Hong Kong Cyberport is rapidly emerging as the preferred springboard for Chinese AI companies expanding internationally, transitioning from a local tech park into a full-service globalization platform connecting mainland innovation with Southeast Asian, Middle Eastern, and global markets. Zhipu AI, Yunji Technology, and CAS Hong Kong Innovation Institute are among the notable success stories that have leveraged Cyberport infrastructure, funding networks, and regulatory familiarity to accelerate their international presence.

The first deck is computing power. The Cyberport AI Supercomputing Center, launched in 2024, provides 3,000 PFLOPS supported by a 3 billion HKD three-year AI funding program covering 70-90% of AI development costs. CAS Hong Kong Innovation Institute used this compute to train CARES Copilot, a surgical video AI model that became the global number-one downloaded model in its category.

The second deck is standards and regulation. Chinese AI companies often discover that products proven in the mainland market encounter unexpected barriers in overseas markets around data governance, contract law, liability frameworks, and cultural expectations. Cyberport provides a controlled international environment where companies can adapt their products to global standards before full-scale international expansion. WaveCloud, a major Chinese cloud provider, adopted a three-step strategy: first establishing joint labs with Hong Kong universities for technology incubation, then formally landing at Cyberport in 2024 where 26 international technologies were incubated, and finally launching the Hong Kong Smart Cloud Platform in 2025. The company gained not just market access but standards leadership, using Hong Kong as a base to participate in international AI standard-setting.

The third deck is market access. Yunji Technology, the service robot company listed on HKEX in October 2024 as the service robot first stock, exemplifies how Cyberport bridges product capability with market readiness. Yunji robots already serve over 40,000 hotels globally with 750 million service executions annually across 20 countries. However, Hong Kong taught the company a different lesson: market acceptance depends not on technical capability alone but on nuanced questions about whether hotel staff and guests accept robots, how fire and property regulations apply, who takes responsibility for failures, and whether service workflows designed for mainland hotels adapt to overseas hotel management styles.

Cyberport CEO Dr. Zheng Songyan described the transformation as moving from isolated company registrations to industrial cluster expansion, with multiple mainland cities including Zhongguancun, Suzhou, and Guangxi establishing dedicated overseas bases at Cyberport. With over 200 venture capital connections, IPO assistance programs, and organized market access trips to Southeast Asia and the Middle East, Cyberport has evolved from a simple Hong Kong landing pad into a comprehensive globalization platform where Chinese AI companies learn not just how to export technology but how to become multinational technology enterprises.

WAIC 2026 has become the hottest ticket in Shanghai literally. The 168 yuan standard day pass was being scalped for over 1,000 yuan online, and the 899 yuan membership pass was sold out weeks before the conference opened. On the morning of July 17, visitors began queuing at 10 AM in near-50 degree outdoor temperatures, waiting nearly two hours for admission. The scene captured a defining moment: China AI enthusiasm has reached fever pitch, and everyone whether AI practitioners, startup founders, investors, or job-seeking graduates wants to be part of it.

The scale is unprecedented. Over 1,100 enterprises present more than 3,000 exhibits across 100,000 square meters spanning three zones and four halls. More than 300 products are planned for global premiere. The conference features a historic three-zone, four-hall layout: World Expo for core technology and industry applications, Zhangjiang for computing infrastructure, and Xuhui West Bund for consumer AI experiences. The scope reflects how AI has expanded from a specialized technology domain into a cross-industry infrastructure layer touching everything from chip design to coffee delivery robots.

The energy inside the halls was palpable. Approximately two-thirds of attendees wore badges color-coded by role: red for VIP guests, blue for media, orange for exhibitors and forum hosts, green for staff. This visual identification system turned the exhibition floor into a dense networking environment where strangers could identify relevant contacts instantly. Beyond the floor, dedicated capital matching zones brought over 200 professional investors to meet with startups, while one-person company pavilions showcased 22 solo founders selected from 711 applicants, reflecting the democratization of AI entrepreneurship.

Beyond the hype, substantive signals emerged. The conference agenda revealed that the three most discussed topics across 200 sessions were no longer large model benchmarks but industry development, talent ecosystem, and computing power. The most crowded exhibition booths were no longer model demos but working robot production lines and supernode computing clusters. Kimi K3, Honor Robot Phone, and StepFun STEPX Neo drew the largest crowds, but the deeper story was how these products connected into a coherent industrial ecosystem.

WAIC 2026 marks a transition point. As Wu Xiaobo noted, China has always been the most aggressive internet experimentation market, and its approach to AI reflects the same spirit combining anxiety, passion, and action. The queues in 50-degree heat, the scalped tickets, and the relentless networking are not just FOMO they represent a collective national response to a technological transformation that China believes it can lead rather than follow.

ZTE Nubia and ByteDance have unveiled the second-generation Doubao phone, the Nubia NaviX Ultra, at WAIC 2026. The device represents the evolution of the agent smartphone concept first introduced with the Nubia M153 Doubao phone in late 2025, incorporating lessons from eight months of co-development between Nubia and the ByteDance team to address four verification gates: user pain point translation, making agent capabilities native to the phone OS, converting single successes into stable experiences, and most critically, security and privacy.

The hands-on experience reveals significant improvements. The NaviX Ultra features a redesigned horizontal large camera module with four color options: blue, dream, black, and white. The glass-like UI integrates Doubao with other apps on the phone, enabling cross-application task orchestration. In a live demonstration, instructing Doubao to find cat videos on Douyin triggered autonomous app operation, with the phone running the app independently and presenting results within approximately 30 seconds. The AI displays real-time execution steps, reasoning processes, and task progress with visual animation feedback throughout.

Multi-tasking represents a major advancement. The NaviX Ultra supports submitting multiple tasks simultaneously: asking Doubao to find videos on Douyin while searching for flights on Qunar and ordering an iced American at Luckin Coffee simultaneously. Tasks that cannot be processed immediately are automatically queued and executed in sequence. This queue-based autonomous execution model transforms the phone from a tool requiring constant user attention into an agent that independently works through task lists in the background.

Proactive memory is a key differentiator. The phone automatically learns user preferences from content saved across apps and from usage patterns, creating a personal memory system that makes the phone feel more intelligent and responsive over time. Unlike first-generation agent implementations that required explicit user commands for every action, the NaviX Ultra aims for anticipatory assistance where the phone proactively offers relevant actions based on context, implicitly drawing on user data with a privacy-first architecture that keeps user ownership of collected information.

Nubia president Ni Fei emphasized that the agent smartphone category has reached industry consensus, with competitors now entering the space, but warned it is not a demo category. The four gates identified during development user pain, native OS integration, stable experience at scale, and privacy represent engineering challenges that differentiate production-ready products from concept demonstrations. With over 10 partner platforms including Douyin, Luckin, Qunar, Meituan, and Didi integrated, the NaviX Ultra positions the Doubao phone ecosystem as the most extensive agent smartphone platform available, capable of executing real-world errands rather than just demonstrating technical capability.

Honor has unveiled the world first Robot Phone at WAIC 2026, a device that fundamentally reimagines the smartphone form factor by integrating a 4-degree-of-freedom titanium alloy mechanical gimbal system into the phone body. The gimbal enables the camera module to move independently, following subjects, responding to music rhythm, and physically expressing AI output through motion, transitioning smartphones from passive touch-screen tools to proactive interactive companions.

The mechanical gimbal system is the centerpiece innovation. Built from titanium alloy with micro-motors 70% smaller than mainstream alternatives, the system achieves CIPA 5.5-level stabilization with 96% walking shake compensation. The camera module can rotate 360 degrees for tracking and supports 90-degree and 180-degree smart motion transitions for cinematic effects. In daily use the system hides entirely within the phone body, deploying with a single button press when needed. The engineering achievement required solving substantial challenges in miniaturization, durability, and power management within smartphone form factor constraints.

The AI capabilities go beyond traditional smartphone assistants. Honor YOYO AI demonstrates proactive task execution, processing continuous commands from a single user request: opening anti-motion-sickness mode, adjusting screen brightness, setting alarms, and hailing a taxi all executed sequentially with minimal user intervention. In a World Cup prediction demonstration, Robot Phone leveraged real-time match data analysis to predict Spain defeating Argentina 3-1, physically expressing its analytical confidence through the gimbal motion system nodding and swaying to emphasize key points during the prediction explanation.

Camera specifications are flagship-grade. The rear triple camera system includes a 200MP F1.6 23mm-equivalent 4D gimbal main camera, a 50MP ultra-wide lens, and a 200MP periscope telephoto lens. The phone is also the first product from the Honor-ARRI strategic partnership, bringing ARRI Log-C encoding and LUT color grading to mobile video, along with professional cinema workflow features including C-Log, LUTs, focus lock, white balance lock, AE lock, and motion data recording, targeting serious video creators.

The Robot Phone represents the first concrete product from Honor Alpha Strategy announced in March 2025, the company three-phase roadmap to transform from a smartphone manufacturer into a leading AI terminal ecosystem company. Rather than incremental hardware upgrades, the device represents a category redefinition: when AI moves beyond software into physical form, the smartphone evolves from a cold electronic slab into an intelligent companion that perceives its environment, connects with the physical world, and proactively serves its user. The mechanical gimbal is not a gimmick but the physical manifestation of Honor thesis that the next competitive frontier in smartphones is how AI capability integrates with device morphology to enable new interaction patterns.

Xiaomi next-generation flagship smartphone series has passed network certification in China, with model number M154FF confirmed as the Xiaomi 18 Pro series. The certification covers GSM, WCDMA, TD-LTE, LTE FDD, and 5G network bands including the 4800-4960 MHz N79 frequency. Multiple sources indicate the lineup will launch on September 24, following the Qualcomm Snapdragon Summit scheduled for September 22-24 in Hawaii, with Xiaomi expected to secure global first-launch rights for the 2nm Snapdragon 8 Elite Gen6 chip.

The processor is the centerpiece. Snapdragon 8 Elite Gen6 is built on TSMC N2P enhanced 2nm process with Gate-All-Around transistor architecture, delivering 30% higher transistor density than the previous 3nm generation. The Adreno 850 GPU includes 18MB dedicated graphics cache and supports LPDDR6 high-speed memory. At equivalent performance, power consumption is reduced by 36%, while at equivalent power, performance improves by 18%, marking one of the largest generational efficiency leaps in recent smartphone processor history.

Camera specifications are equally ambitious. The Xiaomi 18 Pro Max will feature a dual 200MP configuration: a 1/1.28-inch LOFIC ultra-large main sensor and a 1/1.56-inch large-sensor telephoto macro with 3x optical zoom and F2.4 aperture supporting approximately 15cm close-focus macro photography, complemented by an ultra-wide-angle lens. The Xiaomi 18 Pro shares the dual 200MP main and telephoto macro combination. This imaging system positions the 18 Pro series as a direct competitor to dedicated camera systems, leveraging the LOFIC sensor technology for improved dynamic range and the high-resolution telephoto for versatile composition options.

Battery capacity sees substantial upgrades reflecting both the higher efficiency of the 2nm chip and the increasing power demands of on-device AI processing. The Xiaomi 18 Pro, positioned as a compact flagship, packs a 7000mAh-class battery, while the 18 Pro Max reaches 8000mAh-class capacity. Both models support 100W wired fast charging and wireless charging. The combination of 2nm efficiency gains and larger battery capacity promises significant endurance improvements over the Xiaomi 17 series.

However, pricing is expected to increase noticeably. Storage chip prices have been rising since H2 2025 as AI data center demand consumes available memory production capacity, directly affecting smartphone bill of materials. Combined with the 2nm wafer cost premium and the dual 200MP camera system expense, across-industry sources expect the Xiaomi 18 Pro series to carry a meaningful price increase over its predecessor. The final pricing will be revealed alongside the September launch event, where Xiaomi will need to balance component cost pressures against the increasingly price-sensitive Chinese smartphone market that has experienced five consecutive quarters of declining shipments.

Tencent has officially released the WorkBuddy APP simultaneously on iOS, Android, and HarmonyOS, making it the first native AI agent application to land on Huawei HarmonyOS ecosystem. The app brings the desktop WorkBuddy agent capabilities to mobile form factors, supporting both direct connection to local computers and cloud-based independent operation, significantly expanding the reach of Tencent AI productivity tools.

The dual-mode architecture is the key differentiator. In computer-connected mode, the app links directly to the user desktop WorkBuddy instance, providing full access to desktop capabilities including file system operations, application control, and long-running automation tasks. In cloud work mode, the app operates independently with most core desktop features including Skills integration, expert consultations, automated scheduled tasks, output downloading and sharing, and project management, connected through Tencent Docs and ima knowledge base for data access.

WorkBuddy supports multiple input modalities including text, voice dictation, photo upload, local file import, and direct import from Tencent Docs and ima. The user can describe a work objective in natural language and WorkBuddy autonomously breaks down the multi-step task, operates computer files, generates documents and reports, and delivers completed outputs. This represents a fundamental departure from traditional AI tools that only provide suggestions and recommendations, positioning WorkBuddy as an executable digital coworker rather than a conversational assistant.

The HarmonyOS launch is strategically significant. Huawei HarmonyOS has been building its native application ecosystem since leaving Android compatibility, and WorkBuddy becomes the first generalized AI agent application available on the platform. The cross-platform parity ensures user experience consistency across the three major mobile ecosystems, critical for enterprise deployment where organizations use diverse device fleets. The mobile launch follows WorkBuddy establishment as a category leader with 8.85 million monthly visits on desktop according to Analysys data.

Tencent agent strategy positions WorkBuddy as the flagship product in a portfolio that includes Zaohua Gongfang for creative content, Ardot for intelligent design, and specialized agents for WeChat ecosystem operations. The mobile app launch completes the device coverage, enabling continuous AI assistance across phone, tablet, and computer with seamless task handoff. As agent-based AI productivity becomes the mainstream paradigm, WorkBuddy multi-platform coverage and deep integration with Tencent ecosystem apps including WeChat, WeCom, Tencent Meeting, and Tencent Docs creates a defensible position that competing standalone agent products will find difficult to replicate.

Elon Musk called it impressive. The Nasdaq called it another reason to sell. Kimi K3, released by Moonshot AI on the eve of WAIC 2026, has become what analysts call the third shockwave from Chinese AI after DeepSeek R1 and Manus. Each struck a different nerve: DeepSeek challenged compute cost assumptions, Manus raised the agent product ceiling, and K3 now confronts the capability premium that US frontier AI companies rely on for pricing power.

The market reaction was swift. On July 16, the Nasdaq fell 1.47% and the Philadelphia Semiconductor Index dropped 4.29%, with CNN and AP citing K3 as a contributing factor. K3 provided a fresh reason for investors to question AI valuation assumptions built on the premise that Chinese models would remain permanently behind.

The fundamental difference between K3 and earlier Chinese models is pricing. At $3 per million input tokens and $15 per million output tokens, K3 is approximately 3-4 times the price of Moonshot K2.6 and significantly more expensive than DeepSeek or GLM. This breaks the Chinese cheap model narrative that had dominated global perception. Moonshot is betting that performance in complex coding, agent, and knowledge work tasks can command a premium, and early data supports this: in Artificial Analysis testing, K3 average task cost of $0.94 is lower than Claude Opus 4.8 $1.80 and close to GPT-5.6 Sol $1.04, meaning higher token prices still result in competitive total task economics when the model requires fewer attempts or shorter reasoning chains.

The timing is significant. Chinese models had already crossed 30% of US enterprise token usage on OpenRouter by mid-2026, peaking at 46%. DeepSeek share rose from 9% to nearly 20%. But most of this usage was on low-margin tasks. K3 attempts to shift Chinese AI from commodity token supplier to premium capability provider, competing directly with Anthropic Claude Opus and OpenAI GPT-5 series for high-value enterprise workloads where reliability and capability matter more than absolute token cost.

Silicon Valley reactions reflect the changing landscape. Vercel CEO Guillermo Rauch published benchmark results showing K3 achieving the best performance on nextjs.org evaluations, reaching parity with Fable in less time. Ai2 senior research scientist Nathan Lambert argued that distillation arguments no longer explain away Chinese model progress, noting that Chinese teams are genuinely good at building models now. Bernstein analyst Robin Zhu called K3 a home run, stating that Chinese top labs keeping pace with global frontiers is no longer coincidental. The third shockwave may be the most consequential because it attacks not just cost assumptions but the profit pool supporting the entire US AI ecosystem business model.

WAIC 2026 has concluded with over 1,100 enterprises showcasing more than 3,000 exhibits across 100,000 square meters of exhibition space, the largest in the conference history. Beyond the scale, the content has fundamentally shifted. From six days of walking the exhibition floors and attending forums, several clear trends emerge that define where the Chinese AI industry is heading.

Trend one: the strongest model is no longer the biggest winner. While Kimi K3 dominated discussions with its 2.8 trillion parameter open-source release, the conference agenda told a different story. The three most frequent topics across 200 sessions were not large models but industry development, talent ecosystem, and computing power. When models were discussed, the questions had shifted from benchmark rankings to how models enter real production systems. The real competitive battleground has moved beyond model capability to system efficiency, product experience, and industrial deployment.

Trend two: robots have stopped performing tricks and started working. Robot demonstrations at WAIC 2026 overwhelmingly focused on real industrial tasks rather than dancing or acrobatics. Companies brought entire production lines to the exhibition floor assembly, sorting, quality inspection. Family scenarios featured robots understanding long-horizon instructions like organizing a living room. The shift reflects a broader realization that the gap between completing a single action and performing reliably in real environments remains the central engineering challenge for embodied intelligence.

Trend three: domestic computing infrastructure reached a tipping point. Supernode cabinets dominated the infrastructure hall, with every major domestic chip company presenting cluster-level solutions. Huawei Atlas 950, Sugon Dawn 8000, and Dongfang Suanxin DF1000 demonstrated that the conversation has moved from individual chip specifications to system-level computing efficiency. The ecosystem compatibility barrier that once kept domestic chips on the sidelines has eroded as mainstream models complete Day-0 adaptation.

Trends four through six cover the rise of world models as the next data battleground, the proliferation of agent-native consumer terminals from phones to glasses to earbuds, and the globalization of Chinese AI. International participation reached new highs with 1,400 guests from 10-plus international organizations. The conference itself has evolved from a product showcase into a coordination mechanism connecting scientists, model companies, chip makers, robot builders, investors, and policymakers in a shared trajectory toward industrial-scale AI deployment across the Chinese economy.

CXMT ChangXin Memory Technologies has launched its STAR Market IPO at 8.66 yuan per share, with total fundraising expected to reach approximately 66.6 billion RMB, making it the largest A-share IPO of 2026. The IPO attracted 9.43 million effective subscription accounts with a lottery rate of approximately 0.47%, both STAR Market records. Multiple institutions project a post-listing market valuation between 2-3 trillion RMB, which would place CXMT among the most valuable semiconductor companies globally.

The listing comes at a pivotal moment for the global memory industry. CXMT has grown to become the world fourth-largest DRAM manufacturer with 7.67% market share as of Q4 2025, according to Omdia. Annual revenue for 2025 reached 29.1 billion RMB with net profit of 24.8 billion RMB in Q1 2026 alone, representing 1,688% year-over-year growth. The company projects H1 2026 net profit of 50-57 billion RMB, driven by the AI-driven memory super-cycle that has created unprecedented demand for both traditional DRAM and emerging HBM products.

The IPO structure reflects strategic positioning. Alibaba holds approximately 5% of CXMT shares through a 7.6 billion RMB investment that has become worth an estimated 130 billion RMB at IPO valuation, one of the most successful corporate investments in Chinese technology. The funds raised will be directed entirely toward DRAM production expansion. Production capacity is expected to reach approximately 350,000 wafers per month by end of 2026, approaching Micron level output according to Citrini Research.

The market dynamics favor CXMT expansion. Samsung, SK Hynix, and Micron are aggressively shifting production capacity toward high-margin HBM for AI accelerators, voluntarily vacating the mainstream DDR5 and LPDDR5 markets. CXMT has captured this market opening, with LPDDR5, DDR5, and LPDDR5X products achieving volume production in 2023-2025. The company latest LPDDR5X 10667 products have already been sampled, matching the specification level of international leaders.

However, significant challenges remain. Without EUV lithography equipment due to export controls, CXMT trails international peers by approximately two technology generations in process nodes, resulting in higher unit cost. HBM production remains elusive with unresolved yield challenges. Counterpoint Research director MS Hwang estimates that Chinese DRAM manufacturers will not catch Korean leaders in high-end products until at least 2030. The company must navigate intellectual property risks, achieve approximately 15% global market share to generate sufficient EBITDA for self-funded capex, and develop breakthrough innovations rather than following established technology paths. Morgan Stanley recently downgraded Samsung and SK Hynix ratings citing CXMT capacity expansion as a structural risk to Korean memory dominance, validating the growing impact of Chinese DRAM on the global memory industry landscape.

Moonshot AI has priced Kimi K3 at $3 per million input tokens and $15 per million output tokens, approximately 3-4 times the pricing of the previous K2.6 model. This makes K3 the most expensive Chinese large model by API pricing, moving directly into the price range of mid-to-high-end US models. The pricing strategy marks a significant departure from the past year playbook of Chinese AI companies competing primarily on cost advantage, with DeepSeek and other Chinese models often priced 60-90% below comparable US offerings.

The rationale reflects a maturing market. Chinese models have successfully crossed the first adoption threshold through low pricing, with OpenRouter data showing Chinese models accounting for over 30% of US enterprise token usage since February 2026, peaking at 46%. However, usage volume does not automatically translate to revenue for model companies, since open-weight models can be deployed on third-party cloud platforms and developers can switch between providers through routing platforms like OpenRouter. K3 pricing signals Moonshot ambition to prove that Chinese models can command premium pricing for complex task capability rather than competing solely on commodity token economics.

The K3 architecture supports this positioning. With 2.8 trillion parameters using 896 experts with 16 activated per inference, 100 million token context window, and native image understanding, the model targets long-horizon agentic coding, software engineering, and complex knowledge work. Artificial Analysis benchmark data shows K3 achieving a 57 overall intelligence score, approaching Claude Opus 4.8 levels. In frontend coding blind tests, K3 briefly ranked first on Arena, surpassing Fable 5. The average task cost of approximately $0.94 per evaluation is comparable to GPT-5.6 Sol and lower than Claude Opus 4.8, suggesting that higher token prices can still result in lower total task costs when the model requires fewer attempts or shorter chains of thought to complete complex assignments.

The pricing strategy also reflects changing market dynamics. Anthropic plans to raise Opus 4.8 prices by 50% in September 2026, while Chinese computing infrastructure has reached a tipping point where domestic chips can support model training and inference at scale. The combination of improving model capability, competitive task-level economics, and increasingly adequate domestic computing infrastructure enables Chinese AI companies to move beyond the cost leader trap and compete on capability differentiation.

The true significance of K3 pricing may extend beyond the model itself. If Moonshot can sustain premium pricing in the face of open-source alternatives, it validates that Chinese AI companies can build sustainable businesses based on model quality rather than relying on subsidies or cost dumping. This would represent a structural shift in the global AI competitive landscape, where Chinese AI companies transition from being seen as efficient imitators to legitimate premium providers capable of capturing value from frontier AI capabilities.

Moonix, the AI wearables brand under Xinmou Technology, has begun global sales of its AI glasses standard edition at 2299 yuan, claiming the title of world lightest AI glasses at just 14.9 grams. The split-structure design features temple sections only 4mm at their thinnest point, enabling a form factor visually indistinguishable from regular optical glasses. The battery life reaches 16 hours including 6 hours of active recording and 10 hours of smart recording, sufficient for all-day wear from morning to night.

The product strategy reflects a deliberate focus on recording and memory rather than the feature-stacking approach common in the smart glasses industry. Moonix actively removed translation, teleprompter, and navigation functionality that the team determined have low real-world usage frequency. Instead, the glasses focus exclusively on ambient recording with an AI assistant that automatically captures and processes environmental information at regular intervals. A button-triggered manual mode provides continuous recording for important events. The core philosophy, as stated by Xinmou Technology president Guo Yuchen, is that recording is the only AI glasses function that generates compounding value over time, since human information intake is 90% visual and auditory, and glasses occupy the optimal physical position for both.

The hardware capability is supported by six microphones with dual recording modes: active button-triggered recording and algorithm-driven smart recording that dynamically determines when ambient capture is valuable. Privacy protection is built into the product architecture: all user data belongs to the user, raw data is stored exclusively on the user phone, cloud processing is ephemeral with raw data deleted after analysis, and the company has committed to open-sourcing data processing modules for public audit.

Moonix offers 39 frame styles in beta titanium, nano nylon, acetate, and stainless steel materials, with the beta titanium option achieving the 14.9 gram headline weight. The glasses are available through Moonix official mini-program, JD.com flagship store, 100 Dr. Glasses offline stores, and 3 Wujie Optics stores, allowing consumers to try before buying. A Pro version with camera will debut at WAIC 2026 Xuhui West Bund zone for August release, and an Ultra variant with display is in pre-development.

The launch addresses what Moonix identifies as the fundamental challenge of AI glasses: the industry is shouting but not selling because existing functionality is insufficient for all-day wear. By prioritizing weight and focusing on the single high-frequency use case of recording and memory, Moonix aims to achieve what others have not: making AI glasses that users genuinely want to wear every day, creating a continuous personal knowledge base that becomes more valuable over time through accumulated life logging and AI-powered memory retrieval.

WAIC 2026 has become a coming-out party for Chinese AI computing infrastructure. With 108 chips, 261 large models, 208 AI terminal products, and over 300 real robots across three zones spanning more than 100,000 square meters, the scale reflects a domestic computing ecosystem reaching critical mass. The Zhangjiang zone dedicated an entire section to Chinese chips, covering 40-plus domestic computing companies including Huawei, Muxi, Moore Threads, and Tianshu Zhixin.

Supernode cabinets dominated the computing infrastructure exhibits. Huawei Atlas 950 SuperPoD showcased 1024 Ascend card interconnection with full adoption of 800G optical modules and high-density fiber optics, a dramatic upgrade from last year copper-heavy CloudMatrix 384. Other exhibitors including ZTE, Muxi, Biren Technology, and Moore Threads all presented competing supernode solutions. Some companies like Tsingway and Dongfang Suanxin demonstrated reconfigurable and programmable chip architectures that use control circuits to reconfigure computing units for higher utilization without relying on advanced process nodes.

The ecosystem maturation is visible in the reduced concern over software compatibility. As one industry insider noted, last year domestic chips struggled to gain adoption due to software ecosystem concerns and expectations for NVIDIA restricted supply. This year, with NVIDIA unable to supply China and the trillion-dollar US AI infrastructure buildout consuming most advanced GPU capacity, domestic chips have been pushed onto the supply table. When customers face no alternative, the ecosystem compatibility issue naturally diminishes over time as mainstream models complete Day-0 adaptation for domestic computing platforms.

Model-side highlights included Kimi K3, the 2.8 trillion parameter open-source model that became the most discussed AI model at the conference. Beyond benchmark scores, the real story is cost optimization: Kimi K3 uses a Stable LatentMoE framework with 896 experts activating only 16 per inference, achieving competitive performance at accessible pricing points that make AI affordable for the mass market.

The contrast between headline models and the broader ecosystem is revealing. While leading AI labs chase Fable 5 and GPT-5.6 level performance, the majority of WAIC exhibitors focused on vertical applications in healthcare, life sciences, education, and tourism. Bei Dian Shu Zhi demonstrated how AI-enabled family medicine allows one online staff member to support 5,000-8,000 residents. Baichuan showed M4 model applications in oncology and pediatrics. These deployments prove that most enterprises do not require frontier-level models for practical value, making cost-effective domestic computing infrastructure the true enabler of AI industrialization across China economy.

On the opening day of WAIC 2026, nearly 30 entrepreneurs from across China AI ecosystem shared their expectations for the country AI development, reflecting a unified vision that AI must transition from laboratory breakthroughs to real-world productivity transformation. The letters, published by Yicai, span AI model companies, infrastructure providers, robotics startups, healthcare AI firms, and global partners investing in China AI.

Zhipu AI emphasized that AI competition is not a zero-sum game, and real security comes from openness rather than closure. The company pledged to continue converting technological breakthroughs into accessible stepping stones for the global developer community. Sugon senior vice president Li Bin highlighted the shift from visible technological breakthroughs to visible application value, pointing to Dawn 8000 as infrastructure designed for this transition. JD Cloud vice president Gong Yicheng stressed that high-quality physical world data remains the critical bottleneck for embodied intelligence development, with JD leveraging its logistics, retail, and industrial scenarios to build embodied AI data systems.

Multiple entrepreneurs emphasized the importance of AI meeting real human needs. Agricultural AI company Maimai Tech CEO Li Nan called for AI to reach farms and fields rather than remaining a laboratory technology. Shenzhen Wuxin Tech CEO Zhang Jiaming hopes AI elder care products can reach nursing homes and homes more quickly. Liangyuan Xinchuang founder Jiang Xu expects physical AI to evolve through three stages: large-scale pre-training, alignment, and deployment.

Global participation was notable. Insilico Medicine founder Alex Zhavoronkov expressed excitement about China using cutting-edge AI to develop new drugs extending human healthy lifespan. SkAI Intelligence co-founder Lee Jae-cheol emphasized that physical AI needs not just model capability but a complete data loop of digital twins, synthetic data, simulation, and field feedback working together with industrial robot partners like ABB.

The collective message reflects a maturing Chinese AI industry: the model race has produced globally competitive capabilities demonstrated by K3, GLM-5.2, and SenseNova-Vision, but the real test lies in deployment at scale across industries. Entrepreneurs consistently returned to the theme that AI should be an integral part of industrial innovation systems rather than an attached tool or plugin, with standards, open interfaces, and cross-domain collaboration identified as critical enablers for the next phase of growth.

Alibaba Tongyi Qianwen has unveiled its first AI agent earbuds at WAIC 2026, marking the company entry into AI wearable hardware. The clip-on earbuds feature an all-day wearable design and integrate Qianwen AI assistant capabilities including real-time simultaneous interpretation, automated meeting minutes generation, and health monitoring features, representing a new paradigm for always-available AI interaction beyond smartphone screens.

The product design philosophy centers on ambient intelligence. Unlike traditional AI interaction models that require users to look at a screen, type queries, or hold conversations with their phone, the clip-on form factor enables continuous, context-aware AI assistance in the background. The earbuds can listen to meetings and generate automatic summaries, provide real-time language translation during cross-language conversations, and track health metrics through continuously worn sensors. The use case is fundamentally different from smartphone-based AI: instead of pull-based interaction where the user initiates a request, the earbuds enable push-based intelligence where the AI proactively provides relevant information based on the user current audio context.

The real-time simultaneous interpretation capability is particularly notable. The earbuds can process spoken language input and deliver translated output through the ear speaker with minimal latency, supporting cross-language conversations without requiring either party to switch to a common language. This positions the product for business travelers, international conference attendees, and cross-border professionals who represent a significant addressable market for always-on translation hardware.

The Qianwen AI agent earbuds wearables strategy represents Alibaba effort to extend its AI ecosystem beyond cloud and software into physical form factors. The company already supplies Qianwen model capabilities to Apple Intelligence for the China market and has partnered with multiple hardware manufacturers. By launching its own branded wearable, Alibaba gains direct user interaction data, hardware-software integration control, and a showcase platform for demonstrating Qianwen capabilities in a form factor optimized for its model strengths.

Consumer AI hardware remains an experimental but rapidly growing category. The global smart wearable market is expanding with AI integration becoming a key differentiator. Alibaba entry into AI earbuds follows similar moves by ByteDance with Ola Friend and Tencent with potential wearable products, indicating that Chinese AI platform companies see wearable form factors as the next battleground for AI user engagement. The success of Qianwen earbuds will depend not just on AI capability but on battery life, comfort, and whether the use cases are compelling enough to change consumer behavior away from smartphone-based AI interaction patterns.

WeRide has released WITT, a physical AI foundation model that uses minimum physical fact units as its fundamental building blocks for unified multimodal scene understanding. The model is designed to serve as the perceptual and cognitive backbone for WeRide autonomous driving systems and broader robotics applications, representing the company bet that physical AI requires fundamentally different architectural approaches from language or vision-only models.

WITT concept of minimum physical fact units draws inspiration from how humans understand the physical world through discrete, composable observations. Rather than processing raw sensor streams through separate perception modules for detection, tracking, prediction, and planning, WITT unifies these traditionally separate functions within a single framework that reasons about the physical world at a fundamental level. Each minimum physical fact unit represents a quantum of physical understanding that can be combined and composed to represent arbitrarily complex scenes.

The model is designed to handle the diversity of autonomous driving scenarios that traditional modular approaches struggle with. In complex urban environments with unpredictable actors, occluded viewpoints, and varying weather conditions, WITT can leverage its unified representation to maintain consistent scene understanding across different sensor modalities and operating conditions without requiring separate model architectures for each task.

WeRide, founded in 2017 by former Baidu autonomous driving executives, has established itself as one of the leading commercial autonomous driving companies globally, with robotaxi operations in multiple Chinese cities and international markets including the UAE and Singapore. The company went public on Nasdaq in 2024 and has been expanding beyond robotaxis into autonomous vans, minibuses, and sanitation vehicles. WITT represents a strategic pivot toward developing proprietary physical AI technology as a foundation for all WeRide product lines, rather than relying on third-party perception models.

The physical AI foundation model approach positions WeRide for the next phase of autonomous driving competition, where the differentiator shifts from deployment scale to intelligence generalization. As autonomous driving companies transition from geofenced operations to broader operational design domains, the ability to understand and reason about the physical world in a unified manner becomes increasingly critical. WITT unified architecture could provide advantages in handling corner cases, adapting to new environments with minimal additional training, and transferring learned knowledge across different vehicle types and deployment scenarios.

Sugon Dawn 8000, the large-scale fully domestic super-AI fusion cluster, made its public debut in physical form at WAIC 2026 and was selected as a Treasure of the Hall exhibit. The system adopts a super-AI fusion architecture supporting full-precision computing from FP64 to INT8, with each computing unit achieving a 20x improvement in computing density over previous generations. The self-developed scaleFabric interconnect technology enables stable interconnection at the 100,000-card scale, making it one of the largest homogeneous domestic AI computing clusters demonstrated to date.

The system represents the latest evolution of Sugon computing infrastructure strategy. By integrating supercomputing precision with AI computing efficiency in a unified architecture, Dawn 8000 can handle scientific computing workloads requiring FP64 double-precision accuracy alongside AI training workloads that benefit from lower-precision INT8 and FP8 formats. This flexibility is critical for emerging AI-for-science applications that combine traditional numerical simulation with neural network-based approaches.

Sugon has completed over 300 application optimizations for the Dawn 8000 platform covering more than 20 research and industry domains. The application portfolio spans climate modeling, materials science, drug discovery, and other computationally intensive fields. The system has already been deployed at a 100K-card core node where these optimizations were validated at scale, demonstrating that domestic AI infrastructure can support production-grade workloads previously only feasible on imported GPU clusters.

The Dawn 8000 launch at WAIC comes as Chinese AI computing infrastructure enters the system era. Rather than competing solely on individual chip specifications, the focus has shifted to cluster-level performance, interconnect bandwidth, software ecosystem, and application readiness. Sugon approach of building a complete super-AI fusion system with domestic processors and full-stack software optimization positions it as a key infrastructure provider for the expanding Chinese AI data center market, competing with both Huawei Atlas clusters and traditional GPU-based solutions.

Sugon senior vice president Li Bin stated at WAIC that AI is moving from visible technological breakthroughs to visible application value. Making AI accessible and usable for more industries and enterprises is the key challenge for the next development phase. Dawn 8000 represents an exploration toward meeting future AI development demands, alongside similar systems from Huawei, Alibaba, and other Chinese technology leaders who are collectively defining what domestic AI infrastructure looks like at scale.

Fudan University research team led by Professor Zhou Peng and Associate Professor Liu Chunsen at the State Key Laboratory of Integrated Chips and Systems has published a breakthrough in Science: Quantum Flash technology that achieves room-temperature non-volatile single-electron storage. The innovation constructs a coplanar drain-channel-source unified structure called Guiyi, enabling clear observation of single-electron charge storage behavior at 27 degrees Celsius for the first time, overturning the long-held assumption that single-electron non-volatile storage was theoretically impossible at room temperature.

The breakthrough builds on the team previous Nature publication from April 2025, where they demonstrated the PoX flash device achieving 400-picosecond non-volatile storage, the fastest semiconductor charge storage technology ever reported. While PoX solved the speed-versus-nonvolatility dilemma that had persisted since the floating-gate transistor invention in 1967, the density question remained. Quantum Flash addresses the fundamental storage density limit by achieving one electron per bit, the smallest physically possible data unit as electrons are indivisible elementary particles.

The technical challenge was immense. When an electron is confined to increasingly small spaces, quantum effects become significant. Previous attempts in 1997 Science publication showed only 55 mV signals that decayed within 5 seconds at cryogenic temperatures. The Fudan team used two-dimensional semiconductor atomic-level thickness as a natural confinement layer, combined with a self-aligned planar cutting approach, to create the Guiyi structure. The result: a single injected electron produces a 0.5 Volt storage window that remains stable at room temperature, amplifying the quantum signal by nearly an order of magnitude compared to prior work.

The strategic significance extends beyond the laboratory. Traditional memory hierarchy separates high-speed cache from slow bulk storage. Quantum Flash technology, combined with the team PoX speed breakthrough and Changying prototype chip that has already demonstrated CMOS-compatible integration, points toward unified memory that is simultaneously fast, dense, and non-volatile. This would fundamentally change AI chip architecture by eliminating the data migration bottleneck between compute and storage.

The team has completed the full chain from materials to prototype chip verification and plans to commercialize within 1-3 years, including establishing a company, engaging leading AI customers, and bringing strategic partners on board. If successful, the technology could enable mobile devices and servers to run larger local AI models with longer context memory while consuming significantly less power, addressing the memory wall that has become the primary bottleneck in AI compute efficiency.

ModelBest, the edge AI company behind the MiniCPM series, has seen its valuation surge past 20 billion RMB after raising over 5 billion RMB in the first half of 2026. The company has emerged as the highest-valued dedicated edge AI unicorn in China, driven by its density law thesis that the parameters required to achieve a given level of intelligence halve every 3.5 months.

Founded in 2022 from Tsinghua University NLP laboratory incubation, ModelBest chose a counter-consensus edge-first strategy when most AI startups were competing in cloud-based large models. The company density law, published in Nature sub-journal cover two years before comparable insights from Meta and other international institutions, posits that model intelligence density doubles approximately every 3.5 months, meaning tasks requiring cloud-scale computing today will soon be achievable on edge devices with progressively smaller models.

The MiniCPM series validates this thesis. MiniCPM5-1B achieves near GPT-4o-level performance at 1/200th the parameter count. The models span text, multimodal, full-modal, and voice variants from 0.5B to 9B parameters. Cumulative downloads on GitHub and Hugging Face have exceeded 38 million. Samsung Galaxy AI recently received China regulatory filing approval with its on-device AI capabilities powered by MiniCPM series models.

Commercial deployment spans automotive, smartphone, AIPC, embodied intelligence, and smart home sectors. ModelBest has partnered with Changan Mazda and Geely for production-line deployment of on-device models in vehicles. The company investor base includes national funds, state-owned enterprises, automotive manufacturers, and leading financial investors. The latest funding round closed on July 15, 2026, with the company raising over 5 billion yuan in H1 2026.

ModelBest success reflects a broader industry recognition that edge AI represents the next major growth frontier. Frost and Sullivan forecasts the global edge AI market growing from 321.9 billion RMB in 2025 to 1.22 trillion RMB by 2029 at approximately 40% CAGR. With AI agent deployment expanding from cloud-based chat interfaces into real-world devices spanning vehicles, phones, industrial equipment, and home appliances, ModelBest first-mover position in dense, efficient on-device intelligence has made it the defining pure-play edge AI bet in China venture market.

Moonshot AI has officially released Kimi K3, a 2.8 trillion parameter open-source foundation model that ranks as the world third strongest on comprehensive benchmarks, with developer tests describing it as delivering Fable-level user experience. The model, already available with API service and developer documentation, is the largest open-source model by parameter count globally and represents a significant milestone in China AI development toward closing the gap with leading US foundation models.

Kimi K3 introduces Moonshot self-developed Kimi Delta Attention (KDA) mechanism combining linear attention and Attention Residuals (AttnRes) structure to optimize long-sequence information processing and deep network information flow. The model supports 100 million token context windows with up to 6.3x decoding acceleration through KDA, while AttnRes improves training efficiency by approximately 25% at less than 2% additional cost.

The model focuses on long-horizon agentic coding and self-evolving workflow capabilities. With a higher-sparsity MoE design featuring 896 expert modules while activating only 16 per inference, K3 balances expanded model capacity against controlled computational cost. It can understand large codebases, call tools, execute tests, and continuously adjust task plans based on feedback, representing a shift from answering questions to completing tasks.

Benchmark results place Kimi K3 in the global top tier. In the GDPval-AA v2 knowledge work test, K3 scored 1687 points, surpassing Claude Opus 4.8 Max (1600) and trailing only Claude Fable 5 Max and GPT-5.6 Sol Max. In the AA-Briefcase agent knowledge work test, K3 scored 1527 points, ranking second only to Claude Fable 5 Max. The model also achieved 91.2 on BrowseComp for long-context information retrieval, leveraging its 100 million token context window without requiring additional context compression.

Developer community feedback highlights K3 autonomous planning and execution capabilities in complex agent tasks, continuous programming, and multi-tool orchestration. Several hedge fund evaluations rated K3 performance above Opus 4.8 at approximately 60% of its pricing. The open-weight release strategy follows DeepSeek precedent, allowing global developers to freely download, deploy, and modify the model, creating competitive pressure on closed-model strategy companies like OpenAI and Anthropic. At a time when Anthropic plans to raise Opus 4.8 prices by 50% in September, K3 and other Chinese open-source models present a cost-effective alternative that could reshape the global AI model market dynamics.

The 2026 World Artificial Intelligence Conference opened in Shanghai on July 17 across three major zones covering Zhangjiang, World Expo, and Xuhui West Bund, forming the largest WAIC in history. The theme Intelligent Partners, Co-Creating the Future reflects a fundamental shift from model capability demonstrations to industrial application deployment, with domestic chip full-lineup showcases, supernode computing clusters taking center stage, humanoid robots entering real factories, and consumer AI products becoming hands-on experiences.

Zhangjiang Science Hall presents the hardware ceiling of the conference with over 100 computing chip and infrastructure companies displaying 200 exhibits including 67 domestic first launches. A dozen general computing chip companies compete on the same stage. Highlight exhibits include Dongfang Suanxin DF1000, China first large-scale AI chip built on mature 14nm process claiming 4nm-equivalent performance through a third-way architecture innovation. Other chip companies including Muxi, Tianshu Zhixin, and Moore Threads showcase their latest products with live training demonstrations.

The World Expo H2 hall is dominated by supernode systems. Huawei debuts Atlas 950 SuperPoD with 1024 Ascend cards on display, while the Atlas 950 SuperCluster can scale to 500,000 cards. Sugon Dawn 8000 scaleX shows the 100K-card fully domestic AI super-fusion cluster with over 300 optimized applications. Alibaba Zhenwu M890 and ZTE OEX supernodes complete the lineup, enabling direct comparison of competing supernode technical approaches.

H3 Embodied Intelligence Hall features 161 companies and 314 exhibits. Xuanshi highlights include Unitree GD01 world first manned deformable mecha capable of switching between humanoid and quadruped forms. Zhiyuan demonstrates a production line replicated on the expo floor. Fourier debuts GR Nano desktop companion robot. Qianshun Intelligence Moz1 performs long-horizon single-instruction room organization tasks. Multiple companies demonstrate robot pharmacies, coffee bars, and other real-world applications.

Sustainable Xuhui West Bund presents 64 consumer AI products for hands-on experience. Honor Robot Phone, Bubbo 1 companion robot, multiple AR smart glasses from Rokid, Liangliang Vision, Xinmou Technology, and Ant Group are available for user testing. Music generation, video creation, and virtual human interaction stations round out the consumer experience zone. The conference structure itself signals that AI industry has moved beyond model competition into a phase where infrastructure, terminals, applications, and ecosystems compete as integrated systems.