Thursday, September 10, 2026

1. DeepSeek will reduce its V4-Flash series API prices by up to 60% starting today at noon (Beijing Time). Concurrently, it launched the V4.1 Flash beta, introducing a new architecture that processes text, images, and audio within a unified model. Until V4.1 Pro is released, V4.1 Flash will also process Pro requests.

2. Huawei unveiled its new flagship chip, the Kirin 9050 Pro, for the first time in six years. It incorporates "Logic Folding" technology for the first time, vertically stacking logic circuits. GPU performance improved by 142%, and the chip supports on-device inference for a 30-billion-parameter MoE model.

3. The Ministry of Industry and Information Technology (MIIT) released the 15th Five-Year Plan for the Information and Communications Industry. It sets targets of 4 trillion yuan in sector revenue and a 95% 5G penetration rate by 2030, while officially including the phrase "timely launch" of 6G commercialization in a policy document for the first time.

4. Unitree released footage of fully autonomous combat by a humanoid robot powered by its world-action model UnifoLM-X2-1.0. Without teleoperation or pre-programming, the robot predicted subsequent situations in real time to execute offensive and defensive maneuvers.

5. Evidence emerged that Aivres, a U.S. subsidiary of Inspur Group, diverted $5.6 billion worth of advanced semiconductor equipment to Southeast Asia between April 2024 and February 2026. The shipments included $3 billion worth of Nvidia Blackwell chips, with ByteDance and Alibaba as the end customers. This development directly impacts the agenda of the September 24 U.S.-China summit.

DeepSeek Adjusts V4-Flash Pricing and Launches V4.1 Flash Beta (Announced 2026-09-09, Effective 09-10)

DeepSeek will adjust its API pricing for Flash series models starting at noon on September 10 (Beijing Time). During off-peak hours, cache-hit input is priced at 0.02 yuan per million tokens, a 60% reduction from the previous 0.05 yuan. Cache-miss input is priced at 1 yuan (down 33% from 1.5 yuan), and output is priced at 4 yuan (down 11% from 4.5 yuan). Peak-hour rates are double off-peak rates. Peak hours are weekdays from 9:00 AM to 12:00 PM and 2:00 PM to 6:00 PM (Beijing Time).

Following a price increase on August 17, DeepSeek experienced a steep decline in direct API traffic. Traffic dropped by 79% for V4-Flash and 72% for V4-Pro. Within 24 days, the company rolled back off-peak prices to pre-hike levels.

Concurrently, DeepSeek opened a two-day beta for V4.1 Flash starting September 8. While the previous V4-Flash-Vision-Exp added vision capabilities onto a text model, V4.1 Flash processes text, images, and audio within a single unified architecture. During the beta, output generation reached speeds of up to 507 tokens per second. Until V4.1 Pro is released, Pro API requests will be routed to V4.1 Flash and billed at Flash rates.

Enterprises utilizing the DeepSeek API can reduce costs during off-peak hours starting after noon today.

IT之家 https://www.ithome.com/0/999/953.htm

TechNode https://technode.com/2026/09/09/deepseek-v4-1-flash-multimodal-limited-beta/

DigitalApplied https://www.digitalapplied.com/blog/deepseek-v4-1-flash-pro-routing-prices-early-tests

ByteDance Secures $29.6 Billion Loan and Trains 10-Trillion-Parameter Model (Reported 2026-09-03–04)

ByteDance has secured a $29.6 billion (approximately 40 trillion KRW) syndicated loan arranged by Citigroup and JPMorgan. The facility carries a three-year maturity with participation from approximately 30 banks. The initial target was $20 billion, but the size was expanded amid strong bank demand.

The funds will be directed toward constructing AI data centers in Southeast Asia. ByteDance is considering capital expenditures of up to $70 billion this year, more than double last year's level. The company is also establishing a large-scale AI data center campus in Ulanqab, Inner Mongolia.

According to the Financial Times, ByteDance's Seed team is pre-training a 10-trillion-parameter model, the largest ever undertaken by a Chinese enterprise. The 2,000-member team is led by former Google DeepMind researcher Wu Yonghui. ByteDance stated that it is developing the model independently from scratch without relying on distillation from other models.

TechNode https://technode.com/2026/09/04/bytedance-reportedly-lines-up-a-29-6-billion-syndicated-loan/

이데일리 https://www.edaily.co.kr/News/Read?newsId=04349286645576512

Financial Times (Report on ByteDance's 10-trillion-parameter model)

Robotics, Humanoids & Embodied Intelligence

Unitree Unveils Fully Autonomous Humanoid Robot Combat Powered by World Model (2026-09-07)

On September 7, Unitree released video footage demonstrating fully autonomous combat by a humanoid robot powered by its world-action large model, UnifoLM-X2-1.0. Without teleoperation or pre-programming, the model predicted subsequent physical states in real time to autonomously execute attacks and defenses. The company stated that it has achieved "the world's first implementation of fully autonomous humanoid robot combat driven in real time by a world model."

Unitree listed on the Shanghai STAR Market in August, surging 460% on its debut day. Its closing price on September 7 stood at 536.35 yuan, down 50% from its peak. First-half revenue reached 1.152 billion yuan, up 48.5% year-on-year, and the company turned profitable with a net profit of 274 million yuan. Its market capitalization is approximately 216.9 billion yuan ($32.4 billion).

In South Korea, Nota, LG Electronics, and Mobilint were selected for a humanoid robot development project based on domestic NPUs. With Chinese robots accounting for 97% of global shipments, the project aims for differentiation through South Korean on-device AI semiconductors.

IT之家 https://www.ithome.com/0/999/432.htm

澎湃新闻 https://m.thepaper.cn/newsDetail_forward_34025139

AI타임스 https://www.aitimes.com/news/articleView.html?idxno=214972

IFA 2026 Berlin: Chinese Robotics Cohort Dominates the Stage (2026-09-04–09)

At "Robots on the Runway," the first humanoid robot stage showcase in IFA's 102-year history, 11 companies took the stage, more than six of which were Chinese firms. Participants included Unitree, AgiBot, EngineAI, Deep Robotics, Galbot, and XGSynBot. Chinese companies accounted for 932 of the 1,900 total participating companies at IFA, representing nearly half.

AgiBot debuted its emotionally interactive conversational humanoid robot "Lumo" in Europe. Galbot showcased the Galbot G1, a wheeled pharmacy robot deployed in 10-hour shifts at Beijing pharmacies. XGSynBot presented the Z1, a modular industrial robot capable of automatically exchanging tools in six seconds.

The United States banned new imports of Chinese humanoid robots via an FCC order effective July 29. With the EU having no separate restrictions yet, Chinese companies are accelerating their European expansion. However, Bluetooth vulnerabilities in the Unitree G1 (CVE-2026-76639, CVE-2026-76640) were exposed without patches during the IFA exhibition, highlighting the absence of security standards in Europe.

Cryptopolitan https://www.cryptopolitan.com/unitree-agibot-ifa-2026-robot-runway/

CNMO https://ai.cnmo.com/news/817431.html

TechTimes (Unitree vulnerability) https://www.techtimes.com/articles/326651/20260904/unitree-humanoid-robots-hacked-30-meters-eu-buyers-ifa-face-no-import-restriction.htm

AI Semiconductors & Computing Infrastructure

Huawei Unveils Kirin 9050 Pro: First Application of Logic Folding Technology (2026-09-07)

On September 7 in Guangzhou, Huawei unveiled the Kirin 9050 Pro, its first new flagship mobile chip in six years since the Mate 40.

The Kirin 9050 Pro is the world's first Logic Folding Tau chip. By vertically stacking logic circuits within a single die, it shortens signal pathways. Leveraging Lingxi CPU multithreading technology, single-core performance increased by 24% and multi-core performance by 52%. GPU performance improved by 142%. It features a 9-core configuration (1 ultra-large, 2 performance, 4 efficiency, 2 small) alongside separate low-power microcores.

The chip supports on-device inference for 30-billion-parameter MoE models. It debuted in the Mate XT2 tri-fold smartphone, priced starting from 19,999 yuan.

The significance of this chip is profound. Vertically stacking circuits to increase transistor density without EUV lithography equipment is becoming established as a design strategy to circumvent export controls. While not directly competing with the foundry and HBM businesses of Samsung Electronics and SK Hynix, it demonstrates that Huawei's chip design capabilities are advancing rapidly.

IT之家 https://www.ithome.com/0/999/429.htm

IT之家 (Mate XT2) https://www.ithome.com/0/999/337.htm

凤凰网 https://tech.ifeng.com/c/8wE4gXJeH6G

Moore Threads Hits 20% Lower Price Limit Following Lockup Expiration (2026-09-07)

Moore Threads, a leading Chinese GPU enterprise, hit the 20% daily lower price limit on the Shanghai STAR Market on September 7. The lockup expired on 25.77 million shares (5.48% of total equity) allocated at IPO, nearly doubling the circulating share count from 30.22 million to 56 million shares. Its market capitalization fell below 200 billion yuan.

First-half revenue reached 17.36 billion yuan, up 147% year-on-year. However, the company faces assessments that its software ecosystem (MUSA) remains weaker than Nvidia's CUDA. On December 7, an additional 185.91 million shares (39.55% of the total) will be released from lockup.

ChangXin Memory Technologies (CXMT) has commenced low-volume mass production of HBM3E, a first for China. Alibaba's T-Head (PingTouGe) and Cambricon are currently testing the chips. External estimates place yields at around 25%, trailing Samsung and SK Hynix by approximately one generation. CXMT aims for full-scale mass production by 2027.

澎湃新闻 https://www.thepaper.cn/newsDetail_forward_34022807

The Decoder (CXMT HBM3E) https://the-decoder.com/chinas-cxmt-makes-its-first-hbm3e-chips-closing-the-ai-memory-gap/

Data Studios (Moore Threads) https://www.datastudios.org/post/moore-threads-stock-crash-ipo-lockup-china-ai-gpu-race

Telecommunications, Networks & Edge AI

MIIT Releases 15th Five-Year Plan for Information and Communications Industry: First Official Codification of 6G Commercialization (2026-09-07)

On September 7, MIIT unveiled the "15th Five-Year Plan for the Development of the Information and Communications Industry," outlining 13 primary indicators and 26 key tasks.

The core targets are as follows: achieving 4 trillion yuan in information and communications industry revenue by 2030, reaching a 95% 5G user penetration rate, and including the phrase "timely launch (适时启动)" of 6G commercialization in an official policy document for the first time.

Specific tasks include developing 6G base stations, 6G core network equipment, and 6G smartphones. China plans to build an operating system based on OpenHarmony (鸿蒙) that supports agentic communications and ambient sensing capabilities. It will deploy AI computing clusters exceeding 10,000 nodes, designating the convergence of "communications, sensing, intelligence, and computing" (通感智算) as a strategic direction.

This development carries key implications for telecommunications carriers and equipment manufacturers.

中新网 https://www.chinanews.com.cn/cj/2026/09-07/10692127.shtml

光明日报 https://m.gmw.cn/2026-09/08/content_1304560973.htm

新华网 https://www.news.cn/tech/20260612/a0fe3e279bc24fce8ef99f3bb63b70ff/c.html

Policy, Regulation & National Strategy

Entry-Exit Administration Regulations Effective September 15 (D-5): Restrictions on AI Talent Exits Materialize (Effective 2026-09-15)

The State Council's Entry-Exit Administration Regulations will take effect on September 15. The regulations bar individuals implicated in export control or technology transfer violations from leaving the country, with no statutory time limit. The principle of "whoever approves bears responsibility (谁批准谁担责)" will be enforced.

The two co-founders of AI agent startup Manus have been under an exit ban since March. Meta attempted to acquire Manus for $2 billion, but negotiations effectively collapsed due to these restrictions.

Civil servants, state-owned enterprise employees, university professors, executives of publicly traded companies, and even university students must obtain prior approval before overseas business trips or study abroad. RFA reported that key AI talent is now classified as strategic assets, effectively barring their departure from China.

For foreign enterprises seeking to recruit Chinese AI talent or pursue joint research, these regulations may pose a substantial practical barrier.

司法部 https://www.moj.gov.cn/pub/sfbgw/gwxw/xwyw/202607/t20260731_538116.html

France 24 https://www.france24.com/en/asia-pacific/20260818-clamping-down-china-tightens-exit-rules

RFA https://www.rfa.org/mandarin/shehui/2026/09/03/china-entry-and-exit-management-regulations-capital-fleeing-technology-and-talent-outflow-strict-approval-processing/

Generative AI Service Filings Surpass 1,372; Standardization of AI Agents Underway (As of 2026-09)

Cumulative generative AI service filings in China have surpassed 1,372, more than tripling from 406 at the end of 2024. This comprises 861 large model approvals (备案) and 511 large model registrations (登记).

Three government agencies—the Cyberspace Administration of China (CAC), the National Development and Reform Commission (NDRC), and the Ministry of Industry and Information Technology (MIIT)—jointly issued implementation opinions on standardizing AI agents (智能体). The framework establishes a tiered management system across 19 sectors spanning five domains: scientific research, industry, consumer services, public livelihood, and social governance.

The Stanford 2026 AI Index Report assessed that the performance gap between U.S. and Chinese AI models has effectively closed, with models from both nations alternating in the top position since early 2025.

安全内参 https://www.secrss.com/articles/90111

光明日报 https://epaper.gmw.cn/gmrb/html/content/202609/03/content_23946.html

Inspur Aivres Diverts $5.6 Billion in Nvidia Chips via Circumvention Route (Reported 2026-09-07)

According to The New York Times and Asia Times, Aivres, a U.S. subsidiary of China's Inspur Group, exported $5.6 billion worth of advanced equipment to Southeast Asia between April 2024 and February 2026. More than $3 billion of these shipments consisted of servers equipped with Nvidia Blackwell chips, with ByteDance and Alibaba as the ultimate customers.

The background is as follows: The U.S. Department of Commerce's Bureau of Industry and Security (BIS) placed Inspur Group on the Entity List in March 2023. In August of the same year, Inspur renamed its California subsidiary to Aivres. In November 2025, Inspur lowered its equity stake to 33%, sidestepping the 50% threshold under the affiliate rule. While BIS issued guidance in May 2026 requiring export licenses for all entities with parent companies in arms-embargoed countries, transactions prior to that guidance were technically legal.

The U.S. House of Representatives passed the Remote Access Security Act (RASA) by a vote of 369 to 22, though the Senate has yet to act on it. The Chip Security Act has also cleared the House Foreign Affairs Committee. This issue serves as a critical variable for the Trump-Xi summit on September 24 and the preceding U.S.-China AI Safety Dialogue scheduled for mid-September.

As the United States tightens its regulatory net, compliance costs for exports routed through Southeast Asia will rise.

Asia Times https://asiatimes.com/2026/09/nvidia-chip-export-loophole-clouds-us-china-ai-summit-talks/

헤럴드경제 https://biz.heraldcorp.com/article/10865516

이데일리 https://www.edaily.co.kr/News/Read?newsId=02515766645577824

The first bilateral meeting dedicated to AI since the inauguration of the second Trump administration may take place in mid-September. According to Reuters and CNBC, Treasury Secretary Bessent is expected to lead the U.S. delegation, while Vice Premier He Lifeng or Standing Committee Member Ding Xuexiang is being considered to head the Chinese side.

The agenda spans frontier model risk management, monitoring AI-driven cyberattacks, and intellectual property protection. The Inspur Aivres export diversion issue is likely to directly influence the tone of the discussions. Specific dates and venues have not yet been finalized.

At the September 24 Washington summit (D-14), AI chips, electric vehicles, and rare earths will be the focal agenda items. Xi Jinping is expected to be accompanied by a large delegation of corporate CEOs.

CNBC https://www.cnbc.com/2026/09/05/us-china-gear-up-for-mid-september-ai-safety-talks-reuters.html

머니투데이 https://www.mt.co.kr/world/2026/09/05/2026090508550551646

Attorney Kim Kyung-jin's Commentary

The two central themes of today's briefing are pricing and regulation. DeepSeek rolling back its price hike in just 24 days indicates that price elasticity in the Chinese AI model market remains high. An experimental cycle of raising prices and then lowering them when demand drops continues to repeat across the API market.

Huawei's Logic Folding in the Kirin 9050 Pro demonstrates that performance enhancement methods bypassing EUV lithography have reached actual mass production. Viewed alongside CXMT's low-volume mass production of HBM3E, the self-contained loop of China's domestic semiconductor supply chain is narrowing another step closer to completion.

The Inspur Aivres case exposes structural loopholes in export control enforcement. A window where companies could evade restrictions simply by rebranding persisted for nearly two years. Unless the United States closes this loophole legislatively, its leverage to demand concessions from China at the September 24 summit will be weakened.

Several key developments warrant close observation moving forward.

Generated: 2026-09-10 06:55 KST