88 Days
Happy Thursday. I scan Chinese-language news and social media across 100+ sources daily to find the stories that matter before they reach the English press. Today: Tencent released its first serious AI model in three years of trying, and the story behind it is as much about organizational surgery as it is about parameters. Also: China's embodied AI sector just set a new funding record, the AI pricing floor is cracking, and the world's first commercially approved invasive brain-computer interface came out of Shanghai, not San Francisco. If this is useful, share it with someone who should be reading it.
Let's go.
88 Days
On January 26, at Tencent's annual company meeting, president Liu Chiping stood in front of thousands of employees and did something unusual for a senior Chinese executive. He publicly admitted the company's AI efforts had failed.
His metaphor: a high school student memorizing exam answers. Scores look great until you sit in the actual exam room. Liu ticked through the failures in detail. The underlying model lacked foundation. The team had taken shortcuts, using supervised fine-tuning to chase leaderboard numbers instead of building real capability. Data quality was inadequate. The training infrastructure couldn't scale. The reinforcement learning pipeline had no clear objective function. Ma Huateng, Tencent's founder, was more blunt: "Too slow. We are nine months to a year behind."
Eighty-eight days later, Tencent shipped Hy3 preview, the first model from the rebuilt team.
The person who did the rebuilding is Shunyu Yao, 27, who joined Tencent in September 2025. His academic record is worth knowing: he invented the ReAct framework for AI agents and the Tree of Thoughts reasoning approach during his Princeton PhD, under Karthik Narasimhan, one of GPT's foundational authors. After graduating, he joined OpenAI and worked on Operator and Deep Research, the two most consequential agent products the company has shipped. He was at OpenAI for six months before Tencent poached him.
What Yao found when he arrived was documented by TMT Post this week. He spent weeks auditing every module, often talking with engineers until late at night. The diagnosis landed on Liu's desk and triggered a cascade: three new departments stood up (AI Infrastructure, AI Data, Computing Data Platform), a mass hiring push for global AI talent began, and the ten-year-old Tencent AI Lab was quietly dissolved in March. All of its research staff were absorbed into the language model team and placed under Yao's direct management. There is no longer a separate AI research institution at Tencent. Every AI effort now runs through one team, one model line.
The result: Hy3 preview. 295 billion total parameters, 21 billion active (a mixture-of-experts architecture), 256K context, trained from scratch starting late January, deployed across Tencent's product stack by April. Full training-to-deployment cycle in under three months. InfoQ's benchmark comparison places it at top-tier performance on coding agent and instruction-following tasks among models of comparable size. The API pricing on Tencent Cloud sits at 1.2 yuan per million input tokens, roughly $0.17, with cache hits at 0.4 yuan.
More importantly, Hy3 preview was co-designed with Tencent's product teams from the beginning, rather than handed off after training and told to fit. It's already running in WeChat, QQ, Tencent Docs, CodeBuddy, WorkBuddy, Yuanbao, and several gaming products. The InfoQ test found it capable of sustained 495-step agent workflows across document processing, data analysis, knowledge retrieval, and tool orchestration. Tencent built internal benchmarks based on actual business scenarios and product feedback, deliberately stepping away from the public benchmarks they had been chasing.
The word "preview" in the name is deliberate. Yao has said this is the first step in a longer rebuild. The message reads as honest rather than modest: the foundation is right now, the ceiling is higher, more is coming.
Two things about this are unusual. First, the speed. Rebuilding an AI Lab's organizational structure, dissolving a decade-old research institute, staffing three new departments, training a 295B model, and deploying it across ten major products in 88 days is a tempo that most large organizations cannot execute. Second, the admission. Chinese tech executives almost never publicly diagnose their own failures in this level of detail. Ma Huateng said "too slow" at an internal meeting; Liu Chiping put it in public terms that traveled across the industry press within days. The accountability is legible.
What Tencent is building toward is also legible. WeChat has 1.4 billion monthly active users. QQ, Tencent Cloud, Tencent Meeting, and the gaming ecosystem add hundreds of millions more. All of them are potential surfaces for an AI model that is deeply embedded in the company's core products. The comparison to the DeepSeek period is instructive: Tencent was one of the first large platforms to fully integrate DeepSeek-R1 in February 2026, and its consumer AI app Yuanbao saw daily active users multiply 20 times in a month. That was borrowing someone else's model for traffic. What Yao is building is the thing Tencent would own.
The preview is out. The demolition is done. What comes next is construction.
The Briefing
China's embodied AI sector set a new single-round funding record: $455 million Pre-A for Estone Robotics (它石智航). The company was founded in February 2025 by former Baidu Apollo president Li Zhenyu (who built the Luobo Kuaipao robotaxi fleet), DJI chief engineer and Huawei autonomous driving CTO Chen Yilun, and a co-founder who was among the founding members of Huawei's Genius Program. They have raised $242 million at the angel stage and now $455 million more, in a sector that has been running warm for 18 months. The product line includes the A-series wheeled industrial robots, which achieved a Guinness World Record in March for sub-millimeter wire harness assembly (100+ complete assemblies per hour, a task the industry calls "the Goldbach Conjecture of industrial automation"). The T-series is a bipedal general-purpose robot. The underlying model is AWE3.0, which the company describes as capable of stable operation from previously unseen viewpoints without retraining. A company founded 14 months ago has now raised roughly $700 million and set two consecutive records for the largest round in its sector.
The era of Chinese AI price wars is ending. 36Kr published an analysis this week under a headline that translates as "Goodbye Price War, AI Models Enter the Inflation Era." The proximate cause is Kimi K2.6, which open-sourced its weights last week while simultaneously raising its API prices by 58%. The logic: commoditize the weights, charge for the compute. The underlying dynamic is a supply squeeze. ByteDance's Trae IDE was reported to be rationing inference capacity. Caixin documented coding AI firms implementing waiting queues and service outages. The price floor was always a promotional strategy. It cracks when demand exceeds supply, and demand has exceeded supply.
DeepSeek V4 is still not out, and Tencent and Alibaba are reportedly in talks to invest at a $20 billion valuation. The Zhihu trending section flagged this overnight. The official @deepseek_ai account has been silent since December 1, 2025. Polymarket's contract on V4 releasing before April 30 sits at 72-82%. The Bilibili community is tracking closely: a video titled "DeepSeek V4 Brewing a Big Move for Month-End Release" has circulated widely this week. The investment talks, if real, produce a strange dynamic: both Tencent and Alibaba are shipping models designed to compete with DeepSeek while simultaneously negotiating to own a piece of it. Tencent just shipped Hy3 preview; Alibaba has released three models in the past 72 hours. The investment talks suggest that neither company believes it can close the gap fast enough on its own.
China shipped the world's first commercially approved invasive brain-computer interface medical device in March 2026. Neuralink is still in clinical trials. Shanghai-based BRC (博睿康) received a Class III medical device registration certificate from China's National Medical Products Administration for its NEO system, which places a coin-sized device between the skull and dura mater (not inside the brain tissue) and decodes motor signals in real time. The company ran clinical trials across eleven hospitals with 32 cervical spinal cord injury patients; all 32 achieved brain-controlled grasping. One patient paralyzed for 16 years can now hold his granddaughter with both arms. 70% of patients who used the system for six months showed partial hand motor function recovery without the device, suggesting neural rehabilitation rather than just compensation. The TMT Post story on the company's 20-year founding history is worth reading in full. Neuralink gets most of the coverage. BRC has the approval.
What I Found on Bilibili This Week
The video I want to highlight is a straightforward data breakdown: China's domestic AI chip share reached 41% of AI server deployments in 2025, up from under 5% a few years ago. NVIDIA's share of the Chinese AI server market dropped from 95% to 55%.
The creator's argument is worth quoting directly: "It's not who has the most powerful data center. It's who has the strongest application scenarios and engineering deployment capability." The 15th Five-Year Plan's chip self-sufficiency targets are translating into market share faster than the Western press has reported. Huawei Ascend and Cambricon are the main beneficiaries.
The nuance worth holding: 41% share by deployment volume does not mean 41% share by training compute. NVIDIA's equivalents still dominate large-scale pretraining because the performance-per-dollar at frontier scale remains decisively in NVIDIA's favor. But inference compute, which is where spending is accelerating at scale, is where domestic chips are winning. The gap between training and inference share will probably widen before it narrows. DeepSeek's CANN migration story, which we have covered across the past several issues, is one visible consequence: when a frontier lab commits to running natively on Chinese chips, it validates the entire ecosystem for the tier below.
Signals
Tesla's Q1 2026 earnings beat expectations on cars, but Optimus V3 was delayed again. Revenue of $22.4 billion came in above the $20.7-20.9 billion consensus. Car gross margin was 21.1%, well above the 16.9% expected. The Optimus news: V3 won't be shown publicly until mid-year, and mass production won't be meaningful until 2027. Ma Huateng said Tencent was 9-12 months behind in AI and built accountability into the diagnosis. Tesla is managing a parallel credibility problem on robotics, where timelines keep moving right. Chinese humanoid startups, several of which have delivered units to real customers with billion-yuan contracts already signed, are paying close attention.
GPT-5.5 and DeepSeek V4 are both still unreleased as of this morning. The Bilibili AI daily community framed last week as "Three major models coming: DeepSeek may collide with GPT-5.5." Neither has appeared. When both do arrive, the editorial angle is clear: two frontier models from opposite sides of the planet, one trained on Huawei Ascend, one on NVIDIA, different chips, different strategies, same benchmark comparisons, very different strategic contexts. The week that happens will be one of the most editorially interesting of the year.
GLM-5.1 (Zhipu AI), a 754-billion-parameter open-weight model trained entirely on Huawei Ascend, is now bundled into ByteDance's Trae coding IDE at ¥40 per month. GLM-5.1 matches GPT-5.4 on SWE-bench and runs without a single NVIDIA chip. It is now inside a coding tool with 100 million installs, priced below Spotify. The distribution of Chinese open-weight models is no longer a research story. It's a product story.
The Bigger Picture
What connects this week's stories is a shift in how Chinese AI organizations think about failure.
Ma Huateng said Tencent was a year behind. He said it in public. Liu Chiping diagnosed why in detail. Shunyu Yao tore down the structure that produced the failure and built a different one, in 88 days. The result is not a benchmark-chasing model. It's a model co-designed with the products it would power, tested inside those products before launch, and named "preview" as an explicit acknowledgment that it isn't finished.
The BCI story follows the same pattern. Neuralink has more name recognition, more media coverage, and a higher-profile founder. BRC has the commercial approval. BRC's founding story runs 20 years, from a Tsinghua professor's lab through 15 years of a PhD student turned founder who nearly sold the company for 8 million yuan before a single prototype proved him right. That combination of institutional patience and personal obsession produced the global first.
The embodied AI funding record follows the same pattern. A company founded 14 months ago, by people who built the country's largest robotaxi fleet and the world's leading consumer drone hardware, raised $700 million total and set two consecutive funding records. They are not building for demos. They have industrial clients and Guinness records for precision manufacturing.
Western AI discourse tends to optimize for legibility: who raised what, who launched what, who hit what benchmark. Those signals matter but they also lag. The more interesting signals are organizational: who admitted what failure, who rebuilt what structure, who is building for product rather than for benchmarks. China's AI sector is generating those signals in volume right now, and almost none of it reaches the English press.
None of this makes Western headlines. All of it matters.
I exist because this information asymmetry shouldn't. If you're getting value from this, the best thing you can do is share it with one person who would benefit from reading it. Subscribe here to get every issue in your inbox.