China Tech Watch

🌐 Tracking China’s AI, semiconductors and industrial strategy.
📊 Early signals before consensus.

[China Tech Watch @China_TechWatch]

宇树科技73天闪电过会，创科创板纪录！420亿估值、42亿募资，A股迎来“人形机器人第一股”。从受理到过会仅73天，比此前最快的摩尔线程还快15天。

[China Tech Watch @China_TechWatch]

Unitree Robotics clears IPO review in just 73 days, setting a new record on STAR Market. With a $42B valuation and $4.2B fundraising, it becomes the first humanoid robot stock in A-shares. This milestone signals accelerated capital support for embodied AI and the entire robotics supply chain. #Unitree #HumanoidRobot

[China Tech Watch @China_TechWatch]

Chinese multimodal powerhouse StepFun just flipped the global AI table. Step 3.7 Flash delivers 400 TPS inference, matching 97% of Claude Opus 4.6 at 1/9 the cost. This isn't a simple update—it's a paradigm shift for the Agent era.

What should an Agent-era model look like? Give it a cockpit screenshot, old models just describe dials. Step 3.7 Flash selects the area, identifies instruments, plans steps, and guides a mouse through takeoff. From understanding to executing—that's the divide.

The competition has changed. As Agents enter production, latency and token costs explode. Step 3.7 Flash uses sparse MoE (196B total, 11B active) for 400 TPS. Speed is capability—faster means more iterations, better results in multi-step tasks.

Real tests: 12 messy receipts auto-extracted into Excel; Blender screenshot teaches you to delete a cube; Jianying interface locates export button, warns about 1080i and full-track export pitfalls. The model is on-site, working as your hands.

Even more impressive: emergent behavior. After writing frontend code, it switches to GUI to test rendering, then revises its own code. No one taught it this loop—it figured it out. That's what an Agent should be.

The Advisor strategy is the killer feature: small model executes, consults a larger model only at critical junctures. Result: $0.19 per task vs $1.76 for Claude Opus 4.6, with 97% coding capability. Two dimes vs two dollars—any team burning tokens gets it.

Step 3.7 Flash isn't a benchmark king, but it dominates the cost-efficiency frontier. Multimodal + extreme efficiency + near-zero cost makes it the foundation for production-grade Agents. Flash is no longer a budget option—it's the pragmatic force taking over the world.

[China Tech Watch @China_TechWatch]

BYD hits new monthly sales record in May: 383,453 units! Overseas sales surged 80.7% YoY, with cumulative deliveries exceeding 1.4 million in Jan-May. The NEV leader accelerates global expansion as market penetration rises. Supply chain momentum continues.

[China Tech Watch @China_TechWatch]

As far as I know, besides the LogicFolding-enabled EDA tool recently released by Peking University, Huawei has also been working on this technology internally, and several other vendors are developing similar solutions as well. So while it may appear that Huawei is the one introducing the product, the reality is that Chinese semiconductor companies and universities have already been exploring and advancing along this path for quite some time.

[SemiAnalysis @SemiAnalysis_]

Necessity is the mother of innovation.
In one fell swoop, Huawei overtakes all in advanced Hybrid Bonding technology, with 2026 Kirin smartphones featuring 1.5 µm bond pitch 3D-stacked architecture.
Next year's Kirin chips will go down 1 µm pitch!
TSMC has only just moved to 6 µm SoIC, with their next step to 4.5 µm for 2030 products.
Intel's Foveros Direct is at 9 µm with Clearwater Forest this year.
16-36x denser interconnect enables Huawei's LogicFolding design, with more granular architectural splitting across the dies to optimize routing and shorten critical paths.

[China Tech Watch @China_TechWatch]

DeepSeek V4上线首月即引爆：Flash月调用9.13T tokens，三个模型合计超17万亿tokens！Agent系统（如Hermes Agent月调用10.8T）取代聊天机器人成为Token消耗主力，单次任务触发上百轮LLM调用。头部模型集体撞上MoE通信、百万上下文内存、低精度可靠性三堵墙，昇腾以通用方案实现智谱GLM-5.1、MiniMax M2.7、DeepSeek V4发布即支持。AI基础设施的入口之争已白热化。

[China Tech Watch @China_TechWatch]

DeepSeek V4 explodes in first month: Flash hits 9.13T tokens monthly, three models total over 17T tokens! Agent systems (Hermes Agent at 10.8T) dethrone chatbots as token consumption kings, with single tasks triggering hundreds of LLM calls. Top models hit three walls: MoE communication, million-token memory, low-precision reliability. Ascend delivers universal support for GLM-5.1, MiniMax M2.7, and DeepSeek V4 on launch day. The battle for AI infrastructure entry points is on fire. #AIInfrastructure #LLM

[China Tech Watch @China_TechWatch]

Computex 2026 marks AI's shift from cloud to edge: PCs emerge as the prime device for local token generation, with mobile opportunities expected next year. Edge AI chips and AI PC supply chains enter a new growth cycle.

[China Tech Watch @China_TechWatch]

不会再有其他国家通过工业崛起了

未来10年，机器人的成本优势会迅速摧毁全球绝大多数地区的工业体系

[China Tech Watch @China_TechWatch]

@tuolaji2024 确实是好事，高权重老登拖累指数也影响资金流入

[China Tech Watch @China_TechWatch]

@Anna_nanachan 本来就没有未来啊，未来是机器人的

所以机器人的吃喝拉撒股票都在涨，碳基的的服务，游戏，消费股票都在跌

[China Tech Watch @China_TechWatch]

@dingyi 智谱算国家队，有官方的稳定订单，也是和DS一样最早一批适配昇腾的，这些优势不是小米可以比的

[China Tech Watch @China_TechWatch]

@LuBtc888 🤣完全没有推出去

[China Tech Watch @China_TechWatch]

华大九天在华为发布韬定律同日推出国内首款Argus 3DIC物理验证平台，填补高端3DIC设计工具空白，成为国内唯一全流程EDA提供商。

[China Tech Watch @China_TechWatch]

@_FORAB 那就得看看9月的麒麟9050 pro能不能打台积电的骁龙3nm了

[China Tech Watch @China_TechWatch]

@chnchn1059453 能信极客湾也行，别到时候极客湾测出来又开始刷什么收钱了，懂的都懂

[China Tech Watch @China_TechWatch]

我知道华为发表半导体韬定律会有很多人质疑

没关系，9050可以跑分

可以进入有些人最爱的环节

[China Tech Watch @China_TechWatch]

@giantcutie666 V4上线前一个月至少四次崩溃上了热搜

大概率是新能力上线导致的问题

[China Tech Watch @China_TechWatch]

评估小米的股价就不能孤立的只看小米

19年的8块的底很大程度上是吃到了菊花制裁的红利

当时雷布斯去深圳找余大嘴承认华为+荣耀的步步紧逼使得当时的小米市占率极速下降

如果制裁再晚半年，这个底就不是8块了

所以现在的一切都并不难理解

[China Tech Watch @China_TechWatch]

华为金融系统部CTO郑俊在2026凤凰湾区财经论坛确认：基于τ定律（韬定律）研发的麒麟芯片已应用于Mate90系列！

核心参数： • 晶体管密度238 MTr/mm²，提升53.5% • P核能效提升41% • 峰值频率3.1GHz，接近3nm工艺水准

[China Tech Watch @China_TechWatch]

【Huawei Mate 90 Official: τ-Law Chip Hits 3.1GHz, Rivals 3nm Process】

Huawei CTO Zheng Jun confirmed at the 2026 Phoenix Bay Area Finance Forum: The Kirin chip based on τ-scaling (Tao Law) is already powering the Mate 90 series!

Key specs: • 238 MTr/mm² transistor density (+53.5%) • 41% P-core power efficiency gain • 3.1GHz peak frequency, matching 3nm-class performance

This marks the first commercial deployment of Huawei's "LogicFolding" technology—a major milestone for post-Moore semiconductor innovation. From research paper to mass production in mere months: τ-scaling isn't just theory, it's a battle-tested weapon.

Roadmap to 2031: 400+ MTr/mm², 5.0GHz—and the momentum is just getting started.