benchmark, it came fourth. The asterisk is the whole export-control story.

A teardown found part of DeepSeek V4’s architecture was co-designed for Huawei’s Ascend 950DT inference chip. The model and the silicon are now being built for each other.

Cambricon became China’s first pure-play AI chip designer worth over a trillion yuan. The same report shows the distance left: a 2.9% share of the cards China bought in 2025, against Nvidia’s 55%.

A Chinese memory middleman’s half-year profit jumped 743x. AI’s appetite for HBM is being paid for one layer down, by the budget phone in your pocket.