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

China's Jiaolong submersible completed 63 dives across the Atacama and Mohorovičić trenches over 156 days, returning to Guangzhou today with samples that reshape how we understand deep-sea life. The expedition involved 83 researchers from six countries, but the structural win here isn't the international optics, it's that China now has proof chemosynthetic ecosystems exist in the Southern Hemisphere. Previously, the deepest known chemosynthetic life clusters were in the Northwest Pacific (Kuril-Kamchatka and Aleutian trenches). Those discoveries, published in Nature, showed tube worms and bivalves thriving on chemical energy from hydrothermal vents in complete darkness. This new finding in the Atacama trench validates the "global chemosynthetic life corridor" hypothesis: Earth's deepest ocean trenches may all harbor these systems, driven by geological fluids rather than sunlight. The team also documented at least three distinct species of hadal snailfish in the same trench, suggesting regional biodiversity far higher than expected, and identified fault rupture structures linked to historical seismic activity. The Atacama region sits on the boundary where the Nazca Plate slides beneath South America; the 1960 magnitude 9.5 earthquake that triggered trans-Pacific tsunamis left geological signatures now visible in the seafloor. This matters because China's "Global Deep-Sea Exploration Initiative," launched by the Institute of Deep-Sea Science and Engineering in 2022, gained UN endorsement in June 2025 as part of the Ocean Science Decade framework (2021-2030). The Atacama expedition is the flagship operation. What looked like a research cruise was actually a structural test: proving China can run multinational deep-sea operations, manage data sharing across borders, train talent collaboratively, and maintain long-term seafloor observatories. Success here translates to institutional leverage in every Pacific trench survey China conducts next.

The semiconductor rally is showing how extreme the AI storage cycle has become. Enterprise SSD and NAND suppliers are suddenly posting explosive numbers as AI infrastructure demand collides with tight memory supply. Micron, Sandisk, and others have all warned that AI-driven demand is outpacing supply, pushing pricing and margins sharply higher. But the market is increasingly pricing these shortages as permanent. Memory has always been cyclical. NAND pricing looks incredible while supply is constrained, but history suggests elevated margins normalize once new capacity ramps and hyperscaler buying stabilizes. Datapro's 10x IPO rally in 14 trading days masks a harder structural reality: the company just posted its first profitable quarter after four consecutive years of losses totaling $2.6 billion. The math is stark. Revenue grew 4x year-over-year in Q1 2026 to $1.87 billion, and the company swung to $528 million net profit. Constraints? Datapro competes against Samsung and SK Hynix in enterprise SSD controllers, a 20-year-old race Datapro entered a decade late. South Korea controls the market. The company has under 3% market share domestically and sells almost nothing internationally except to Google. Even with Nvidia and xAI validation coming in 2025, international brand credibility remains a structural gap that one good quarter doesn't close. AI has made semiconductors the hottest sector in the world right now, but a lot of today’s valuations depend on scarcity lasting longer than the industry usually allows.

In a telling case, digital games maker HoYoverse won a case against two livestreamers who claimed game characters were not carefully constructed fictional personas, but disguised representations of real company employees. One defendant had a pattern of livestreaming insults about HoYoverse games. What makes HoYoverse unusual among games makers is that it mainly sells emotional attachment to fictional characters. That emotional connection drives spending on “gacha” merchandise, livestreams, concerts, cosplay, and community participation. For many fans, the rumor reset their relationship with game characters. Instead of engaging with mythic or idealized fictional figures, players now felt manipulated by corporate insiders hiding behind avatars. Awarding the company 430,000 yuan (~$61,000) in damages plus mandatory public apologies, the court said this kind of rumor had an “industry-wide destructive impact." The company’s win shows that Chinese courts will protect the emotional substrate of fandom itself. And this is something Chinese studios have depended on. Modern entertainment economies are built on emotional trust. In digital fandom economies, the valuable asset is often belief itself. This includes belief that characters are coherent fictional entities, and belief that creators act in good faith. However, all this seems so pre-AI. Character-driven IP is hardly going to keep its value when its creation is two-a-penny.

ByteDance's domestic ad market is saturated. TikTok faces regulatory uncertainty in major Western markets. The only growth vector left is being the infrastructure backbone for China's AI ecosystem, which means outspending everyone on chips and compute. That probably explains why ByteDance is planning to put $28.5 billion into AI infrastructure this year, up 25% from its November plan. ByteDance is also shifting more of that $28.5 billion budget toward domestic AI chips, which typically cost more and have longer lead times than buying from NVIDIA. The company's huge (70%) net income drop in 2025 is more a feature of its funding model than any business model. ByteDance is likely using profitability as a valve to fund infrastructure. Something competitors are hard pressed to come close to.

Zhang Ruxiang, SMIC founder and China's semiconductor elder statesman, has just offered a convenient reframing the chip narrative, telling everyone obsessed over 2/3nm fabs that they're looking at the wrong 20% of the market. Mature nodes and specialty processes are what the other 80% need, he says. These service ultrasound equipment and other medical devices, are radiation-hardened for aerospace purposes, and power industrial robotics. And dwell in niches where Chinese manufacturers can actually break through without needing $20 billion fabs. Every tier-2 and tier-3 city is copying Shanghai's "full-stack" playbook and drowning in it. Zhang's formula cuts deeper: pick one niche, dominate it globally, then expand from there. The human capital insight is one nobody wants to hear. China's semiconductor bottleneck is the absence of master craftspeople on production floors and a value system that treats factory work as beneath engineers.

China's April investment data reveals where capital is moving. AI and humanoid robots saw 175% year-over-year growth in funding. Infrastructure projects in data, compute, and networks grew 61.7%. The headline is "economic momentum." The structure is capital rotation away from real estate into frontier tech stacks. This shows policy intent translating into allocation. When the state's own metrics highlight AI funding growth alongside infrastructure spending, institutional capital follows. The pivot isn't gradual, it's compressed into one month of data. Consumption recovered too. Spring travel searches jumped 118% month-over-month. Instant retail orders near parks and attractions grew 10.5%. But the investment story,175% into AI, 61.7% into compute infrastructure, is where structural reallocation happens. Consumption bounces with holidays. Investment reallocation sticks.

Ford, Boeing, and every major US industrial manufacturer have filed zero announcements about entering the space of humanoid robots, so why is China mad keen on them? On first sight, it's hype, but actually 60% of a humanoid's core components overlap with EV supply chains. Motor controllers, joint actuators, sensors from electric vehicles need minimal modification to fit robot limbs. What used to cost 50,000 yuan now costs 500 yuan. Core component costs have dropped up to 80%. A supply chain that was saturated with automotive and phone components found an outlet by reprising those parts into robots. Unitree's G1 humanoid at $85,000, smaller models under $10,000. 18,000 units shipped in 2025, up 508% YoY. Deep's 90% of hardware prototyping happens in Huaqiangbei in one day. Robots like these need 3-5 years before profitability. Investors don't fund long-burn hardware when AI software reports earnings this season. China's humanoid explosion mainly just shows overcapacity in two supply chains finding an adjacent outlet. Without this dynamic, building a humanoid industry in the US destroys unit economics.

Zhipu AI just bought a 22,700 sqm office tower in Zhongguancun for 361 million yuan (~$51 million). The real move isn't the real estate. It's what the move says about cash position and confidence in 2026. Zhipu went public on HK in January at 116.20 HKD per share. Stock is now at 918 HKD, an 8x return in four months. Revenue grew 131.9% YoY to 724 million yuan in 2025, but losses expanded to 4.72 billion yuan. Despite the burn, they're deploying capital into fixed assets instead of burning it on R&D subsidies or price wars. API pricing up 83% in Q1 2026. Call volume up 400%. CEO Zhang Peng says the market is supply-constrained, not demand-constrained. That shifts everything. When you're rationing access instead of fighting for users, you can afford to hold real estate and take losses on paper while the unit economics compress. ByteDance dropped 6.1 billion yuan on Beijing land in Q1. JD.com is building regional HQs in Hangzhou and Guangzhou. But those are growth plays. Zhipu's move is different: it's a capitalization play. A unicorn that went public four months ago doesn't lock 361 million in illiquid assets unless it believes it's solved the scarcest constraint, inference efficiency at scale. The building isn't the strategy. The willingness to buy the building is.

The most important export from Huawei is likely its organizational DNA. In 2026, former Huawei employees are filing IPOs across energy storage, automotive radar, AI infrastructure, and robotics. Sigenergy, founded by former Huawei solar executive Xu Yingtong, surged on its Hong Kong debut after attracting cornerstone investors like Temasek and Goldman Sachs Asset Management. Cheng-Tech, founded by former Huawei R&D engineers, has rapidly scaled in millimeter-wave automotive radar. Peng Zhihui, Huawei’s famous “genius youth,” left to build Zhiyuan Robotics and is now seeding a broader embodied-AI ecosystem. Huawei spent decades building one of the most disciplined execution systems in Chinese industry. Process control, supply-chain resilience, systems engineering, and the ability to operate under extreme constraints are capabilities that very few organizations have replicated at scale. So when Huawei veterans leave, they transplant proven industrial operating procedures into sectors where investors were previously skeptical hardware startups could execute reliably. These are founders who already know how to scale under pressure in strategically important sectors like semiconductors, EV supply chains, robotics, and AI infrastructure. Hubble Investment, Huawei’s investment wing, reinforces the same logic. Its portfolio spans chips, EDA, packaging, materials, and industrial software. Huawei’s people now people China’s entire hard-tech stack.

Toyota's March sales dropped 7.3% year-over-year, its second consecutive monthly decline. For a company with 10+ million annual sales and a decades-long grip on global No. 1, this stings. But Toyota is willing to tolerate all this while it waits for everyone else to exhaust themselves. BYD on the other hand shipped 454,300 overseas vehicles in the first four months of 2026, up 70.9% year-over-year in April alone. This isn't just normal growth. BYD behaves like a company that knows the rules right now are actually favorable. By this, we mean that Chinese industrial policy, export markets, EV enthusiasm, and overcommitted Western incumbents are the wave on which it is accelerating. So it is racing to lock in global scale before tariffs rise, markets harden, and legacy automakers regain balance. Toyota behaves like a company that assumes every window eventually closes. Its institutional memory was forged through oil shocks, trade wars, currency crises, and multiple technology cycles. Toyota’s bet is that EV intelligence will commoditize faster than expected, and that the real long-term moat will revert to manufacturing discipline, supply-chain leverage, dealer networks, capital efficiency, and the ability to survive low-margin environments. Toyota optimizes for endurance. Toyota and BYD are often framed as opposites in the EV race. But they are really optimizing for different phases of industrial competition. One company races because the rules still work in its favor. One company waits because it built the machine that survives when the rules change.

China's embodied AI companies are racing toward scale, and the bottleneck isn't the robots anymore, it's real-world data. At GEIS, a robotics conference in Silicon Valley, the consensus became clear: synthetic data alone doesn't work because it can't capture friction coefficients, latency, tactile feedback. So the industry has converged on hybrid training. Magic Atoms collects about 16,000 data points daily from real deployments, then synthesizes that 10,000x. Unitree posted 5,500 units shipped in 2025 ($17.07 billion revenue, over 50% from overseas). The companies winning at scale aren't the ones with the best algorithms, they're the ones with the most efficient data loops from actual deployment sites. The hard part isn't building better robots. It's getting robots into messy real environments fast enough to find failure modes labs never see, wet floors, rust, bright light, multiple systems running. OpenMind founder Jan Liphardt made this explicit: deploy early or fail late. That's why new energy vehicle manufacturing is now the primary data mine for training. The factories are already moving parts; robots just need to learn by doing it alongside humans who can correct them. On embodied "brains," VLA (vision-language-action) dominates because touch sensors are still immature. Amazon AI researcher Haozhi Qi noted the architecture choice is really just engineering pragmatism: vision sensors work, so use them to compensate for weak tactile systems. Meanwhile, dexterous hands are splitting into three paths: linkage (cheap, simple), tendon (fine manipulation, expensive), direct drive (balanced, heat management unsolved). The emerging winner is hybrid: tendon structure for precision plus AI for control. The bottleneck in dexterity isn't physics, it's learning efficiency. A robot that can't learn from failure stays expensive.

China's Hainan first Hualong One unit just loaded fuel for the first time, moving toward grid connection. This matters less as a nuclear milestone and more as evidence of how Beijing is solving the clean energy sequencing problem that has stalled rivals. The Changjiang unit 3 is a 1.2 gigawatt pressurized water reactor jointly built by China National Nuclear Corporation and China Huaneng Group. Fuel loading on April 30 marks transition from construction to commissioning. Once operational, the unit will bring in 9+ billion kilowatt-hours annually. The full two-unit site is projected to hit 18+ billion kWh/year by 2027, and this will mean coal savings of 6.3 million tonnes annually. The constraint Beijing solved here is cost and speed alignment. For rivals like EDF or GE, the bottleneck is regulatory fragmentation and project-specific financing. Hualong One units take 5-6 years from first concrete to grid. The economics only work if you can deploy multiple units at scale within a decade without refinancing cycles. Hainan's designation as a special economic zone simplified permitting and allowed accelerated investment schedules. China has now deployed Hualong units across Guangdong, Zhejiang, and Fujian on parallel timelines.

The embodied AI market just repriced platform value. Zhiyuan Robot landed a 15 billion yuan valuation. Qingtian Rental, its logistics and deployment platform, just raised at 3 billion yuan, 20% of the hardware maker's valuation. In mature SaaS this ratio is standard. In embodied AI, where hardware dominates and supply chains are fractal, it's new. Capital is betting that the ability to connect hardware to end-use cases is scarcer than the ability to manufacture the robots themselves. It looks like equity dilution is accelerating, a smart signal of governance shift from founder control to multi-stakeholder checks. On top of this, non-Zhiyuan robots are joining up fast. Competing OEMs are seeing that it's cheaper to plug into Qingtian's 400-city partner network than build their own. Third, the partner training curriculum is already abstracted, it teaches "embodied AI operator" standards, not Zhiyuan-specific tactics. Europe and North America may have information-matching brokers but they don't have operational platforms. Qingtian has bundled hardware + deployment + insurance + field service into one go-to-market. What happens when a dozen other hardware teams race to build their own fleets? The 80%+ conversion rate at partner workshops suggests they think yes. If the next 18 months show sustained multi-brand logistics velocity and insurance underwriting scale, the 3 billion valuation anchors a genuinely new category, not rental, but infrastructure.

China's offshore revenue threshold just jumped 10 percent year-over-year. The bar for making the top 50 overseas earners moved from 400 billion yuan to 442 billion yuan ($58.56 Billion to $64.71 Billion). In short, China’s export story is consolidating at the top, and less fragmented than before. In 2025, just 50 listed firms drove over half of all overseas revenue, with 22 clearing ¥100B each. The leaders span energy, electronics, EVs, and mining, hinting at where China’s global leverage is strongest. The top five are China National Petroleum, Foxconn Industrial Internet, BYD, Lishen Precision Electronics, and Zijin Mining. Offshore revenue concentration is rising faster than the offshore revenue pool itself. Companies without scale in manufacturing, resources, or supply chain integration are losing overseas footing to those that have. We'll have to see if this trend reverses or locks in.

In this BBC World Service interview, Rui shares context on China’s role in major tech deals and why understanding the ecosystem matters. youtu.be/MEwMD7-prbA?si… via @YouTube

CATL is evolving from a component supplier into a system-level infrastructure company. In Q1 2026, the battery maker generated around RMB 20B in profit, with margins far exceeding BYD despite similar revenue scale. Chinese media site TMTPost explained that CATL is positioning itself as a neutral supplier across the entire auto ecosystem, unlike BYD which is integrating vertically. In April, according to TMTPost, CATL moved to expand its battery swapping alliance (BAIC, Chery), promote its “Chocolate” standard (interoperable swapping), publicize its next-gen batteries (semi-solid, sodium-ion, fast charge) and its accelerated factory building in Germany, Hungary, Spain Battery swapping is the key pivot. Unlike one-off battery sales, it creates recurring revenue, data ownership, and ecosystem lock-in across automakers. If CATL succeeds, it will gain control of the entire system cars run on. Moreover, with its 350Wh/kg-class batteries, CATL is signaling future tech leadership while competitors are pushing what’s already in mass production. The goal is clear: control the field and be known as more than just a battery supplier.

The shift to physical AI isn't a marketing narrative for algorithm vendors, but a survival requirement. Zhuo Yu, a driving algorithm firm, has a new native multimodal foundation model that it showed off at the Beijing auto show last week. Its VP Beibei Yu framed the transition bluntly: "If you don't take this technical route, you'll likely be unable to compete later." 30% of training comes from vehicle-collected data, 30% from robotics, and 40% from internet-gathered videos. But physical AI must deploy to hardware. That's the bottleneck. Zhuo Yu is already moving beyond the traditional Tier 1 model (sell hardware, collect licensing fees). They're building Robotaxi and RoboVan L4 services on subscription and profit-share revenue, plus per-use fees for cloud inference on low-risk robotics.

Alibaba Damo Academy just published their colorectal cancer screening results in the Annals of Oncology, showing that their model, DAMO COCA, detected 5 missed cancers from 27,433 plain CT scans with 86.6% sensitivity and 99.8% specificity. The biggest advantage is that it uses CT scans instead of colonoscopies, which means uptake increases dramatically. Far less uncomfortable. This is the third cancer screening model DAMO has published (pancreatic, gastric, colorectal), so it isn't a one-off lab result. Hospital imaging departments are now opportunistic screening systems and there's no need for workflow redesign. DAMO ran real-world validation on 27,433 patients and found one case was missed for two consecutive years before a colonoscopy caught it in year three. But it only works if radiologists use the model at clinical decision-making time. These results should give clinicians more trust in using AI.

Bytedance's AI assistant Doubao is launching a freemium model with three paid tiers (68/200/500 yuan monthly, 688/2048/5088 yuan annually) or $10/$28/$70 per month, targeting production workflows. The move reveals a structural shift in how Chinese AI companies are thinking about monetization: complexity as the pricing signal, not model capability. The paid tiers unlock PPT generation, data analysis, and video editing, tasks that demand higher inference costs and longer compute time. This allows Bytedance to isolate high-compute workloads into a separate economic model. Synthesis tasks (generating a 50-slide deck, analyzing a 10GB dataset) are chewing up compute in ways that break the free model's unit economics. It signals how ByteDance will compete with OpenAI's subscription model in China. OpenAI charges for "smarter" reasoning. Doubao charges for "heavier" tasks. In the Chinese context, where inference infrastructure is capital-intensive and domestically fragmented, Doubao's model aligns better with the ecosystem's actual bottleneck.

Doubao launched "Help You Choose" inside its app on April 29, letting users search, order, pay, and manage after-sales without leaving, all routed through Douyin Shop. The test results were solid. It correctly inferred camping speaker requirements, and it calculated that a 3,334 yuan (~$476) phone exactly maximizes China's 500 yuan government subsidy cap. The no-app-switching experience is genuinely frictionless. The ceiling is the same as the floor: everything routes through Douyin Shop. ChatGPT pulls from multiple platforms, presents comparisons in tables, and gives plain-language tiebreakers. Doubao has no path forward for cross-platform price comparison or brands not carried on Douyin. Users who already shop on Douyin will find this useful. Users who compare prices across platforms will hit the wall immediately.