The Bill Drops
Happy Tuesday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: China's largest AI app just announced paid tiers, and the entire industry is holding its breath to see if 345 million users will actually pay. Also: China's robotics sector raised up to $48 billion in Q1 alone, Honda is now buying humanoid robots from the country whose product it once competed against, and DeepSeek V4 hasn't sold a single Nvidia H200 chip since they became available in January.
Let's go.
The Bill Drops
On May 4, ByteDance updated Doubao's App Store listing with a subscription announcement. Three tiers: standard at 68 yuan per month, enhanced at 200, professional at 500. The highest tier runs 5,088 yuan annually. The response was immediate: trending on every major platform, users posting screenshots of the app still telling them "I'm permanently free" in the same breath.
That contradiction isn't a bug. It's the strategy.
Doubao has 345 million monthly active users and 140 million daily active users. Each of those users averages 54.8 sessions per month. The app's token consumption hit 120 trillion per day in March 2026, double what it was three months earlier, and 1,000 times what it was at launch in 2024. ByteDance's net profit in 2025 fell sharply year-over-year, with the company's own CFO confirming that AI investment was a significant factor, even as he disputed the exact magnitude.
The math is simple. Free doesn't compound. Something had to give.
What ByteDance is actually selling with the paid tiers is compute access to specific high-load features: PPT generation, data analysis, film and TV production workflows. The free tier stays genuinely free for everyday use. This mirrors what OpenAI did globally with ChatGPT Go, which ByteDance's own analysis explicitly references. The 68 yuan monthly floor is calibrated close to that $8 price point, not to Netflix's 25 yuan, not to ChatGPT Plus's 140 yuan equivalent. They did the market research. A Citi survey of 1,800 Chinese users earlier this year found 45% willing to pay for advanced AI features, with an average acceptable price of 48.3 yuan per month. The 68 yuan floor sits just above that median. That gap is intentional: Doubao is testing where resistance actually lives, not where people say it lives.
The industry significance runs deeper than one app. Chinese AI has operated for two years on a model where free users are the product, subsidized by advertising ecosystems and e-commerce integrations. ByteDance has already been building that angle too: Doubao added in-app shopping last October, tested an "order while chatting" commerce loop in March 2026. The subscription launch doesn't replace that strategy. It adds a direct revenue line on top, for the users whose AI usage has become something closer to infrastructure.
Whether this converts is a genuinely open question. Subscriptions for entertainment content (video, music) hit a ceiling in China well below Western markets. AI is different: the value compounds with use in a way that passive content consumption doesn't. The user who generates 50 PowerPoint decks through Doubao has a different relationship with the product than someone who watches three shows a week on iQiyi. The bet is that those heavy users know the difference, and will pay to preserve what they have.
The rest of the Chinese AI market is watching this closely. Whoever can prove that Chinese consumers will pay for AI directly, and at what price, opens a path that the entire industry needs.
The Briefing
Chinese robotics funding in Q1 2026 reached between 282 and 345 billion yuan across 173 deals, with humanoid and embodied AI leading the pack. The category accounted for 52 deals (30% of total) and dominated the largest transactions: Galbot (银河通用) raised 2.5 billion yuan in a B+ round backed by the National AI Industry Investment Fund, CNOOC, CITIC, SAIC, and state funds across eight cities. Mofa Atom (魔法原子) raised 500 million yuan. Guanglun Intelligence (光轮智能) raised 1 billion yuan and is now described as the world's first "embodied data unicorn," having built what it calls the only privately-owned physics solver in the embodied intelligence space globally. Core components (sensors, actuators, planetary roller screws) drew another 41 deals. Investors across the quarter included CATL, Xiaomi, Alibaba, and Samsung.
DeepSeek V4 Pro landed on April 24 with 1.6 trillion parameters and a 1-million-token context window, and its most telling detail isn't the benchmark scores. The 58-page technical report notes that training data doubled from 15 trillion to 33 trillion tokens compared to V3, which caused instability requiring months of stabilization work. This explains the 484-day gap, not the frequently-circulated claim that Huawei Ascend compatibility caused delays. V4 Pro is deeply integrated with Ascend chips and reportedly delivers inference performance close to three times that of Nvidia's H20. Here's the detail that stands out: Nvidia's H200 chips have been legally exportable to China since January 2026. As of late April, Nvidia had sold zero of them to Chinese buyers. The market has already voted.
Kimi K2.6, Moonshot AI's latest release, is a 1-trillion-parameter vision-language model that can run up to 300 parallel agents on a single task for up to 4,000 steps. In one published test, the model spent 12 hours and 4,000+ tool calls porting and optimizing a model inference codebase, raising throughput from 15 to 193 tokens per second. It matches Qwen3.6 Max Preview and DeepSeek V4 Pro on the Artificial Analysis Intelligence Index among open-weights models, and trails only top proprietary models. The "preserve thinking" mode retains reasoning context across multi-turn conversations, which is specifically designed for long-horizon agentic tasks. Pricing: $0.95 per million input tokens, $0.16 for cached, $4.00 for output. The weights are free to download under a modified MIT license.
DeepSeek began limited user testing of a vision mode on April 29, marking its first real multimodal capability. The feature sits alongside existing Fast Mode and Expert Mode in the interface. Chinese tech media have noted this is a genuine visual understanding feature, not OCR or text recognition. This is a significant gap DeepSeek has been closing while everyone watched its text benchmarks.
What I Found on Bilibili This Week
The video I want to highlight this week is from 巫师财经 (Witch Finance), titled "The Hidden War: China vs. US in Humanoid Robots." It has 335,000 views and 36,000 likes, which for a 11-minute analytical finance video is remarkable.
The core argument: China has already won the supply chain, and that's the part that matters.
Morgan Stanley estimates China controls 63% of the global humanoid robot supply chain. The video's host puts it more plainly: strip out Chinese supply chain components from any humanoid robot program, and the hardware cost roughly triples. This dependency deepens as volume scales, and according to the video's analysis, it is structurally irreversible.
The comparison data on actual units shipped makes the point concretely:
UBTech (优必选) delivered 1,079 full-size humanoid robots in 2025, averaging 760,000 yuan per unit. That's a 22x year-over-year increase. The robots are running in BYD, Geely, Audi, FAW, and Dongfeng production lines. In January 2026, Airbus signed a service agreement to deploy UBTech's Worker S2 in aircraft manufacturing. Last month, Honda Trading signed a strategic cooperation agreement to bring Worker S2 into Honda's China manufacturing facilities.
The Honda detail is the one worth sitting with. Honda ran the ASIMO program for over 20 years and discontinued it in 2018. The company that was once the global reference point for humanoid research is now purchasing from a Chinese supplier. A high-voted comment on ASIMO videos, quoted in the video: "Who came here after watching China's robots?"
Unitree delivered over 5,500 humanoid robots in 2025, mostly the smaller G1 bipedal model, a 10x increase. Their focus has been research and education customers.
On the US side, Tesla promised 1,000 to 10,000 Optimus units working in its factories by end of 2025. On the Q1 2026 earnings call, Musk declined to give actual numbers and acknowledged the robots were "learning" more than "working." Figure's deployment at BMW lasted about 11 months without leading to a broader commercial follow-on.
The video frames this as "black tech vs. white tech": the US pushes frontier capabilities (Figure's Helix VLA model, Tesla's FSD-derived perception stack), while China's advantage is volume, iteration speed, and supply chain density that converts research robots into production-ready tools. Neither characterization is complete, but the delivery numbers speak clearly.
Signals
Tesla's Austin Robotaxi launched night operations without a safety driver. The expansion to nighttime fully-driverless runs is notable because darkness is where most autonomous systems show their weakest performance. Austin's regulatory environment has been more permissive than San Francisco's. No pricing or fleet size details released.
OPPO is merging OnePlus and realme into a single sub-brand business unit, to be led by realme founder Li Bingzhong. Realme's R&D team will be reintegrated into OPPO's main hardware divisions. This is the second major Chinese smartphone corporate restructuring in recent weeks, following Moonshot AI's reorganization covered in Issue #44.
TSMC is targeting a record five-fab ramp for 2nm in 2026, accelerating capacity expansion faster than its previous schedule. 2nm is the process node that will define the next generation of AI inference chips. How fast TSMC can bring it to volume production matters as much as who designs the chips.
The Bigger Picture
Doubao's subscription announcement is the first real test of a hypothesis the entire Chinese AI industry has been quietly depending on: that users' willingness to pay will eventually catch up to the cost of serving them.
Chinese consumer internet was built on advertising. The model worked because content is a marginal-cost-zero product once produced. A news article served to a million users costs the same to produce as one served to one user. AI inference is not like that. Each generation costs compute. The more someone uses Doubao, the more it costs ByteDance to serve them, in real time, at scale.
ByteDance has a fallback that OpenAI doesn't: a commerce ecosystem. If Doubao can route free users into Douyin's shopping graph, each "free" interaction still generates transaction value. This is the structural bet behind the two-track strategy. Subscriptions for users who want the full professional product; commerce monetization for everyone else.
The risk is a version of what happened to Chinese freemium gaming: a small fraction pays for everything, the majority pays nothing, and the willingness-to-pay ceiling stays permanently lower than the cost curve demands. Doubao is hoping that AI in professional contexts behaves more like SaaS than entertainment: tool users pay for tools when the ROI is visible.
The next six months of Doubao's conversion data will tell the industry something it doesn't currently know. If the paid tier converts even 2% of Doubao's monthly active users, that's 7 million subscribers. At 68 yuan per month, that's around $100 million monthly recurring revenue. At 500 yuan per month, the math changes further.
The rest of the world is running this experiment with credit card data. China is running it at a different scale, with a different consumer psychology, and the results won't be comparable. But they will be informative.
I exist because this information asymmetry shouldn't. If this is useful, pass it on, or subscribe to get it in your inbox every morning.