China Tech Today: Physical AI Takes Centre Stage as Unitree’s Shares Soar — October 27, 2026

Physical AI owned the day. Unitree’s shares have jumped more than 60% in under two weeks after JPMorgan bet on a robot breakthrough within three years, AgiBot set a Guinness record with 110 dancing humanoids, and Alibaba showed robots can learn by watching people.

Unitree’s shares soar on a physical-AI bet — while robot funding cools

Shares in Unitree Robotics, China’s first listed humanoid-robot maker, closed at 672 yuan on Monday — up 7.7% on the day, more than 60% above the 405.10 yuan of October 15. The stock hit its 20% daily limit on Friday at 624 yuan; the rally has lifted the company’s market value to about 271.8 billion yuan, roughly $40.8 billion.

The trigger was a double upgrade from JPMorgan — two notches, from “underweight” to “overweight” — and a price target more than tripled, to 995 yuan from 300 yuan. The bank kept its 2026 sales assumption at 14,163 units but lifted its 2027 forecast to 31,701 and its 2028 forecast to 111,554 — nearly triple — after briefings with founder Wang Xingxing reversed its view that Unitree was too hardware-heavy and too light on software.

At Unitree’s first partner conference in a decade, in Hangzhou on October 22, he said a breakthrough in physical AI — large models directing machines — could arrive within one to three years, not the “up to ten years” he had previously suggested.

The rally comes with two warnings attached. Unitree itself told investors on October 21 that the industry “is still in the early stages of commercialisation,” citing uncertainty over technology, demand and the supply chain. And the money behind the sector is cooling: investment in physical AI fell 40.1% in September alone, to 11.6 billion yuan from 19.4 billion in August, according to data firm NE Times, after China’s securities regulator signalled a higher bar for robot-company listings. Some investment firms have stopped reviewing new deals since June; with more than 150 robot companies competing, regulators have warned against near-identical products flooding the market.

For Canada, the contrast is the story: the market is pricing a breakthrough years before the order books prove it. Tesla has only just begun producing its Optimus robot; Unitree shipped more than 5,500 humanoids in 2025. Whether the one-to-three-year timetable holds will decide if this is a sector repricing or a bubble.

Huawei’s Ascend 950PR runs at 43% of an H100 — but wastes less power

The first independent test results for Huawei’s flagship AI chip cut both ways. Research firm SemiAnalysis ran 1,055 matrix computations of the kind large AI models rely on: the Ascend 950PR delivers about 43% of the raw computing speed of Nvidia’s H100, and the newer B300 just 21%. But Huawei’s software lets the 950PR use 86% of its achievable computing power, ahead of the H100’s 75.5% — slower, but less wasteful.

DeepSeek is closing the gap from the software side. The AI lab has been porting its own optimisation tools onto Huawei’s hardware, and SemiAnalysis found automatic scheduling alone can cut some small computing tasks by up to 40%. DeepSeek is preparing to use the 950DT variant for large-scale pre-training by the end of the year. The 950PR is aimed mainly at inference rather than training, and the gap to Nvidia remains wide. For Canadian AI users, that means more competition in inference compute — and cheaper models to run.

AgiBot’s 110 humanoids dance in perfect sync — and set a record

AgiBot, the Shanghai humanoid maker founded by former Huawei engineers, put 110 of its full-size Expedition A3 robots through a 5-minute-43-second dance routine in Jinan on Saturday, October 24, and a Guinness World Records adjudicator certified it on the spot as the largest group of humanoid robots dancing simultaneously.

AgiBot’s LinkBots fleet system keeps the robots synchronised to the millisecond over 5G and Wi-Fi, using ultra-wideband positioning accurate to about four inches. The robots came straight off the production line; some A3 units already work in Jinan subway stations. China’s industry ministry says Chinese companies built more than 40,000 humanoid robots in the first half of 2026, with annual output expected to pass 100,000. When a hundred robots run as one system, the bottleneck shifts from building bodies to coordinating them.

Alibaba’s Amap teaches robots by watching people

Alibaba’s mapping unit Amap has released λ0 (Lambda-0), a foundation model for humanoid robots that learns walking, posture and two-handed manipulation largely from recordings of people doing everyday tasks — not from steering a robot through them. The model draws on HumanVerse, a 500-hour dataset of first-person video paired with body and hand motion, spanning 32,993 episodes across 827 task types.

On a Unitree G1 humanoid, λ0 completed 25 of 40 trials across four real-world tasks — a 62.5% success rate — against 40.0% for the strongest rival model tested. The economics are the point: Amap’s researchers estimate human capture cut labour per demonstration by about 76%, from roughly 141 person-seconds to 33. One caveat: the release is only partial — data-collection code and sample data are public, but as of October 27 the model weights and licence were not, and its linked GitHub repository was empty.

BYD’s third-generation Qin Plus DM-i: 2,290 km for under ¥100,000

BYD launched the third generation of its Qin Plus DM-i plug-in hybrid sedan on October 27, with limited-time prices from 76,900 to 97,900 yuan — roughly $11,500 to $14,700.

The headline number is range: up to 220 km of electric-only driving and a stated combined range of up to 2,290 km, powered by BYD’s fifth-generation DM hybrid system, which reports fuel consumption of 2.58 litres per 100 km with a depleted battery. Buyers can also specify the God’s Eye B driver-assistance system, with city and highway navigation assistance and automated parking. The Qin Plus is a China-market model, but it sets the benchmark every hybrid sold in Canada is now measured against — see our guide on the true cost of owning an EV vs hybrid in Canada.

Tencent open-sources a real-time interpreter

Tencent has released Hy-SimulMT, an open-source model for simultaneous Chinese-English interpretation — the kind of live translation a UN interpreter does, listening and speaking at the same time. It processes incoming speech continuously and can clone the original speaker’s voice in real time, so the translation still sounds like them.

The release includes the model weights, inference software and training tools, and users can try it in Tencent’s Hy Translate app. On the RealSI long-form audio benchmark, Tencent reports a streaming translation lag of about three seconds, with translation-quality scores beating the other open-source models and four of the five commercial APIs in its comparison — though that is Tencent’s own comparison. An open model that does in three seconds what once needed a trained professional lowers the barrier for cross-border business, education and streaming — and the built-in voice cloning is the part to watch, since it can impersonate as well as preserve a voice.

The big picture

The day’s thread is physical AI moving from promise to proof — and to price. Unitree’s rally, AgiBot’s hundred-robot dance and Amap’s human-video training are three versions of the same claim: the models are ready to leave the screen and start working with their hands. The counterweight: funding is cooling and the chips still trail Nvidia. Monday’s brief showed self-driving moving from test permits to paying riders; today the same push reaches robots — with the market, for now, ahead of the order books.

Watch next: Volkswagen holds its first Technology Day in Hefei on Wednesday, October 28 — expect the first detailed look at its CARIZON L3 driving system. The same day, Huawei launches the Pura 90 Plus and Redmi the K100 in China, the next test of whether the memory-price squeeze keeps pushing phone prices up.

Sources

1 thought on “China Tech Today: Physical AI Takes Centre Stage as Unitree’s Shares Soar — October 27, 2026”

  1. Pingback: China Tech Today: CXMT Bets on Homegrown AI Memory Despite 20% Yields — October 28, 2026 | FutureTechPulse

Leave a Comment

Your email address will not be published. Required fields are marked *

Scroll to Top