|
|
|
|
Bottom Line Up Front Big Fund leads DeepSeek funding round at a $45 billion pre-money valuation, with the National IC Industry Investment Fund anchoring DeepSeek's first external raise of roughly 70 billion yuan alongside Monolith Capital and Hillhouse, more than doubling its $20 billion valuation from a month earlier. MOFCOM quantifies tariff rollback scope at $30 billion or more of products on each side, with spokesperson He Yadong saying the two sides have agreed in principle under the Trade Council framework tied to the May 14 Xi-Trump summit and will finalize specific arrangements. MOFCOM defends Japan entity listings covering 20 firms on a control list and 20 on a watch list, with He framing the measures as preventing Japanese remilitarization, attributing bilateral strain to Prime Minister Takaichi Sanae, and noting that civilian use license applications for rare earths and other critical minerals are under review. |
| Big Fund leads DeepSeek's first external round at $45 billion pre-money valuation China's National IC Industry Investment Fund (the "Big Fund") is leading DeepSeek's first external funding round, Caijing reported via CnBeta, with the round size at about 70 billion yuan (roughly $10 billion) and pre-money valuation at $45 billion. DeepSeek's valuation stood at $20 billion one month earlier, per a May 22 Bloomberg leak that first put the round at 70 billion yuan. The investor list reportedly also includes market firms Monolith Capital and Hillhouse Capital and was substantially locked by the end of May. |
| MOFCOM puts post-summit reciprocal tariff cut at $30 billion or more per side Commerce Ministry spokesperson He Yadong said at the May 28 regular press conference that both sides have agreed in principle under the Trade Council to a framework for reciprocal tariff reductions covering $30 billion or more of products on each side, Yicai reported relaying Xinhua. He said the two economic and trade teams will maintain close communication, agree on specific arrangements and push for swift implementation. The figure is the first specific dollar scope publicly attached to the May 14 Xi-Trump summit framework. |
| MOFCOM defends 40-entity Japan export control package on anti-remilitarization grounds He Yadong said at the same May 28 press conference that China's placement of 20 Japanese entities on a control list and 20 on a watch list is "entirely justified, reasonable, and lawful," framed as preventing Japan's "re-militarization and nuclear ambitions," per CCTV News relayed by Yicai. He said Chinese export controls on rare earths and other critical minerals are implemented under domestic law and that civilian use license applications are being reviewed. He attributed the strain in China-Japan relations to Prime Minister Takaichi Sanae and urged Tokyo to "reflect and correct" its course. |
| Mianbi releases AI written training framework with claimed 10% gain over Megatron on Ascend Mianbi (ModelBest) released ForgeTrain, described as the first production-grade large model training framework written entirely by AI agents with zero human coding, InfoQ China reported. The framework was validated training MiniCPM5-1B on Huawei Ascend chips with a claimed 10% speed advantage over Nvidia's Megatron training framework. MiniCPM5-1B pretraining took 3 to 5 days on Ascend, against two days for the smaller MiniCPM4-0.5B on Nvidia GPUs, per Mianbi's disclosure. Independent reproduction of the speed claim has not been published. |
| Tencent's Canghai V2 codec ASIC sweeps MSU encoding benchmark and enters mass production Tencent said its self-developed Canghai codec chip series took first place across the 30 to 240 fps speed tiers on SSIM, PSNR and VMAF metrics at the May 26 Moscow State University hardware video encoding competition, leading by more than 30% on multiple measures against more than 40 entries from AMD, Nvidia, Intel and other vendors. Canghai V1 entered mass production in 2023 with over 100,000 units deployed across Tencent Cloud workloads including live streaming, video on demand, 4K transcoding and cloud gaming. The next-generation Canghai V2 has powered on and entered the mass production cycle, with H.265 compression more than 10% better and H.266 more than 30% better than the prior generation, and full service planned for the second half of 2026. |
|
| Foreign Ministry frames rare earth consultations as proceeding within May 14 summit framework MFA spokesperson Mao Ning said at the May 28 regular press conference that the Chinese Embassy in the United States has forwarded preliminary outcomes of China-U.S. economic and trade consultations, Xinhua reported via Yicai. She said both sides should jointly implement the consensus reached by the two heads of state at the May 14 summit and maintain stable development of bilateral economic and trade relations. The remarks came in response to questions about Chinese rare earth export controls and place MFA alongside MOFCOM as a public voice on the rare earth dimension of the bilateral framework. |
| CAS National Time Service Center opens first Hong Kong node for national standard time The Chinese Academy of Sciences National Time Service Center and the Hong Kong Observatory unveiled the "National Standard Time and Combined Atomic Time Hong Kong Node" on May 28, ITHome reported relaying CMG Greater Bay Area Sound. The Observatory feeds raw measurement data through the new node into the production and maintenance of China's combined atomic time. Center director Zhang Shougang called time "an important strategic national resource" and described the Hong Kong deployment as the first practice of drawing the territory into national time-base production. China's national standard time keeps within 2 nanoseconds of UTC, per the National Time Service Center. |
|
| III | AI & FOUNDATION MODELS | |
| DeepSeek-V4 first adapts to Huawei Ascend at 2026 Kunpeng/Ascend Developer Conference DeepSeek's V4 model series has been first adapted to Huawei Ascend chips, with the full Ascend super-node lineup supporting the V4 family, InfoQ China reported from the 2026 Kunpeng/Ascend Developer Conference. Huawei is positioning Ascend as the AI compute base and the Kunpeng CPU as the runtime layer for agent orchestration, tool calls, sandbox execution, memory retrieval and security isolation. The disclosure pairs Chinese frontier model and domestic silicon co-design rather than a model release in isolation. Independent benchmarks for V4 on Ascend versus Nvidia GPUs have not been published. |
| SuperCLUE 21-model evaluation places top Chinese models near 5th globally SuperCLUE's latest evaluation covers 21 leading domestic and foreign large language models across 492 questions in six task families: math reasoning, science reasoning, code generation, agent planning, instruction following and hallucination control, CnBeta reported. Google's Gemini, OpenAI's GPT-5.5, Anthropic's Claude-Opus and Gemini-Flash hold the top four positions; DeepSeek-V4-Pro, Qwen3.7-Max and Doubao Seed 2.0 Pro cluster near fifth as the domestic top tier. Qwen3.7-Max came within two points of leading models on code generation, per SuperCLUE. The methodology has been disclosed but is not externally audited. |
| Xiaomi locks in permanent MiMo-V2.5 API price cut of up to 99% Xiaomi's MiMo team announced on May 27 a permanent price reduction for the MiMo-V2.5 series API, with the largest discounts reaching 99% and unified pricing replacing the prior context window tiers, CSDN reported. For MiMo-V2.5-Pro, the cache-hit price drops to 0.025 yuan per million tokens and the output price to 6 yuan per million tokens, with Xiaomi saying some scenarios see 86% to 99% reductions versus the prior card. The Token Plan billing system was redesigned to deliver five to eight times the usable tokens at the same price. The cut parallels DeepSeek's May 22 permanent 75% V4-Pro reduction. |
| Alibaba DAMO ships first GPU build of MindOpt solver with hundred-million-variable LP support Alibaba's DAMO Academy released the GPU version of its MindOpt optimization solver on May 28, ITHome reported. Testing on about 2,000 general linear programming benchmark problems showed MindOpt-GPU solving more than 99% of problem types to high precision and supporting linear programs with hundreds of millions of variables. DAMO said the GPU build raises success rates by more than 14% over leading industry products on large-scale problems and is 2.67 times faster on average. The release describes solvers as "the core of industrial software," underlying compute in power dispatch, flight scheduling, advanced manufacturing and financial workloads. |
|
| IV | ROBOTICS & AUTONOMOUS SYSTEMS | |
| XPeng's He Xiaopeng says company killed prior autonomous driving stack to pivot to "Physical AI," targets L4 in 18 to 24 months XPeng chairman He Xiaopeng said in a Sina Finance interview that the company halted its prior autonomous driving system after multi-billion-yuan investment to pivot to a new "Physical AI" architecture, ITHome reported. He projected that the new architecture will let XPeng reach L4 autonomous driving within 18 to 24 months, putting the target around the end of 2027. He named humanoid robots as the carrier of "Physical AI" and said high-grade humanoid commercial mass production is likely to begin in 2027. He listed autos, robots and global business as XPeng's three growth curves for the next decade. |
| Agibot subsidiary AGILINK closes fourth round in five months at over $1 billion valuation Agibot (Zhiyuan), Unitree's chief humanoid rival, has spun its dexterous hand team into AGILINK, which recently closed a several-hundred-million-yuan financing round pushing its post-money valuation past $1 billion, Huxiu reported. AGILINK was formed in January 2026 when Agibot carved out the dexterous hand team as an independent entity, and the latest round is the fourth since the spinoff. Agibot crossed 10,000-unit annual humanoid output and 1 billion yuan in revenue over the past year. The AGILINK round precedes the June 1 STAR Market listing committee hearing on Unitree's IPO. |
|
| CAS Space breaks ground on Guangzhou reusable rocket and liquid engine production base CAS Space, the Chinese Academy of Sciences-affiliated commercial launch provider, broke ground on May 28 on Phase I of its reusable launch vehicle, spacecraft and liquid rocket engine production base on Xiji Island in Guangzhou's Huangpu District, ITHome reported. The base is the first anchor project of the Guangzhou Aerospace Town and is planned for completion and entry into operation in 2027. CAS Space's existing launch services cover China, Europe, North America, South Asia, the Middle East and North Africa. The project sits in a riverside, port-adjacent location to support transport of large aerospace equipment. |
| LandSpace's Wuxi base begins operations, moves Zhuque-3 reusable rocket into batch manufacturing LandSpace's Wuxi production base entered operations on May 27, supporting subsequent R&D testing, batch production and commercial missions for the Zhuque-3 launch vehicle, ITHome reported relaying a Wuxi municipal announcement. The 112-mu (about 7.5-hectare) site received 2.3 billion yuan in Phase 1 investment and is designed to support 20 launches per year in assembly, testing and logistics. Zhuque-3 is China's first stainless steel liquid launch vehicle and reached orbit on December 3, 2025. LandSpace said launch costs are expected to fall by more than 80% versus expendable rockets once reuse is fully mature. |
|
| Moonshot AI begins dismantling VIE and red chip structures to clear Hong Kong IPO path Moonshot AI (Kimi) has notified shareholders that it is dismantling its VIE and red chip structures to clear the way for a Hong Kong listing, Huxiu reported relaying Bloomberg. Founder Yang Zhilin had said six months ago that there was no rush to go public near term; Moonshot's valuation has moved from about $300 million at its 2023 founding to above $20 billion as of May 2026. Huisheng International Capital president Huang Lichong said the unwinding window could run six months to a year if regulatory, shareholder, tax and foreign investment arrangements move smoothly. The structural unwinding follows Zhipu and MiniMax onto Hong Kong over U.S. listings as the preferred venue for Chinese AI champions. |
| Nvidia CEO Jensen Huang reportedly accepts Tsinghua advisory board seat Nvidia chief executive Jensen Huang has agreed to join Tsinghua University's School of Economics and Management advisory board, CnBeta reported, citing the Financial Times and two people familiar with the matter. The 65-member board is chaired by Apple CEO Tim Cook and counts Elon Musk, Satya Nadella, Mark Zuckerberg, Michael Dell, Jamie Dimon, Larry Fink and Jane Fraser among its members. The appointment follows Huang's recent trip to China alongside the Trump delegation. Nvidia had not responded to a request for comment as of CnBeta's filing. |
|
|
|