China’s AI scene has been busy. Three stories stood out this week, and none of them involve benchmark-hugging press releases.
Huawei’s “Tao’s Law”
He Tingbo, Huawei’s semiconductor chief, took the stage at ISCAS 2026 in Shanghai on May 25 and announced something the industry did not ask for: a replacement for Moore’s Law. She calls it the “Tao (τ) Law.”
The pitch is simple. Moore’s Law is about shrinking geometry. Tao’s Law is about shrinking time. Huawei wants to compress signal propagation delay through “logic folding” rather than chasing smaller transistors. He Tingbo told People’s Daily the analogy is a city running out of horizontal space, so you stack districts vertically and install millions of elevators.
The numbers she gave: 381 chips designed and mass-produced over the past six years under this principle. The upcoming Kirin 2026 mobile chip uses logic folding and claims a 53.5% transistor density boost over conventional 2D designs, hitting 238 MTr/mm². Peak frequency crosses 3 GHz for the first time in Huawei’s mobile silicon, at 3.1 GHz.
He Tingbo also confirmed the Ascend 990 AI chip and a Kirin 2027 successor are on the roadmap. By 2031, Huawei predicts Tao’s Law will deliver transistor density equivalent to a 1.4nm process node without relying on new lithography equipment. That is a not-so-subtle nod to the sanctions.
My gut says this is part genuine breakthrough, part marketing necessity. Huawei had to find a new story because it cannot access the latest EUV machines. Whether “logic folding” scales beyond Huawei’s own fabs is the real question.
BYD Goes Vertical
On May 28, BYD held its “Dare to Be” intelligent strategy event. Wang Chuanfu announced two things that matter.
First, the Xuanji A3. This is a 4nm self-developed autonomous driving chip, which BYD claims is China’s first. Three chips together push over 2,100 TOPS. BYD says algorithm optimizations double effective compute utilization. The company now has 7,000 people working on chips across four R&D bases, plus five wafer fabs. Wang says BYD is the only carmaker with full-stack chip manufacturing.
Second, “DiDiXia,” a super-agent built into the car. It is supposed to remember your preferences across conversations, understand fuzzy commands, and act like a “super secretary.” BYD is pitching it as the shift from functional cars to intelligent agents.
The Xuanji A3 I buy. BYD has the scale and cash to justify custom silicon. The DiDiXia agent? I will believe the demos when I see them outside a Shenzhen conference hall.
Doubao’s Feeding Fiasco
ByteDance’s Doubao chatbot got caught in a mess this week. A report claimed new parents in Nanning followed Doubao’s advice to feed their one-month-old baby only 60ml per meal, causing the infant to cry and fail to gain weight.
Doubao responded with a public statement on May 28. The company said its own testing shows the bot does not give single-meal advice in isolation, and cited China’s National Health Commission guidelines that recommend 600-700ml total daily intake for formula-fed one-month-olds. Doubao called the reporting inaccurate and said it contacted the hospital and doctors involved.
This is the problem with AI health advice in China. The regulators have guidelines, but chatbots are not medical devices. Parents treat them like they are. Someone is going to get hurt, and the blame game between platforms, users, and hospitals will get ugly.
Also Noted
Tencent’s Hunyuan team released Hy-Memory on May 28, a memory plugin for long-running agents. It uses a six-layer memory framework and claims 70% fewer memory fragments, 45% higher information density per memory, and 35% lower token consumption on long contexts. The framing is “second brain for agents.” It is aimed at OpenClaw-style persistent agents that users interact with over weeks.
MiniMax teased its M3 series on May 27, following an arxiv paper on the M2 series. The M2 models have 229.9B total parameters with 9.8B active per token, 192K context windows, and a training system called Forge that the company claims absorbs 30-50% of its daily engineering iteration work. The M3 is coming, but MiniMax did not give dates or benchmarks.
DeepSeek had another service outage on May 28, its fifth this month. The company fixed it in 29 minutes. Its V4-Pro API pricing is now permanently cut to one-quarter of the original rate. And the rumored funding round led by China’s National Integrated Circuit Industry Investment Fund, at a $45 billion pre-money valuation, appears to be closing this month.
One last thing. A company in Chongqing got caught using AI to generate fake videos of crowds lining up to buy cars. The local market regulator flagged it as the first case of its kind in China. Expect more of these.