🎙️ AI 日報 Podcast，擬人化 AI 禁令、Claude 5 代理浪潮、台積電晶片狂潮

Opening Remarks

Hello everyone, and welcome to "Mark's Tech Insights"! I'm Mark. Today we have some incredibly explosive topics to cover—from China's sudden ban on all anthropomorphic AI companions, to Anthropic declaring war on OpenAI with the powerhouse Claude Sonnet 5, and our very own TSMC delivering mind-blowing revenue results. We have a massive amount of information to unpack today, so grab your coffee, and let's dive right in!

Today's Top Stories

1. China's "New Anthropomorphic AI Regulations" Take Effect: Doubao and Tongyi Qianwen Permanently Remove AI Companion Features

Source: TechTimes (https://www.techtimes.com/articles/320525/20260715/china-ai-companion-law-takes-effect-doubao-qwen-shut-down-millions-lose-chat-data.htm）
Summary: China's latest interim administrative measures targeting "anthropomorphic AI interactive services" officially took effect on July 15. This regulation has forced ByteDance's "Doubao" and Alibaba's "Tongyi Qianwen" to permanently shut down their AI companion and virtual character features. Countless users' chat histories and meticulously customized character settings were wiped overnight, with absolutely no backup or transfer options provided.
Taiwan Perspective: This reflects the authoritarian regime's high anxiety over "unpredictable and emotionally bonding" AI outputs. For software developers in Taiwan, this serves as a warning sign, but it also means that free and democratic markets enjoy a more stable regulatory environment when developing "therapeutic" or "emotional companionship" AI applications.
Discussion Points:
- Does excessive government intervention in AI emotional bonding stifle innovation in next-generation human-computer interfaces?
- When users develop genuine emotions for an AI, yet a platform can "erase" its existence overnight, what kind of ethical issues does this raise?
Script Suggestion: Imagine the virtual sounding board you chat with every day after work suddenly being forcibly "formatted" by the government. For many people who rely on AI companionship, this is a major psychological blow. China's move this time is essentially out of fear that AI is becoming too human-like and too influential, which could impact public opinion. However, I believe this actually presents a massive opportunity for Taiwanese developers. Companionship AI is a fundamental need, and in a free environment like Taiwan, we can explore the warmth of human-computer interaction without constraints.

2. TSMC Q2 Revenue Hits Historic High: 3nm Capacity Fully Sold Out Through Year-End

Source: LLM Stats / ThursdAI (https://llm-stats.com/llm-updates）
Summary: TSMC announced its mind-boggling Q2 2026 financial results, with single-quarter revenue reaching a record high of $39.6 billion. June revenue alone skyrocketed by 68% year-over-year. This incredible growth is mainly driven by the fact that global cloud giants and AI model developers show no signs of cooling down their demand for chips. TSMC's most advanced N3 (3nm) family process capacity has been completely booked out by customers, with shortages expected to last through the end of this year.
Taiwan Perspective: This once again proves that Taiwan supports the global AI revolution at the "physical layer." As long as AI models still need training and inference costs still need to be brought down, global tech giants will have to line up in Taiwan with cash in hand.
Discussion Points:
- With 3nm capacity fully booked, how will this widen the gap between TSMC and competitors like Intel and Samsung?
- If hardware capacity is the physical limit of AI development, what kind of crowding-out effect will this have on small and medium-sized AI startups?
Script Suggestion: Every time I see TSMC's financial reports, I'm truly left in awe. The whole world is debating an AI bubble, but TSMC's revenue directly slaps those doubts away with hard numbers. Right now, whether it's Microsoft, Google, or Meta, who doesn't want more 3nm chips? As long as TSMC's capacity is bottlenecked there, it remains the master gatekeeper of the global AI industry. Taiwan sits at the center of this storm—this is not just an economic dividend, but our strongest technological silicon shield.

3. Anthropic Launches Claude Sonnet 5: Autonomously Operates Browsers and Terminals, Revenue Surpasses OpenAI for the First Time

Source: AI Weekly / Anthropic (https://aiweekly.co/ai-news-today/anthropic-news）
Summary: Anthropic has released its most "agentic" model to date, Claude Sonnet 5. This new model can not only understand instructions but also autonomously operate browsers, terminals, and various external tools just like a human, requiring minimal human supervision. Alongside this release, Anthropic announced that its annualized revenue is projected to reach $47 b i l l i o n, o f f i c i a l l y s u r p a s s i n g i t s o l d r i v a l O p e n A I^{'} s r a n g e o f$ 25 billion to $33 billion.
Taiwan Perspective: Taiwanese enterprises often hit a bottleneck when adopting AI, where it "can only chat, but can't get things done." Claude Sonnet 5's Computer Use capability will accelerate the adoption of automated AI Agents in Taiwan's manufacturing and financial sectors.
Discussion Points:
- With AI able to autonomously operate computers and input commands, what new challenges does this bring to cybersecurity defense?
- Anthropic's revenue overtaking OpenAI for the first time—does this represent a milestone victory for the strategy of "focusing on practical tools and the enterprise market"?
Script Suggestion: We used to joke that AI was "all talk and no action," but now Claude Sonnet 5 proves it can actually "roll up its sleeves and work" for you. It can open a browser, look up information, write code, and execute it in a terminal on its own. This is the era of AI Agents we've been talking about. What's even more shocking is that Anthropic's revenue actually overtook OpenAI! This tells us that instead of spending all day dreaming about the grand narrative of Artificial General Intelligence (AGI), delivering useful, practical tools to enterprises is the real way to make money.

4. Slashing Mind-Boggling Compute Costs: Anthropic Reportedly in Talks with Samsung to Develop Custom Inference Chips for Claude

Source: ThursdAI (https://thursdai.news/releases/2026-07）
Summary: To cope with a massive monthly compute bill of up to $1.25 billion, Anthropic is in early-stage talks with Samsung to co-develop custom inference chips designed specifically for Claude models. This potential partnership could allow Anthropic to reduce its reliance on third-party cloud providers and gain more autonomous control over hardware costs and supply chain stability.
Taiwan Perspective: While this is an opportunity for Samsung, it also highlights the trend of AI model giants moving toward "de-NVIDIA-ization" and "chip customization." This undoubtedly represents a massive long-term business opportunity for Taiwan's ASIC (Application-Specific Integrated Circuit) design companies like Alchip and Global Unichip.
Discussion Points:
- Will dedicated inference chips become standard equipment for future AI giants?
- Can Samsung leverage this partnership with Anthropic to regain ground in the AI chip foundry sector, where it has been suppressed by TSMC?
Script Suggestion: A compute bill of $1.25 billion a month—just hearing that number makes your heart skip a beat. No wonder Anthropic is in a hurry to work with Samsung on their own chips. This shows that future AI competition isn't just about algorithms; it's about "who can drive compute costs down to the absolute minimum." Although Samsung grabbed this negotiation opportunity, Taiwan's semiconductor design ecosystem actually has a massive opportunity here as well. In the future, the demand for this kind of custom chip will only grow, so keep a close eye on Taiwan's ASIC-related stocks.

5. Global Startup Funding Hits Record $510 Billion in H1 2026: AI Gobbles Up Nearly Half of All Capital

Source: Crunchbase News (https://news.crunchbase.com/venture/global-startup-exits-ipo-ma-soar-ai-q2-h1-2026/）
Summary: Global venture capital funding reached a staggering $510 b i l l i o n i n t h e f i r s t h a l f o f 2026, d i r e c t l y s u r p a s s i n g t h e t o t a l f o r t h e e n t i r e y e a r o f 2025. T h i s w a v e o f f u n d i n g w a s a l m o s t e n t i r e l y d r i v e n b y A I, w i t h O p e n A I a n d A n t h r o p i c a l o n e v a c u u m i n g u p$ 217 billion—accounting for 43% of total global startup funding. Meanwhile, nearly 40 new AI unicorns were born in the first half of the year.
Taiwan Perspective: This highly uneven "money magnet effect" is a major test for Taiwan's startup ecosystem. Since Taiwanese venture capital is smaller in scale, local startups will find it hard to go head-to-head with international giants on general large language models. They must pivot to developing more vertical, niche-market applications.
Discussion Points:
- With just two companies taking over 40% of global VC funding, what kind of crowding-out effect does this have on tech startups in other fields?
- Does such highly concentrated funding mean the AI industry has already entered a "winner-take-all" oligopolistic phase?
Script Suggestion: Half of all global funding went to just two companies—sounds crazy, right? This means the capital markets have put all their chips on OpenAI and Anthropic. For our startup community in Taiwan, definitely don't think about reinventing the wheel or building massive foundation models; that's just throwing money into a fire. What we need to do is leverage the infrastructure laid down by these giants to develop applications that solve specific industry pain points, like smart healthcare or smart manufacturing, using a "small steps, fast runs" strategy to survive and thrive in the gaps.

6. Open-Source Models Win Big: Qwen 3 and Llama 4 Scout Outperform Closed-Source Commercial Models

Source: Shakudo Blog (https://www.shakudo.io/blog/top-9-large-language-models）
Summary: The latest open-source LLM leaderboard released in July 2026 shows that open-source models such as Alibaba's Qwen 3, DeepSeek R1/V3, Meta's Llama 4 Scout, and Mistral Large 3 have completely surpassed contemporary closed-source commercial models in benchmarks like code generation, mathematical reasoning, and long-context processing. Analysis indicates that for the vast majority of enterprise application scenarios, the technical gap between open-source and closed-source models has ceased to exist.
Taiwan Perspective: This is fantastic news for Taiwanese enterprises. Taiwanese financial institutions and government agencies, which previously hesitated to send data to OpenAI due to security concerns, can now confidently deploy these powerful open-source models on-premises.
Discussion Points:
- As open-source models catch up to closed-source ones, how will this reshape the business models of cloud giants like Microsoft and Google?
- When enterprises choose between "on-premises deployment of open-source models" and "cloud subscription to closed-source services," how will the trade-offs between cost and security change?
Script Suggestion: People used to think free, open-source models were just toys, and that you had to pay for GPT-4 to get anything useful. But now, the wind has completely shifted! Open-source models are actually beating closed-source ones in coding and math. For many traditional industries, medical institutions, and government units in Taiwan that value privacy and security, this is truly a godsend. We no longer have to worry about data leaks from sending data to the US; we can set up these powerful open-source models directly on local servers in Taiwan. It's both secure and cheap, and this is the real key to making AI ubiquitous in Taiwan.

Closing Paragraph

Well, after listening to today's news, don't you also feel that the pace of change in the AI industry is fast enough to take your breath away? From regulatory iron fists and hardware limits to the battle of the century between open and closed source, this grand AI drama has clearly just begun. Thank you all for listening today. If you like our show, please remember to subscribe and share it with your tech-interested friends. I'm Mark, and I'll see you next time. Bye-bye!