Alibaba pushed two AI updates in two days this week, deliberately splitting consumer and enterprise audiences. On August 7, the Qwen App shipped five new features — reasoning research, scheduled tasks, a workspace assistant, an agent plaza and voice calling — all running on the Qwen3.8-MAX flagship model and all offered free of charge. On August 6, Alibaba Cloud opened public testing for Wan 3.0, the third generation of the Tongyi video model, with API access through Alibaba Cloud Bailian, the Wanxiang portal and the Qwen PC creation client.

The Qwen App changes are the bigger consumer story. The new workspace assistant can decompose a goal, call tools, and deliver output across phone and PC, which makes the app behave more like a lightweight agent than a chatbot. Free access is a deliberate land grab: Alibaba wants scale and entry-point dominance first, with paid quotas reserved for professional workspace users later.

Wan 3.0's differentiator is input modality. For the first time in the family, the model accepts standard documents — doc, xls, ppt, pdf and md files — as direct input. A user can feed in a pitch deck and ask for a 30-second video in return. Single-shot generation runs to 30 seconds per clip. API pricing is set at 0.3 yuan per second at 480p, 0.6 yuan at 720p, and 1.2 yuan at 1080p, a deliberate attempt to land SME budgets rather than chase premium studio accounts.

The combined bet is structural. As model capabilities converge, the next competitive moat is how quickly raw capability gets packaged and shipped into daily use. Alibaba's strategy pairs a free-to-use consumer entry point with a metered enterprise bill, and lets the consumer funnel eventually subsidize the enterprise book. The Qwen App handles the funnel; Wan 3.0 and the Bailian API handle the bill.

Whether the strategy runs depends on three variables: how soon Wan 3.0 leaves public beta, whether the Qwen App's workspace assistant actually retains professional users past the first 13 free queries, and whether ByteDance's Doubao and Tencent's WorkBuddy hit the same document-input video capability before Wan 3.0 stabilizes its API. For now, Alibaba is the first Chinese hyperscaler to ship a document-input video model at SME prices. The window will not stay open for long.

A new kind of certificate started circulating across Chinese smartphone marketing teams in late July. The label, L3, comes from the Ministry of Industry and Information Technology's national grading standard for AI-enabled terminals, published in May. The first batch of test results, released on July 17, listed 11 qualifying devices, including nine smartphones and two tablets from Huawei, Motorola, Honor, vivo, OPPO, Xiaomi and Stepfun.

The certification is not a regulatory gate. A phone does not need an L3 mark to ship, and models outside the first batch are not failing any test. The standard is a yardstick that helps carriers, app developers and procurement teams reason about which devices can actually plan, call tools and remember user preferences, and which ones are still little more than voice-activated front-ends to a chatbot.

L3 is also not the top of the ladder. The MIIT document defines five dimensions (perception, cognition, execution, memory and learning) and a total of 14 sub-capabilities, with explicit levels from L1 to L4. The first batch only measured up to L3. Vendors have nevertheless leaned hard into the wording, calling L3 the current highest grade. The framing is technically correct but missing the second half: L4 is intentionally left blank because the rules for cross-device collaboration, error attribution and authority transfer between agents have not been written yet.

This is what separates AI-agent phones from autonomous driving, despite the shared L1 to L4 scale. Automotive grades are defined around responsibility transfer between car and driver. Phone grades are defined around task complexity, tool orchestration, and memory. A phone that handles a multi-day trip plan with payment, calendar and hotel bookings still has to hand authority back to the user for anything irreversible. The empty L4 slot is therefore not a sign that models are not ready. It is a sign that collaboration rules are not ready.

For consumers, the practical answer is straightforward. Most flagship phones already reach L2, which behaves like a chatbot with good tool-calling. L3 devices are on shelves now, including the Huawei Pura X Max and Lenovo's Motorola line. None of them require a new purchase for the L3 features to be useful. The next jump, when it arrives, will not be a hardware purchase decision. It will be a rules decision.

A new kind of worker has been quietly reshaping Chinese factory floors over the past year. The role does not require writing a single line of code. The job is to stand next to a humanoid robot, repeat the same motion hundreds of times, and slowly teach the machine how a human body actually behaves when it picks up a metal bar, turns, and drops it into a box. Average salaries range from 500,000 to 800,000 yuan per year, and the cohort of people willing to do this work is still tiny.

The role now has an official name. The Ministry of Human Resources and Social Security published in late July a proposed addition of 12 new occupations, including the embodied-AI robot application technician. It is the first time China has formally classified a humanoid-robot training job as a digital occupation, and it codifies the salary and skill expectations of an industry that has been hiring in the dark.

Wei Xingfeng, who moved into the role after years as a traditional industrial robot engineer, describes the rhythm of the work as closer to a driving instructor than to a programmer. A human demonstration is captured, trimmed to about eight seconds of useful motion, repeated thirty times, then fed back into the model. The expectation is patience and muscle memory rather than algorithms. The work also produces something that pure coding cannot: a continuous stream of human motion data that no synthetic pipeline can replicate at scale.

The hiring data confirms the gap. Recruitment index numbers for embodied-AI roles rose 15-fold year-on-year between January and April 2026, reaching 579 against a base of 36. Average monthly pay in the segment hit 62,000 yuan. New positions tied to humanoid robots grew 215.8%, and the average annual offer reached 406,100 yuan. Minth Group, the Ningbo-headquartered auto-components supplier, says its embodied-AI business grew 400x in a single year and still has 40 unfilled trainer seats.

Universities are catching up, but slowly. Nine top engineering schools, including Harbin Institute of Technology, Beihang University and Beijing Institute of Technology, opened embodied-AI majors in 2026. The industry's view is that production demand is doubling every quarter, while graduate throughput is linear. The shortage is therefore not a passing bottleneck but the defining labor-market story of the embodied-AI decade.

China added a new category of satellite to its orbital fleet in late July. Chenguang-1 lifted off from the Dongfeng commercial space innovation test zone carrying an eight-card compute server designed to run AI inference on orbit. The mission is the country's latest proof point for space-based compute, a category that turns satellites from dumb cameras into autonomous agents that filter, recognize and decide before downlinking.

Liu Yaoqi, chairman and CEO of Zhongke Tiansuan, framed the shift in plain terms. Traditional satellites capture raw imagery and stream it to ground stations for processing. Less than five percent of the data ever reaches the ground, and what does make it down can take days to be turned into something useful. Satellites that carry their own compute can extract high-value answers from the imagery they capture and send only those answers, dramatically shrinking the bandwidth bill.

The economic story extends beyond remote sensing. If a global orbital compute layer becomes a real piece of infrastructure, it will also offer a path to compute equity for regions that lack fiber and data centers. Zhongke Tiansuan is preparing a POPS-class, fully domestic on-board computer for delivery by the end of 2026, and has finished an eight-card space-based training server prototype based on domestic CPU and GPU architectures. The longer-term ambition is a 10,000-card space supercomputer the company calls the Tiansuan Plan.

China's stack reaches well beyond one company. The National Supercomputing Center in Tianjin showed the public the Tianhe Space SuperIntelligence Fusion Facility at the 2026 World Intelligence Industry Expo in May. AI satellite-internet firm GoSharpened Aerospace has been shipping its StarCompute Plan, a 2,800-satellite orbital compute constellation. The Institute of Computing Technology at CAS has built a domestically-sourced liquid-cooled high-performance compute payload and an anti-radiation RISC-V multi-core processor. Universities in Shenzhen and Shanghai have already added space-based compute courses, and Hangzhou Institute for Advanced Study at UCAS opened a dedicated space-based compute R&D center.

China is far from alone. SpaceX has flown AI-chip-equipped Starlink test satellites, Google has sketched plans for space data centers, and other US players are pursuing their own orbital compute modules. What makes China's position distinct is the breadth of the stack it has assembled in parallel, and the speed at which it is converting engineering experiments into named orbital infrastructure.

A team at the Institute of Automation under the Chinese Academy of Sciences (CASIA) has proposed a world model, PhiZero, that separates the task of understanding physical dynamics from the task of painting frames. The paper, titled "PhiZero: A World Model Built Around Physical Language," was published on arXiv and has begun circulating through the embodied-AI community this week.

World models that predict future video frame by frame tend to entangle two very different problems inside one prediction space. The appearance of a scene, its textures, lighting and background, shares parameter capacity with the dynamics that move objects through time. The CASIA group argues that this conflation is why pixel-domain world models often produce visually convincing frames that, when watched in sequence, drift away from physical plausibility.

PhiZero splits the prediction pipeline into two stages. The first stage takes the current world state and an action, and emits a short sequence of discrete tokens that the authors call a "physical language." The second stage uses those tokens as a compressed motion plan, and only then renders the next frames. A tokenizer module, built around a transition-level Q-Former and finite scalar quantization (FSQ), produces the discrete tokens. A diffusion decoder reuses the first frame's appearance as a high-frequency prior, so the language tokens can spend their capacity on motion rather than texture.

The compression is significant. For a four-second, 33-frame clip, PhiZero uses 256 physical-language tokens drawn from a roughly 25,000-entry vocabulary. The Wan2.2 VAE, by comparison, needs 44,800 continuous visual tokens for the same clip. That is about a 175-fold reduction in the number of tokens the world model has to consume and predict. Reconstruction quality remains competitive with other highly compressed video tokenizers.

The implication is structural. If world models can reason over short, discrete state trajectories rather than long, continuous pixel streams, robotics, autonomous driving and embodied-AI agents inherit a representation that is closer to a programming language than to a video codec. PhiZero's contribution is less the exact 175x figure than the bet that physical language, like natural language, can be the substrate on which world models compose.

Alibaba's Qwen App released a major update on August 7, bundling office assistant, scheduled task and flagship Qwen3.8-MAX model access into a single version of the consumer AI app. For the first time, the release carries an explicit pricing signal: core features stay free, but the new office assistant will run on a quota that requires payment once exhausted.

The pattern is the same one that ByteDance's Doubao set in June, when it launched a three-tier subscription ranging from 68 to 500 yuan per month. Tencent followed with its own personal tiers between 99 and 999 yuan, and Alibaba itself quietly shipped a 98-yuan-per-month standard subscription for its corporate Qwen Office product on August 3. Office workflows have emerged as the segment where Chinese AI vendors believe users will actually pay.

Qwen App sits inside a more awkward position than its peers. It was built as Alibaba's AI-to-consumer flagship, integrated with Taobao, Alipay, Fliggy and Amap, and fronted a 30-billion-yuan New Year marketing campaign that pushed its daily active users from 7 million to a peak of 73.5 million. Once the campaign ended, DAU fell by nearly half. Per-user session time also collapsed, from 6.3 minutes before the campaign to a low of about 3 minutes after, a 51% drop. Customer acquisition cost ran roughly 144 yuan per DAU, well above ByteDance's Doubao and Tencent's Yuanbao.

Internally, Qwen App also collides with the Qwen Office product that Alibaba launched publicly on August 3 under the new DingTalk CEO Chen Yusen. The office product consolidated three previously competing agent products, QoderWork, Wukong and MuleRun, into a single entry point. With DingTalk handling enterprise sales and Qwen Office absorbing the productivity roadmap, Qwen App is left to argue for a distinct consumer identity it has not yet built.

The paid-office bet is therefore less about monetization than about narrative. Whether Qwen App can clear enough quota usage to qualify as a real office product will determine whether it ends up as the consumer front-end to Alibaba's AI stack or simply as another chatbot window in a crowded market.

China's state broadcaster rolled out a sweep of late-summer 2026 technological milestones, each of which removes an industrial ceiling the country has been pressing against for years. The package spans heavy industry, astronomy, semiconductors, displays, energy and space.

The faster pace is also a defensive move. Several of the categories where China has now closed the gap, including radio-astronomy-grade receivers, ultra-thin precision strip steel, and large AMOLED evaporators, were the subject of export-control tightening by foreign suppliers over the past two years. Domestic supply in each category changes the negotiation dynamic in subsequent licensing rounds. Domestic supply also gives downstream Chinese system integrators a more predictable cost curve at a time when several of their end markets, including new energy vehicles, premium consumer electronics and space-ground link hardware, are simultaneously scaling up.

Foremost is a new high-end steel grade rolled in batches thinner than an A4 sheet of paper, with thickness and surface quality previously monopolized by a handful of foreign mills. The same announcement carries a hardware refresh for the Five-hundred-meter Aperture Spherical Telescope (FAST) in Guizhou, China's flagship single-dish radio telescope. Operators say the upgrade substantially boosts the antenna's ability to capture weak cosmic signals, a capability previously held by fewer than three radio observatories worldwide.

Two experimental satellites carrying AI payloads were launched from the Dongfeng commercial space innovation test zone. They are designed to do most of their data filtering, recognition and inference on orbit, sending back actionable answers rather than raw imagery. China ranks among the first countries to put a satellite-to-satellite compute cluster into orbit, and to demonstrate that orbital compute can be more than a single-satellite demo.

Other items in the broadcast include a domestic replacement for a niche semiconductor material that was previously imported, a new generation of large-area AMOLED display panels, a low-cost high-density sodium-ion battery cell, a high-temperature superconducting cable network, and progress on a fourth-generation nuclear reactor design.

Read together, the announcements suggest an industrial policy that keeps stacking small wins across multiple sectors at once, instead of placing single bets on a few flagship industries. The pattern matters as much as the individual products.

China's export story has been told in two acts. First the "old three": garments, furniture, home appliances, built on cheap labor. Then the "new three": electric vehicles, lithium batteries, photovoltaic products, built on supply-chain density. The third act is being written now, and the headline is not another commodity. It is a triad the state press has started calling the "new new three": industrial and service robots, AI large models, and innovative drugs.

The latest export data and use-case trail tell the same story. Chinese-made wall-climbing cleaning robots are replacing human spider workers on high-rise facades in Australia. Chinese-built large language models are running grid-maintenance workflows for Brazilian utilities. Shanghai-discovered fruquintinib, marketed globally as FRUZAQLA, has been folded into the world's authoritative treatment guidelines. None of these exports pass through a container port. The units of trade are no longer tons; they are model weights, API calls and licensed molecules.

The export pattern has three distinctive features. The first is value density per kilogram. OpenRouter's most recent weekly token-volume number, 36.11 trillion tokens, of which the top four spots all went to Chinese models, suggests Chinese AI infrastructure is now metered in global traffic rather than shipping manifests.

The second is intellectual property gravity. The new new three is built on patents, model weights, drug pipelines, and proprietary data loops. Hengzhi Future's four-legged "Xiaotian" robot dog, deployed in industrial inspection, is one example. Runsheng Pharmaceutical's small-molecule pipeline, anchored by Chongqing-based manufacturing, is another. The export unit is now know-how, not hardware.

The third is standard-setting. China is no longer waiting for international standards to be set elsewhere and then aligning. It is sitting on AI working groups, drug-pharmacopoeia committees, and robotic safety standards bodies. The shift from passive follower to active definer is the most important narrative change.

The challenges are real. The United States and Europe still own the upstream: foundation-model architectures, industrial robot core components, and the front-end drug targets no domestic pipeline can yet fully replace. Friend-shoring and re-shoring policies are creating friction. National drug-regulatory systems still do not accept each other's approvals.

The signals from OpenRouter's traffic rankings and from foreign-deployment case studies in Australia, Brazil and Africa are that the climb is happening. Whether it accelerates depends less on China's industrial policy and more on whether the rest of the world accepts Chinese innovations as global defaults. The old three asked the world to buy cheap goods. The new three asked it to buy green goods. The new new three is asking it to buy the underlying intelligence, and that is a much harder sell.

For the first half of the AI infrastructure era, the assumption was simple: bigger chips win. The 2026 World Artificial Intelligence Conference in Shanghai marked the moment that assumption officially died. The dominant motif is no longer single-board teraflops but supernodes, rack-scale systems that link hundreds to thousands of accelerators into one shared-memory fabric. The race is now about system-level efficiency, not chip-level peak.

Biren Tech anchored the optical-interconnect camp. Its 1,024-card NPO optical supernode uses the company's in-house BLink 2.0 protocol and near-package optics to push "copper out, light in" through the scale-up domain. Up to 1,024 GPUs share one unified memory space. The product matrix runs from a 16-card electrical baseline to a 128-card high-density rack to the 1,024-card optical supernode.

Yixin Intelligent went the other direction. Its WAIC announcement was the first RISC-V AI supernode in the industry, anchored by the company's in-house Epoch cloud AI chip, the first mass-produced RISC-V high-end AI accelerator in China. Epoch already supports block-quantized FP8 precision and will move to EXFP4 and MXFP4 next generation. The supernode uses an orthogonal backplane-free interconnect that cuts interconnect hardware cost by 80 percent and tightens latency into the hundred-nanosecond band.

The rest of the show floor was a tour of competing philosophies. Infinigence CoreX pushed dense general-purpose training silicon. Kunlun Chip, Baidu's spun-off accelerator unit, presented its third-generation architecture. SingularMOL demoed a chiplet-based interconnect. Infinigence, the cloud-side orchestrator from Infinigence Cloud, showed the software story: a dispatch layer that can weave fragmented domestic hardware into a single virtual supercluster.

The economic point underneath is uncomfortable. GPUs in a poorly wired cluster spend most of their time waiting for data. Bandwidth, not compute, is now the binding constraint for training and inference. Communication walls mean stuffing a thousand cards into a rack without a system-level interconnect strategy yields a "token conversion rate" that can be a tenth of theoretical. Chinese supernode vendors are competing on usable tokens per dollar of capital expenditure.

Hardware choice is no longer a single-vendor decision. A reasonable Chinese AI lab in 2026 will buy at least two different accelerators and wrap them in a software layer that schedules tiles across vendors. The labs that build that orchestration layer will own the next bottleneck. WAIC 2026 will be remembered as the year Chinese AI infrastructure vendors stopped selling chips and started selling systems.

For most of the post-ChatGPT era, Chinese AI was framed as the open-source counterweight. The framing was polite. The reality was that Chinese models were a tier below the frontier. The latest Hugging Face spring report changes that. China-origin open-weight models now account for 41 percent of all downloads on the world's largest model hub, surpassing US-origin models for the first time. Cumulative downloads have crossed 10 billion.

On OpenRouter, the four highest-token-volume models in the most recent week are all Chinese. Kimi K3, released July 27 by Moonshot AI with 2.8 trillion total parameters, is the largest open-weight model ever shipped. Three days later, Alibaba released Qwen3.8-Max, a 2.4-trillion-parameter base model with an open-weight plan in motion.

The political economy is straightforward. US export controls make the largest Nvidia boards unavailable to Chinese labs, forcing algorithmic efficiency that would have been nice-to-have elsewhere. Investors reward open distribution because open downloads are the cheapest moat against closed competitors. Andreessen Horowitz partner Martin Casado said last year that roughly 80 percent of Silicon Valley AI startups pitching for funding were building on Chinese open models. Hugging Face CEO Clement Delangue said China is likely to catch US frontier labs by end of 2026.

The technical picture is shifting. Ion Stoica, the UC Berkeley professor who co-authored Spark and Ray, told Chinese press the gap between Chinese open models and global frontier closed models has narrowed from six to nine months down to two to three months. Nvidia founder Jensen Huang publicly endorsed the open route.

Capital is following. DeepSeek closed its first external round at more than 50 billion yuan. Moonshot AI closed an F round that put its post-money valuation at 35 billion dollars and has started a Series G pre-IPO round. The hot money is no longer chasing who can ship the largest model fastest. It is chasing which open model becomes the substrate that everyone else builds on, the way Linux became the substrate of cloud.

Government industrial policy is a tailwind, not the cause. The 15th Five-Year Plan pushed open-source system building as a national priority. The national AI open-source community now hosts more than 11 million users and over 70,000 models. The infrastructure exists because the work was already happening; the policy accelerates distribution, not invention.

The next phase is the global phase. Chinese open models have become the default starting point for developers in Southeast Asia, Africa and Latin America. The United States still leads in raw frontier capability, but the question for the next two years is whether the rest of the world builds its AI on top of Chinese open weights the way it built cloud on top of Linux.

At an internal all-hands on August 6, ByteDance CEO Liang Rubo, CFO Zhou Shouzi, and the heads of Doubao, e-commerce, monetization, and the creative services platform stood on the same stage and told employees that AI strategy had changed. The consumer side will keep Doubao competitive and protect Seedance's video-model lead. The real weight will shift to enterprise productivity, where ByteDance has historically been the weakest of China's internet giants.

The week before, ByteDance folded Lark's product team into Doubao and pushed Lark's sales, marketing and customer-success functions into Volcano Engine. The Lark brand that spent five years trying to beat Tencent and DingTalk in workplace collaboration is now a thin layer inside a much larger productivity stack. Zhao Qi, who runs Doubao, becomes the most powerful AI product executive in the company. Former Lark head Xie Xin now reports up to Zhao rather than directly to Liang, a sharp demotion that sources inside the company read as deliberate.

The trigger was a market shift that happened almost without warning. In June, Tencent's desktop-native AI agent platform drew 20.97 million monthly visits in China, ranking first in the country and ahead of ByteDance's TRAE IDE plus Alibaba's QoderWork combined. Enterprise buyers have stopped paying for collaboration tools. They are paying for AI agents that can independently decompose tasks and execute them. Lark's product foundation, collaboration plus chat plus docs, was structurally wrong for this shift.

Doubao now carries the weight. As of March 2026 it had 345 million monthly active users and a peak daily active count of 150 million. Doubao models process 180 trillion tokens per day, up 1,500-fold since the model's 2024 launch. In June the company launched Doubao Pro at 68, 200 and 500 yuan per month. Behind it, a Lark-integrated Doubao Enterprise edition entered closed beta with knowledge-base retrieval, data isolation, permission controls and audit logging.

The architecture is now a clean three-layer enterprise stack. Volcano Engine at the bottom, providing model-as-a-service and cloud. Doubao in the middle, owning the model and the user experience. Lark on top, owning office workflows and customer penetration. Liang framed this as "high priority, thick trunk, long horizon": pull non-core bets back into the trunk business, let the trunk feed adjacent businesses.

The bet underneath the move is that AI-native office will be won by whoever controls the model and the agent surface, not whoever owns the legacy workflow chart. ByteDance is putting the model layer at the center and accepting a period where Lark looks diminished. If Doubao becomes the trunk business that drags search, e-commerce, content distribution and collaboration forward the way Douyin once did for monetization, the strategy will look obvious in hindsight.

When China's State Council elevated "low-altitude economy" into the official work report in 2024, provincial governments rushed to set trillion-yuan targets. Three years later the bottleneck is no longer airframes but who pays. DJI is winning the early commercial race with a grey box most tourists will never see: the unmanned drone airport.

On July 29, the Wuhan Surveying and Mapping Research Institute, working with DJI Enterprise, showed off a network of 146 uncrewed docks across the megacity. Outside no-fly zones, a drone can now reach any point inside Wuhan within five minutes. State Grid Taizhou has installed 207 docks inside a 5,800-square-kilometer service area and now scans 6,434 kilometers of transmission line every three days. DJI's Sun Mingming told 21st Century Business Herald that DJI docks have flown more than 8 million sorties and 2 million hours across 500-plus cities and 80-plus countries since 2022.

The routes that survived share a single trait: the mission is "cannot-not-fly." Traffic control, fire prevention, grid inspection, and surveying all require airborne data every week, and the cost of stopping the flight shows up immediately in operations.

Wuhan's "Smart City Eye" system has been adopted by 16 municipal departments. Wuchang traffic police alone handle nearly 1,000 incidents a year with the help of drones, and about 30 percent of elevated highway crashes now end with a remote megaphone. Guangzhou Nansha took a similar path under a single government-led buyer and is flying 36 docks with 100 percent online uptime.

DJI is clear about what it will not do. Its enterprise business director Song Tianqi told the newspaper the platform layer must be owned by embedded system integrators in each industry, and DJI's job is to keep the hardware substrate open. More than 1,000 third-party software platforms have already integrated with the dock's cloud APIs.

The next phase is whether infrastructure-as-a-service can replace one-off subsidies. The Wuhan and Nansha models show a viable answer when one department anchors procurement and the rest piggyback. DJI's white paper calls the next frontier IoD, Internet of Drones, where hundreds of docks coordinate flight plans like a router schedules packets. The mid-game of China's low-altitude economy is being played in uncrewed box-shaped rooms, and DJI has spent the last three years quietly filling them.

July 2026 turned up the heat on China's low-altitude economy. On July 1 the newly revised Civil Aviation Law took effect, formally writing low-altitude economy into national law for the first time. At the end of the month, the 2026 International Low-Altitude Economy Expo closed in Shanghai with 452 exhibitors, 570 aircraft and 65 debut products and technologies. From three years of 'new growth engine' to 'emerging industry' to 'emerging pillar industry' in successive government work reports, low-altitude has now moved from demonstration flights into routine commercial operation, said Wang Fan, deputy director of CCID Research Institute's Science, Technology and Standards research office.

CAAC data puts China's 2025 low-altitude economy market at 1.5 trillion yuan, with nearly 40,000 drone operators and core technologies moving from 'following' to 'selective leadership'. The Yangtze River Delta, Pearl River Delta and Chengdu-Chongqing regions have built end-to-end industrial chains. Shenzhen surpassed one million cargo drone flights last year with 310 cargo routes and 8 cross-city logistics corridors. Shanghai's early-2026 cluster plan targets 'World eVTOL Capital'. Local practice is turning blueprints into daily routine.

Safety has become the gating concern. CAAC's May establishment of a dedicated Low-Altitude Safety Department closes the regulatory loop from airspace planning to airworthiness certification to market supervision. The 'Three Firsts, Three Lasts' principle now governs the rollout: cargo before passengers, segregated airspace before integrated airspace, suburban operations before urban. United Aircraft Group, whose TD550 industrial unmanned helicopter is China's first CAAC Type Certified industrial rotorcraft, applies flight-control redundancy, navigation redundancy, flight-envelope protection, emergency power, no-fly-zone avoidance, and autonomous fault detection and isolation.

AutoFlight, the eVTOL company behind the V2000CG cargo aircraft and the five-ton V5000 Sky-Dragon, is moving ahead on the same playbook. Its large cargo eVTOL completed a 12.8 km resupply flight over Wenzhou Cangnan during the Dragon Boat Festival, carrying nearly 200 kg of goods and cutting transit time tenfold against traditional shipping.

Despite the policy and product momentum, the industry is still in its 'scale-up infancy'. Five core pain points named by zerog aircraft industry in Hefei cut across the sector: scarce and fragmented airspace, incomplete ground infrastructure, immature rules and standards, weak public trust that requires many routine flights to build, and a serious talent shortage. The 'Three Firsts, Three Lasts' principle keeps regulatory risk bounded, but the harder challenge is industrial: turning a working aircraft into a profitable route, a working route into a network, and a working network into a service layer that the rest of the economy actually pays for.

Video generation is the loudest AI story of 2026. ByteDance Seedance 2.0, MiniMax H3 and Seedance 2.5 have put Chinese teams on the global leaderboard, and use cases keep multiplying across film, short drama and e-commerce. Underneath the loudness, the math is getting worse. Inference cost is still high, multi-stream video understanding is exploding, and general-purpose GPUs keep wasting compute, power and latency on workloads that look nothing like what they were designed for.

SmarCo HT Tech, one of the earliest Chinese teams to industrialize the dataflow architecture, joined the gap with its new Jingang GC3. The chip carries 12 RISC-V cores, 200 TOPS of INT8 compute, and 128GB of unified LPDDR5 ECC memory. CTO Wu Dongdong framed it as 'born for video': same silicon that can decode and encode multiple streams can also run video-generation models, with a memory subsystem designed for the scale of intermediate tensors today's video transformers demand.

Wu broke down why general GPUs lose on video. For video understanding, raw streams hit a server, get decoded to pixel frames, and then run through GPU inference. The bottleneck is rarely raw FLOPS, he said. It is the latency of moving each frame between memory and the compute units across many parallel streams. For video generation, the GPU runs a diffusion or transformer model that produces frames, which an encoder then compresses. The model is compute-hungry, but every intermediate result still has to be read and written, and those tensors can run into the hundreds of gigabytes. Read them too often and the silicon waits.

The architectural mismatch sits underneath both stories. CPUs and GPUs are von Neumann machines with a program counter that walks through instructions. That control-flow design is wonderfully general, which is exactly what makes it mediocre at video. Pixel-level and matrix-level parallelism outrun what a central controller can schedule. Memory traffic saturates. Multiple cores spend more time hitting barriers than doing useful work.

Dataflow architecture flips the model. The program becomes a graph of operations; data dependencies are the edges; an instruction fires the moment its inputs are ready, with no program counter deciding its turn. Wu's metaphor is a factory line: each station starts when its parts arrive and immediately passes the result downstream, with no central dispatcher and almost no idle time. Jingang GC3 fuses dataflow as the compute engine with RISC-V as the control core and unified memory across CPU, VPU, GPU and NPU paths. The argument the industry is testing in production is whether dataflow can replace control flow as the default substrate for AI compute. Jingang GC3 is SmarCo HT Tech's answer to that question, with video as the proving ground.

Physical AI is the largest investment consensus in tech this year. From autonomous driving to world models to AI for science, capital keeps circling back to the same question: how does AI act on the physical world? Autonomous driving and embodied AI have been sprinting. AI for science and AI for drug discovery, despite a flashy IPO or two, look stuck. The most recent cold water came from BCG: 60 billion dollars invested in AI pharma globally, zero approved drugs.

Insilico Medicine's late-2025 Hong Kong IPO marked the peak of optimism. Almost simultaneously, Google DeepMind disbanded the team behind AlphaFold after eight stable years, redirecting AI4S into a Gemini-driven general research system. Several biotechs have slipped back toward CRO work, serving incumbent pharma with AI rather than inventing new drugs. The gap between digital hypothesis generation and physical hypothesis verification is widening.

The bottleneck is no longer compute or algorithms. It is data. AlphaFold worked because half a century of experiments had produced hundreds of thousands of high-quality protein structures. Beyond structure, the data wall returns at every layer: structure to efficacy, efficacy to toxicity, toxicity to clinical response. As Tsinghua's Global Health Drug Discovery Center director Ding Sheng put it, model architectures and compute are mature enough, but our grasp of underlying life mechanisms is still incomplete.

The deeper problem is that such data cannot be accumulated by human experimenters. A researcher can run tens of experiments a day. AI needs tens of millions. The only path forward is to redesign the experiment itself: tools that produce data, accumulate data and close the loop on their own. MegaRobo, founded in 2016, has spent the last decade building toward exactly this.

MegaRobo started by automating life-science lab workflows and gradually built infrastructure connecting physical experiments to AI. Founder and CEO Huang Yuqing frames automation as 'freeing humans from repetitive low-value work and impossible high-precision work', then adding learning and reasoning. The company's full-stack PCE architecture links Perception, Conception and Execution. A pharma client ran three full design-verify-iterate cycles in six months, bringing per-cycle cost to roughly one-twentieth to one-thirtieth of a traditional cycle.

The infrastructure was built in three stages. Stage 1.0 automated manual workflows and gave MegaRobo deep understanding of life-science processes. Stage 2.0 stress-tested the loop in semiconductor manufacturing, where precision and yield demands are unforgiving. Its Manavis wafer-defect inspection system now runs below 0.05 percent miss rates and 0.1 percent over-detection, with classification accuracy above 98 percent. Stage 3.0 rebuilds every product AI-native from the silicon up, designing tools for AI consumption rather than for human use. The company that controls the perception-to-execution loop in the lab becomes the layer every AI-for-science model eventually pays to call.

DeepSeek announced on August 6 that it plans a significant across-the-board price increase for its API services, with the new pricing to be detailed in a formal notice. The news lands just weeks after V4 Flash climbed to the top of OpenRouter's weekly global token-usage ranking at 7.22 trillion tokens. V4 Pro and V4 Flash remain the two main offerings, with V4 Flash priced at 0.02 yuan per million tokens for cached inputs, 1 yuan for non-cached, and 2 yuan for output. The price hike is framed as a response to surging compute costs.

The pattern of price moves this year explains why. In April DeepSeek trimmed the V4 Pro cached input price and ran a limited 75 percent-off promotion through May 5. On May 23 it announced that, after the promo ended on May 31, the price would move to a quarter of the original list rate. The same week Xiaomi's MiMo-V2.5 API announced a permanent price cut of up to 99 percent, the moment that turned the domestic model price war into a rout.

The pivot came at the end of June. A DeepSeek upgrade email told users that V4's official version was scheduled for mid-July, and that the new release would bring 'peak and off-peak' pricing. During weekday peak hours, API prices would double. For V4 Pro, cached input moves from 0.025 to 0.05 yuan per million tokens, non-cached input from 3 to 6 yuan, and output from 6 to 12 yuan. V4 Flash doubles the same way.

The new round is a flat raise. According to industry analyses, the peak-and-off-peak structure introduced earlier was less a price hike than a standardization tool for rationing scarce compute, with predictable times and predictable multipliers. The next round lifts the floor. DeepSeek's compute bill has grown in lockstep with demand. On August 1 alone, the OpenCode platform recorded 8 trillion tokens of V4 Flash traffic. On August 4 the V4 Flash API hit a capacity outage from unprecedented inbound volume.

DeepSeek is not alone. At Zhipu AI's 2025 results briefing, CEO Zhang Peng disclosed that Q1 2026 API pricing rose 83 percent year over year, with API call volume growing 400 percent. In June Doubao Pro introduced a three-tier subscription at 68, 200 and 500 yuan per month. In July Kimi paused new consumer sign-ups after K3 launch traffic exceeded its cluster capacity within 48 hours. The whole industry is rotating from a 'lowest price wins the user' playbook toward a 'sustainable cost structure' playbook. Training and serving frontier models on Chinese compute is no longer a loss leader that the rest of the stack subsidizes forever.

Zhang Yiming, founder of ByteDance, walked into a Seed all-hands meeting a month ago and ended a two-year internal debate. ByteDance will not use other people's models to accelerate its own LLMs, he said. Not closed-source frontier models, and not open-weight ones either.

The rule dates to 2023. Seed's early team had run experiments using GPT API outputs as training data. After ByteDance introduced GPT API call audits in April 2023, the team wrote a hard rule: no GPT-generated data goes into ByteDance training sets. The guardrail predated anyone outside China arguing about distillation.

The first internal argument came in January 2025 after DeepSeek-R1 dropped, and Seed researchers asked the obvious question: if other labs are quietly using frontier outputs to catch up, why can't we? Some proposed a careful dose of distilled outputs to patch the gap without abandoning Seed's own pretraining. Leadership rejected it.

A year later, with Nvidia Blackwell GPUs deployed at scale inside US labs, the debate returned. Chinese labs cannot buy top-tier B-cards. Seedance 2.0 was trained on a cluster of H20s, whose training performance sits around one-fiftieth of a B200. GPT-5.5 and Claude 4.8, especially Claude Fable 5, made new jumps in reasoning, coding and science that coincided with Blackwell. Closed models can't be distilled, but the best open-weight models can.

What tipped the debate into an all-hands was Kimi K3. Moonshot's open-weight K3 reached ranges that put it in the same competitive neighborhood as frontier closed models. For Seed, a smaller Chinese lab punching at global tier-one weight class made the question unavoidable: why has Seed, with bigger talent density and more compute, not produced a language model of equivalent standing? Distillation would let Seed funnel scarce training resources into directions that have already proven effective elsewhere.

Zhang's answer was short. Seed can accept temporary lagging, he said. It cannot accept closing that gap by distilling competitors. The position is consistent with the 2023 policy, but the stakes are higher. Anthropic has publicly accused Tongyi Qianwen, Moonshot and MiniMax of harvesting Claude outputs at scale. Whether those allegations hold or not, the geopolitical, legal and commercial risk has risen sharply, and Seed is keeping the line even at the cost of years of lagging in benchmarks.

The interesting bet is the meta-bet underneath. Seed's strategy assumes that the frontier labs pulling ahead on today's reasoning benchmarks will not pull ahead on the intelligence Seed is actually trying to build, and that borrowing their data now would lock Seed into their architecture of intelligence rather than its own. It is a bet Zhang is willing to underwrite with real model-position losses, one other Chinese labs may find harder to make.

Embodied AI is the loudest robotics race of 2026. At the WAIC show floor last month, humanoid robots punching and dancing gave way to demos of robots running real production lines, picking, sorting and assembling on 1:1 factory replicas. Investors followed: in the first half of 2026, more than 300 deals in China's embodied-AI and adjacent supply chain moved over 90 billion yuan. Headcount in embodied-AI roles grew fifteen-fold in the first four months alone.

Behind the noise sits one shared worry: data. Compute once throttled model scale; algorithms once throttled capability. Now data is throttling the ceiling. Robots cannot inherit the internet's accumulated text, images and video. They need a different corpus, the friction between materials, the forces from different weights, the causal loop between a gripper and an object, all the subtle physics that no website already produces. Industry consensus puts high-quality manipulation data globally at a few hundred thousand hours; supporting a general-purpose robot is likely to need tens of millions.

Teleoperation captures high-precision data but is expensive. First-person video scales cheaply but lacks joint, force and trajectory detail. Pure simulation breaks under sim-to-real transfer. WUWENAI's founder and CEO Liu Shengxiang, a former Baidu autonomous-driving data-and-test lead, treats these not as alternatives but as a single pipeline. Multimodal collection rigs in real environments keep harvesting physical-interaction data; a generative world model expands scenarios and edge cases in simulation; both feed back into training and into the next round of data production. WUWENAI calls it a 'virtual-real fused closed-loop data system,' and it is built around three components: a Data Factory, a World Model, and a World Simulator that runs training, evaluation and feedback at scale.

Liu sees this as more than a data-supply platform. Once a high-fidelity world simulator can reliably reproduce real physics, a robot can do the AlphaGo move, run millions of low-cost trial-and-error cycles in simulation, and shift from imitation learning toward reinforcement learning. WUWENAI already began shipping this Real2Sim2Real stack months before World Labs' recent acquisition of SceniX. Liu frames the parallel as an industry signal: world models become infrastructure only when connected to real data, robot simulation, scaled training and real-world feedback. Video and 3D generation alone are not enough.

The company's bet is structural rather than product-level. Embodied-AI models are constrained less by parameters than by the data they can consume, and the data industry is shifting from raw capture to orchestrated infrastructure. WUWENAI's next test is whether a Chinese startup can scale into a world-class physical-AI infrastructure company while giants like World Labs, Tesla with Optimus, and Google DeepMind run their own stacks in parallel. The road is being paved either way. WUWENAI just chose to be the one laying the asphalt.

When Yu Kaicheng first started talking about world models in 2023, the response he kept getting was that the field was too far away, even fantasy. Three years later, with Fei-Fei Li, Yann LeCun and a stream of new entrants raising billions, world models have become the physical-AI narrative of the moment. Yu, by then already a world champion on ECCV's autonomous-driving track and number one on the Fail2Drive benchmark, did the least intuitive thing a researcher can do: he rejected the very technique that had won those titles.

Yu's path runs through Alibaba's DAMO Academy, where he joined as an Alibaba Star after his PhD on AutoML. His BEVFusion work there, which fuses LiDAR and camera into a unified physical space, became one of the most cited papers in multi-sensor perception and a NeurIPS 2022 highlight. After DAMO he moved to Westlake University in 2023 to set up the AutoLab, building generative world models for autonomous driving. Deployment work with a top-tier Chinese automaker cracked the assumption open. If a model can already generate realistic data, Yu reasoned, it already partially understands the world. Continuing to render pixels is just an expensive detour.

In late 2024 the team pivoted to an implicit route. The model no longer renders pixels; instead it builds representations of physical concepts in an abstract mathematical latent space, decomposes scenes into a finite set of composable concepts, and recombines them to handle situations that never appeared complete in training data. Yu places this in the same implicit camp as LeCun's JEPA, but with a different answer to the question 'what lives in the latent space': composable physical concepts, learned end-to-end. Internal experiments in 2025 showed notable gains in success rate on complex tasks over imitation-learning and reinforcement-learning baselines, while keeping prediction cost lower.

In January 2026 Yu spun the research out of the lab into Awomo, based in Hangzhou. The founding team was just Yu and chief scientist Tong; five months later a group Yu describes as 'consistently self-validated geniuses' had joined. Awomo has now closed a seed and angel-series funding package worth more than 100 million yuan, backed by InnoFund, Dongfang Jiafu Fund, Zhengxuan Investment, Tianqi Capital, Westlake Innovation Fund and Jinma Investment.

Awomo's bet is that the scaling-law ceiling Yu once helped build does not extend to physical intelligence. A VLA-style model fine-tuned on more demonstrations will not suddenly grasp a new factory layout. A model that has learned to decompose the scene into concepts can instead recompose them on the fly: catching a falling cup while simultaneously closing a door, instead of mechanically finishing one subtask before the next. The first release, Awomo-v0.1, has already been evaluated on the RoboTwin 2.0 testbed.

A rumored FCC move on high-speed optical modules has put the spotlight back on the supply chain behind every AI data center. LightCounting's global optical-module supplier ranking now places seven Chinese companies inside the top ten, with Innolight and Eoptolink at numbers one and two. In the 400G, 800G and 1.6T tiers that hyperscalers actually need for AI training and inference, Chinese vendors are no longer challengers. They are the majority of supply.

That position has become structurally hard to displace. Omdia senior analyst Lyu Mingyang argues that even as Coherent, Lumentum, Nokia, AOI and Corning expand US-based capacity over time, the China and Asia-Pacific optical-communications manufacturing and delivery base remains effectively irreplaceable in the near term. The division of labor is now mature: Chinese vendors own module-level scale manufacturing, packaging, test and rapid delivery, while high-speed DSP and analog chips largely come from Marvell and Broadcom, and premium EML lasers plus parts of the silicon-photonics stack come from Coherent and Lumentum.

The bottleneck is moving down the stack. Module makers including Accelink, Source Photonics, HGTECH and CIG round out the Chinese top tier. Innolight, Eoptolink and TFC run significant capacity in Thailand; DSBJ runs 69.7 percent of its capacity in mainland China, with the rest in Taiwan, and is steering US-bound output through Thai and Taiwan plants toward a 35-million-unit 800G and 1.6T target by 2027.

Indium phosphide is the next stress point. The substrate is critical for 400G, 800G and 1.6T modules. Industry estimates put 2026 demand at 2.6 to 3 million wafers against effective capacity of roughly 750,000, and prices for 2-inch optical-grade InP substrates have risen sharply this year. EML lasers, silicon photonics, DSP chips and InP substrates are now the four chokepoints everyone is trying to pre-book.

The race is widening beyond modules. Ciena just printed 40 percent revenue growth in its fiscal Q2 with 7.7 billion dollars in backlog stretching into 2027; Nokia's optical-networks segment grew 20 percent year over year in Q2. In China, Huawei, ZTE and FiberHome dominate domestic backbone and compute-center interconnect, with AI-OTN upgrading from single-wavelength 400G to 800G and 1.6T. China Mobile and China Telecom have already kicked off fresh optical-transport equipment tenders for 400G and 800G OTN.

As AI moves from model competition to infrastructure competition, optical communications is no longer one product line. It is a system that spans pluggable modules, switches, optical chips, coherent optics, CPO, DCI and long-haul transport. Whoever controls the upstream materials, the core devices and a multi-year supply plan will define who wins the next decade of AI infrastructure.