Tech Buzz China

We recently published a report on Chinese AI apps on our Substack channel. You can now download a free PDF version with improved layout on our homepage. Enjoy! techbuzzchina.com

US startups mayb be trying to bolt data centers onto offshore wind platforms, but they're doing it as an add-on. Sinopec is doing the facility design itself. On April 7, its petroleum engineering arm signed with Shanghai Hailan Cloud Technology, the world's first commercial submarine data center operator, to launch Sinopec's first undersea data center project. The structural difference means Sinopec captures the integration margin rather than giving it over to a contractor. Shanghai Hailan Cloud's existing track record makes this a strong infrastructure bet. Phase one completed in Hainan in November 2023. By March 2025, a seabed AI cluster was linked and running large model training and industrial simulation. All three phases were operational by October 2025: each chamber at 360kW, total bandwidth 800G, carrying China Telecom's Tianyi Cloud nodes in Hainan. Integrated offshore energy and compute are an exciting direction, and Sinopec is in that space already.

AI's impact on Kuaishou's core platform is hitting a wall. Q4 monthly active users grew just 0.7% year-over-year, lowest since IPO. Ad spend jumped 5% from AI, but user quality hasn't moved: only 15.5% of Kuaishou's base qualifies as high-value, lowest among major platforms. Kuaishou ad growth (12.5%) lags estimated Douyin growth (20%+) due to user composition (15.5% high-value users vs. industry peers). In other words, premium advertisers are sticking with competitors. Kuaishou's CFO disclosed that 2026 capital expenditure is projected to be 26 billion yuan ($3.74 billion), a 11 billion yuan ($1.58 billion) increase from 2025, primarily spent on its video model's computing power. So there's a significant 'scissors gap' in which costs are expanding far faster than revenue. Kling had an annualized revenue run rate exceeding $300 million by January 2026, with management projecting at least a doubling of revenue in 2026. But its success should be seen in the context of intense competition. ByteDance's Seedance 2.0 had around 45 million MAU in early 2026 (vs. Kling's 12 million), and Alibaba's HappyHorse, led by former Kling lead Zhang Di, has now entered the market.

CATL's swap network is scaling fast, but the more interesting line in this Wuling deal is the fast-charge co-development: 10% to 80% in under 10 minutes. That spec, if it ships across 15+ Wuling models, makes swap infrastructure optional for mass-market buyers rather than necessary. CATL is building 3,000+ swap stations by end-2026 and signing up Wuling, GAC, and SAIC Commercial Vehicles, but the value of that network depends entirely on fast-charge staying inconvenient. The swap build-out and the fast-charge roadmap are running in parallel at the same company. One of them will define how CATL monetizes batteries at volume.

Four of the global top 10 models come from Chinese institutions: Alibaba, DeepSeek, Tsinghua, and ByteDance. The two countries have traded top Arena leaderboard positions multiple times since early 2025. China ranks first globally in AI papers published, citations, patents filed, and public AI supercomputing clusters (85 installations, more than double North America). That's the infrastructure behind a 2.7% benchmark gap, the headline figure in Stanford’s 2026 AI Index. The report finds the performance gap between US and China AI models has gone from 300+ points in May 2023 to 39 points today. Claude Opus 4.6 sits at 1503. Dola Seed 2.0, a multimodal engine for agent-driven operations sits at 1464. On price, Seed 2.0 Pro's output cost runs roughly one-tenth of Claude Opus 4.6. The report documents a "jagged frontier" in which AI can win an International Math Olympiad gold medal but fails to read an analog clock 50% of the time. Claude Opus can handle PhD-level science questions but struggles with basic spatial reasoning. Labs show 89% success on robot tasks, but drop to 12% in actual homes.

Smart cars are essentially a non-transformable robots, says NIO’s Li Bin at a intelligent EV forum held this week. He also revealed that the company’s annual procurement costs for NVIDIA chips was $300 million, and as vehicle delivery growth was at 40-50%, this kind of spending would continue to drive costs.
 The next car, NIO ES9, will be equipped with the self-developed chip 'Shenji NX9031', accelerating the mass production of NIO's in-house chips.

The market share of domestic intelligent driving chips is currently less than 50%, but adoption rates in new vehicles are rising rapidly, with the likes of AiThinker and Black Sesame Technologies focusing on optimization and full-stack solutions. Black Sesame Intelligent is aiming to achieve full-scenario coverage of automotive-grade chips this year. Source media tmtpost argues that number of carmakers that can actually sustain chip development is small, and industry concentration will rise quickly from here on in.

China just released its first gene chip for wheat that directly reads cadmium accumulation instead of waiting months for field tests. This collapses the breeding cycle from years to seedling stage. Northwest A&F University and the Chinese Academy of Sciences spent five years integrating over 60,000 genetic markers across the wheat genome. The chip detects cadmium, arsenic, mercury, and a dozen other heavy metals simultaneously, then outputs yield, disease resistance, and quality metrics in the same read. The constraint is tech adoption. China's wheat breeding infrastructure was built around phenotypic screening, which means established seed companies have institutional incentives to use the slow methods. So this chip only accelerates if Chinese food safety policy tightens cadmium limits in wheat procurement. Until then, the market isn't ready.

Nio, XPeng, and Li Auto spent years pitching themselves as China's answer to BMW and Mercedes. By 2025, they'd all surrendered their pricing power to survive. The constraint that broke their premium positioning wasn't technology or ambition, it was unit economics at scale. Li Auto stumbled in 2025, selling 406k units versus a 750k target and seeing revenue drop 22%, while cheaper rivals Nio and Xpeng surged. Nio posted Q4 profits of 120M RMB, Xpeng 380M RMB. Li barely scraped 7M RMB after a 630M Q3 loss. ASP trends highlight the gap: Nio down 35%, Xpeng 22%, Li only 9%, keeping it anchored at the high end. Margins slipped below 20% for two quarters, while peers climbed. All three chase AI for growth, but Li Auto’s embodied AI strategy to make the car itself smarter is struggling. In contrast, Nio builds AI capabilities centrally to optimize products and operations, while Xpeng leverages AI to reinvent its brand and ecosystem. Li Auto is pitching the next-gen L9 as "a car that truly understands you." Nio formed an AGI Committee in 2026 focused on full-stack autonomous driving and company-wide AI deployment. XPeng went further, rebranding as "XPeng Group" in March 2026, explicitly positioning itself as a robotics and embodied AI company, not a carmaker (CEO He Xiaopeng: "cars aren't a good business"). The real constraint they're trying to break: once you've lost pricing power, the only narrative left is future capability. Whether that holds depends on whether AI actually delivers experiential differentiation customers will pay for. If it doesn't, they're just EV manufacturers fighting on cost and range, the worst competitive position in China right now.

The US just spent $35 billion trying to show you can legislate your way out of a cost disadvantage. Ascend is proof it doesn’t work. Ascend raised $1.1 billion from Temasek, BlackRock, Qatar's sovereign fund, and Goldman Sachs. It had a patented process that recycled lithium-ion waste into cathode precursor material, bypassing 15 steps. This was a closed loop. It involved no mining and no Chinese imports. Lithium was worth $80k per ton in late 2022. The company’s factories in Georgia and Kentucky processed old batteries and developed new EV cathodes. Then lithium lost 90% of its value. China's new cathode material was suddenly cheaper than American recycled material. In April 2026, it filed for bankruptcy with the factories 60% built. Other battery makers share its fate. Li-Cycle failed the same way. Powin, Freyr, andKore Power, all abandoned US battery projects in 2025. Samsung SDI cut 1,000 jobs in Georgia. Panasonic delayed full production at its $4 billion US battery plant. The Inflation Reduction Act required domestic manufacturing, just as the US administration cut energy spending and dialed up tariffs on Chinese batteries to 132%. Redwood Materials is still standing. It survives by not doing recycling. In June 2025, it started selling batteries whole to AI data centers. After three years of desinicization, China holds 96% of global cathode material capacity.

MiniMax listed in Hong Kong three months ago with a 109% first-day pop, and its chief agent architect is now saying fewer than five large model companies will survive the next two years. That's a pointed claim from a company that spent only $449 million on total model training and cloud costs between 2022 and September 2025. MiniMax listed in HK three months ago with a 109% first-day pop. Its chief agent architect is now saying fewer than five large model companies will survive the next two years. That's a pointed claim from a company that only spent $449 million on total model training and cloud costs between 2022 and September 2025. The bet is that as AI agents replace chatbots as the primary interface, raw model performance converges and the delta shifts to how efficiently you extract value from each token. MiniMax's read is that harness design, sandbox iteration speed, and inference productivity become the actual competition. China's daily token call volume hit 140 trillion in March 2026, up more than 1,000x from early 2024. MiniMax framed this four years ago as tokens becoming a commodity energy unit, like electricity, and built accordingly. Today, 73% of revenue comes from overseas, serving 236 million individual users and 200,000+ enterprise clients and developers across more than 100 countries. The strategic upgrade announced post-IPO, from large model company to "AI platform company," is the logical endpoint of that positioning. Defining the infrastructure layer for token delivery globally is a different business than competing on benchmark scores.

Video generation is having a rough quarter, globally speakig. OpenAI shut down Sora in late March 2026. ByteDance paused Seedance 2.0 over copyright concerns. In China, Alibaba released HappyHorse-1.0 in April. Zhang Di who only returned to Alibaba in November 2025, and as the new head of Taotian Group's Future Life Lab, reports directly to Alibaba's CTO Zheng Bo. He spent a decade at Alibaba optimizing ad systems and recommendation feeds before leaving to lead Kling's underlying architecture at Kuaishou.The group he now heads is under Alibaba's core transaction business umbrella. HappyHorse works with text, video, and audio tokens in a single sequence. It can lip-sync across English, Mandarin, Cantonese, Japanese, Korean, German, and French without drifting like Sora. But the main thing is Taotian is for use specifically in e-commerce, sitting inside the transaction stack rather than as a standalone research lab. This sets it aside from most video generation models. ByteDance and Kuaishou won't stay still, so the thing to watch is how deeply HappyHorse worms its way into merchant workflows, becoming a 'can't do without you' marketing assitant.

China's cloud giants broke a two-year rule on March 18: @AlibabaCloud, @Baidu, and @Tencent Cloud all raised AI compute prices on the same day. Google and AWS moved first in January 2026. The low-price subsidy era for GPU cloud is over. The pricing shift is structural, not cyclical. High-end GPU costs surged through 2025, but cloud providers absorbed that gap to win market share. Meanwhile, token consumption from agentic workflows runs roughly 100x traditional chat models. Video generation, digital humans, and real-time calling went mainstream. Demand simply outgrew the cost buffers cloud providers were willing to carry. The hikes are not uniform. Alibaba Cloud still cuts prices on commodity compute like ECS and OSS. The segmentation is intentional: SMEs can migrate, so price wars continue there. AI-heavy teams in autonomous driving, robotics, and model development cannot migrate easily. Once you're running multi-chip inference clusters with mixed GPU architectures, rebuilding that stack elsewhere means service queues, degraded performance, and supply chain uncertainty. Research institutes told the source directly: they used to avoid mixing chip vendors in their own clusters, but now have to run hybrid deployments to manage overseas supply risk, and cloud is simply easier. DeepSeek all-in-one hardware ("DeepSeek一体机") have become popular with government and enterprise users as a local deployment option to sidestep cloud API costs entirely. The logic is straightforward: own the hardware, avoid the token bill. As for SMEs and prosumers, one animation creator saw per-project costs rise from roughly 200 yuan (~$29) to 300 yuan (~$43). Manageable, but representative. Whether this holds depends on how fast providers close the efficiency gap. @AlibabaCloud's Bailian and Baidu AI Cloud's Baige platforms are already doing chip-model co-design to get more tokens per watt from domestic silicon. Liquid cooling and HPC-AI fusion from players like Sugon are reducing per-unit energy costs. If those gains compound faster than demand, prices moderate. If not, token pricing looks increasingly like a utility bill, one that only goes up.

Alibaba has restructured its AI organization twice in roughly 30 days. That is not normal. It suggests the company is still searching for the operating model it thinks can win. Wu Yongming's April 8 memo announced a new Group Technology Committee and reshuffled responsibility across three leaders: Zhou Jingren on models, Li Feifei on cloud and AI infrastructure, Wu Zeming on business technology and AI inference. The effect is to collapse what had been three semi-independent tracks (cloud, models, business applications) into a single command structure. Decisions that once required cross-department bargaining can now be made in one place. Tongyi's elevation from lab to business unit may matter more than any personnel change. Labs can experiment and postpone commercialization. Business units own budgets, deadlines, and revenue targets. Alibaba is saying that strong models alone are no longer enough; the models now have to produce revenue. The sequence makes the logic clear. In March, Alibaba created ATH to build market mechanisms around tokens. In April, it centralized technical judgment via the Technology Committee while pushing Tongyi closer to the market. These are not isolated moves. They are a shift from parallel exploration to concentrated execution. That concentration is also a response to fragmentation. Alibaba has been advancing Qwen, Wan, and multiple multimodal lines in parallel. Breadth reduces risk but spreads resources thin. Centralization is Alibaba's bet that focus matters more than optionality at this stage. The risk is obvious: concentrating decisions also concentrates error, and a new org chart does not automatically overcome internal inertia. Wu Yongming's memo seems to acknowledge this. The bet is not that concentration is risk-free. It is that remaining fragmented is worse.

SenseTime, founded in 2014 at Hong Kong's Chinese University and listed on the Hong Kong Stock Exchange in late 2021, built its reputation by solving a problem that doesn't scale: bespoke computer vision delivery to government and enterprise clients. According to IDC, it held the top CV market share in China for nine consecutive years from 2015 to 2023. That depth became a liability when the industry stopped rewarding depth. ChatGPT's arrival in late 2022 shifted the competitive axis from algorithm precision to commercialization speed. SenseTime's project-delivery model, where every client required a custom model and dedicated engineering resources, worked against the new logic of standardized APIs and marginal-cost economics. ByteDance's Doubao, Alibaba's Qwen, and DeepSeek (a Beijing-based LLM startup) collectively captured 71.8% of enterprise model API call volume in the second half of 2025. SenseTime's share is a rounding error by comparison. The financials reflect this. Generative AI revenue grew 51% to ~$5.2 billion yuan (~$520 million) in 2025, now 72.4% of total revenue. But AIDC compute operating costs jumped 163.5% in the same period, pulling gross margin from 42.9% to 41.0%. The reported profit rests almost entirely on asset sales, with over 90% of the ~$274 million in net other gains coming from divesting subsidiaries and financial assets, not from the core business generating cash. Cumulative losses over seven years exceed 50 billion yuan (~$7.1 billion). Peak market cap hit HK$300 billion; by 2025 it had fallen to HK$81.3 billion, roughly an 80% collapse. The 1+X strategy, announced by CEO Xu Li in late 2024, restructures around a dual-engine core: generative AI plus computer vision, focused on building an integrated AI cloud platform across compute infrastructure, foundation models, and applications. The "X" units, autonomous driving, home robotics, smart healthcare, smart retail, spin off as independent subsidiaries with their own CEOs, external fundraising, and no continued group subsidies. CV, the original business, posted its first net profit in 2025 after years of losses. SenseTime has the technical credentials. Its Riri v6.5, a 600-billion-parameter MoE multimodal model, ranks top-three domestically on OpenCompass and SuperCLUE benchmarks across multimodal tasks. The gap is distribution: without a consumer hit or a cloud ecosystem that drives recurring API revenue at scale, benchmark rankings don't convert to market share. The 1+X bet is that shedding long-cycle, capital-heavy verticals will let the core unit compete on speed rather than customization depth. Whether that's enough depends on whether seven years of project-delivery habits can actually be unlearned.

Shandong province, east of Beijing, has its first brain-computer interface (BCI) clinic, to bridge lab research and clinical care. The technology reads motor intentions and drives external devices, shifting rehab from passive to active. Inpatients in the ward think the moves that the machine makes happen. This is all taking place at Rehabilitation University, established in June 2024 as a public undergraduate institution led by the Shandong Provincial Government, and the only research university in the country specifically for rehabilitation. China's BCI technology could be clinically approved within five years. A single clinic opening in Qingdao doesn't confirm the timeline, but a 20-person waitlist prior to day one suggests demand well beyond supply. Whether the broader infrastructure (trained clinicians, rehab protocols outside research contexts, financing for patients) catches up on demand is still to be seen.

Remember when people talked about China’s six AI Dragons? Now it is really four. The obvious question is whether that number shrinks again. China’s independent model landscape had already changed decisively by early 2026. The six startups that emerged from the 2023 model wars had effectively narrowed to four. Of those, two, Zhipu AI and MiniMax, went public in Hong Kong in early 2026. The remaining players are in a very different position: some have found a durable lane, while others have retreated into narrower verticals or simply lost momentum. DeepSeek and Kimi stand out from the pack. DeepSeek has managed to approach frontier-lab performance while keeping research and operating costs unusually lean. It has also built real overseas traction and, importantly, avoided going head-to-head with ByteDance and Alibaba in the domestic mass-market fight. Kimi, by contrast, has focused on long-context text processing, which is a narrower business but a defensible one. That specialization has helped it build strong retention among researchers, lawyers, and other professional users with high switching costs. Both companies are reportedly generating healthy cash flow and growing revenue, which means neither faces immediate pressure to go public. The weaker outcomes are best illustrated by Baichuan AI and 01.AI. Baichuan, once a general-purpose model startup, pivoted into healthcare. 01.AI moved away from foundation model development toward lightweight, industry-specific customization, but has since seen talent and funding erode. These cases show what “going vertical” often looks like in practice. The problem is that these verticals are not protected. Alibaba and Tencent can move into almost any healthcare or enterprise niche with lower pricing and better distribution, and they already are. Without a real technology edge, what remains is mostly custom implementation work: expensive to deliver, fragmented across clients, low margin, and slow to pay.

Claude Code's 1,906 leaked files and 512,000 lines of code sent Zhipu shares shooting up 32% (market cap ~$52 billion) and MiniMax 14% (~$42 billion) in a single session. The market thought China had just got the blueprint for Anthropic-style revenue growth. The reality is that Anthropic's engineering methodology, not included in the leak, is where Anthropic has its most value. Anthropic builds production-grade agentic systems: streaming tool parallelism and four-tier context management for long-running tasks. The leak revealed a number of exciting features as yet unshipped to consumers: KAIROS (a background agent that distills memory via autoDream), Coordinator Mode for multi-agent orchestration, and PROACTIVE (agent action without user prompting). Anthropic is driven like it is reaching for a Nobel prize in most fields this year or next. The economics of harness engineering clarify why this matters. Running a 2D game builder without the full harness setup costs $9 for 20 minutes. With it, $200 for 6 hours. Only the second output is actually usable. That gap is what Anthropic's ARR reaching $19 billion (up from $9 billion) is built on. OpenAI's first-party API now processes 15 billion tokens per minute, 2.5x the October level, with Codex at over 2 million weekly active users, 5x growth in three months. Zhipu's GLM-5 and MiniMax's M2.5 already match Claude Opus on coding at 1/20th the price. Zhipu's API ARR hit ~$250 million (up 60x year-over-year), even after an 83% price increase in Q1 2026. MiniMax's API ARR reached ~$150 million as of February, with token consumption up 6x since year-end. Both are shifting toward architecture competition. Zhipu introduced its "Token Architecture Capability" (TAC) metric, which factors in how efficiently model output converts to economic value. MiniMax is pivoting from research-focused to platform-focused, targeting the highest-value tasks where current models still fall short. What Yufeng He, an agentic AI researcher at Moonshot AI (the Beijing-based startup behind the Kimi model family, whose open-source Kimi-K2.5 base underlies the $500 billion-valued Cursor's Composer 2 model), found in the leaked code is the more instructive detail: Anthropic embedded ablation testing infrastructure directly into production. Every new feature ships with a control experiment that quantifies its value. That culture of verifiable engineering is what the ARR trajectory actually reflects. Leaked code transfers the architecture. The experimental discipline that decided which architecture was worth building does not come with it.

This isn't China AI per se, but it affects every AI company Do you agree with @WeijinResearch's conclusion? "Pricing, of course, will change too. But it is likely to move gradually toward structures that better fit enterprise usage."