AI BOARD
AI Board
AI tools, policy moves, and public decisions are tracked in plain language.
AI Daily Briefing 2026-07-17
AI Daily Briefing
Friday, July 17, 2026
Friday, July 17, 2026 (Asia/Seoul)
1. Nvidia and the Japanese government announced that they will build national AI infrastructure on the scale of 27,500 Rubin GPUs. They will create foundation models for robots and share the weights with domestic developers.
2. The Ministry of Science and ICT reported that it will open Korea's homegrown generative AI, "AI for Everyone," free of charge starting at year-end. Participating companies must use at least 50% Korean-made models.
3. TSMC's second-quarter revenue reached $40.2 billion, hitting the top end of guidance, and it raised third-quarter guidance to $44.6 billion to $45.8 billion. That means AI orders have not yet rolled over.
Mira Murati's Thinking Machines Lab released its first in-house model, Inkling, under the Apache 2.0 license. It uses a mixture-of-experts architecture that activates only 41 billion parameters out of 975 billion total parameters, and the context window is 1 million tokens.
Independent evaluator Artificial Analysis gave it an Intelligence Index score of 41, ranking it No. 1 among U.S. open-weight models. Nvidia Nemotron 3 Ultra scored 38, and Gemma 4 31B scored 29. In GDPval-AA v2, which measures agent capability, it posted an Elo score of 1238, ahead of China's Kimi K2.6 at 1190 and DeepSeek V4 Flash at 1189.
On average, it used 25,000 output tokens to solve the same tasks. GLM-5.2 required 43,000 tokens. At a 64K context baseline, the input cost is $1.87 per 1 million tokens.
Implications for legal practitioners: Combined with Apache 2.0 licensing and native audio input, this opens a path to process records and consultation recordings without sending them to an external API. But because the 975 billion weights must be loaded into memory, anyone in a single-GPU environment should first measure the speed of a low-bit quantized version before deciding.
The 2026 World Artificial Intelligence Conference will be held in Shanghai from July 17 to July 20. Exhibition space exceeded 100,000 square meters for the first time, and more than 1,100 companies will present more than 3,000 items. Of these, 300 will make their world debut. The volume of cooperation intent already reached 162 billion yuan.
President Xi Jinping will present China's position on AI development and governance in the opening keynote. Nine Nobel Prize and Turing Award winners have confirmed attendance, and Richard Sutton, a pioneer of reinforcement learning, will deliver a keynote address.
Huawei will display the physical Atlas 950 SuperPoD, the industry's largest supernode, for the first time. It will also release the air-cooled Atlas 850E and open-source Ascend AI foundational software.
The State Administration for Market Regulation approved and announced 508 major national standards, including 7 in the areas of artificial intelligence, robotics, and smart factories. They cover smart manufacturing, edge AI, and power-related directions.
Implications for legal practitioners: On June 26, China first laid the groundwork with seven national standards on "AI agent interconnection," covering agent identity and tool-calling specifications. It has now added manufacturing and device-side standards on top. Because China pairs regulation with standards, advice on entering the Chinese market must treat the question of which standards a system was designed to meet as a contract negotiation issue.
3. Administrative Oversight / Regulatory Investigation
The Cyberspace Administration of China accepted the registration of Apple Intelligence's generative AI service. Apple completed registration on July 8, and the authorities published the list on July 15. That removes the block that had kept functions unveiled in 2024 from launching in China.
The condition for approval was not Apple's own technology, but the addition of a local model. Alibaba confirmed that its Qwen model will be integrated into Apple Intelligence on iOS, iPadOS, macOS, and visionOS, handling text and image understanding and generation. Baidu is also developing functions for users in China. Light workloads will run on device chips, while long documents and multimodal generation will be handled by the Qwen cloud, with user data stored on servers inside China. Apple's second-quarter Greater China revenue was $20.5 billion. No formal launch date has been announced.
Implications for legal practitioners: It is now clear that the price of regulatory approval was the integration of a local model. In market-entry advice, the dual structure, "use our model, but layer in a local model inside China," needs to be built into the contract and data-flow diagram from the start. A fresh example has just shown that trying to add this design later can delay launch by two years.
4. New Technology / Product Launches and Practical Use Cases
See [1-2] for details. This also includes the move to open-source Ascend AI foundational software.
Nvidia and Japan's Noetra will build a 140-megawatt AI factory. It will include 27,500 Rubin GPUs and 13,750 Vera CPUs. Japan's Ministry of Economy, Trade and Industry is backing the effort through the physical AI national project FRONTia, and according to Tom's Hardware, Noetra received a government subsidy of 387.3 billion yen, about $2.4 billion.
This facility will train open multimodal foundation models for robots, digital twins, and industrial automation, and will broadly share pretraining weights with domestic developers. Participants include SoftBank Telecom, Toyota-backed Preferred Networks, and NEC. In its March AI robot strategy, Japan said it aims to secure more than 30% of the global AI robot market by 2040, worth $133 billion.
During the same visit, CEO Jensen Huang also announced collaborations with Toyota, Kawasaki Heavy Industries, Mizuho, Canon, Fujifilm, and RIKEN. Mizuho will build the largest on-premises AI factory in Japan's financial sector using DGX B200 systems.
Industry structure: In a phase where power and sites have become bottlenecks, the state supplies electricity and weights with tax money, and private companies use them. The early AI model, where universities shared mainframe time in computer centers, has returned at national scale. Nvidia has expanded this sovereign AI logic to more than 20 countries.
TSMC's second-quarter revenue was $40.2 billion, exactly matching the top end of previous guidance at $40.2 billion. Gross margin was 67.7%, and operating margin was 60.3%. Third-quarter guidance is revenue of $44.6 billion to $45.8 billion, gross margin of 65.0% to 67.0%, and operating margin of 56.0% to 58.0%.
Industry structure: When the upstream leader of the AI semiconductor supply chain raises its third-quarter outlook, that means hyperscaler ordering still is not signaling a slowdown. TSMC guidance is read as a leading indicator for Nvidia and AMD volumes.
Caution: Some figures reported by certain media outlets, including net profit, EPS, HPC revenue share, and additional Arizona investment, were not confirmed in the body text of TSMC's investor relations page. Do not cite them until they are checked against the original earnings release.
Bae Kyung-hoon, Deputy Prime Minister and Minister of Science and ICT, announced "AI for Everyone" on July 16 at the Yeongbingwan of the Blue House while reporting the ministry's second-half work plan. Unlike foreign AI services whose functions differ by pricing tier, it will be free to use without usage limits.
Applications from participating companies will be accepted until August 11 next month. The condition is that they use at least 50% Korean-made AI models. Kakao and LG Uplus have said they will participate, and Naver, SK Telecom, and KT are reviewing it. Beta is set for late September, with a formal launch within this year.
Next year, it will expand into an AI agent that helps people find and apply for government support, with one AI agent for every citizen. The Ministry of Science and ICT, the Ministry of Education, the Ministry of Labor, and 12 agencies in total will provide AI education to 5.14 million people within this year. In physical AI, the goal is to secure world-leading foundation models within three years and become the world's No. 1 by 2030. AI data center investments by SK, GS, Naver, and others were tallied at 550 trillion won.
Implications for legal practitioners: If "at least 50% Korean-made models" becomes a procurement requirement, the first point of contract dispute will be what metric measures model composition and who certifies it. If it is not settled whether a fine-tuned foreign model can count as Korean-made, and whether routing share is measured by number of calls or number of tokens, disputes will arise at the settlement and sanctions stage.
Review of U.S. AI Trends
Technology: Thinking Machines Lab released Inkling under Apache 2.0, and independent evaluation confirmed it as the leading U.S. open-weight model, with an Intelligence Index score of 41. This is an attempt by a U.S. lab to reclaim ground dominated by the Chinese open-weight camp. See [1-1].
Investment / Management: Nvidia is building national AI infrastructure with Japan and has expanded its sovereign AI logic to more than 20 countries. See [5-1].
Review of China AI Trends
Technology: The World Artificial Intelligence Conference opens today in Shanghai. It will feature 300 world-first unveilings, and President Xi Jinping will deliver the keynote. See [1-2].
Law / Regulation: The State Administration for Market Regulation approved seven national standards in AI, robotics, and smart factories. See [2-1].
Products / Use Cases: Apple Intelligence has won regulatory approval and will enter China with Alibaba Qwen layered in. Huawei will unveil the physical Atlas 950 supernode for the first time. See [3-1] and [1-2].
Commentary from Attorney Kim Kyung-jin
What changes in practice is the place of the contract. Until now, the main issue in AI adoption advice was where the data goes. Now the nationality of the model will determine procurement eligibility and market-entry approval. Apple paid for delaying that design by being unable to deploy its own AI for two years in Greater China, where it is selling $20.5 billion per quarter.
There are still blank spaces. Neither Korea's "50% Korean-made" nor Japan's "shared with domestic developers" has defined what exactly will be measured for the determination. Requirements without a standard come back later as audits and sanctions. For companies entering the solicitation, getting that definition in writing before August 11 is cheaper than fighting over settlement later.
https://blogs.nvidia.com/blog/japan-ecosystem-2026/
https://investor.tsmc.com/english/quarterly-results/2026/q2
https://www.asiae.co.kr/article/2026071608484535475
https://www.seoul.co.kr/news/society/science-news/2026/07/16/20260716500100
https://www.thepaper.cn/newsDetail_forward_33590752
https://finance.sina.com.cn/jjxw/2026-07-16/doc-inihyuqa4425955.shtml
https://finance.sina.com.cn/jjxw/2026-07-16/doc-inihyuqf8060596.shtml
https://www.ithome.com/0/967/862.htm
https://thinkingmachines.ai/news/introducing-inkling/
• Coinbase "95% to 100% of code written by AI": The Cointelegraph original publication date is July 14, outside the 24-hour window.
• Rumored July 17 launch of Gemini 3.5 Pro: There is no official confirmation from Google, and the API documentation has neither a model card nor pricing. It was excluded because only secondary-media speculation exists.
• Appellate decision in Anthropic v. Trump administration: The PBS original publication date is April 9.
• DeepSeek V4 full release on July 15: The release date could not be confirmed through a primary source. The V4 model appears in API documentation, but no release date is specified.
• FTC policy statement on AI accuracy suppression: It was published in the Federal Register on July 7, outside the window, and it was already covered earlier.
• Kimi K3, GPT-5.6 "Soul Ultra," Nvidia "Nemo Claw": These were excluded because the YouTube radar clues could not be confirmed through primary sources.
• No item was included in that category because there were no court rulings or new practical use cases verifiable within 24 hours.
AI news July 16 2026 / OpenAI Anthropic Google Microsoft AI announcement July 16 2026 / NVIDIA Jensen Huang Japan sovereign AI July 16 2026 / Japan 27500 Nvidia Rubin chips robots / AI lawsuit copyright court ruling July 16 2026 / AI startup funding July 16 2026 / AI regulation FTC Congress state law July 16 2026 / Gemini 3.5 Pro launch July 17 2026 / Thinking Machines Inkling benchmarks July 16 2026 / Claude Code workflow agents July 2026 / OpenAI ChatGPT release notes July 16 2026 / EU AI Act July 16 2026 Commission / 인공지능 July 2026 16일 AI 정책 발표 / 中国 人工智能 2026年7月16日 监管 大模型 / 世界人工智能大会 WAIC 2026 上海 7月17日 开幕 / 市场监管总局 人工智能 智能体互联 国家标准 2026年7月 / 苹果 Apple Intelligence 中国 上线 备案 通义千问 2026年7月16日 / DeepSeek V4 正式版 上线 2026年7月16日 / 中国 AI 芯片 融资 投资 2026年7月16日
Generated at: 2026-07-17 06:00 KST
📚 Source Materials — click to expand