{
 "title": "China's Vertical AI Play: Robots, Chips, and Digital Companions",
 "description": "Today's essentials: Zhipu's GLM-5.2 open-sourced under MIT license with 1 million token context and simultaneous launch of ZCode 3.0 agent coding tool.\n\nKey stories covered:\n- China's tech dominance in five critical areas\n- Zhipu GLM-5.2 and China's AI model landscape\n- Humanoid robot price war concerns\n- Unitree's IPO and NVIDIA partnership\n- Robotera's real-world logistics deployment\n\nLeave a comment with topics from today's news you'd like us to dig deeper into. If you see the Hype button within 7 days of upload, hitting it helps us reach new viewers.\n\nThis broadcast is brought to you by lawyer Kyungjin Kim.\n\n#AI #ArtificialIntelligence #KyungjinKim",
 "link": "https://kimkj.com/%EC%98%A4%EB%8A%98%EC%9E%90-%EB%89%B4%EC%8A%A4-%EC%98%81%EC%83%81/?mod=document&uid=5974",
 "segments": [
  {
   "id": "S001",
   "slide": 1,
   "speaker": "Host",
   "text": "This broadcast is brought to you by lawyer Kyungjin Kim."
  },
  {
   "id": "S002",
   "slide": 1,
   "speaker": "Host",
   "text": "There's a market where 140 companies are churning out 330 types of humanoid robots. Caixin Global did an in-depth report on the price war fears in this market. Even the development and reform commission issued an overproduction warning, and that same week, Unitree's IPO was approved. Robotera has already deployed humanoid robots in actual logistics operations. Today, we're covering five threads running through all of this: the boom in the robot market, Zhipu's GLM-5.2 open source, Huawei's Ascend 950PR mass production, and regulations on AI emotional companions."
  },
  {
   "id": "S003",
   "slide": 1,
   "speaker": "Guest",
   "text": "There's so much here—from robots to chips to regulation—but is there a connecting thread?"
  },
  {
   "id": "S004",
   "slide": 1,
   "speaker": "Host",
   "text": "There's one thread running through it all. China's building its AI ecosystem vertically. Models get open-sourced, chips are ramped up through domestic production, robots move from factory showcases into real logistics operations, and regulation fills in every application layer. When telecom pricing for tokens shows up at next week's MWC Shanghai, the infrastructure picture comes completely into focus."
  },
  {
   "id": "S005",
   "slide": 2,
   "speaker": "Guest",
   "text": "So Zhipu released GLM-5.2 under the MIT license. That's the third model update in three months, right?"
  },
  {
   "id": "S006",
   "slide": 2,
   "speaker": "Host",
   "text": "It came out on June 13th. It supports a 1 million token context, and they released the ZCode 3.0 coding agent tool alongside it. ZCode 3.0 cuts the dependency on external frameworks and runs on its own agent kernel. Since it's MIT licensed, anyone can use it commercially. You're seeing a clear pattern: Chinese AI companies are bundling open source with their own developer tools to expand into global ecosystems."
  },
  {
   "id": "S007",
   "slide": 2,
   "speaker": "Guest",
   "text": "I heard Baidu's ERNIE 5.1 is also doing pretty well."
  },
  {
   "id": "S008",
   "slide": 2,
   "speaker": "Host",
   "text": "On LMArena's leaderboard, it hit global rank 4 with a score of 1,223. It's the only Chinese model in the top 10. Using a mixture-of-experts structure, they reduced parameters to one-third and trained it at just 6 percent of the previous model's cost, yet achieved this performance. In an environment where high-end chips are scarce, that's becoming China's survival strategy in AI: build competitive models with minimal computing."
  },
  {
   "id": "S009",
   "slide": 2,
   "speaker": "Guest",
   "text": "And global token consumption keeps climbing, right?"
  },
  {
   "id": "S010",
   "slide": 2,
   "speaker": "Host",
   "text": "According to OpenRouter, weekly call volume is 44.6 trillion tokens. That's 8 weeks in a row of growth. If inference demand keeps climbing at this pace, we'll see simultaneous intensification of chip demand, power consumption, and token pricing competition. Huawei Cloud just released dedicated agent infrastructure. They unveiled an AICS compute cluster, agent memory storage solutions, and the AgentSphere runtime environment all at once—and this is the first time any major Chinese tech company has systematically laid out dedicated infrastructure for agents."
  },
  {
   "id": "S011",
   "slide": 3,
   "speaker": "Host",
   "text": "This broadcast is brought to you by lawyer Kyungjin Kim."
  },
  {
   "id": "S012",
   "slide": 3,
   "speaker": "Guest",
   "text": "Caixin did quite a deep dive on the humanoid robot price war. Is it really that serious?"
  },
  {
   "id": "S013",
   "slide": 3,
   "speaker": "Host",
   "text": "We've got over 140 companies shipping 330+ products at this point. Anji Bot's VP Yao Maoqing said there are already clear signs of oversupply in both the humanoid and entertainment robot segments. Morgan Stanley is warning that production will vastly outpace sales, predicting industry consolidation, and even China's National Development and Reform Commission cited risks of product duplication and R&D overcapacity. When a government agency explicitly signals oversupply, that's a clear market signal."
  },
  {
   "id": "S014",
   "slide": 3,
   "speaker": "Guest",
   "text": "Didn't we see something similar with electric vehicles?"
  },
  {
   "id": "S015",
   "slide": 3,
   "speaker": "Host",
   "text": "In 2018, China had over 500 EV brands. Today, maybe 10 of them have meaningful revenue. Robots are likely to follow the same path. When the EV market consolidated, suppliers lost their customers first—and that pattern could repeat here."
  },
  {
   "id": "S016",
   "slide": 3,
   "speaker": "Guest",
   "text": "So what does that mean for Korean component suppliers?"
  },
  {
   "id": "S017",
   "slide": 3,
   "speaker": "Host",
   "text": "For Korean companies supplying reducers, motors, and sensors, it's a double-edged sword. With 140 companies ordering simultaneously, short-term volumes are definitely fat. But the moment price war kicks in, you get pressure to cut component prices. During consolidation, customers vanish overnight—EV suppliers already lived through that."
  },
  {
   "id": "S018",
   "slide": 4,
   "speaker": "Guest",
   "text": "Unitree passed their ChiNext IPO. How big is the funding round?"
  },
  {
   "id": "S019",
   "slide": 4,
   "speaker": "Host",
   "text": "42 billion yuan, roughly $616 million in funding. It's the first humanoid robot IPO on the A-share market. In 2025, they had 1.699 billion yuan in revenue and shipped 5,500 humanoid robots—number one globally. That same day, NVIDIA's Jensen Huang announced the Unitree H2 Plus as the reference robot for Isaac GR00T. It's fitted with a Jetson Thor powered by Blackwell GPUs, delivering 75 degrees of freedom and 2,070 FP4 TFLOPS of on-device compute."
  },
  {
   "id": "S020",
   "slide": 4,
   "speaker": "Guest",
   "text": "Isn't it unusual for NVIDIA to partner with a Chinese robot company amid US-China tensions?"
  },
  {
   "id": "S021",
   "slide": 4,
   "speaker": "Host",
   "text": "Robotics isn't yet a direct target of AI chip export controls. So for NVIDIA, robotics is a channel to maintain China market access. But if restrictions expand to robotics, this partnership ends immediately. Stanford, ETH Zurich, and UC San Diego are adopting this reference robot for research—you can see NVIDIA's play to lock in both academia and industry."
  },
  {
   "id": "S022",
   "slide": 4,
   "speaker": "Guest",
   "text": "And Unitree's not the only IPO, right?"
  },
  {
   "id": "S023",
   "slide": 4,
   "speaker": "Host",
   "text": "EngineAI filed confidentially for a Hong Kong IPO. They're valued at around 1 billion yuan, and they're rolling a humanoid robot off the line every 15 minutes at their Shenzhen factory. Linkbot is pushing for a US listing too. Capital markets are beginning to take humanoid robots seriously as an investment class."
  },
  {
   "id": "S024",
   "slide": 5,
   "speaker": "Guest",
   "text": "So Robotera actually put humanoid robots into real logistics operations?"
  },
  {
   "id": "S025",
   "slide": 5,
   "speaker": "Host",
   "text": "They deployed the Synbot M7 humanoid robot across 10+ logistics hubs for China Post and S.F. Express. At the Guangzhou Changgao facility, it processes 1,200 packages per hour. Human workers do 1,392 per hour, so there's still a 192-package gap. Humanoid robots that used to just walk around trade shows are now actually sorting cargo."
  },
  {
   "id": "S026",
   "slide": 5,
   "speaker": "Guest",
   "text": "What stands out is that the investors are logistics companies."
  },
  {
   "id": "S027",
   "slide": 5,
   "speaker": "Host",
   "text": "They've raised a total of $280 million, hitting a $1.4 billion valuation—unicorn status. S.F. Group, HSG, and IDG Capital led the funding. When logistics companies invest directly, it's not a tech demo—they're deploying this in production. Starting June 24th, MWC Shanghai is hosting a humanoid robot soccer penalty challenge, and Chinese companies like Anji Bot are entering en masse."
  },
  {
   "id": "S028",
   "slide": 6,
   "speaker": "Host",
   "text": "This broadcast is brought to you by lawyer Kyungjin Kim."
  },
  {
   "id": "S029",
   "slide": 6,
   "speaker": "Guest",
   "text": "Huawei's Ascend 950PR is entering mass production. What's the market reaction?"
  },
  {
   "id": "S030",
   "slide": 6,
   "speaker": "Host",
   "text": "They started mass production in April with a target of 750,000 units shipped in 2026. At 70,000 yuan each, that's less than one-third the price of NVIDIA's H200 at 250,000 yuan. ByteDance has ordered $5.6 billion worth, and Tencent and Alibaba are discussing additional orders. It packs 112GB of proprietary HBM, and the CANN Next architecture lowered the cost of migrating CUDA code."
  },
  {
   "id": "S031",
   "slide": 6,
   "speaker": "Guest",
   "text": "What's the actual market share of Chinese AI chips in China right now?"
  },
  {
   "id": "S032",
   "slide": 6,
   "speaker": "Host",
   "text": "As of Q1 2026, it's 52.3 percent. First time over 50 percent, and Ascend alone accounts for 37 percent of that. That's a measurable inflection point—breaking free from NVIDIA dependency. CXMT delivered 16nm HBM3 samples to Huawei, but there are reports the mass production timeline might slip. Their Shanghai back-end packaging facility is targeted to go live by end of 2026."
  },
  {
   "id": "S033",
   "slide": 6,
   "speaker": "Guest",
   "text": "And there are moves in materials too, I hear?"
  },
  {
   "id": "S034",
   "slide": 6,
   "speaker": "Host",
   "text": "SMIC and Huahong jointly launched a semiconductor materials supply platform in May. It's a structural push to reduce dependence on US-sourced materials. Building their own supply chain—not just foundry, but materials and equipment too—is playing for the long game on export restrictions. Chips, memory, and materials are all on parallel localization tracks."
  },
  {
   "id": "S035",
   "slide": 7,
   "speaker": "Guest",
   "text": "China just passed a regulation on AI emotional companion services?"
  },
  {
   "id": "S036",
   "slide": 7,
   "speaker": "Host",
   "text": "Five government bodies worked on it together: the Cyberspace Administration, the National Development and Reform Commission, the Ministry of Industry and IT, the Public Security Bureau, and the Market Supervision Administration. It takes effect July 15th. It's a dedicated rule covering AI emotional companions, virtual romantic partners, and virtual family members. Minors are banned from virtual romantic and family services, and under-14s need parental consent. For services targeting seniors, they require health usage guidance and safety warnings. AI emotional companionship has been a global regulatory blind spot, and China just became the first to write dedicated rules for it."
  },
  {
   "id": "S037",
   "slide": 7,
   "speaker": "Guest",
   "text": "I heard there's also a separate regulation on agents?"
  },
  {
   "id": "S038",
   "slide": 7,
   "speaker": "Host",
   "text": "On May 8th, those same three bodies released implementation guidelines on intelligent agent standards and applications. They defined AI agents as intelligent systems capable of autonomous awareness, memory, decision-making, and execution, and outlined 19 application scenarios. Creating dedicated governance for agents separate from 2023's generative AI rules is a global first. Registered generative AI services hit 868 by late April, and the algorithm registration process effectively functions as a market barrier to entry."
  },
  {
   "id": "S039",
   "slide": 7,
   "speaker": "Guest",
   "text": "What's happening on the telecom side?"
  },
  {
   "id": "S040",
   "slide": 7,
   "speaker": "Host",
   "text": "The Ministry of IT is implementing a three-year action plan linking AI and telecom. By 2028, they want to lift 1-millisecond latency coverage for computing power networks to 75 percent. It's physically linking telecom infrastructure with AI compute—and it's the foundation for edge AI and remote robot control. Next week's MWC Shanghai's two big pillars are telecom companies demonstrating token pricing and 6G standards talks. China's three telecom giants are consuming 140 trillion tokens a day—a 1,400x increase in just two years."
  },
  {
   "id": "S041",
   "slide": 8,
   "speaker": "Guest",
   "text": "Korean semiconductor exports to China surged. How much?"
  },
  {
   "id": "S042",
   "slide": 8,
   "speaker": "Host",
   "text": "Semiconductor exports to China in May jumped 243 percent year-over-year. Korea's trade balance with China swung from a $764 million deficit in December 2025 to a $3.8 billion surplus in May. Morgan Stanley data shows DDR5 prices up 682 percent, NAND up 807 percent, driving the export surge. Right now, China's on a buying spree for Korean memory—it's a boom."
  },
  {
   "id": "S043",
   "slide": 8,
   "speaker": "Guest",
   "text": "The key is how long this boom lasts, right?"
  },
  {
   "id": "S044",
   "slide": 8,
   "speaker": "Host",
   "text": "The answer depends on when CXMT's HBM3 goes into mass production. If it slips, Korea's memory exports to China stay hot, and if they hit the timeline, substitution begins. NVIDIA's also a variable. Jensen Huang refused to appear before a Senate hearing, and the US Commerce Department officially confirmed it's requiring licenses for AI chip exports to overseas subsidiaries of Chinese companies. CNAS called this a strategic inconsistency: acknowledging security needs while leaving loopholes open."
  },
  {
   "id": "S045",
   "slide": 8,
   "speaker": "Guest",
   "text": "What should we check first thing tomorrow morning?"
  },
  {
   "id": "S046",
   "slide": 8,
   "speaker": "Host",
   "text": "Two things. Is CXMT's HBM3 timeline on schedule, and is Huawei's Ascend hitting its 750,000-unit quarterly target? When those two numbers move, Korea's memory export trajectory shifts. In a market where Chinese AI chips already exceed 50 percent, NVIDIA's footprint shrinks, and Korean equipment makers and material suppliers in that ecosystem need to reassess their positions."
  },
  {
   "id": "S047",
   "slide": 8,
   "speaker": "Host",
   "text": "This broadcast is brought to you by lawyer Kyungjin Kim."
  }
 ]
}