AI Board

AI BOARD

AI Board

AI tools, policy moves, and public decisions are tracked in plain language.

All posts AI Briefs Tools Policy Education Industry

China AI Robotics Semiconductor Telecom Intelligence Briefing 2026-08-15

Author
김 경진
Date
2026-08-15 07:15
Views
141
China AI Robotics Semiconductor Telecom Intelligence Briefing 2026-08-15
2026-08-15 · KIMKJ.COM CHINA AI INTELLIGENCE BRIEFING
China AI, Robotics, Semiconductor, and Telecom In-Depth Briefing

Saturday, August 15, 2026

Top 5 Highlights

1. DeepSeek will significantly raise its API prices tomorrow (August 16, 16:00 UTC). The V4 Pro output token price will increase 4.5-fold, from $0.87 to $3.96 per million tokens during peak hours. This marks a turning point where the low-price offensive, which pushed the market below cost, slows down.

2. Alibaba is releasing Qwen3.8-27B under the Apache 2.0 license today (August 15). It is a 27.8-billion-parameter dense multimodal model that can run on a single personal GPU.

3. SMIC's Q2 net profit reached $479 million, a 3.6-fold increase year-on-year. Quarterly revenue exceeded $3 billion for the first time, and backed by strong AI demand, the company announced further wafer price hikes for Q3.

4. In the first half of 2026, Chinese companies accounted for 97% of global humanoid robot shipments. A total of 19,100 units were shipped, and AgiBot took the top spot, Overtaking Unitree with 8,400 units.

5. Unitree's IPO allocations have been finalized. With 19,414 winning allocation numbers, it is expected to list on the STAR Market around August 19. On the same day, the World Robot Conference (WRC 2026) will open in Beijing.

Chinese AI Models & Platforms

[Follow-up] DeepSeek Confirms August 16 Price Hike and Official Launch of V4 Pro 0813

DeepSeek’s API price hike will take effect tomorrow (August 16, 16:00 UTC). V4 Pro's cache-miss input price will increase from $0.435 per million tokens to $1.32 during peak hours and $0.66 during off-peak hours. Output prices will jump from $0.87 to $3.96 during peak hours and $1.98 during off-peak hours. Peak hours are defined as 01:00–04:00 and 06:00–10:00 UTC. The price increase ranges from 50% to 1,100% depending on the specific item.

Even after the increase, it remains cheaper than Western competitors. However, the price-driven market consolidation is set to slow down. The change is driven by severe server loads, reaching 8 trillion tokens per day. DeepSeek resumed its $8 billion secondary funding round this week at a valuation of $74 billion. The company is also building a 1GW data center in Ulanqab, Inner Mongolia, targeting partial operations by late 2027 or early 2028. It is considering chips from both Nvidia and Huawei.

In the same week, DeepSeek released Harness v0.1, an agentic coding tool, under the MIT license. It is designed to allow developers to swap all components, such as models, tools, sessions, and sandboxes, as plug-ins. Built on TypeScript, it gained 33,000 GitHub stars within a single day, establishing itself as a Chinese alternative to Claude Code.

South Korean Implications: South Korean firms using the DeepSeek API must adjust their systems today. Since peak-hour costs will jump more than fourfold, shifting non-urgent workloads to off-peak hours will help avoid the price spike.

(Published: 2026-08-13)

Alibaba Releases Qwen3.8-27B Today (August 15) Under Apache 2.0 License

Alibaba is uploading Qwen3.8-27B to Hugging Face and ModelScope today. This 27.8-billion-parameter dense model processes text, images, and video, with a context window of 262,144 tokens. Released under the Apache 2.0 license, it has no commercial use restrictions.

This is different from Qwen3.8-Max (a 2.4-trillion-parameter MoE model) released on August 12–13. Max is under a proprietary license requiring revenue sharing for companies with annual revenues exceeding $20 million. This is the first time a Chinese AI company has imposed revenue-sharing terms on open-weight models. The open-weight version of Max only processes text and lacks the 1-million-token context window and image recognition capabilities of the API version.

Since the 27B model is small enough for inference on a single personal GPU, it targets the on-premise installation market.

South Korean Implications: Since the 27B model is licensed under Apache 2.0, South Korean companies can use it freely. In contrast, Max-class open-weight models carry revenue-sharing clauses, meaning companies must verify licensing terms before large-scale commercial deployment.

(Published: 2026-08-15)

[Follow-up] Tencent’s Q2 AI Capex Reaches RMB 51.8 Billion, Up 190% YoY

Tencent’s Q2 capital expenditure surged to RMB 51.8 billion (approximately $7.1 billion). As funds were channeled into AI infrastructure, free cash flow recorded negative RMB 13.8 billion. The average daily token usage of the Hunyuan 3 model grew sixfold, entering OpenRouter's global top three. While the new AI product segment incurred a quarterly loss of RMB 10.5 billion, core business profits excluding AI grew by 19%. Q2 revenue reached RMB 204.8 billion, representing an 11% year-on-year growth. The next-generation Hunyuan 4 is scheduled for release within this year.

(Published: 2026-08-12)

[Follow-up] Moonshot AI to Close Pre-IPO Round at $50 Billion Valuation on August 27

Moonshot AI will close its pre-IPO funding round at a $50 billion valuation (approximately KRW 67 trillion) on August 27. The company aims to file for listing on the Hong Kong Stock Exchange by September 30. Institutional investment demand flooded in following the release of Kimi K3 (2.8 trillion parameters). The company's annualized revenue stands at $300 million. This represents a rapid follow-on raise immediately after securing $3.5 billion in a Series F round at a $34.9 billion valuation.

(Published: 2026-08-05)

Robotics, Humanoids, and Embodied AI

Chinese Humanoid Robots Account for 97% of H1 Global Shipments

According to market research firm Smart Analytics Global, global humanoid robot shipments in the first half of 2026 reached approximately 19,100 units, a 3.7-fold increase from the 5,100 units shipped in the same period last year. Chinese enterprises captured 97% of this market.

Rankings have also shifted. AgiBot climbed to the global number one spot (44% market share) by selling approximately 8,400 units, while Unitree fell to second place with 5,900 units. Industrial and commercial applications accounted for over 70% of the shipments, indicating that humanoid robots are moving past the exhibition phase and entering actual factories and retail spaces. China itself absorbs more than 85% of global demand.

Annual shipments are projected to reach 60,000 units this year and 500,000 units by 2030. However, because the US FCC has banned new imports of Chinese humanoid and quadrupedal robots, this growth could remain confined to the Chinese domestic market and non-US markets. In H1, China's robot exports reached RMB 6.29 billion across 141 countries, representing a 359% export growth rate.

South Korean Implications: South Korean robot companies will face direct competition with Chinese counterparts in the global market. While China leads in mass production costs and supply chain integration, South Korea must differentiate in key components—such as reducers, motors, and sensors—and software. The US import ban may open a window of opportunity for South Korean companies, but that window will remain narrow without price competitiveness.

(Published: 2026-08-11)

[Follow-up] Unitree IPO Allocation Finalized, STAR Market Listing Around August 19

Unitree's 19,414 winning allocation numbers have been finalized. The oversubscription ratio reached 8,289x for retail investors and 2,618x for institutional investors, resulting in a winning rate of 0.0181%. Based on the IPO price of RMB 150.80, the market capitalization stands at RMB 61 billion (approximately $9 billion). Strategic investors include the National Social Security Fund, DeepSeek (RMB 140.8 million), Tencent, and PetroChina (Kunlun Capital). The listing is expected around August 19.

On the same day, the World Robot Conference (WRC 2026) will open in Beijing. Running from August 19 to 23, the event spans 52,000 square meters and features 774 participating companies, with 49 state-owned enterprises setting up exhibition halls for the first time. The listing of the STAR Market's first humanoid robot company overlaps with the opening of China's largest robotics exhibition on the same day. Next week is expected to be a watershed moment for the Chinese robotics industry.

(Published: 2026-08-14)

[Follow-up] The 4th Shanghai Embodied AI Robot Industry Exhibition Closes

The embodied AI robotics exhibition, held at the Shanghai New International Expo Center from August 12 to 14, has concluded. More than 400 brands participated, with consultations heavily focused between humanoid robot manufacturers, joint/sensor suppliers, automotive OEMs, and component makers. It is a UFI-certified exhibition spanning five categories: key components, service robots, special robots, industrial robots, and humanoid robots.

(Published: 2026-08-14)

AI Semiconductors & Computing Infrastructure

SMIC Q2 Net Profit Surges 3.6-Fold, Signals Q3 Wafer Price Hike

SMIC announced Q2 revenue of $3.01 billion and net profit of $479 million. Revenue rose 36% year-on-year, surpassing the quarterly $3 billion mark for the first time. Net profit was 3.6 times higher than the previous year's $133 million, nearly doubling the market consensus of $253 million. Gross profit margin rose to 25.3% from 20.1% in the previous quarter.

During the earnings call, CEO Zhao Haijun revealed that the company renegotiated prices with clients in Q1 and will apply additional hikes to processed wafers in Q3. Q2 wafer shipments reached 2.9 million units (in 8-inch equivalents), up 14% quarter-on-quarter, while average selling price (ASP) rose by 5.7%. He explicitly stated, "There is no possibility of price cuts." He explained that a "spillover effect" is occurring, with AI driving up the prices of peripheral chips as well as AI chips themselves.

SMIC is the only Chinese foundry mass-producing 7nm chips. The mass production of Huawei's Ascend 910B is considered a core pillar of its revenue growth. The company expects Q3 revenue to grow by 2% to 4% quarter-on-quarter.

South Korean Implications: SMIC has simultaneously demonstrated 7nm mass-production capability and pricing power. This signals that the Chinese foundry ecosystem has reached a stage where it can grow solely on domestic demand. While Samsung Foundry's opportunities in the Chinese market are shrinking, the technological gap in advanced processes below 3nm remains.

(Published: 2026-08-14)

[Follow-up] Deployment of Huawei Ascend 950DT on Huawei Cloud Underway in August

Huawei is currently deploying the Ascend 950DT on Huawei Cloud, on schedule for August. Featuring 144GB of proprietary HBM (HiZQ 2.0), a memory bandwidth of 4TB/s, and an interconnect speed of 2TB/s, it is expected to be 2.5 times faster than the previous generation 910C. It supports FP8 and MXFP4 low-precision operations. The next-generation CloudMatrix aims to combine more than 512 NPUs into a single supernode to deliver over 500 PFLOPS of total compute power. DeepSeek is being mentioned as a candidate adopter.

(Published: 2026-06-08)

[Follow-up] Cambricon's H1 Revenue Reaches RMB 5.996 Billion, Up 108%

Cambricon recorded H1 2026 revenue of RMB 5.996 billion (up 108% YoY) and net profit of RMB 2.311 billion (up 123%). Cloud products accounted for 99.98% of total revenue. Porting for major Chinese AI models—including DeepSeek, Qwen, Kimi, and MiniMax—has been completed, and order books are full through 2027. Although the growth rate slowed from 159% in the previous quarter, it exceeded the total revenue of the previous fiscal year in just six months.

(Published: 2026-08-07)

[Follow-up] Moore Threads Pursues Dual Listing in Hong Kong

Chinese GPU developer Moore Threads is pursuing a dual listing on the A-share market and in Hong Kong. H1 revenue reached RMB 1.736 billion (up 147% YoY), while net loss narrowed by 96% year-on-year to RMB 11.6 million. Similar to Cambricon, it surpassed its full-year revenue of the previous fiscal year in just six months. Chinese AI semiconductor firms are rapidly scaling up by raising funds in capital markets.

(Published: 2026-08-09)

Telecom, Networks, and Edge AI

[Follow-up] China Mobile's 5G-A High-Speed Rail Premium Service Commercially Active in 31 Provinces
China Mobile, in partnership with Huawei, is operating its 5G-A-based high-speed rail premium service commercially across 31 provinces nationwide. It delivers downlink speeds of 2Gbps and uplink speeds of 150Mbps, supported by over 3,000 base stations. The service initially launched on the Beijing-Shanghai High-Speed Railway (Jinghu Line) and will expand to all routes by the end of the year. By offering exclusive badges to VIP subscribers, it represents a model for telecom monetization of 5G-A.

(Published: 2026-06-25)

[Follow-up] China Unicom Launches World's Largest Commercial 5G-A Uplink Network in Beijing
China Unicom, in partnership with Huawei, has deployed a commercial 5G-A uplink network of 10,000 base stations in Beijing. It features a peak speed of 1Gbps and an average speed of 397Mbps. This is the world's largest commercial deployment using the Supplementary Uplink (SUL) scheme. The telecom infrastructure is designed to prepare for the edge AI era, targeting the uplink demand of AI glasses and agentic services.

(Published: 2026-07-24)

Policy, Regulation, and National Strategy

[Follow-up] CAC Generative AI Service Registrations Reach 988, Nearing 1,000 Milestone

Generative AI services registered with the Cyberspace Administration of China (CAC) have reached 988. The pace of registrations has accelerated ahead of the September 30 deadline. Apple Intelligence has also been included in the registry. Since the provisional measures for managing AI anthropomorphic interaction services took effect on July 15, companies like ByteDance and Alibaba have reduced emotional chatbot features.

(Published: 2026-08-04)

[Follow-up] DeepSeek Pursues Construction of 1GW Data Center in Ulanqab, Inner Mongolia

According to Bloomberg, DeepSeek is building a 1GW data center in Ulanqab, Inner Mongolia. Targeting partial operation by late 2027 or early 2028, the company will utilize both self-owned facilities and leased capacity. The location was chosen to save on cooling costs due to its low average annual temperature. This highlights a trend where Chinese AI firms are expanding investments beyond model development to physical infrastructure.

(Published: 2026-08-01)

US-China Tech Hegemony and Impact on South Korea

[Follow-up] US FCC Drafting Ban on Imports of Chinese Optical Transceivers for Data Centers

The US Federal Communications Commission (FCC) is drafting a proposal to ban imports of new models of Chinese-made optical transceiver modules. Optical transceivers are critical components that transmit signals using light within data centers. The ban targets implementation within this year. Approximately 60% of optical transceivers used in AI data centers are manufactured by Chinese companies, with Zhongji Innolight accounting for 27% of global revenue.

If implemented, this ban will increase AI infrastructure costs for US hyperscalers like Amazon, Microsoft, and Google, as Western alternatives like Lumentum and Coherent cannot fill this volume within one to two years. Even Western-made transceivers rely on Chinese-sourced Indium Phosphide (InP) substrates, creating a deeply entangled supply chain. Following the announcement of the draft, Zhongji Innolight's stock price fell 14%, while Lumentum (7%), Coherent (11%), and AOI (18%) saw gains.

South Korean Implications: This presents both an opportunity and a risk for South Korean optical communication component manufacturers. While a US ban on Chinese modules could boost demand for South Korean alternatives, South Korean firms also rely heavily on Chinese raw materials, meaning supply chain restructuring must precede any major gains.

(Published: 2026-08-04)

[Follow-up] Nvidia Shifts Resources to Next-Gen Feynman; TSMC Accelerates Advanced Packaging Expansion

According to DigiTimes, even during the mass-production phase of Vera Rubin, Nvidia is already shifting R&D and supply chain resources toward its next-generation Feynman platform. Feynman integrates TSMC's 2nm-class A16 process, 3D chiplets, SoIC, and Co-packaged Optics (CPO). TSMC is accelerating construction at Chiayi AP7 and Southern Taiwan Science Park AP8 to prepare production capacity for SoIC, CoWoS-L, and next-generation CoPoS. Mass production of CoPoS is scheduled for the second half of 2028.

South Korean Implications: As the advanced packaging competition intensifies, demand for SK Hynix's HBM and Samsung's back-end packaging technologies will grow. However, TSMC's move toward vertical integration—including packaging—poses a threat to South Korean back-end packaging firms.

(Published: 2026-08-14)

South Korea's MSIT Hosts Public Debate on 2nd Draft of AI Ethics Principles

The Ministry of Science and ICT (MSIT) and the Korea Information Society Development Institute (KISDI) held a public debate on AI ethics principles at the President Hotel in Seoul on August 14, unveiling the second draft. The framework consists of three core values (human dignity, social public good, and human sustainability) and seven principles (human-centricity, privacy protection, fairness/inclusion, accountability, safety, reliability, and transparency). These serve as non-binding voluntary regulations, with finalization expected within August. Civil society demanded improvements to damage relief procedures, while the industry requested detailed guidelines for companies.

South Korean Implications: China has already implemented rules on AI anthropomorphic interaction services (effective July 15), a generative AI service registry (988 services), and an algorithm registry. How South Korea's voluntary framework differs from China's mandatory regulations will be a key reference point when drafting the Basic AI Act.

(Published: 2026-08-14)

Attorney Kyung-jin Kim's Commentary

DeepSeek's price hike signals a structural shift in the Chinese AI industry. After months of locking in the market at prices below cost, the industry is entering a monetization phase. A severe server load of 8 trillion tokens per day exposed the physical limits of its low-price strategy. The $8 billion follow-on raise and the construction of a 1GW data center follow a textbook platform strategy: scale cheaply first, then charge.

Alibaba’s revenue-sharing license points in the same direction from a different angle. Under the banner of "open weights," it effectively imposes licensing fees on large-scale commercial users. Chinese AI firms are transitioning from the ecosystem-building phase via free models to the harvesting phase.

In robotics, the fact that industrial and commercial applications accounted for over 70% of shipments is more notable than the total of 19,100 units. It indicates that these machines are shifting from exhibition prototypes to working assets in factories and retail spaces. While the South Korean robotics industry is competitive in component technology, it faces difficulties keeping up with China's supply chain speed and mass-production costs. The US import ban might open a niche, but that door remains narrow without price competitiveness.

SMIC's earnings show that the Chinese semiconductor ecosystem has established a self-sustaining loop despite sanctions. Possessing both 7nm mass-production capability and pricing power signals that Chinese AI chip demand is not just growing individual firms, but increasing the density of the entire industry.

[Source Links]

DeepSeek API Price Hike

https://api-docs.deepseek.com/quick_start/pricing

https://qz.com/deepseek-api-price-increase-v4-peak-off-peak-081326

https://fortune.com/2026/08/13/deepseek-increases-prices-for-ai-services-by-multiple-times/

DeepSeek V4 Pro 0813 Official Launch

https://www.techtimes.com/articles/324241/20260813/deepseek-v4-pro-0813-goes-ga-benchmark-claims-await-independent-proof.htm

https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813

DeepSeek Harness v0.1 MIT Open Source

https://thenewstack.io/deepseek-harness-open-source-plugins/

https://technode.com/2026/08/14/against-claude-cowork-deepseek-opens-its-open-source-harness-to-developers/

Alibaba Qwen3.8-27B Released

https://finance.biggo.com/news/b3b5cb0c-d942-401f-ba61-2923b0c81857

Alibaba Qwen3.8-Max Open-Weight and Revenue-Sharing License

https://huggingface.co/Qwen/Qwen3.8-2.4T-A95B

https://qz.com/alibaba-qwen-open-source-revenue-sharing-080726

Tencent Q2 Earnings

https://www.cnbc.com/2026/08/12/china-tencent-earnings-q2-2026-gaming-ai-advertising.html

https://finance.biggo.com/news/US_TCEHY_2026-08-12

Moonshot AI Pre-IPO Round

https://kr-asia.com/moonshot-ai-targets-august-27-closing-for-pre-ipo-round-ahead-of-hong-kong-filing

Chinese Humanoid Robots Take 97% of Global Shipments

https://www.globaltimes.cn/page/202608/1367923.shtml

https://technode.com/2026/08/11/chinese-makers-accounted-for-more-than-97-of-global-humanoid-robot-shipments-in-h1-2026/

Unitree IPO Allocation Results

https://m.gasgoo.com/news/70469135.html

https://finance.sina.com.cn/jjxw/2026-08-06/doc-inimkwsx4538737.shtml

World Robot Conference (WRC) 2026

https://english.beijing.gov.cn/beijinginfo/sci/event/202607/t20260710_4756514.html

The 4th Shanghai Embodied AI Robot Industry Exhibition

https://sh.eirobotshow.com/

SMIC Q2 Earnings and Price Hikes

https://qz.com/smic-chip-prices-ai-demand-revenue-earnings-081426

https://asia.nikkei.com/business/tech/semiconductors/china-s-smic-says-ai-spillover-effects-boosting-peripheral-chip-prices

Huawei Ascend 950DT

https://www.huaweicentral.com/huawei-confirms-ascend-950dt-ai-chip-to-debut-in-august/

https://www.trendforce.com/news/2026/06/08/news-huawei-brings-forward-ascend-950dt-deployment-to-august-deepseek-v4-2-seen-as-potential-early-adopter/

Cambricon H1 Earnings

https://www.scmp.com/tech/big-tech/article/3363351/cambricon-posts-108-surge-first-half-revenue-amid-chinas-massive-ai-chip-drive

Moore Threads Hong Kong Dual Listing

https://www.scmp.com/tech/tech-trends/article/3363448/moore-threads-plans-hong-kong-listing-after-posting-147-jump-first-half-revenue

China Mobile 5G-A High-Speed Rail Premium

https://www.huawei.com/cn/news/2026/6/mwcs-chinamobile-high-speed-rail

China Unicom Beijing 5G-A Uplink

https://www.developingtelecoms.com/telecom-technology/wireless-networks/20550-china-unicom-and-huawei-boost-5g-a-uplinks-for-over-10-000-sites-in-beijing.html

FCC Drafting Ban on Chinese Optical Transceivers

https://www.bloomberg.com/news/articles/2026-08-04/us-drafting-ban-on-china-data-center-components-reuters-says

https://www.caixinglobal.com/2026-08-05/us-drafts-ban-on-chinese-optical-modules-exposing-mutual-supply-chain-risks-102471268.html

Nvidia Feynman and TSMC Advanced Packaging

https://www.benzinga.com/markets/tech/26/08/61201714/nvidias-next-gen-feynman-ai-platform-is-ramping-up-forcing-tsmc-to-race-ahead-report

https://www.digitimes.com/news/a20260814PD215/nvidia-tsmc-cpo-rubin-development.html

DeepSeek Inner Mongolia Data Center

https://www.datacenterdynamics.com/en/news/deepseek-planning-1gw-data-center-in-inner-mongolia-china-report/

South Korea's MSIT Public Debate on AI Ethics Principles

https://zdnet.co.kr/view/?no=20260814161240

https://www.fnnews.com/news/202608141218124650

CAC Generative AI Service Registration Status

https://www.stcn.com/article/detail/4011897.html

Search Keywords

DeepSeek V4 Pro price increase August 16, DeepSeek Harness MIT open source, Qwen3.8-27B Apache 2.0, Qwen3.8-Max revenue share license, Tencent Q2 AI capex RMB 51.8B, Moonshot AI pre-IPO $50B August 27, 中国 人形机器人 97% global shipments H1 2026, AgiBot Unitree humanoid rankings, 宇树科技 IPO 中签结果, WRC 2026 北京 August 19, SMIC Q2 profit triple wafer price hike, 华为 昇腾 950DT August cloud, 寒武纪 H1 2026 earnings, Moore Threads Hong Kong dual listing, China Mobile 5G-A high-speed rail premium, China Unicom Beijing 5G-A uplink, FCC Chinese optical transceiver ban, Nvidia Feynman TSMC advanced packaging, DeepSeek Inner Mongolia 1GW data center, 과기정통부 AI 윤리원칙

Generated at: 2026-08-15 07:00 KST

📚 근거 자료 (Sources) — 클릭하여 펼치기
KIMKJ.COM
#김경진 #김경진변호사 #중국AI #중국인공지능 #AI패권전쟁 #로봇 #휴머노이드 #구현지능 #반도체 #AI반도체 #화웨이 #DeepSeek #Qwen #통신 #5G #6G #미중기술패권 #한국산업전략 #AI정책 #AI규제 #kimkj #kimkjcom
kimkj.com Home
Scroll to Top
kimkj.com Home kimkj.com Home