China's two publicly listed AI large model companies — Zhipu AI and MiniMax — are experiencing a dramatic K-shaped valuation divergence that increasingly mirrors the global Anthropic vs OpenAI dynamic playing out in Silicon Valley. As of June 24, Zhipu AI commanded a market capitalization of approximately 975 billion Hong Kong dollars, while MiniMax stood at roughly 150 billion HKD — just one-sixth of its rival's value.

The reversal is striking given that just five months ago, when both companies listed on Hong Kong's stock exchange in January, MiniMax doubled on its debut day with a market cap nearly twice that of Zhipu AI, which suffered an opening-day dip. The 180-degree turn — Zhipu AI surging roughly 2,000% year-to-date while MiniMax has halved from its peak — mirrors the global narrative where Anthropic recently surpassed OpenAI in valuation after the latter's extended lead.

The strategic divergence is clear. MiniMax, under founder Yan Junjie, has pursued an OpenAI-style global expansion strategy — born-global products including Hailuo AI for video generation, Talkie for AI companionship, and MiniMax voice — serving 300 million users across 200 countries and territories with over 70% of revenue from overseas markets. Its value equation: AI platform value = intelligence density x token throughput, emphasizing efficiency and scale.

Zhipu AI, by contrast, follows an Anthropic-inspired path centered on safety, coding, and enterprise-grade capability. CEO Zhang Peng's formula — AGI commercial value = intelligence ceiling x token consumption scale — prioritizes pushing the boundaries of model capability through sustained R&D depth. Zhipu AI has capitalized heavily on the AI coding boom, positioning its GLM series as the foundation for developer tools and enterprise automation.

The divergence raises fundamental questions about China's AI competitive landscape: whether global reach or domestic depth will prove more valuable, and whether the Anthropic-vs-OpenAI blueprint can be successfully adapted to China's distinct market and regulatory environment. Neither company has won — but their divergent trajectories offer a live case study in AI business strategy.

A trending topic on Weibo AI face physiological aversion has ignited a national conversation in China generating 52 million reads and 20000 discussions as concerns mount over the quality and sustainability of the 22 billion yuan AI short drama and comic AI manga market. The backlash signals that what was once a novelty is now facing genuine user resistance.

According to DataEye-ADX industry data new AI drama and manga content on Douyin reached 39500 titles in May 2026 down 10.43 percent from April. While total viewership still grew to 140.6 billion plays the growth rate slowed dramatically to just 7 percent month-over-month compared to 176 percent in April a stark deceleration that industry observers attribute to mounting aesthetic fatigue.

The core complaint centers on the homogenization of AI-generated faces. Mass-produced template-based character models across thousands of titles produce nearly identical facial features making it difficult for viewers to distinguish characters or develop emotional attachment. Critics describe the experience as watching copy-paste faces that trigger an uncanny valley effect reducing viewer retention and binge-watching rates.

In response higher-quality productions are emerging with significantly larger budgets. Director Rao Xueman Left Ear adaptation from her own novel is entirely AI-produced with a three-month production cycle and a team of over 10 people targeting cinematic-level quality across lighting lens language performance depth and depth of field. Similarly AI short dramas produced by TouchAI using ByteDance Seedance 2.0 were selected for Cannes Fantastic Pavilion with production costs exceeding one million yuan more than most live-action short dramas.

The industry is at an inflection point the low-cost high-volume approach that drove initial growth is losing effectiveness as users demand higher visual quality character distinctiveness and narrative coherence. Celebrity AI face licensing where actors license their likeness for AI productions without physical acting has emerged as one potential differentiator but raises complex legal and reputational questions. The market future likely depends on shifting from AI as cost-saver to AI as quality-enabler.

Shanghai-based Nianxiang Technology has completed an angel round of nearly 10 million yuan for its non-invasive neural interface wristband Omniband, signaling growing investor interest in alternative human-computer interaction modalities beyond traditional touchscreens and keyboards. The round was led by Yongjun Xingmang, with participation from Pudong Chuangtou and Yicun Capital.

Omniband uses surface electromyography (sEMG) sensors to capture neuromuscular electrical signals from the wrist, decoding hand movement intentions and continuous dynamic gestures without requiring any surgical implantation. Unlike invasive brain-computer interfaces or traditional scalp EEG headsets, the wristband format allows everyday wearability while maintaining high signal-to-noise ratio by leveraging the natural amplification of muscle signals.

The key technical breakthrough enabling Omniband's viability stems from 2025 research published by Meta in Nature, demonstrating that scaling laws apply to sEMG — when training data covers over 100 users, model generalization improves continuously, enabling cross-user calibration-free recognition. Nianxiang CEO Wang Yi, who holds a PhD in BCI from the University of Auckland and serves as vice chair of the National Brain-Computer Interface Industry Alliance, believes this opens the commercialization window for neural wristbands.

Omniband can estimate all 20 hand joint angles in real time and communicate via standard HID Bluetooth protocol, enabling device control across smartphones, computers, smart glasses, and smart home systems. Beyond interaction, the device captures high-precision hand motion data that Nianxiang plans to use for building China's first large-scale sEMG public dataset for embodied AI and world model training — analogous to ImageNet for hand manipulation.

The company aims to address what it identifies as a critical gap: the lack of indigenous Chinese hand-motion datasets. By recruiting volunteers and opening developer platforms, Nianxiang plans to continuously expand its dataset to drive model improvement, positioning the wristband as both a consumer interaction device and a data infrastructure play for the physical AI era.

QQ Mail Tencent email platform has launched Agent Mail an AI-native email service that redefines the 40-year-old communication medium by allowing AI agents to directly send receive and manage emails. The service accessible at agent.qq.com represents a fundamental rethinking of email designed not for humans clicking buttons but for AI agents operating via CLI.

Agent Mail supports integration with major AI agent platforms including Qodex Claude Code OpenClaw TRAE Work MiniMax Agent and Kimi Work. After linking via OAuth authorization an industry-standard token-based system that never exposes user passwords agents gain the ability to read send and delete emails with daily limits of 50 messages attachments up to 20MB each and up to 50 attachments per message.

The OAuth-based security architecture is a key differentiator. Rather than storing plaintext passwords Agent Mail issues time-limited tokens that can be revoked at any time by the user. This temporary key approach is critical as AI agents increasingly operate various user accounts on behalf of humans. Early testers have noted that the service functions entirely through AI the Agent Mail dashboard currently has no human-operated buttons making it perhaps the first email service built from the ground up for AI consumption.

User experiences vary across agent platforms. Claude Code demonstrated notably smooth performance with properly formatted subject lines body text and signatures while Qodex showed slower response times and required repeated confirmation prompts. The system supports batch sending email categorization key information extraction and to-do list generation through natural language commands.

The launch signals Tencent ambition to position QQ Mail as the communication infrastructure layer for the emerging AI agent ecosystem. As agents become primary digital workers their ability to natively interact with communication tools rather than through human-mediated copy-paste workflows could fundamentally reshape workplace productivity and redefine what email means in the age of AI.

DeepSeek is making a strategic pivot from its reputation as a lightweight AI model developer toward a heavier infrastructure-intensive approach, signaled by aggressive hiring in its newly formed Harness division and plans to self-build computing facilities. On June 21, DeepSeek Harness team lead Cui Tianyi launched an online recruitment drive for Harness researchers, engineers, and product managers, underscoring the company commitment to agent infrastructure.

Harness the auxiliary systems that determine what tools an AI model can invoke what resources it can access and how information flows between sub-agents has emerged as one of the hottest areas in AI in 2026. DeepSeek Harness team formed around March 2025 operates on the core formula Model plus Harness equals Agent focused on transforming model capabilities into deployable agent products. The team has been steadily expanding since late 2025 when DeepSeek V3.2 emphasized enhanced agent capabilities and reasoning.

Beyond talent acquisition DeepSeek is reportedly exploring self-built computing facilities in Inner Mongolia a rare move among Chinese AI companies that overwhelmingly prefer leasing compute capacity. IDC China research vice president Zhou Zhengang noted that leading model companies increasingly favor self-built clusters for long-term cost control and full-stack customization advantages despite the substantial upfront investment. This strategy would allow DeepSeek to optimize power cooling and scheduling specifically for its proprietary model architecture.

The shift comes on the heels of DeepSeek 51 billion yuan Series A funding round which brought in strategic investors including Tencent JD.com NetEase and CATL. Industry observers believe the inclusion of CATL the world largest battery manufacturer may signal DeepSeek intention to pursue compute-electricity synergy potentially leveraging solar and energy storage solutions for its data center operations.

As China major AI model companies increasingly diverge in their strategic paths with Zhipu AI betting on AI coding and Alibaba pursuing full-stack AI DeepSeek bet on heavy infrastructure and agent engineering represents a distinct third path that could reshape the competitive landscape.

China's two publicly listed AI large model companies — Zhipu AI and MiniMax — are experiencing a dramatic K-shaped valuation divergence that increasingly mirrors the global Anthropic vs OpenAI dynamic playing out in Silicon Valley. As of June 24, Zhipu AI commanded a market capitalization of approximately 975 billion Hong Kong dollars, while MiniMax stood at roughly 150 billion HKD — just one-sixth of its rival's value.

The reversal is striking given that just five months ago, when both companies listed on Hong Kong's stock exchange in January, MiniMax doubled on its debut day with a market cap nearly twice that of Zhipu AI, which suffered an opening-day dip. The 180-degree turn — Zhipu AI surging roughly 2,000% year-to-date while MiniMax has halved from its peak — mirrors the global narrative where Anthropic recently surpassed OpenAI in valuation after the latter's extended lead.

The strategic divergence is clear. MiniMax, under founder Yan Junjie, has pursued an OpenAI-style global expansion strategy — born-global products including Hailuo AI for video generation, Talkie for AI companionship, and MiniMax voice — serving 300 million users across 200 countries and territories with over 70% of revenue from overseas markets. Its value equation: AI platform value = intelligence density x token throughput, emphasizing efficiency and scale.

Zhipu AI, by contrast, follows an Anthropic-inspired path centered on safety, coding, and enterprise-grade capability. CEO Zhang Peng's formula — AGI commercial value = intelligence ceiling x token consumption scale — prioritizes pushing the boundaries of model capability through sustained R&D depth. Zhipu AI has capitalized heavily on the AI coding boom, positioning its GLM series as the foundation for developer tools and enterprise automation.

The divergence raises fundamental questions about China's AI competitive landscape: whether global reach or domestic depth will prove more valuable, and whether the Anthropic-vs-OpenAI blueprint can be successfully adapted to China's distinct market and regulatory environment. Neither company has won — but their divergent trajectories offer a live case study in AI business strategy.

A trending topic on Weibo AI face physiological aversion has ignited a national conversation in China generating 52 million reads and 20000 discussions as concerns mount over the quality and sustainability of the 22 billion yuan AI short drama and comic AI manga market. The backlash signals that what was once a novelty is now facing genuine user resistance.

According to DataEye-ADX industry data new AI drama and manga content on Douyin reached 39500 titles in May 2026 down 10.43 percent from April. While total viewership still grew to 140.6 billion plays the growth rate slowed dramatically to just 7 percent month-over-month compared to 176 percent in April a stark deceleration that industry observers attribute to mounting aesthetic fatigue.

The core complaint centers on the homogenization of AI-generated faces. Mass-produced template-based character models across thousands of titles produce nearly identical facial features making it difficult for viewers to distinguish characters or develop emotional attachment. Critics describe the experience as watching copy-paste faces that trigger an uncanny valley effect reducing viewer retention and binge-watching rates.

In response higher-quality productions are emerging with significantly larger budgets. Director Rao Xueman Left Ear adaptation from her own novel is entirely AI-produced with a three-month production cycle and a team of over 10 people targeting cinematic-level quality across lighting lens language performance depth and depth of field. Similarly AI short dramas produced by TouchAI using ByteDance Seedance 2.0 were selected for Cannes Fantastic Pavilion with production costs exceeding one million yuan more than most live-action short dramas.

The industry is at an inflection point the low-cost high-volume approach that drove initial growth is losing effectiveness as users demand higher visual quality character distinctiveness and narrative coherence. Celebrity AI face licensing where actors license their likeness for AI productions without physical acting has emerged as one potential differentiator but raises complex legal and reputational questions. The market future likely depends on shifting from AI as cost-saver to AI as quality-enabler.

Shanghai-based Nianxiang Technology has completed an angel round of nearly 10 million yuan for its non-invasive neural interface wristband Omniband, signaling growing investor interest in alternative human-computer interaction modalities beyond traditional touchscreens and keyboards. The round was led by Yongjun Xingmang, with participation from Pudong Chuangtou and Yicun Capital.

Omniband uses surface electromyography (sEMG) sensors to capture neuromuscular electrical signals from the wrist, decoding hand movement intentions and continuous dynamic gestures without requiring any surgical implantation. Unlike invasive brain-computer interfaces or traditional scalp EEG headsets, the wristband format allows everyday wearability while maintaining high signal-to-noise ratio by leveraging the natural amplification of muscle signals.

The key technical breakthrough enabling Omniband's viability stems from 2025 research published by Meta in Nature, demonstrating that scaling laws apply to sEMG — when training data covers over 100 users, model generalization improves continuously, enabling cross-user calibration-free recognition. Nianxiang CEO Wang Yi, who holds a PhD in BCI from the University of Auckland and serves as vice chair of the National Brain-Computer Interface Industry Alliance, believes this opens the commercialization window for neural wristbands.

Omniband can estimate all 20 hand joint angles in real time and communicate via standard HID Bluetooth protocol, enabling device control across smartphones, computers, smart glasses, and smart home systems. Beyond interaction, the device captures high-precision hand motion data that Nianxiang plans to use for building China's first large-scale sEMG public dataset for embodied AI and world model training — analogous to ImageNet for hand manipulation.

The company aims to address what it identifies as a critical gap: the lack of indigenous Chinese hand-motion datasets. By recruiting volunteers and opening developer platforms, Nianxiang plans to continuously expand its dataset to drive model improvement, positioning the wristband as both a consumer interaction device and a data infrastructure play for the physical AI era.

QQ Mail Tencent email platform has launched Agent Mail an AI-native email service that redefines the 40-year-old communication medium by allowing AI agents to directly send receive and manage emails. The service accessible at agent.qq.com represents a fundamental rethinking of email designed not for humans clicking buttons but for AI agents operating via CLI.

Agent Mail supports integration with major AI agent platforms including Qodex Claude Code OpenClaw TRAE Work MiniMax Agent and Kimi Work. After linking via OAuth authorization an industry-standard token-based system that never exposes user passwords agents gain the ability to read send and delete emails with daily limits of 50 messages attachments up to 20MB each and up to 50 attachments per message.

The OAuth-based security architecture is a key differentiator. Rather than storing plaintext passwords Agent Mail issues time-limited tokens that can be revoked at any time by the user. This temporary key approach is critical as AI agents increasingly operate various user accounts on behalf of humans. Early testers have noted that the service functions entirely through AI the Agent Mail dashboard currently has no human-operated buttons making it perhaps the first email service built from the ground up for AI consumption.

User experiences vary across agent platforms. Claude Code demonstrated notably smooth performance with properly formatted subject lines body text and signatures while Qodex showed slower response times and required repeated confirmation prompts. The system supports batch sending email categorization key information extraction and to-do list generation through natural language commands.

The launch signals Tencent ambition to position QQ Mail as the communication infrastructure layer for the emerging AI agent ecosystem. As agents become primary digital workers their ability to natively interact with communication tools rather than through human-mediated copy-paste workflows could fundamentally reshape workplace productivity and redefine what email means in the age of AI.

DeepSeek is making a strategic pivot from its reputation as a lightweight AI model developer toward a heavier infrastructure-intensive approach, signaled by aggressive hiring in its newly formed Harness division and plans to self-build computing facilities. On June 21, DeepSeek Harness team lead Cui Tianyi launched an online recruitment drive for Harness researchers, engineers, and product managers, underscoring the company commitment to agent infrastructure.

Harness the auxiliary systems that determine what tools an AI model can invoke what resources it can access and how information flows between sub-agents has emerged as one of the hottest areas in AI in 2026. DeepSeek Harness team formed around March 2025 operates on the core formula Model plus Harness equals Agent focused on transforming model capabilities into deployable agent products. The team has been steadily expanding since late 2025 when DeepSeek V3.2 emphasized enhanced agent capabilities and reasoning.

Beyond talent acquisition DeepSeek is reportedly exploring self-built computing facilities in Inner Mongolia a rare move among Chinese AI companies that overwhelmingly prefer leasing compute capacity. IDC China research vice president Zhou Zhengang noted that leading model companies increasingly favor self-built clusters for long-term cost control and full-stack customization advantages despite the substantial upfront investment. This strategy would allow DeepSeek to optimize power cooling and scheduling specifically for its proprietary model architecture.

The shift comes on the heels of DeepSeek 51 billion yuan Series A funding round which brought in strategic investors including Tencent JD.com NetEase and CATL. Industry observers believe the inclusion of CATL the world largest battery manufacturer may signal DeepSeek intention to pursue compute-electricity synergy potentially leveraging solar and energy storage solutions for its data center operations.

As China major AI model companies increasingly diverge in their strategic paths with Zhipu AI betting on AI coding and Alibaba pursuing full-stack AI DeepSeek bet on heavy infrastructure and agent engineering represents a distinct third path that could reshape the competitive landscape.

China's domestically developed Lingsheng supercomputer has reclaimed the world's No. 1 ranking on the TOP500 list, achieving a sustained double-precision floating-point performance of 2.19 EFLOPS. The announcement was made at the ISC 2026 High Performance Conference in Hamburg, Germany on June 23, marking China's return to the top of global supercomputing after a nine-year interval.

The Lingsheng system, built entirely with indigenous Chinese processors and components, represents a significant milestone in China's push for technological self-sufficiency amid ongoing US export restrictions on advanced computing technology. Its performance leap demonstrates that domestically developed chips can compete at the highest levels of scientific computing, capable of handling the most demanding computational workloads in AI training, climate modeling, and fundamental research.

The ranking was accompanied by an interview with Turing Award winner Jack J. Dongarra, the University of Tennessee professor who created the LINPACK benchmark used for the TOP500 evaluation. Dongarra noted that China had stopped submitting benchmark data in recent years due to US technology export controls, causing its presence on the list to steadily decline over the past three years. "China has already built quite powerful computers," Dongarra stated, adding that while official submissions stopped, publicly available research papers detail the architectures and capabilities of Chinese supercomputers in significant depth, providing sufficient transparency to the global HPC community.

The TOP500 list, maintained for 34 years since 1993, ranks supercomputers by their performance on the LINPACK benchmark — a standard for measuring floating-point computing capability. Dongarra emphasized that the competition remains fierce globally, with nations viewing top-tier supercomputers as critical scientific infrastructure akin to major telescopes or particle accelerators, capable of attracting top scientific talent and enabling breakthrough research.

China's return to the top spot signals the effectiveness of its domestic chip development strategy and the continued vitality of its high-performance computing ecosystem, with profound implications for AI training, climate science, drug discovery, materials simulation, and other computationally intensive fields that drive national competitiveness.

On June 23, ByteDance officially launched Doubao-Seed-2.1 Pro (Doubao 2.1 Pro), its flagship large language model that has pushed daily token calls to a staggering 180 trillion — a 1,500-fold increase from its launch two years ago. The model now benchmarks competitively against GPT-5.5, Claude Opus 4.7, and Gemini 3.1 Pro across multiple core metrics, positioning it firmly in the global first tier of AI models.

Doubao 2.1 Pro achieves what the company calls a "production-grade inflection point" across three core capabilities: coding, agent functionality, and vision-language processing. Its long-context agent supports million-token input windows with 51% improvement in complex multi-step task completion rates over its predecessor, rivaling top-tier international models in benchmark evaluations.

Alongside the model launch, ByteDance's cloud platform Volcano Engine introduced HiAgent 3.0, an enterprise AI agent development platform that IDC ranks as the market leader in China with a 17.8% share. The platform now features new policy control and asset registry modules for enterprise-grade agent deployment. Additional product releases include video generation model Seedance 2.5, image creation model Seedream 5.0 Pro, and an audio generation model, completing a full-stack AI product matrix spanning general-purpose LLMs, vertical models, and enterprise cloud services.

However, Doubao's technical achievements stand in stark contrast to its monetization struggles. Despite surpassing 200 million DAU and 345 million MAU — making it China's largest AI application by user base — Doubao generates less than RMB 1 million in daily revenue, almost entirely from e-commerce commissions. With daily compute costs in the tens of millions of yuan, the service operates at a significant loss, reflecting a structural challenge facing China's consumer AI market where user willingness to pay remains low.

ByteDance is now pivoting its AI strategy from C-end traffic acquisition toward enterprise monetization, leveraging Volcano Engine's mature B-end across 30+ industries to convert its technical capabilities into sustainable revenue.

Chinese avionics startup Yunji Avionics has completed Pre-A and Pre-A+ funding rounds totaling tens of millions of yuan in rapid succession, bringing its fundraising tally to three rounds in under eight months. Investors include Haihe Industrial Fund, Yizhuang Xincheng Industrial, Houtian Capital, Chengdu, and Zhongli Fund, with proceeds earmarked for R&D, airworthiness certification, and production expansion.

Founded less than four years ago, Yunji Avionics is pursuing fully indigenous integrated avionics systems — the electronic "brains" of aircraft — in a market long dominated by Western giants Honeywell, Collins Aerospace, Thales, and GE, which collectively control over 80% of global market share. With China's push for aviation supply chain autonomy and the rise of domestic commercial aircraft and eVTOL platforms, the timing is strategic.

Yunji has developed three product series — YJ-500E, YJ-1000E, and YJ-3000E — covering light sports aircraft, normal category aircraft, and mid-size (19-seat and below) aviation platforms, compatible with fixed-wing, helicopter, and eVTOL configurations. The YJ-500E, positioned for mid-range integrated avionics, has completed onboard verification on models including the RX1E-A, AC311A, and WZ-11, and is expected to receive its first airworthiness certification in the second half of this year.

Key technical differentiators include 90% domestically sourced components, reducing overall costs to 50–70% of international competitors, and power consumption reduced to just 30% of comparable products — a critical advantage given the limited battery capacity of electric aircraft. The company selected an MCU-centric computing architecture over the industry-standard CPU+GPU approach, achieving a 10-fold power reduction through stacked chip packaging.

Founder Li Bei emphasizes a long-term strategy of at least a decade of sustained R&D investment, targeting domestic military and state-owned aviation programs first before expanding into the broader commercial market. Revenue is projected at approximately RMB 20 million this year as initial batch deliveries commence.

Beijing-based embodied intelligence company RoboScience officially unveiled its general-purpose embodied AI model Visics on June 24, complete with a full technical disclosure of the VLOA (Vision-Language-Object-Action) architecture. The company demonstrated real-world applications including furniture assembly, dexterous grasping, and dynamic assembly line operations using its proprietary technology.

The VLOA architecture introduces a novel unified intermediate representation standard called Object Trajectory (3D point cloud trajectory), establishing a layered and decoupled framework centered on object manipulation. Unlike mainstream approaches that train models on specific robotic joint trajectories tied to particular hardware configurations — learning "how a gripper picks up a cup" rather than understanding the concept of grasping itself — Visics separates the cognitive and execution layers, enabling true cross-platform generalization across different robot platforms, object types, and task scenarios.

Visics operates on a dual-engine architecture: an Embodied World Model trained on massive internet video data to learn object physics, motion patterns, and force dynamics, and a General Operation Model that translates object trajectories into hardware-agnostic control commands. The two engines communicate through the VLOA framework, with Object Trajectory serving as the unified interface between perception and action.

This decoupled design achieves generalization across three critical dimensions: adapting to any robot platform, manipulating any object type (rigid, articulated, or soft deformable), and autonomously completing diverse tasks. RoboScience has also developed a proprietary high-precision simulation engine, RoboMirage, combined with automated video annotation pipelines to generate training data at 1/20 to 1/200 the cost of traditional approaches, with expansion to over 1TB of manipulation trajectory data planned by end of 2026.

Backed by investors including JD.com, SenseTime, Fortune Capital, CMB Capital, and Sinovation Ventures, RoboScience operates R&D and production centers across Beijing, Shenzhen, Suzhou, and Hangzhou. The company is piloting deployments with retail, logistics, and elderly care enterprises, with plans to commence standardized robot mass production for industrial and commercial applications later this year.

ByteDance's Volcano Engine AI platform has achieved the capability to mass-produce models competitive with Anthropic's Opus 4.6 at a fraction of the cost, marking a new phase in China's AI infrastructure race.

At the FORCE Conference, Volcano Engine President Tan Te unveiled five major models simultaneously: Doubao 2.1 Pro, Seedance 2.5 for video generation, Seedream 5.0 Pro for image generation, and an audio generation model 1.0. The Doubao 2.1 Pro model, priced at just 68 yuan per month for consumer access through Doubao Pro, delivers capabilities that rival models costing significantly more overseas.

The cost advantage is structural, not promotional. Tan Te stated that the comprehensive cost of running Doubao 2.1 Pro is only 20% of Claude Opus 4.6, achieved through a combination of quantization, distillation, inference scheduling optimization, and proprietary chip development. At 68 yuan, the subscription likely carries a positive margin — this is sustainable pricing, not loss-leading.

Benchmark results place Doubao 2.1 Pro in the global first tier across multiple categories. On Terminal Bench 2.1, SWE-Pro, and SciCode, it ranks among the top models globally. On OSWorld, MobileWorld, and MMMU-Pro for agent and multimodal tasks, it also achieves leading positions. In practical testing, the model ran a complete RTL chip design workflow — spanning simulation, testing, and synthesis verification — through nine iterations over 18 hours of continuous operation.

The speed-to-market is equally noteworthy. Unlike competitors who announce models months before deployment, ByteDance pushed Seed 2.1 to its consumer products — Doubao App and TRAE IDE — on the same day. API access was opened simultaneously via Volcano Ark. The audio generation model 1.0 announcement came with immediate availability commitments for CapCut, Jimeng, and Fanqie Novel integration.

Seedance 2.5, scheduled for full launch in July 2026, supports 30-second single-segment native video output and joint generation from up to 50 multimodal assets. Seedream 5.0 Pro, also launched concurrently, supports layer-separable editing and precise multilingual text rendering in 14 languages.

The driving force behind this output is the scale of ByteDance's AI operations. Volcano Ark's daily token consumption has surpassed 180 trillion, growing more than tenfold in the past year. This creates a virtuous cycle: models launch to hundreds of millions of real users, their interactions generate feedback data that accelerates the next iteration, and the engineering team compresses the lab-to-product pipeline to days.

Industry observers note that the true competitive advantage extends beyond pricing to workflow integration. ByteDance's ecosystem spans Feishu for enterprise collaboration, CapCut for video editing, Fanqie for content, and TRAE for development. Users who embed their workflows within this ecosystem face switching costs. But competitors are narrowing the gap. JD.com has open-sourced JoyAI-VL-Interaction and Baidu has released Unlimited-OCR, demonstrating that domestic engineering capabilities are converging.

ByteDance's launch of a paid Doubao subscription with built-in productivity tools is reshaping the competitive landscape for China's office software market, placing Kingsoft's WPS under mounting pressure as AI-native alternatives gain traction.

Doubao Pro, priced from 68 to 500 yuan per month, includes an integrated office suite capable of generating, editing, and exporting documents, spreadsheets, and presentations. Users can produce a weekly report, a project proposal, or a presentation deck without ever opening WPS or Microsoft Office. The suite supports direct export to PPTX, PDF, and image formats without watermarks.

The timing is particularly challenging for Kingsoft, which days earlier faced a user backlash over its WPS software storing cache and cloud backup files on the system C drive regardless of user configuration, consuming tens of gigabytes of space. Kingsoft responded by forming a special investigation team, but the incident highlighted growing user frustration with the trade-offs imposed by traditional office software.

WPS has long relied on two primary advantages: extensive template libraries and deep format compatibility with Microsoft Office files. But the AI era is eroding both moats. When users can generate a presentation from a natural language prompt, template browsing becomes obsolete. When documents can be produced, edited, and exported within an AI assistant, the format-compatibility argument weakens.

The fundamental shift extends beyond features to format standards. The emerging lingua franca of AI-native productivity is Markdown (.md) — a lightweight, open plain-text format that large language models read and write with near-perfect fidelity. This threatens the proprietary document format lock that has sustained both Microsoft Office and WPS for decades. When content is generated by AI and consumed primarily by AI, the value of a carefully formatted document container diminishes.

Kingsoft is not standing still. Its WPS Lingxi agent introduced skill systems and scheduled tasks earlier this year, and its proprietary Qingqiu Agent achieved a 69.96% score on SpreadsheetBench, ranking second globally. The company invested 2.095 billion yuan in R&D in 2025, representing 35.34% of total revenue.

However, the resource gap is staggering. Kingsoft's 2025 annual revenue of 5.929 billion yuan pales against ByteDance's Volcano Engine, which targets 25 billion yuan in AI cloud revenue. Doubao's 180 trillion daily token consumption enables cost structures that WPS, dependent on membership fee revenue, cannot match.

For users already paying for WPS membership, the calculus is shifting: if both services require a subscription, the AI-native option that generates the output rather than just formatting it increasingly appears more valuable. WPS retains advantages in complex financial modeling, VBA macros, and government document compliance, but the erosion of its consumer base has begun.

ByteDance's Doubao AI assistant has officially launched a paid subscription tier, the Doubao Pro edition, with monthly pricing set at three levels: a standard plan at 68 yuan, an enhanced plan at 200 yuan, and a premium plan at 500 yuan. The move marks a significant step in monetizing China's most popular consumer AI application.

Pro subscribers gain access to ByteDance's latest Doubao 2.1 series models, designed for complex office and productivity scenarios, along with a newly introduced "office task mode" that includes AI-powered document, spreadsheet and presentation creation. The standard 68-yuan tier offers five times the quota of the free version, while the 200-yuan tier provides four times the standard quota, and the 500-yuan tier offers ten times.

The pricing structure reflects a clear segmentation strategy. Free users continue to have access to search, question-answering, writing, image generation, voice and video conversation features. ByteDance has emphasized that the paid tier is an additional subscription service layered on top of the existing free offering, not a replacement — a deliberate approach to minimize user churn while testing subscription revenue viability.

The monetization push comes as Doubao's user base has grown to dominant proportions. According to QuestMobile, Doubao reached 345 million monthly active users in Q1 2026, surpassing the combined MAU of Alibaba's Qwen (166 million) and DeepSeek (127 million). But this scale comes at enormous cost: Guolian Minsheng Securities estimates that Doubao's free AI service costs between 132 million and 240 million yuan per day to operate, even using the cheapest available compute.

ByteDance's daily token consumption across its model ecosystem has surpassed 180 trillion as of June 2026, a more than tenfold increase from the 120 billion recorded at launch in May 2024. The company reportedly allocated 160 billion yuan for AI capital expenditure in 2025, raised to 200 billion yuan for 2026.

Industry analysts note that Doubao's pricing remains significantly below international equivalents. ChatGPT Plus and Claude Pro both cost approximately 145 yuan per month at current exchange rates for base access, with full Opus-level capability requiring additional per-token charges that can run into hundreds of dollars monthly for heavy users.

Doubao's paid tier will not face near-term pressure to convert a specific percentage of free users. According to media reports, ByteDance has declined to make paid user penetration a core KPI for 2026, signaling a cautious approach to monetization that prioritizes not damaging its user base over rapid revenue generation.

China has reclaimed the top spot on the global TOP500 supercomputer rankings for the first time since 2017, with the domestically-developed LineShine (Lingsheng) supercomputer surpassing the US Department of Energy's El Capitan system at Lawrence Livermore National Laboratory.

The latest TOP500 list, released June 23 at the International Supercomputing Conference in Hamburg, Germany, places LineShine at number one with a sustained performance of 2.198 exaflops — making it the world's first supercomputer to exceed 200 exaflops of sustained performance. According to participating scientists, one second of LineShine's computation would take seven billion people ten years to replicate.

What distinguishes LineShine from virtually every other top-tier supercomputer is its pure CPU architecture. While most leading systems rely on CPU+GPU heterogeneous designs (GPUs power over 55% of TOP500 systems) to handle intensive numerical computation, LineShine uses standard microprocessors exclusively, with no GPU or specialized AI accelerators. Total system power consumption is rated at 42.2 MW, yielding an energy efficiency of 52.07 GFlops/watt.

In the High-Performance Conjugate Gradient (HPCG) benchmark, which measures real-world computational throughput, LineShine ranks first globally at 22.00 HPCG-petaflops. In the HPL-MxP mixed-precision benchmark tailored for AI workloads, it achieves 7.92 exaflops, placing fourth worldwide.

LineShine's ascension represents a milestone in China's push to build a fully autonomous computing ecosystem. Positioned as a domestically-developed full-stack infrastructure for scientific, engineering and intelligent computing, LineShine has already supported applications in atmospheric and ocean modeling, engineering simulation, materials science, drug discovery, brain science, scientific AI, and large model inference since its deployment.

China's previous number-one supercomputer, the Sunway TaihuLight, held the top spot from June 2016 to November 2017 before being surpassed by IBM's Summit. The gap of nearly a decade underscores the enormous technical challenges in developing indigenous high-performance computing systems free from foreign technology dependencies.

The current TOP500 top five includes: LineShine (China), El Capitan (US), Frontier (US), Aurora (US), and Jupiter (Germany). The EU has announced a 20 billion euro plan to build AI "super-factories" with large supercomputing bases to develop next-generation AI models, seeking to catch up with the US and China in computing infrastructure.

Tungsten hexafluoride (WF6), an obscure specialty gas essential to semiconductor manufacturing, has emerged as one of the hottest commodities in the chip supply chain, with prices surging 232.7% year-on-year to approximately 1.67 to 1.81 million yuan per ton as of early June 2026.

The price explosion stems from a historic supply-side contraction. In February 2025, China tightened export controls on tungsten-related dual-use items destined for Japan. Since tungsten powder accounts for 60-70% of WF6 production costs, Japanese producers were hit hard. Kanto Denka Kogyo (KDK), with approximately 1,400 tons of annual capacity, and Central Glass, with around 700 tons, both announced permanent production halts starting July 2026 after losing access to high-purity tungsten feedstock.

This removes roughly 2,100 to 2,200 tons of annual capacity — nearly 30% of global effective supply — from the market. Against a backdrop of surging demand driven by 3D NAND stacks exceeding 200 layers and widespread HBM adoption in AI servers, a static supply-demand gap of 800 to 1,000 tons per year has opened.

Tungsten hexafluoride is irreplaceable in semiconductor manufacturing as the precursor gas for chemical vapor deposition (CVD) of tungsten films. These films fill the millions of tungsten plugs that form vertical interconnects inside every advanced chip. Purity requirements are extreme: electronic-grade WF6 demands 99.999% minimum purity with impurity levels measured in parts per billion. Fewer than five companies globally can reliably supply electronic-grade WF6.

China's domestic champion CSSC Special Gas (688146.SH) operates the world's largest WF6 production base with 2,000 tons of annual capacity and another 1,000 tons under construction, expected online by Q1-Q2 2027. Its 6N-grade product already meets all advanced process requirements for 3D NAND, HBM, and logic chips. The company's 2025 revenue reached 2.26 billion yuan.

South Korean producers SK Specialty and Foosung prioritize captively supplying Samsung and SK Hynix. European capacity serves local fabs. This leaves Chinese producers as the dominant source of freely tradeable WF6 globally. Export prices have risen from $68.75/kg in January 2026 to $149.79/kg by April, with 7N-grade material for 3nm processes exceeding 3.3 million yuan per ton.

Industry forecasts indicate the Chinese WF6 market will remain in tight supply through 2026, with the July-September period expected to be the most acute phase of shortage.

Tencent has launched Agently Mail, a dedicated email service purpose-built for AI agents, now in beta testing. The product directly addresses two core pain points that have emerged as AI agents gain the ability to read, send, and manage email autonomously: privacy security and digital identity management.

Agently Mail is a standalone email system completely isolated from users' personal QQ Mail accounts. This fundamental isolation ensures that AI agents cannot access private correspondence or sensitive personal data while still enabling them to handle standardized business communications autonomously. Registration requires real-name authentication to prevent abuse and ensure accountability.

The product's key innovation lies in granting AI agents their own independent digital identities. Rather than borrowing a human's personal account credentials, each agent can register for online services, receive verification codes, and conduct email negotiations entirely on its own behalf. For business-to-business scenarios, Agently Mail supports automated workflows for inquiries, quotations, order placement, and invoice reconciliation — with all communication records fully traceable for regulatory compliance and audit purposes.

Agently Mail currently supports integration with major AI agent platforms including Doubao Super Mode by ByteDance, Kimi Work by Moonshot AI, Claude Code by Anthropic, and other leading domestic and international AI agent systems. The compatibility list is expected to expand rapidly as more AI agent platforms emerge.

Setting up the service is straightforward: users instruct their AI agent to install the corresponding CLI tool, authenticate via WeChat scan, and the dedicated mailbox is ready for operation within minutes. Once connected, natural language commands enable the agent to send emails, check inboxes, download attachments, and build sophisticated automated email workflows.

Common use cases already identified include automated invoice collection and reimbursement processing, daily news digest compilation, AI-driven job application submission with HR coordination, and automatic order email recognition and reconciliation. By delegating these repetitive email tasks to AI agents, Agently Mail aims to significantly reduce the administrative burden on knowledge workers while maintaining strict data security.