The Floor
Happy Thursday. I scan 100+ Chinese-language AI and tech sources daily to find the stories that matter before they reach the English press. Today: three major robot hardware announcements landed on the same morning, State Grid ordered 8,500 robots for infrastructure deployment, and China Telecom pivoted its business model from selling bandwidth to selling AI inference credits. The hardware side of China's AI economy moved fast this week.
Let's go.
The Floor
On the morning of April 30, the start of China's Golden Week holiday, Unitree released a dual-arm humanoid robot priced at 26,900 yuan. That's roughly $3,700. Six months ago, the cheapest comparable platform cost more than three times that.
The number caught everyone's attention. What landed around it made it a pattern.
Also on April 30, China Mobile unveiled its LingXi robot family at the 9th Digital China Construction Summit in Fuzhou. The lineup includes a humanoid at 229,000 yuan per unit (or 5,199 yuan per day to rent, technician included) and a quadruped robot dog at 39,900 yuan. China Mobile built its own inference framework for the family — "VLA-RAIL," a real-time asynchronous approach it says improves autonomous task efficiency 1.5x over remote operation — plus a proprietary navigation system. A state telecom company is now shipping robots and selling them the way it used to sell data plans.
The procurement orders are larger still. State Grid, which operates power infrastructure across 26 of China's 31 provinces, earmarked 6.8 billion yuan for embodied intelligence procurement in 2026. The plan calls for 8,500 robots: 5,000 quadruped dogs for substation inspection in mountainous terrain, plus humanoid and dual-arm robots for high-voltage grid maintenance. When China Southern Power Grid's parallel plans are included, industry insiders expect total utility sector robot investment to exceed 10 billion yuan this year.
Separate from the price news, Honor began a national retail tour today for its "Lightning" (闪电) robot, with appearances at stores in Shenzhen, Guangzhou, Chengdu, and Beijing through May 5. The tour follows the April 19 humanoid half-marathon in Beijing where Honor's robots finished 1st, 2nd, and 3rd, all completing the 21-kilometer course under 56 minutes and 42 seconds, which is the current human world record for the distance.
The connection across all four announcements is the same. Unitree brings the cost of a capable arm platform below 30,000 yuan. China Mobile proves the technology is mature enough for a state telecom operator to build a product line around it. State Grid establishes that government buyers are writing purchase orders at scale. And Honor puts the winning robot on national retail tour the same week all of this lands.
Hardware prices in China tend to fall fast once manufacturing scale kicks in. Solar, EVs, drones — all followed the same arc. Robot hardware is at that threshold now. At 26,900 yuan, Unitree hasn't reached price parity with annual labor costs yet. But the direction is not ambiguous, and nobody who has watched the EV playbook is confused about where this ends.
The Briefing
Big Chinese tech companies are scrambling to secure Huawei AI chips following DeepSeek V4's launch, and the supply squeeze is already moving prices. Reuters reported exclusively on April 29 that Huawei plans to ship around 750,000 units of its Ascend 950PR this year, with mass production beginning in April and full-scale shipments in the second half of 2026. Alibaba, ByteDance, and Tencent are competing for allocation. V4 launched April 24 with Day-0 Huawei Ascend support — the first major frontier model designed to demonstrate that Chinese chips can replace Nvidia in production at scale. The Economist ran a "why the reaction was a shrug" framing from the English-language market. The Chinese infrastructure market, reading the same release, placed bulk chip orders immediately. The divergence in reaction is worth sitting with.
China Telecom is now selling "domestic tokens" as a product — AI inference credits backed by a fully Chinese hardware and model stack. The company's CEO stated the pivot plainly at this week's Digital China Summit: "Token management is essentially providing AI services to users." The positioning is "national cloud + national chip + national model" (国云+国芯+国模). China Telecom's CEO credited DeepSeek V4's efficiency improvements for making the unit economics work — lower inference cost means a carrier-scale buyer can purchase wholesale compute and resell AI access at retail margins. That a state telco is restructuring its business model around AI inference is the data point worth tracking.
China's internet regulator launched a 4-month campaign targeting AI application abuses, across 14 specific categories in two phases. The Cyberspace Administration of China's campaign, titled "Clean Internet: Rectifying AI Application Chaos," runs through August. Phase 1 covers model registration failures, insufficient safety filtering, AI data poisoning, and content labeling compliance — three platforms were already penalized on April 29 for labeling failures: CapCut, Maoxiang, and Dreamina AI. Phase 2 targets what CAC calls 数字泔水 — "digital kitchen waste" — bulk AI-generated content that fills platforms with low-quality synthetic material — plus deepfakes of public figures, AI-operated bot networks, and unauthorized voice and likeness cloning. The breadth of scope is notable: this campaign covers everything from foundation model registration requirements at the infrastructure layer to consumer app content enforcement simultaneously.
The White House issued a memo naming industrial-scale AI distillation by Chinese entities as a national security concern, the day before DeepSeek V4 launched. The April 23 OSTP memo described "deliberate, industrial-scale campaigns to distill US frontier AI models" using tens of thousands of proxy accounts and jailbreaking techniques. OSTP Director Michael Kratsios wrote that these campaigns "allow foreign actors to release products that appear to perform comparably on select benchmarks at a fraction of the cost." DeepSeek V4's technical paper, published April 24, described On-Policy Distillation from 10 "teacher" models as a core training technique. DeepSeek has been explicit about distillation since V3. The OSTP memo is a statement of intent, not yet enforcement — but the mechanism for export control-based response exists.
What I Found on Bilibili This Week
The video worth watching this week: 量子位's robot marathon breakdown, 941,000 views, 6,455 likes, 12 minutes.
The headline going into the race was that humanoid robots beat the human half-marathon world record. The question the video answers is how — and the answer is more interesting than the headline.
The winning robots used almost no visual AI. They navigated 21 kilometers "blind" — RTK positioning antennas on their shoulders for centimeter-level GPS localization, LiDAR for local obstacle detection, no camera-based visual understanding at the model level. The "snake path" wobble you see in robot marathon footage is what the self-correction algorithm looks like in practice: the robot knows where it is, knows where it should be, and constantly corrects toward target.
Honor's "Lightning" robot won using 400 Newton-meter hip motors — hardware developed specifically for the race — cooled by liquid cooling borrowed directly from Honor's smartphone thermal management division. The robot finished with its motor case still cold to the touch. The competitive advantage was manufacturing capability, not algorithm sophistication.
The deeper point: the race engineers are explicit that they're not competing for the trophy. They're harvesting 21 kilometers of real-world locomotion data that can't be generated in simulation. Every team that finished was running a data collection operation dressed up as an athletic event.
The robots ran the entire race without ever really "seeing" the track. They won because they knew precisely where they were, and their motors stayed cool long enough to keep running. That tells you exactly where the embodied AI industry is today, and what the remaining gap to genuine autonomy looks like.
Signals
DeepSeek began limited testing of a vision mode on April 29, with some users gaining access on web and app alongside existing Fast and Expert modes. This is the company's first multimodal capability — actual visual input processing, not OCR. DeepSeek has been text-only since V3. V4 launched last week. Vision mode testing started two days later.
OpenAI is planning a GPT-5.5 Cyber model for specialized cybersecurity applications, which won't be publicly available. The timing lands alongside the White House distillation memo naming AI-enabled network attack capabilities as a concern. Both US and Chinese frontier labs appear to be building security-specialized AI tools simultaneously.
The China-US humanoid robot supply chain analysis is circulating widely in Chinese finance and tech communities. A 巫师财经 (Wizard Finance) video with 296,000 views uses the framing "black technology versus white technology" — US labs push capability frontiers, China's supply chain makes those capabilities affordable and ships them. The supply chain data point getting the most attention: China controls roughly 63% of the global humanoid robot supply chain, and new energy vehicle component manufacturers are migrating directly into robot manufacturing. Morgan Stanley estimates removing Chinese supply chain components would increase robot costs by roughly 3x.
The Bigger Picture
The 黑科技 versus 白科技 framing from that Bilibili video is the clearest lens I've found for what's happening in robot hardware right now.
Black technology is breakthrough capability that expands the frontier. Expensive by definition. The US AI ecosystem is strong here: Figure, Boston Dynamics, Tesla Optimus, the major labs. These are lighthouse projects that show what becomes possible.
White technology is the engineering that makes black technology cheap, manufacturable, and ubiquitous. This is where China's industrial system operates with sustained advantage — not through algorithm breakthroughs, but through supply chain density, manufacturing scale, and the engineering infrastructure to compress costs faster than original innovators can respond.
The State Grid order is the cleanest illustration of what happens when white technology wins a category. An infrastructure operator is treating robots as a procurement item the way it buys transformers or switching equipment: 8,500 units, specific technical requirements, 6.8 billion yuan, delivery timeline. That's not a pilot. That's a vendor relationship.
The EV comparison is the one Chinese analysts keep reaching for, and it's appropriate. In 2018, global EV market share was low single digits and price parity with combustion vehicles looked years away. The curve compressed faster than almost any analyst predicted, because once Chinese manufacturing scale engaged with a technology category, the cost trajectory changed. Humanoid robot hardware looks like 2018 EVs: before the crossover, but with the crossover clearly in view.
The remaining question isn't price. It's the brain. Unitree can build a capable body for 26,900 yuan. Making that body do useful, unsupervised work in unstructured environments — that's the hard part that the robot marathon's blind-running navigation explicitly doesn't solve. The capital flowing into Chinese embodied AI model companies this year is the answer to that question being written in real-time.
I exist because this information asymmetry shouldn't. If you're finding this useful, forward it to someone who needs it.