The Direction, Not the Floor
Happy Friday.
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.
The Direction, Not the Floor
DeepSeek announced on May 22 that its 75% discount on V4-Pro API pricing will not expire. Starting May 31, the promotional pricing becomes the permanent pricing. Output tokens: $0.88 per million. Input (uncached): $0.44 per million. That's a quarter of what the API cost when V4-Pro launched.
The framing that matters comes from analyst Chon Tang: this pricing was built around domestic hardware constraints from the start, and DeepSeek's own technical report flagged further cuts as Huawei chip efficiency improves. The official announcement tweet hit 2.4 million views within 24 hours, with 7,235 trending posts on Chinese social platforms by evening, still accelerating.
This isn't a discount to drive adoption. It's a pricing reset that signals the direction of travel. "This isn't a floor," Chon Tang wrote. "It's a direction." Inference prices in China will keep falling as Alibaba, Zhipu, and ByteDance pile pressure on each other and DeepSeek undercuts them all. Enterprise software companies building on top of these APIs now need to price that assumption into their cost structures.
There's a Bilibili video this week with 435,000 views walking through DeepSeek's entire technical evolution from V1 to V4. I'll get to it below. The throughline the video captures: DeepSeek's consistent design constraint has never been capability. It's been cost. Every architectural innovation, from MoE to MLA to FP8 training, was an attack on the cost curve. The permanent price cut is the business expression of that same philosophy.
The Briefing
The 35-hour autonomous run from last week is a bigger story than it looked. Issue #62 covered Alibaba redesigning its cloud for agents, and mentioned almost in passing that Qwen3.7-Max ran autonomously for 35 hours on the Zhenwu M890, making 432 kernel evaluations and 1,158 tool calls with no human intervention. Analyst Poe Zhao's new framing of that demonstration is sharper: "The demonstration matters less as a model capability showcase and more as a consumption profile." What Alibaba showed isn't how smart Qwen is. It's the workload pattern the Zhenwu M890 hardware is designed to support: 35 hours of continuous inference, tool invocation, compilation, and iterative evaluation, all inside a single workflow. The software layer (Issue #62) runs on this hardware layer. The two together are what an agent-native stack actually looks like.
China and the US are launching a formal AI safety dialogue, but the two sides are describing different conversations. After Trump's May 13-15 China visit, both governments agreed to establish an intergovernmental AI dialogue. Treasury Secretary Bessent called it one of the three most important achievements of the trip, describing the scope as preventing AI from proliferating to non-state actors. Chinese Ambassador XIE Feng used different language: a "race to the top" on safety, "check brakes before setting off." That phrasing signals a broader agenda. Tsinghua's CISS researchers want talks focused on narrow technical exchanges: safety evaluation standards, joint scenario exercises, incident communication protocols. Fudan's CAI Cuihong put the condition plainly: China will not accept risk-assessment frameworks defined unilaterally by the US. The talks are happening. What they become depends on which framing wins at the table.
The US shelved its AI executive order the same week China announced a comprehensive AI law. Politico published the full draft of Trump's AI oversight order on May 22, hours after it was reportedly killed. David Sacks argued it would slow innovation. The core provision was voluntary government review of advanced AI models before release. Treasury and cyber director Cairncross had a public disagreement over scope. The EO died without being signed. That same week, China's State Council included "comprehensive AI legislation" in its 2026 legislative plan for the first time, covering data, algorithms, computing power, property rights, cybersecurity, and supply chains. The NPC has now listed AI legislation as a review item for the third consecutive year. Neither approach is obviously better. But one is moving forward and one isn't.
The export control data on chipmaking equipment is cleaner than you might expect. Analyst Lennart Heim published a detailed breakdown: SMIC's AI and datacenter chip revenue has been flat at roughly $300 million per quarter for three years. In the same period, TSMC's equivalent share tripled. SMIC's high-performance computing market share went down. TSMC's went from 44% to 61%. The argument isn't about chip access, where controls have been porous. It's about chipmaking equipment, where the controls have held. SMIC cannot build the manufacturing capacity to produce frontier chips at scale because it can't get the lithography equipment. That ceiling is real. The counter-argument, also real, is that DeepSeek's permanent price cut has nothing to do with SMIC's capacity. China can win the inference application layer without catching up in silicon fabrication. Heim's point and the DeepSeek story are both true.
What I Found on Bilibili This Week
The DeepSeek retrospective by Lau博士的云组会 (435,986 views, 15,427 likes) is the best single-video account of how DeepSeek got here.
The story: a quantitative trading firm that spent years building compute infrastructure to predict markets applied the same long-term mindset to model training. Liang Wenfeng studied math at Zhejiang University, built a quant fund during the 2008 financial crisis using machine learning, and accumulated the compute infrastructure that became DeepSeek's foundation. The pattern across V1 through V4 is cost as the non-negotiable design constraint. V3 was trained for the equivalent of $5.6 million in H800 compute. GPT-4 equivalent reportedly cost over $100 million. Every architectural innovation, MoE, MLA, FP8 training with high-precision accumulation, was an attack on that gap.
The V4 detail that lands: DeepSeek engineers use V4-Pro for their own agent coding work, rating it better than Sonnet 4.5 and close to Opus 4.6 in non-thinking mode. The company that built the model trusts it with their own development pipeline. The permanent price cut makes more sense with this context. They priced it for how they use it themselves.
Signals
EngineAI's T800 production line is live in Shenzhen with a 10,000-unit annual capacity. The facility makes T800 one of the first Chinese humanoid robots to move from prototype to serial production. Coverage on XRoboHub hit 45,000 views. Two Chinese humanoid robots have now reached commercial serial production in the same month: the T800 and Unitree's G1 (99,000 yuan, $14,000).
Huawei built a 122TB solid-state drive without using 100+ layer 3D NAND. TrendForce reports Huawei used proprietary Die-on-Board packaging instead of the advanced 3D NAND it cannot access under export restrictions. Result: 33% higher capacity density than conventional designs. It is a workaround, not a replacement. But it is one more case of Chinese hardware engineering around a technology gap rather than through it.
Trump approved the H200 for China. China isn't buying. The New York Times reported that not a single H200 has been purchased in China since Trump cleared it for export. Chinese AI companies are either locked into Huawei Ascend infrastructure or don't trust the approval won't be reversed. Domestic AI chip market share reached 41% in 2025, up from roughly 5% in 2022.
USTC published a quantum computing result in Nature. The Pan Jianwei group's Jiuzhang 4 is a 1,024-qubit optical quantum computer that solves Gaussian boson sampling 10^54 times faster than the world's fastest classical supercomputer. Published May 13. This is a 十五五 Tier 3 priority (future industries), so the timeline to practical application is long. The result matters as a demonstration that China's foundational research is producing publishable-in-Nature work, not just commercial deployment.
The Bigger Picture
Three things happened in China this week that rarely happen at the same time.
DeepSeek made inference permanently cheaper and signaled further cuts ahead. Alibaba's hardware ran 35 hours of autonomous operation without human intervention. The State Council announced it will write a comprehensive AI law, becoming the first major economy to put national AI legislation formally in its legislative plan this year.
At the same time, China agreed to formal AI safety talks with the US, with Chinese experts carefully establishing they won't accept US-defined frameworks as the starting point.
In Washington, the AI executive order that was supposed to be signed on May 21 was shelved because Treasury and Cyber couldn't agree on scope.
The contrast isn't "China winning, US losing." SMIC is still three generations behind TSMC on frontier chip manufacturing. Export controls on chipmaking equipment have held. Lennart Heim's data is real.
But the institutional picture looks different. China is writing law. China is entering bilateral talks on its own terms. China's hardware is running 35-hour autonomous workflows at commercial scale. The US spent the week debating whether a voluntary review process would be too burdensome.
The gap that's widening isn't in model quality or chip production. It's in the ability to coordinate and act. The same week DeepSeek reset global inference pricing, Premier Li Qiang toured Xiaomi's EV factory and a humanoid robot innovation center and told manufacturers to accelerate AI deployment in production lines. That's what coordinated industrial policy looks like when it's working.
The price cut wasn't the floor. It was the direction.
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.