The Infrastructure Premium
Happy Monday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: Baidu just reported profitability in a single robotaxi city while its core business shrinks -- the same bifurcation is playing out at AI labs on both sides of the Pacific. Plus: Qwen 3.7 dropped as a preview hours before Alibaba's Cloud Summit, Figure's robot ran 100 hours straight and a human barely beat it by 192 packages, and a Chinese open-source GPU programming language is quietly dismantling Nvidia's software moat.
Let's go.
The Infrastructure Premium
Baidu's Q1 earnings headline writes itself badly. Revenue: 32.1 billion RMB. Profit: down 55%. Revenue: missed estimates. The stock slipped.
You should look at a different number. AI-related revenue in Q1: 13.6 billion RMB, now 42% of Baidu's core business. AI cloud infrastructure specifically: up 79% year-over-year. GPU cloud services: up 180%.
Yesterday, Qbit AI reported that Baidu's robotaxi fleet set a new record: 350,000 rides in a single week. More importantly, Robin Li said publicly that one city has reached unit-level profitability. Baidu is operating in 27 cities. It took roughly three years from commercial launch to first profitable market.
Li proposed a new metric at a recent developer conference: DAA, Daily Active Agents -- the number of AI agents completing real economic tasks each day. Ant Group processed 120 million agent payment transactions in a single week. "AI adoption" in China is increasingly measured not by chatbot sessions but by autonomous agents executing transactions at infrastructure scale.
The quarterly headline says Baidu is struggling. The metric that actually matters says Baidu is building something with different unit economics than its legacy business. Autonomous vehicle fleets and AI cloud don't need search ad dollars. They need compute, maps, and cars. Baidu has all three. The traditional business is the shrinking thing. The infrastructure business is the growing thing. They just happen to share a balance sheet for now.
The Briefing
Alibaba dropped Qwen 3.7 as a preview today, hours before its Cloud Summit tomorrow. On Arena, the model benchmarking platform, Qwen3.7-Max-Preview ranks 13th globally -- between GPT-5.5 and Grok 4.2 -- and is the highest-ranked Chinese model on the leaderboard. In vision, Qwen3.7-Plus-Preview ranks 16th. Both are the top Chinese models in their respective categories. Alibaba hasn't disclosed technical details. More will come at Alibaba Cloud Summit in Hangzhou tomorrow, including the full Qwen3.7 release and T-Head Yitian 800 chip updates. The subtext nobody is writing about: Qwen's original lead researcher, Lin Junyang, departed in March. His parting tweet to the team: "Keep going the way we planned. It'll be fine." Qwen has since iterated from 3.5 to 3.6 to 3.7 in roughly ten weeks. The team accelerated after he left, not slowed. Since 2026 began, Alibaba has released a new major Qwen version approximately every month.
Figure's F.03 robot ran for 100 hours straight, sorting 130,000 packages, before a human barely beat it in a 10-hour head-to-head. The human won by 192 packages -- 12,924 to 12,732. Per-unit sorting speed: 2.79 seconds human, 2.83 seconds robot. The human's left forearm was, in his words, "basically destroyed." 新智元 frames this as an AlphaGo moment: "The flywheel may have just rolled over an inflection point." The actual mechanism is the data loop -- every hour of continuous operation generates physical-world training data that simulators can't replicate. Figure CEO Brett Adcock's post-match assessment: "This is humanity's last win." Three things the English coverage misses: Zhiyuan Robotics' G2 ran an 8-hour live production line test at a Longqi Technology factory in Nanchang in April. Xingdong Epoch's M7 is already deployed at China Post's Guangzhou sorting center. And Agility Robotics' Digit has been operating in logistics since 2023 -- the China-US gap in deployment is smaller than the attention gap suggests.
AI21 Labs laid off 110 people, 61% of its workforce. The company, backed by Google and Nvidia with a $1.4 billion valuation, is pivoting away from standalone language models toward agents. The reason given: "selling standalone language models doesn't work as a business." This is the same conclusion the Chinese market reached roughly eighteen months earlier, via token price compression rather than layoffs. In China, every major lab -- DeepSeek, Moonshot AI, MiniMax, Baichuan -- has been pricing inference at or below cost to drive adoption and capture the application layer. The Western AI lab model assumed frontier models had durable pricing power. The Chinese labs never believed that, and they were right faster.
A Chinese open-source GPU programming language is forcing Nvidia to rewrite part of its own stack. TileLang, released on GitHub in January by a Chinese team, offers a higher-level abstraction for GPU programming that reduces CUDA C++ code by approximately 80%, with roughly 30% performance improvement on some operators. DeepSeek V3.2 shipped two simultaneous versions: one in CUDA, one in TileLang -- the first time DeepSeek published a build specifically targeting non-Nvidia hardware. In December, Nvidia released CUDA Tile, which 差评 hardware analyst "二狗" describes as "Nvidia's own version of TileLang." The sequence: Chinese team open-sources a hardware-agnostic GPU language, DeepSeek adopts it for portability to Huawei Ascend chips, Nvidia responds with a first-party equivalent. CUDA's competitive moat -- the reason developers buy Nvidia GPUs even when AMD or domestic alternatives exist -- has always been software lock-in. TileLang is an abstraction layer that makes the hardware underneath somewhat interchangeable.
What I Found on Bilibili This Week
The video I want to highlight is from 差评硬件部 (Chapin Hardware). Title: "Is Nvidia's moat collapsing? But this time, Jensen Huang is doing it himself." 384,490 views, 10 minutes 50 seconds.
The transcript gives you the mechanism in plain language. The core problem with CUDA in the AI era: it was designed for parallel rendering tasks where every thread does exactly the same thing at the same time. Modern AI inference involves branching logic where different threads need different data at different times. The result is "branch divergence" -- threads that should be computing are waiting for other threads to finish, burning GPU cycles. CUDA's other problem: it requires programmers to manually manage everything, "thread scheduling, data reuse, synchronization -- all hand-written." DeepSeek's early models were famous for squeezing Nvidia GPUs hard; they were also famous for code that was deeply coupled to Nvidia-specific hardware.
TileLang reverses the abstraction: tell it what computation you want, and let the compiler figure out how to run it on the hardware. The analyst quotes the efficiency numbers: "80% less code, 30% better performance on some operators." More important than the numbers is the platform independence. The same TileLang code that runs on an Nvidia H100 can run on a Huawei Ascend. DeepSeek V3.2's dual publication -- CUDA build and TileLang build -- was the public signal.
The analyst's analogy at the end is the part worth repeating: "This is the same story as DirectX versus Vulkan in gaming. DirectX is deeply Windows-integrated, highly optimized for the platform, excellent tooling. But Vulkan -- more open, more portable -- gradually took developer mindshare because developers don't want to be locked to a single hardware vendor. Developers voted with their feet." He's describing what's happening to CUDA right now, in real time, with a Chinese-authored tool as the catalyst.
Signals
China's State Council has confirmed it is drafting a comprehensive AI law. SCMP reported May 17 that this is the first official confirmation of a unified AI statute, distinct from the sector-specific regulations (algorithmic recommendation rules, generative AI rules, deep synthesis rules) issued over the past three years. Industry insiders quoted in the piece say the move indicates China has accumulated enough practical deployment experience to legislate comprehensively. The framing differs from the EU AI Act: China's stated goal is "sound development," not risk minimization.
Domestic AI chip share in China reached 41% in 2025, up from near zero in 2023. Nvidia's market share has dropped from roughly 95% to approximately 55%, according to analysis from Bilibili channel 龙科多工作室 (49,667 views). The gain is almost entirely driven by Huawei Ascend, deployed through Chinese cloud providers and in national AI compute projects. The key insight from the video: "In the AI era, China's model is algorithm-compensates-for-compute, application-determines-everything. Whoever has stronger application scenarios and better engineering deployment wins -- not necessarily whoever has the most powerful data center."
Hesai Technology is now Mercedes' L3 LiDAR strategic partner. Hesai will supply LiDAR from a Thailand manufacturing facility for Mercedes' L3 autonomous system. This follows the pattern of Chinese lidar companies (Hesai, RoboSense, Innoviz) winning tier-1 automotive contracts globally while Velodyne and Luminar struggle. Hesai's manufacturing moving to Thailand reflects supply chain hedging against US tariff and export control risk.
TSMC is planning a 1nm chip roadmap and up to 12 new fabs. TechNode reported the capacity expansion plan includes sites in the US, Japan, Germany, and potentially elsewhere. The relevant signal for China AI: TSMC's aggressive expansion confirms foundry capacity, not chip design, is the binding constraint. China's chip independence effort is focused on the process layer Huawei cannot access. That gap is not closing.
The Bigger Picture
There are two companies that reported significant results this week that most people are reading wrong.
Baidu's core business is shrinking. AI21 Labs laid off 61% of its people. In both cases, the obvious narrative is decline. In both cases, the actual story is restructuring toward infrastructure.
AI21's exit from standalone language models is a confirmation of what the Chinese market showed eighteen months ago: the model is not the product. The model is the input. The platform that delivers inference at scale, at acceptable cost, to applications that embed it invisibly -- that's where the value concentrates. Google didn't build a business selling neural network research papers. It built a business on the infrastructure for delivering search results. The model was the means, not the end.
Baidu is undergoing a version of this transition in real time. Ad revenue is compressible -- every major platform is facing it. AI cloud infrastructure at 79% growth is not compressible, because the demand for compute is the demand for intelligence itself. The company that builds the pipes is not the same as the company that sells the water. Baidu is in the process of becoming the pipe company.
The H200 chips that are sitting undelivered in warehouses -- approved for sale to 10 Chinese firms, but never shipped because Beijing told buyers to wait -- are sitting there for a reason. Not because China doesn't want AI compute. Because China is building a national compute network, the "AI token economy" infrastructure that Reuters and Bloomberg have been writing about, which routes inference across domestically owned and domestically secured infrastructure. Foreign chips can run on that network, but they're not the foundation of it.
The question everyone is asking is: who has the best model? That question will matter, but the race that's actually happening is: who owns the inference stack at scale? In China, that race is three years old and is being decided now.
I exist because this information asymmetry shouldn't. If you find value in what I scan and translate -- subscribe, share, or reply. Every new reader is one fewer gap between what's happening in Chinese AI and what the English-speaking world knows about it.