NeuroAI

Wang Xingxing says embodied AI's real bottleneck is generalisation, and proposes a concrete threshold — a robot that works in 80% of unfamiliar places on 80% of tasks. That is 64%, and the industry is not there yet.

At JD's 2026 tech conference the company committed to 3 million robots, 1 million unmanned vehicles and 100,000 drones over five years — and, in the same breath, 100,000 new service-engineer jobs.

China's BCI market is around RMB 3 billion today and is forecast to reach RMB 12 billion by 2030. A Tsinghua professor's shoe metaphor explains why the number is small — and why that is the industry's real problem.

Ling-3.0-flash-VL activates only 5.5B of 124B parameters, reads 256,000 tokens at once, and runs an observe–act–verify–correct loop. Open weights are not charity; they are a land grab.

AI comic-drama uploads went from roughly 20,000 in January to 47,000 by March. The hit rate is 0.16%. The scarce thing is no longer generation — it is the pipeline that keeps a character's face stable across a hundred episodes.

Tencent's WeChat vision team released WeMM-Embedding in 2B, 4B and 9B sizes. The 9B version tops the MMEB-v2 benchmark at 80.6; the 2B beats the previous 8B open model. It already runs at billion-calls-per-day scale inside WeChat.

DeepSeek opened about 150 roles targeting engineers with 2–10 years of experience — server-side development, agent elastic computing. Zero AI research positions. The next day, Reuters reported it had hired CITIC Securities to prepare a STAR Market listing.

From 10 September 2026, DeepSeek's Flash series gets off-peak cache-hit inputs at ¥0.02 per million tokens. The cut targets repeated, high-volume work — a subsidy for AI that works all day, not for AI that chats.

Beijing-backed NEUCYBER says its semi-invasive BeiNao-1 system has now been implanted in more than 40 patients, with the longest-running device past 18 months — the clearest sign yet that Chinese neurotechnology is becoming a real medical product.

A vertical short-drama app has more daily users than China's four biggest streaming platforms combined. An AI episode costs a tenth of a live-action one — and the industry's own producers say where it still fails.

Udeer Robotics listed in Hong Kong on 9 September 2026 with more than 114,600 commercial delivery robots sold to over 5,100 customers. Not humanoid — and that is exactly the point.

Chinese makers now ship more than 400 humanoid robot models — over half of the world's total — and account for roughly 80% of global humanoid and quadruped sales. Scale, not spectacle, is the new story.

In 2026, brain–computer interfaces entered China's Government Work Report and its five-year plan. Behind the clinical headlines, regulators are building the classification rules, ethics guidelines, standards and reimbursement pathways that decide whether neurotechnology becomes medicine.

Huawei's Kirin 9050 Pro supports on-device MoE multimodal models of 30B total parameters with 2B active. The number that matters is not the parameter count: it is the 66% cut in power draw for key tasks.

Supplier guidance went from 150 units a week in July to 300 in August to 1,000 in September — then a single batch order of about 5,000 Optimus units. That is not linear growth; it is the steep part of an S-curve.

China's brain–computer interface work is not confined to paralysis. Demonstration deployments now run from mining helmets that watch for fatigue to a visual prosthesis that lets a blind patient recognise letters, to an interventional device delivered through blood vessels.

On 6 September, an anti-fraud app guided by the Ministry of Public Security's criminal investigation bureau and built by Shanghai police went live. Three functions, one of which — read-aloud — matters more for elderly users than it sounds.

Volume AI short drama went from RMB 5,000 per minute to several hundred in under twelve months. Meanwhile Shanghai studios charge RMB 200,000–300,000 for a 20-minute branded piece. Under 1% of creators earn real money; most earn under RMB 10 a month.

iFlytek's Spark X2.5 was trained and served end-to-end on Chinese accelerators — and shipped with open on-device models small enough to run on a laptop. The constraint moved from silicon to electricity.

Three billion downloads, permissive licences, and a developer default that no longer points West. China's open model labs did not win an argument — they changed what developers reach for first.

Class III is China's highest medical device risk class — long-term implant, strictest review. Since early 2026, multiple BCI devices have cleared it, shifting the industry's question from 'is this possible?' to 'how fast can it reach patients?

On 6 September, Qwen released Qwen-Drive-1.0-4B — described as the first vision-language foundation model built for autonomous driving scenarios. It folds 3D perception, visual question answering and motion planning into one 4-billion-parameter model, weights free on GitHub, Hugging Face and ModelScope.

NEO won regulatory approval in March 2026 after 32 cases across 11 hospitals with a 100% clinical response rate on assisted grasping — and it is now covered by Shanghai's medical insurance. Overseas patients are emailing weekly.

The Hangzhou company says cumulative production of its bipedal humanoids has passed 18,000 units, shipped more than 5,500 in 2025 alone, and turned profitable — a rare set of numbers in an industry still accused of selling videos.

The MIIT's SME AI Entrepreneurship Support Plan (2026–2028) commits to building 10 national inclusive compute centres and issuing compute vouchers. Three predictable pitfalls apply to any voucher scheme: pay-first-reimburse-later, hidden eligibility clauses, and long settlement cycles.

Video generation company Runway introduced Solaris, which predicts and renders the next screen directly instead of generating code first. In testing with 250 evaluators, 61% preferred its interfaces on instruction-following and 71% on natural feel.

Global humanoid shipments passed 22,000 units in the first half of 2026, up nearly 300%. Manufacturing took 13% and logistics 5%. The rest went to stages, showrooms and research labs.

China has built tens of thousands of humanoids. The hard part is the next one: teaching them to do jobs. A national 'real-scenario training' programme, robot training fields and data centres are being built to close the last kilometre.

CloudMatrix 384 links 384 Ascend 910C accelerators into a single logical computer. Independent analysis puts it near double the dense BF16 compute of NVIDIA's GB200 NVL72 — at roughly four times the power.

The MIIT's SME AI Entrepreneurship Support Plan explicitly encourages local governments to accommodate 'one-person companies' and 'super-individuals' doing agile entrepreneurship with intelligent tools — the first national-level recognition of the form.

Reuters reported on 3 September that Moonshot AI, developer of Kimi, has confidentially filed for a Hong Kong IPO targeting around US$3 billion, at a valuation near US$50 billion — two and a half times its valuation at the start of the year.

Kuaishou's Kling passed 100 million global users across 224 countries in June 2026, with quarterly revenue above 850 million yuan and investors valuing the spun-out business near $18 billion.

Vertical integration, cheap actuators and brutal iteration speed have made China the place where humanoid hardware gets built. The strategic question is whether that advantage survives the shift from hardware to intelligence.

On OpenRouter, the share of tokens US companies spend on Chinese AI models has sat above 30% every week since February 2026 and peaked at 46%. Engineers are not switching for ideology — they are switching because the models are good and roughly fifty times cheaper.

The “AI+” initiative is not a research programme. It is an adoption programme with dated targets — and it is the single most under-covered story in Chinese technology.

On 1 September the Nubia NaviX Ultra received network access approval — developed with ZTE and ByteDance, running the Doubao assistant, positioned as the first 'AI agent phone'. The technology works. The question is whether you trust it with payment.

On 2 September, Yuanidian Robotics launched Zeroth Bridge ('Xiaoqiao'), a personal high-performance humanoid at RMB 8,888 for the geek edition, with secondary development supported. Three curves converged at once to make the price possible.

TrendForce forecasts the nine largest cloud providers will spend over US$886.7 billion in 2026, up about 90%. Increasingly the money is not going to GPUs — it is going to liquid cooling, advanced packaging, interconnect, power and memory.

Enflame Technology raised RMB 6.119 billion at RMB 142.18 per share. The listing is a beginning, not an achievement: a high-end AI chip costs hundreds of millions to tape out, and the real moat is software ecosystem, not silicon.

From 1 September, Huawei, Xiaomi and Honor raised prices across current models, some by RMB 200–1,000. The cause is memory: procurement prices reportedly went from around RMB 500 to RMB 2,200. AI data centres took the capacity.

Forget the demos. In Chinese factories and clinics, AI is being measured in changeover minutes, defect rates and diagnosis times — and the numbers are concrete.

Shao Tianlan founded Mech-Mind in 2016. On 1 September the company listed in Hong Kong: 254,000 retail subscribers, 3,843 times oversubscribed, market value around HK$12.4 billion. Sequoia China invested in four consecutive rounds and holds 13.63%.

Running the same workload for a fiftieth of the price is not a discount. It is a different product. China's AI labs turned a compute constraint into a cost structure that is now reshaping global model economics.

A hospital in Chengdu listed a non-invasive BCI adaptation fee of over RMB 900 — among the first official prices for the technology in China. Without a price, hospitals cannot charge; without charging, they cannot procure.

The least glamorous and most consequential AI story in China is not in Beijing's labs. It is in rice breeding, county hospital radiology and the local social-security desk.

OpenRouter logged 113 trillion tokens of global model calls in one week. Chinese models accounted for 55.16 trillion — up 36% week on week, and ahead of the US for eighteen consecutive weeks. The top three were all Chinese.

China shipped over 40,000 humanoid robots in H1 2026 — 97% of the global total. Morgan Stanley raised its full-year forecast from 28,000 to 50,000. At the World Robot Conference, an Indian retail association delegation came to buy.

Unitree listed on the STAR Market in August 2026. Kling is raising at an $18 billion valuation. The question for China's AI sector has shifted from “can it be built” to “does it earn.”

If you cannot buy the fastest chip, buy more chips and wire them together better. China's answer to export controls is the supernode — and in 2026 it started shipping them by the hundreds.

Cambricon's revenue more than doubled. Hygon shipped into a 100,000-card cluster. Moore Threads grew 147% and headed to Hong Kong. China's AI silicon story is no longer a single company holding the line.

Behind China's AI and robotics surge is a particular kind of founder — engineers in their thirties who treat hardware cost curves and open licences as strategy, not afterthoughts.

Alibaba open-sourced 460 models. Huawei open-sourced its entire compute stack. This is not charity — it is the fastest way to become the layer everyone else builds on.

When Sora debuted, generating one second of video cost about US$0.50. Today Kling, Jimeng, Vidu and Wan 3.0 quote under RMB 0.01 per second — a 99% fall in 24 months. What survived is not the cutting.

From brain-interface ethics to generative content labelling, China has shipped a dense body of AI rules. Whatever one thinks of the approach, the rulebook exists — and now it is being exported as standards.

A 100% AI-generated feature film made by a team of seven is heading to Korean theatres. Broadcast networks are using generative video in prime-time programmes. The tooling question has shifted from 'can it?' to 'who operates it?

Henan switched on the first all-domestic 100,000-card AI supercluster, and Anhui and Hubei are building their own niches. China's compute map is being redrawn province by province — with self-sufficiency as the purchasing rule.

Forget the demos. In a crane factory in Xuzhou, a production system replans thirty days of work on its own and retools a line in ten minutes. In Wuhan, an AI inspector catches defects twenty times finer than the human eye at 200 metres a minute.

Beyond the clips: native 4K generation, AI gifts sent 1.1 million times in a quarter, and AIGC material at 10% of marketing spend. Chinese generative media is now ordinary production infrastructure.

Small models, permissive licences and offline deployment are what most of the world actually needs. Those are exactly what Chinese open-weight labs shipped — and that is why they won.

A profitable humanoid maker, a chip company doubling revenue, a video model approaching a billion-dollar run rate. After years of valuations priced on narrative, Chinese AI is being judged on revenue, cost and efficiency.