ChinaBiz Briefing: CapCut Enters ChatGPT, China's Chip Race Heats Up, and AI Glasses Hit a Wall
China's technology sector delivered a week of sharp contrasts: a ByteDance creative tool embedded itself into the world's dominant AI interface, a cohort of automotive chip startups attracted strategically loaded capital, and the AI glasses boom revealed the unit economics of a category still running on narrative rather than margin. Taken together, the week's developments reinforce a single underlying theme — in China's tech economy, distribution and hardware control are increasingly more defensible than software features alone.
• CapCut Plants Its Flag Inside ChatGPT — Before Creators Open an Editor
ByteDance announced at its CapCut World LA event on October 24 that CapCut's AI video editing and generation capabilities will integrate directly into ChatGPT as a plugin, with open beta across mobile, desktop, and web scheduled for November 2026. The architecture is deliberate: ChatGPT handles intent parsing and task orchestration; CapCut executes asset handling, timeline construction, effects, and rendering; the creator validates and publishes. Projects can be passed back into CapCut's web editor for granular refinement before distribution.
The strategic logic runs deeper than a feature announcement. A creator who drafts a concept inside ChatGPT has already formed creative intent before touching any editing tool — leaving CapCut at risk of becoming a rendering endpoint rather than a creative environment. The plugin directly neutralizes that displacement risk. The 2023 version of this integration generated videos from text prompts; the 2026 version processes a user's own raw footage into a usable cut. That distinction is structural: the earlier product reduced the entry cost of creation; the current product enters the production pipeline itself, where professional time is actually spent. For Adobe, Runway, and DaVinci Resolve, the competitive moat is shifting — feature depth matters less if users never reach the feature menu.
• China's Automotive Chip Startups Are Raising Capital Like It's 2021 — Because the Window May Close by 2028
Physical AI chipmaker XHEART disclosed a cumulative RMB 5 billion in funding, with Renault Group, Chery Automobile, and DRAM manufacturer CXMT joining its cap table. The investor mix reveals the real story: a European OEM securing a third supplier option beyond NVIDIA and Horizon Robotics, a Chinese export-focused automaker buying co-development leverage over its silicon roadmap, and a memory producer converting a vendor relationship into equity-aligned supply chain optionality. XHEART's X7 chip is already in series production on the SAIC Volkswagen ID.ERA 9X, with pipeline coverage extending to SAIC MG, GM Cadillac, and select Chery models.
XHEART is not an isolated case. In 2026 alone, NIO's chip subsidiary Anhui Shenjue closed a RMB 2.257 billion Series A at a RMB 10 billion valuation; Li Auto's Mach Heart Technology closed a RMB 1.625 billion Series A at a RMB 17.1 billion post-money valuation; Geely-incubated Ecarx raised over US$200 million. The structural driver is margin math: smart-driving chip gross margins run 40%–50%, against chronic compression for algorithm-only software vendors. Urban NOA penetration reached 20.9% in H1 2026, and Horizon Robotics' Journey chip series surpassed NVIDIA's automotive chip market share in China for the first time. The demand threshold that makes independent chip commercialization rational has been crossed — and industry participants expect the supplier landscape to consolidate within two to three years.
• NIO's Li Bin Reframes Smart Driving Competition Around Two Metrics No One Has Yet Proven
At the 2026 World Smart Connected Vehicle Conference, NIO founder Li Bin identified the industry's next competitive axis: verifiable safety performance and genuine consumer willingness to pay. NIO's system — active across 950,000 vehicles with 9.14 billion kilometers of assisted driving mileage — has reportedly driven a 30%-plus year-on-year decline in accident-related insurance loss costs across four unnamed carriers. Two mandatory national standards take effect in early 2027: GB 47955—2026 governing L2 systems from January 1, and GB 44721—2026 governing L3 and L4 systems from July 1. The July deadline functions as a revenue gate — automakers whose systems qualify for L3 authorization in defined operational domains gain the right to sell a qualitatively different product with higher pricing power and shifted liability allocation.
The monetization arithmetic, however, remains uncomfortable. An OEM with one million subscription-eligible vehicles at RMB 300 per month generates RMB 360 million annually at a 10% take rate — insufficient to cover training infrastructure and cloud compute costs at scale before the feature itself commoditizes. The deeper tension: competitive pressure has already migrated urban navigation-on-pilot from premium to RMB 100,000 entry-level vehicles, compressing the addressable software premium precisely as R&D cost structures remain elevated.
• Unitree's 465x PE Ratio Is a Forward Bet on a Model Capability It Has Not Yet Demonstrated
Unitree Robotics, China's first humanoid robot A-share listing, surged 65.89% over seven sessions beginning October 16 before management issued an "abnormal trading volatility" disclosure citing a trailing PE of 465.22x against a sector average of 38.54x for general equipment manufacturing. The company's own disclosure named perception accuracy, motion control stability, data accumulation, and hardware-software integration as unresolved engineering challenges. The October 21 unveiling of the H2-X humanoid — 170 cm, 65 kg, 31 degrees of freedom, eight-hour uptime — generated viral demonstrations, but the G1 model's cumulative production of 10,000–20,000 units illustrates how early-stage the addressable market remains.
Founder Wang Xingxing made two public statements within ten days calling insufficient belief in AI "the biggest mistake of the past decade" — a deliberate signal that the company is pivoting strategic emphasis from hardware performance toward embodied intelligence model development. Unitree's September 2026 open-source release of UnifoLM-WLA-1.0 is the most direct expression of that pivot. A venture capital partner interviewed by Sci-Tech Innovation Board Daily was explicit: without model capability, there is no path to a valuation premium — and what determines model quality is not headcount but the knowledge density of researchers and how tightly they integrate with hardware engineers.
• Cambricon's RMB 11.7 Billion Inventory Pile Turns a Clean Growth Story Complicated
Cambricon Technologies posted nine-month revenue of RMB 9.526 billion, up 107% year-on-year, with net profit up 131% to RMB 3.715 billion. Yet investor attention at the October 26 earnings call focused on a balance sheet detail management declined to fully explain: inventories reached RMB 11.665 billion at end-Q3, a 41% sequential increase, alongside RMB 502 million in inventory write-down provisions on aging raw material stock. Management framed the build-up as demand-hedging but disclosed neither the inventory's structural breakdown nor the proportion tied to confirmed purchase orders. The company must generate approximately RMB 3.974 billion in Q4 to hit its restricted stock incentive plan revenue threshold of RMB 13.5 billion for the full year — a 12.5% sequential increase from Q3, with no order backlog or delivery schedule provided.
Two underreported positives merit investor attention. Cambricon has achieved compatibility with China's full tier-one domestic LLM roster — DeepSeek, GLM, Qwen, Kimi, and MiniMax — with active optimization work on DeepSeek, Qwen, and Tencent's Hunyuan. The Triton-MLU compiler bridge, enabling GPU-agnostic Triton kernels to execute natively on MLU architecture, lowers porting friction for model teams already working in tools widely used by the DeepSeek community. The company also disclosed R&D funding for switch chips targeting multi-node large-model interconnect — a move that, if productized, would replicate the NVLink playbook of controlling the fabric layer to raise switching costs and per-cluster revenue.
• China's AI Glasses Boom: 4.2 Million Units Shipped, 60% Returned
Global AI glasses shipments reached 4.2 million units in H1 2026, up 127% year-on-year, with China's 618 shopping festival generating 300%-plus growth on JD. New product launches averaged one every three days. The metrics that matter more: brands outside the top five command 54% market share — versus 17% in mature smartphone categories — signaling commodity fragmentation rather than market health. Return rates run at 60%–70%, unchanged year-on-year. Display-category gross margins hover around 20%, with the majority transferred to distribution channels, leaving manufacturers near zero. Xiaomi distributed AI glasses discount vouchers to EV owners; a significant proportion of those vouchers subsequently appeared on Xianyu secondhand platforms priced at a few dozen renminbi.
Nearly 400 brands are active in the Chinese market, with approximately 70% of H1 2026 funding concentrated in upstream optical and micro-display supply chain components rather than terminal brands. A gray market of data agencies selling bespoke "number-one" market-share rankings to paying clients — selecting favorable metrics and category definitions to manufacture leadership positions — has emerged as a structural feature of the category. XREAL's April Hong Kong IPO filing has compressed the timeline for other brands toward secondary markets. The obstacle: public investors will see profitability and cash flow data that private backers could defer. By near-universal assessment among industry participants, those are the two weakest numbers in every AI glasses company's financials.
What to Watch
The November 2026 CapCut-ChatGPT beta launch will be the first real test of whether embedding AI editing inside a chat interface converts intent into measurable daily active user retention — a metric ByteDance has not disclosed separately. In automotive chips, the next 18 months will determine whether OEM-captive chip units can scale fast enough to amortize R&D costs before the supplier landscape consolidates. Unitree's Q4 earnings will offer the first market test of whether UnifoLM-WLA-1.0 is shifting commercial conversations, or whether the hardware showcase is still doing all the work. For Cambricon, full-year results will answer the inventory question the Q3 call deliberately left open.