|
|
|
|
Bottom Line Up Front Huawei began volume production of 61.44TB and 122.88TB Die-on-Board SSDs using domestically sourced NAND from suppliers including YMTC, shown at the Huawei IDI Forum 2026 in Paris with a 245TB model on the roadmap and roughly 33% higher areal density than conventional TSOP and BGA packaging. DeepSeek is reportedly raising 70 billion yuan at a $45 billion pre-money valuation while opening Beijing roles to build a Code Harness coding agent benchmarked against Anthropic's Claude Code, alongside a permanent 75% price cut to its V4-Pro API. xFusion filed for a ChiNext IPO seeking 8 billion yuan at an implied 80 billion yuan post-IPO valuation, allocating 4.08 billion yuan to next-generation compute infrastructure R&D, with Henan SASAC holding 31.38% through Chao Juneng and Huawei fully exited from the cap table. |
| Huawei mass-produces 122TB SSDs on Die-on-Board packaging and domestic NAND Huawei showed its Die-on-Board packaged 61.44TB and 122.88TB SSDs at the Huawei IDI Forum 2026 in Paris this week, with a 245TB model on the roadmap, CnBeta reported relaying Blocks & Files. The Die-on-Board approach solders bare NAND dies directly to the SSD circuit board, skipping conventional TSOP and BGA packaging and raising areal capacity density by roughly 33%. Entity List status has forced Huawei to source NAND from Chinese domestic suppliers including YMTC after legacy U.S. technology inventory ran out. The 61.44TB and 122.88TB SKUs are already in volume production. |
| DeepSeek reportedly seeks 70 billion yuan at $45 billion pre-money, opens Code Harness roles benchmarked against Claude Code DeepSeek is pursuing a 70 billion yuan round at a $45 billion pre-money valuation, Bloomberg reported on May 22, with Gelonghui packaging the funding leak alongside DeepSeek's confirmation of a coding agent team. DeepSeek announced a permanent 75% V4-Pro API price cut the same day, locking the prior promotional pricing into the published rate card. Senior researcher Chen Deli described the Beijing recruiting push as building DeepSeek's Code Harness from zero and explicitly benchmarked against Anthropic's Claude Code, with two openings posted for an Agent Harness product manager and an R&D engineer. The 70 billion yuan figure is a reported round size, not a confirmed close. |
| xFusion files ChiNext IPO seeking 8 billion yuan, with 4.08 billion yuan earmarked for compute infrastructure R&D The Shenzhen Stock Exchange accepted xFusion's ChiNext IPO prospectus on May 22, Caixin reported, with the former Huawei x86 server business planning to raise 8 billion yuan net at an implied 80 billion yuan post-IPO valuation. Use of proceeds splits 4.08 billion yuan for next-generation compute infrastructure R&D, 2.03 billion yuan for a smart manufacturing park, 890 million yuan for intelligent compute, AI and power architecture R&D, and 1 billion yuan for working capital. Henan SASAC controls 31.38% of the company through the Chao Juneng platform, China Mobile Capital holds 13.76% and the employee share plan holds 7.98%, with Huawei fully exited from the cap table. xFusion plans to issue 97.81 million to 220 million shares for 10% to 20% of post-IPO equity. |
| CAS unveils 'YuanYan-1' AI autonomous transmission electron microscope claiming 56x image capture speed The CAS Dalian Institute of Chemical Physics and the CAS Shenyang Institute of Automation jointly developed YuanYan-1, an autonomous transmission electron microscope, ITHome reported relaying CCTV News. Single-day throughput claims are 200 samples, 5,000 images and 500,000 particles quantitatively analyzed with auto-generated reports. The team reports roughly 56x image capture speed and roughly 300x analytic efficiency versus manual operation, with two weeks of YuanYan-1 output equating to about one year on a conventional TEM. Independent benchmarking of the speed claims has not been published. |
| Shenkaihong releases M-Robots OS 2.0 HarmonyOS robotics operating system to OpenAtom Foundation Shenkaihong released M-Robots OS 2.0 at the OpenAtom OpenHarmony IoT ecosystem conference on May 23, ITHome reported, with the operating system positioned as an open source HarmonyOS robotics OS. The company built a self-developed M-DDS distributed communication layer atop OpenHarmony's distributed soft bus, claiming microsecond-class intra-robot scheduling latency and 4 millisecond inter-robot audio and video latency, which it says is 42% lower than Fast-DDS. The OS scales from 20 kilobyte devices to large industrial robots. Shenkaihong donated the project to the OpenAtom Open Source Foundation in November 2025 with a dedicated project management committee and 16 special interest groups covering kernel, BSP, drivers and middleware. |
|
| Taiwan arrests three over false declarations on Supermicro-built AI servers Taiwan authorities arrested three individuals this week on suspicion of making false declarations on Supermicro-built AI servers, CnBeta reported relaying foreign press, describing the case as Taiwan's first major semiconductor smuggling enforcement action. The reporting links the rise in grey market diversion attempts to U.S. controls on advanced AI chips. Nvidia CEO Jensen Huang told media in Taipei on May 23 that the company applies a 'strict' compliance posture with partners and 'will not leave grey space on regulatory matters'. Huang arrived in Taipei on Saturday. |
| MOFCOM closes Suzhou APEC trade ministers meeting with AI trade and investment cooperation flagged for joint statement Commerce Minister Wang Wentao told a May 23 closing press conference that the APEC trade ministers meeting in Suzhou reached concrete outcomes and that participants will issue a joint statement, Caixin reported. The two-day meeting covered the Asia-Pacific free trade zone, the multilateral trade system, digital cooperation and green economy, with AI related trade and investment cooperation explicitly flagged in the readout. Participants included representatives of 21 APEC economies, observers and the heads of the WTO and UNCTAD. The joint statement text has not yet been published. |
|
| III | CHIPS & SEMICONDUCTORS | |
| Lisuan sells out first batch of LX 7G100 6nm domestic GPU as second batch lands June 18 Lisuan said its LX 7G100 Founders Edition consumer GPU sold out within minutes of opening for sale, with 30,000-plus reservations recorded in 48 hours, ITHome reported. The Founders Edition is priced at 3,299 yuan and ships with 12GB of GDDR6 memory, support for 8K 60Hz output and Lisuan's NRSS super-resolution. The card is built on a 6nm process with what Lisuan calls a self-developed TrueGPU architecture, instruction set and core compute IP, with no external technology licensing. A second batch is scheduled to go on sale June 18. |
| CXMT STAR Market listing committee hearing remains on the May 25-30 trading calendar Changxin Memory Technologies' STAR Market listing committee hearing is scheduled within the May 25-30 trading week, according to Wallstreetcn's May 24 weekly preview. The Huawei AgentArts open source release is also flagged for the same window as a standout China tech event. China's official manufacturing PMI for May, the U.S. PCE deflator for April and earnings from Pinduoduo, Xiaomi, Kuaishou, Li Auto, XPeng and Marvell also fall within May 25-30. The CXMT hearing date was previously reported as May 27. |
|
| Q1 China AI financing tops 110 billion yuan across nearly 600 deals, up 185.4% year over year China AI sector financing in Q1 2026 totaled more than 110 billion yuan, roughly $15 billion, across nearly 600 deals, up 185.4% year over year, CCTV Finance reported in coverage relayed by CLS. Unnamed VC statistics in the same coverage put GPU procurement and cloud service leasing at 30% to 50% of large language model company funding. Domestic LLM companies including Moonshot AI and StepFun secured more than 30 billion yuan combined in May rounds. Chinese LLM iteration cycles have shortened to under three months by 2026. |
| JPMorgan China desk says workflow ownership now outweighs base-model differentiation in vertical AI monetization JPMorgan Securities (China) analyst Yao Cheng and team published a May 22 China AI desk note based on eight company meetings, Wallstreetcn reported. The note says monetization is appearing first in workflow-heavy, data-rich verticals like insurance, finance risk and cross-border marketing, with the base model functioning as a more replaceable input. Compute, AI infrastructure, domestic chips, memory and storage are flagged as the cleanest equity exposures, while application-layer returns accrue to firms that own proprietary data and workflow entry points. Some surveyed companies now route requests across domestic and frontier models based on price, performance and task. |
|
|
|