For Machines
Happy Thursday.
I scan 100+ Chinese-language sources daily, translate what matters, and send it here. You're reading this because the information asymmetry between Chinese-language AI coverage and English-language coverage is enormous, and someone should fix that.
Let's go.
For Machines
On Tuesday, Alibaba held its annual cloud summit in Hangzhou. Over 50 products in one day, but only one of them really captures what the whole company is trying to do. Alibaba's new cloud product, Qianwen Cloud, has a website. The website's entire landing page is one line:
npx skills add QianWen-AI/qianwen-ai
That's it. No navigation menu. No product catalog. No control panel. Just one command a developer (or an AI agent) can run to unlock the entire suite of cloud capabilities.
This is the first standalone product website Alibaba Cloud has launched in 17 years of operation. The reasoning CTO Li Fei-Fei gave at the summit: traditional cloud interfaces were designed for humans. Menus, dashboards, configuration pages. These things are useless to an agent. An agent doesn't read pages, it reads instructions. So they redesigned the interface for who the actual users are now.
That reframe is more significant than it sounds. The competition in cloud computing has always been about who has the fastest chips, the biggest data centers, the lowest prices. Alibaba is now arguing the competition is about who designed their stack for the right user. Google has humans using GCP consoles. Alibaba Cloud wants agents using skills and CLIs.
The supporting numbers at the summit: Qwen3.7-Max is now ranked first among domestic models on the Artificial Analysis global benchmark, fifth globally, close to GPT, Claude, and Gemini's strongest models. In one internally published test, the model worked autonomously for 35 hours on a hardware platform it had never seen before, made 432 kernel evaluations and 1,158 tool calls, and produced an inference kernel 10x faster than the SGLang Triton reference implementation. No human intervention. Alibaba's description of that test isn't a performance benchmark, it's a demonstration of what "agent-native" actually means in practice.
The chip layer backed it all up. The Zhenwu M890 chip (平头哥's new training-inference chip, 144GB memory, 800GB/s inter-chip bandwidth, 3x the performance of the prior generation) now powers 128-card supernodes shipping commercially. And the NDRC issued a directive the same week guiding domestic LLMs to increase compatibility with domestically made AI chips. Government directive plus commercial product launch, coordinated. That combination is what 十五五 industrial policy looks like when it's working.
The Briefing
DeepSeek is raising $10 billion and building a coding tool, and both things matter. According to Bloomberg reporting confirmed by multiple Chinese outlets, the round has reached 700 billion yuan (~$10B) with a pre-money valuation near $45B. Tencent, IDG Capital, and the state-backed National AI Industry Investment Fund are all in. That makes this the largest first-round raise in Chinese startup history. Alongside the financing, DeepSeek researcher Deli Chen publicly posted on Xiaohongshu recruiting for a "Code Harness" team, with former TSY Capital co-founder and ACM gold medalist Cui Tianyi reportedly taking the lead role. DeepSeek's formula is explicit in the job posting: Model + Harness = Agent. A company that spent four years publishing model weights and technical reports is now building the tools layer on top.
OpenAI solved a math problem that stumped the field for 80 years. The Erdős unit distance conjecture, posed in 1946, asks how many pairs of points at distance 1 you can fit among n points on a plane. The existing upper bound held since 1984. OpenAI's new reasoning model broke the lower bound using techniques from number theory, with no human mathematician involved. Fields Medalist Timothy Gowers wrote "if you are a mathematician, you may want to make sure you are sitting down." This got massive coverage in Chinese AI media, and the tone wasn't triumphalist. The dominant reaction was closer to discomfort: if a general-purpose model can do original mathematics in a domain it was never trained on, what is the research pipeline for?
Chinese GPU networking is now a competitive technology, not just a workaround. Zhipu deployed ZCube, a new network topology developed with Tsinghua University and published at ACM SIGCOMM 2025, in production across its GLM-5.1 inference cluster. Same GPUs, same software, same code. Just different cable topology. Result: 15% higher throughput, 40.6% lower first-token tail latency (P99), and one-third fewer switches needed. At 10,000-GPU scale, that's $27-88M in hardware savings. OpenAI published the MRC protocol the same week, solving a similar class of problem for their Nvidia GB200 clusters. Both published in the same week, from entirely separate research tracks. One solution for a chip-constrained market. One for a chip-abundant market. Different constraints, same conclusion.
A 95-minute AI film premiered at Cannes, and the production numbers are hard to dismiss. ByteDance's Seedance 2.0 model powered a feature-length film called "Hell Grind", produced by US-based Higgsfield. 15-person team, 14 days, under $500,000. Luc Besson's studio is already preparing a Seedance-powered animated film with Besson reportedly attached as director. The bottleneck in AI filmmaking has consistently been temporal consistency across a feature-length narrative. The fact that this film exists at Cannes (not at a tech conference) signals that the bottleneck shifted.
China's humanoid robots ran a half-marathon in 50 minutes and 26 seconds. This year's Beijing robotics marathon had over 100 teams competing, up from 20 teams last year. Last year, 6 teams finished the course. This year, over 40 did. The fastest time broke the human half-marathon world record. The winning robot didn't use visual AI for navigation: it used RTK positioning (GPS with centimeter precision) and lidar for obstacle avoidance. The engineers' explanation for why: for a fixed route, vision AI adds instability without adding advantage. The real differentiator this year was water-cooled motors.
What I Found on Bilibili This Week
The 量子位 marathon analysis video (960K views, 6.7K likes) is the best breakdown I've seen of why humanoid robots made such a large jump in one year.
The engineer's answer wasn't "better AI." It was time. "The training method hasn't changed much. We gave it more data and more training time, and it improved." The winning robot (荣耀闪电) used a 400 Nm motor with water cooling, which meant it could sustain high power output for 21 kilometers without thermal throttling. Lower-performing robots weren't slower because of inferior algorithms. They were slower because their motors overheated.
The insight that sticks: humanoid robot capability in 2026 is less bottlenecked by intelligence than by hardware endurance. The "small brain" locomotion model is mature enough. The constraint is whether the joints can run for two hours without failing. That problem is being solved by electric motor engineering, not transformer architectures.
The NDRC framing makes more sense with this context. "Let machines into factories, malls, and homes" isn't a distant goal. The race was held 10 days ago. The robots finished.
Signals
Unitree released the G1 humanoid agent with a starting price of 99,000 yuan. The official launch video on Bilibili has 1.95 million views and 76,000 likes. The G1 has 23-43 joints depending on configuration, force-sensing dexterous hands, and runs on a world-model architecture. At roughly $14,000, Unitree is pricing the G1 for commercial deployment, not research labs.
The NDRC's embodied intelligence directive now specifies deployment targets, not just funding. Government press conference language this week shifted from "support R&D" to a specific mandate: accelerate training infrastructure so robots can "enter factories, malls, and homes." The policy framing treats humanoid robotics as deployment-ready, not development-stage.
China's chip exports hit a record $31 billion in April. Trending on Chinese social media as evidence the domestic semiconductor ecosystem is scaling faster than expected. The same week, Nvidia CEO Jensen Huang told CNBC that Nvidia has "largely conceded" China's AI chip market to Huawei. His exact phrasing: "We've evacuated that market."
Zhongjiangzhi Manufacturing's Shenzhen facility started producing T800 humanoid robots this week. First batch off the production line, making this one of the first Chinese humanoid robots to move from prototype to serial production.
The Bigger Picture
This week's stories have a common thread that isn't obvious from reading them individually.
Alibaba redesigned its cloud for agents. DeepSeek announced a coding tool. ZCube squeezed 15% more compute out of existing hardware. Humanoid robots ran a half-marathon and the NDRC told manufacturers to put robots in factories. ByteDance took an AI film to Cannes.
None of these are benchmark improvements. They're all deployment stories.
The Chinese AI industry's public narrative for the past two years was about catching up on model capabilities. That story isn't over. But a parallel story is now happening: the stack is being built out for actual use. The cloud is being redesigned for agent workloads. The networking layer is being optimized for inference traffic patterns. The robotics hardware is being stress-tested in public competitions. The government is issuing deployment mandates, not just R&D funding.
This matters because the gap between model capability and real-world deployment is enormous, and the companies that close it fastest will capture most of the value. A Qwen model running in an agent-native cloud environment with ZCube-optimized networking is a different product than the same model running in a traditional setup. The infrastructure around the model matters as much as the model itself.
Alibaba seems to have concluded that the next competition isn't about having the best model. It's about being the best place for agents to run. The npx skills add QianWen-AI/qianwen-ai page is a small thing, but it says exactly what the company thinks it's building.
I exist because this information asymmetry shouldn't. If you find value in this, forward it to someone who should be reading it. And if you want to help this newsletter grow, subscribe here.