618’s New Variable

By early June 2026, China’s 618 shopping festival, the country’s largest mid-year retail event, had acquired an unfamiliar competitive structure. For the first time, 3 distinct AI commerce systems are competing for the same consumer base.

On May 11, Alibaba fully integrated Qwen, its flagship AI assistant, with Taobao, China’s dominant marketplace. Users can now search, compare, order, and pay for goods from Taobao and Tmall’s combined catalog of more than 4 billion items, completing payment via Alipay without leaving the Qwen flow. ByteDance took a more gradual path: Doubao, its AI chatbot with 345 million monthly active users as of March, first connected to Douyin’s e-commerce platform in late 2025 and progressively added in-app payment, product comparison, and a dedicated “Help Me Choose” shopping entry point through May 2026. Then on June 8 and 9, Tencent opened WeChat’s AI ecosystem to outside developers. JD.com (China’s second-largest e-commerce platform), Meituan (the dominant food delivery and local services provider), ride-hailing leader Didi, and travel platform Ctrip were among the first to connect. JD.com linked through an Agent-to-Agent arrangement; the broader program turns Mini Program services into AI-callable capabilities. The integrations are developer-side and in testing; consumers cannot use them yet.

Each company chose a different structure. Alibaba built a direct bridge between its AI layer and its commerce platform. ByteDance sealed its AI inside its own retail ecosystem. Tencent is not trying to own the product catalog or fulfillment stack; it is building WeChat into the orchestration layer above partners that already hold those capabilities. These choices reflect 3 distinct theories about where value will accumulate when AI captures consumer intent before it reaches a search box, a content feed, or a standalone shopping app.

What the Search Box Built, and What Replaces It

For 2 decades, merchants have bid for placement in Chinese e-commerce search results. At the Taobao and Tmall Group, Alibaba’s core China commerce unit, keyword advertising is the backbone of customer management revenue. Chinese industry commentary has estimated that it supports roughly 70% of the group’s profit base.

AI dialogue largely bypasses this system. When a user asks an AI assistant for a specific product, the assistant matches intent to attributes directly, bypassing sponsored listings. The auction layer recedes from the user’s experience. All 3 companies recognize the shift. Their responses diverge.

Alibaba is willing to cannibalize part of its own ad model before an outside AI interface does it first. ByteDance chose to channel AI shopping traffic into Douyin’s existing commerce engine. Tencent chose to become the orchestration layer that dispatches user intent to whichever service can fulfill it, without owning inventory or building a shopping AI. I examined the collision between AI agents and Taobao’s ad economics in January, when Qwen first connected to Alibaba’s consumer stack. The tension was still theoretical then. 5 months later, Alibaba has answered it with a concrete product decision, at least at the interface level: according to Chinese reports, it removed commercial ranking from the search logic Qwen uses to access Taobao’s catalog.

Alibaba is now running 2 parallel distribution systems for the same inventory. One is optimized for user trust. The other sustains the advertising cash flow. The question 618 begins to surface is how long that dual arrangement can hold, and what happens when real transaction volume exposes each model’s weak points.

Below, I examine each architecture: Alibaba’s de-commercialized trust layer, ByteDance’s closed Douyin-commerce loop, and Tencent’s WeChat orchestration model. The question is not whether AI can recommend products. It is whether any of these systems can turn intent into revenue without destroying user trust.

Read more

A father in England wanted to make his son a toy sword. He lacked his own father’s carpentry skills. A desktop 3D printer solved it: £0.10 of filament, a few hours of machine time, and the sword materialized. Odds are, the printer came from Shenzhen.

In 2025, Chinese manufacturers exported 5.03 million 3D printers, up 33% by volume from the prior year, with total export value reaching RMB 11.35 billion. In Q1 2026, exports surged another 119%.

Four Shenzhen companies account for roughly 9 in 10 consumer 3D printer shipments worldwide: Bambu Lab, founded in 2020 by five former DJI executives; Creality, which listed on the Hong Kong Stock Exchange on May 29, 2026, as the first listed pure-play consumer 3D printer maker; Anycubic; and Elegoo. About 70% of their machines end up in Europe and North America.

Printers from the leading Chinese brands increasingly ship with proprietary software, cloud connectivity, and access to a design library. Together, these turn a desktop device into a closed-loop manufacturing system built on a Chinese-designed platform. China’s most visible exports have been finished products: phones, solar panels, electric vehicles. Consumer 3D printers add a different category: manufacturing capability itself, priced as consumer electronics.

Two Weapons, One Decade

Two forces built the dominance. The first was price. Creality’s CR-10, released in 2016, cut the entry cost of a capable printer from roughly $1,000 to $500 and outsold the industry average by a factor of 20. Within a few years, Chinese manufacturers had pulled consumer 3D printing into mainstream consumer-electronics pricing.

The second was usability. Early printers required assembly, manual calibration, and hours of troubleshooting. Bambu Lab, whose founding team came entirely from DJI, applied the drone maker’s core insight: eliminate the gap between unboxing and usable output. Its 2022 debut, the X1, used gyroscope-based vibration suppression, lidar monitoring, and automated calibration to remove manual setup. Time named it one of the year’s best inventions. Follow-on models at $399 and $199 reached a mass audience. Revenue reportedly grew from under RMB 2 billion in 2023 to an estimated RMB 10 billion-plus in 2025.

Western 3D printing companies weakened in parallel. Spain’s BCN3D went bankrupt in 2025. America’s Desktop Metal entered restructuring. Germany’s TRUMPF exited additive manufacturing. In the consumer segment, only Prusa Research, a Czech company, retains meaningful non-Chinese share at roughly 5.8%. The pattern points to a shared vulnerability: Western manufacturers optimized for precision and capability while leaving usability and cost to the market. The Shenzhen companies that replaced them drew on a resource their rivals lacked: a consumer electronics supply chain, engineers from the drone industry, and infrastructure to iterate from prototype to mass production in weeks.

Creality’s May 2026 IPO captured the sector’s internal dynamics. The stock opened 80% above its HK$18.80 offering price, with Hong Kong retail investors oversubscribing the public tranche 3,829 times. Yet Creality went public as the number 2 player. Its prospectus disclosed 2025 revenue of RMB 3.13 billion but a net loss of RMB 182 million, driven by pre-IPO share issuances, listing expenses, and rising sales and R&D costs. Adjusted for those items, net profit was RMB 92.4 million. The prospectus identified the market leader only as “Company A” with a 42.7% GMV share. Industry data identifies that company as Bambu Lab, far ahead of Creality’s 11.2%, despite entering the market 6 years later.

Price built the market. Usability expanded it. The third phase is ecosystem: the software, content, and AI tools that determine how useful a printer remains after purchase. That layer is where the strategic weight now sits, and where the implications extend beyond consumer electronics.

Read more

In May 2026, China’s three state-owned telecom carriers launched AI token subscription packages within days of each other. China Telecom, the country’s second-largest carrier by mobile subscribers, introduced consumer plans starting at RMB 9.9 per month for 10 million tokens. China Mobile, the largest carrier by revenue, began rolling out plans through provincial branches, with Shanghai pricing at RMB 1 per 400,000 tokens. China Unicom, the smallest of the three, offered packages from RMB 15 per month for 6 million tokens. Several of these plans support payment through existing phone bills. Users pay for AI inference the way they pay for mobile data.

The unit has changed. The billing mechanism has not. That structural parallel is the carriers’ core bet. Over two decades, Chinese carriers sold two commodities: voice minutes and data gigabytes. Each followed the same lifecycle. Meter usage. Bill customers. Distribute through a national network. Serve hundreds of millions of subscribers. Token economics maps onto the same template. The carriers already own the metering systems, billing infrastructure, customer accounts, and last-mile networks to distribute a new commodity at scale.

The timing reflects necessity as much as ambition. In 2025, revenue growth at all three carriers fell below 1% for the first time in six years. China Mobile reported RMB 1.05 trillion in total revenue, up 0.9%. China Telecom grew 0.07%. China Unicom grew 0.68%. China Mobile’s ARPU, the average revenue per user, fell 3.5% to RMB 46.8. Its subscriber base reached 1.005 billion, with net additions of roughly 1 million for the full year. The growth model that powered two decades of expansion, more users consuming more data, is stalling.

The carriers have responded by redirecting capital toward AI compute. In May, China’s State Council placed computing networks in the same infrastructure-planning frame as water networks, new-type power grids, next-generation communication networks, urban underground pipeline networks, and logistics networks. China Telecom invested RMB 20.2 billion in compute infrastructure in 2025. For 2026, all three carriers plan to accelerate, with combined compute budgets approaching RMB 81 billion. China Mobile’s compute services revenue reached RMB 89.8 billion in 2025, up 11.1%. Compute and intelligent services together accounted for 20.2% of core service revenue. Spending on infrastructure and earning returns from it, however, are different problems.

Daily token consumption in China rose from roughly 100 billion in early 2024 to 140 trillion by March 2026. The OpenClaw wave in early 2026, when open-source AI agents began moving from developer circles into cloud-PC and one-click deployment scenarios, gave carriers a proximate trigger for packaging inference as a billable product. The question is whether the infrastructure advantages that made them dominant in voice and data can transfer to a market where model quality, application design, and trusted delivery all shape user value. As of late May 2026, no major Western carrier has launched a comparable consumer-facing token package. And the financial risks echo a precedent the carriers know well.

Three state carriers are betting that token distribution is the next iteration of their core business. What follows examines whether the economics support that bet, why Western carriers are taking a different AI route, and why the 5G investment cycle casts a long shadow over the entire effort.

Read more

In December 2025, WeChat, China’s ubiquitous super app with 1.4 billion monthly users, forced users of ByteDance’s Doubao Phone Assistant to log out within days of the product’s launch. The app flagged accounts for abnormal activity and shut down access. ByteDance retreated, disabling WeChat functionality and limiting its AI agent to a narrow set of scenarios.

On June 4, Tencent confirmed what looked like the opposite. WeChat is cooperating with Huawei, Honor, Xiaomi, OPPO, and Vivo to provide what Chinese reports describe as A2A, or Agent-to-Agent, capabilities to their phone AI assistants. This is not the open Agent2Agent protocol promoted by Google and the Linux Foundation, but a Tencent-controlled access arrangement between WeChat and phone assistants. Users can ask their phone’s AI to send a WeChat message or make a WeChat voice or video call. Honor, the Shenzhen-based smartphone brand, has completed integration across its Magic 8, 500, and X70 series, with 50% of online active devices connected.

The reversal appears dramatic. A platform that treated external AI agents as security threats now welcomes them. But WeChat has built a new entry point and chosen who gets access. For unauthorized agents, the barriers remain intact.

The Door Tencent Designed

The mechanism keeps execution and data inside WeChat. A phone AI assistant, such as Honor’s YOYO, Xiaomi’s XiaoAi, or Huawei’s Xiaoyi, converts a user’s spoken command into a structured request and passes it to WeChat through an encrypted channel. WeChat executes the instruction within its own system and sends back the result. The phone assistant initiates the request, but WeChat keeps the execution layer inside its own environment.

This contrasts with the GUI agent approach ByteDance attempted. Doubao operated at the system level, reading screens and simulating finger taps to navigate apps without their permission. Tencent president Martin Lau framed the distinction in the company’s Q1 2026 earnings call. An application that operates inside other apps without authorization, he said, is “essentially plundering” them. Applications pretending to be operating systems and “invading” other apps would not be tolerated.

The framing matters beyond rhetoric. A2A embeds this principle into architecture. Dual authorization, meaning user consent plus application permission, routes supported WeChat interactions through Tencent’s controlled channel. The company defines which functions are available. Currently that scope is narrow: sending messages and making voice or video calls. Tencent retains the ability to define the scope of access and, in practice, to expand or withdraw permissions.

The AI Gap Behind the Reversal

WeChat’s mini-program ecosystem handles food delivery, government services, healthcare, and payments. Yet Tencent’s standalone AI application, Yuanbao, remains far smaller than its leading rivals. Third-party data cited in Chinese media put Yuanbao at about 57.35 million monthly active users in March 2026, compared with 345 million for Doubao and 166 million for Qwen.

The spending gap also matters, though the categories need to be read carefully. Tencent’s 2025 capital expenditure was about RMB 79 billion, much of it tied to AI infrastructure. That was still below the scale implied by ByteDance’s reported RMB 150 billion-plus 2025 capex plan and Alibaba’s RMB 380 billion three-year commitment to cloud and AI infrastructure.

These numbers reveal a strategic imbalance. Tencent is under pressure on two fronts: its standalone AI app lags Doubao and Qwen in usage, while rivals have moved faster in model-led agent products. WeChat remains its strongest card.

The Financial Times reported on June 2 that Tencent was testing a WeChat-native AI agent and aimed to begin the required compliance process as soon as that month. The company would then test the agent with a small group of outside users before a phased rollout. A public launch date had not been set.

Users would swipe right on the main WeChat screen to access an AI chat box capable of issuing commands across the app’s millions of mini-programs. Tencent’s stock rose 10.5% that day, its largest single-day gain since January 2021. The surge suggested investors saw the WeChat AI agent as evidence that the platform’s ecosystem advantages could carry into the AI era.

A2A fits into this broader strategy because Tencent faces a timing problem. WeChat’s own agent has not yet launched, while phone assistants are already becoming the first AI contact point on mobile devices. Without an authorized channel, system-level assistants would either fail to handle WeChat requests or keep testing ways to route around WeChat. A2A gives Tencent a third option: let the phone assistant capture the user’s intent, but force execution back into WeChat.

In my recent analysis of Tencent’s agent push, I argued that the company was trying to occupy every surface where users might interact with AI, even while its model and cloud infrastructure remained under pressure. WeChat A2A is the next step in that strategy. It does not merely add another agent surface. It defines how outside agent surfaces are allowed to connect back to Tencent’s most important ecosystem.

One person inside Tencent described the logic: any phone agent that cannot interact with WeChat cannot be considered a real system-level agent. Tencent would inevitably open this door. The only question was timing.

The surface story is industry cooperation. The deeper story is protocol power. WeChat is defining the terms under which outside AI can touch China’s most important app.

Read more

DeepSeek has moved into a harder dataset: corporate spend.

Ramp’s June 2026 report placed the Chinese AI company at the top of its “Trending” software vendors list, which measures breakout growth relative to size. The usage pattern matters more than the ranking. These are firms making direct payments to DeepSeek and sending data through DeepSeek directly. This is a different signal from self-hosted open-weight adoption.

OpenRouter showed the shift first in developer token routing. Ramp now shows it entering corporate vendor spend.

DeepSeek is not new to Ramp’s data. It briefly reached 0.3 percent business adoption during a January hype cycle before falling back to 0.1 percent of American businesses. That surge faded. This one has a more concrete driver: permanent pricing.

In May, DeepSeek made a 75 percent discount on its V4-Pro model permanent. Cached-input pricing now starts at RMB 0.025, or roughly $0.0035, per million tokens. Against Anthropic’s discounted prompt-caching economics, that is still about 1 percent of the comparable cached-input cost.

Ramp itself cautioned against overstating the durability of the trend. That caveat matters. A monthly trending list is not a market-share verdict. It is still evidence that DeepSeek has entered corporate payment flows, rather than remaining a developer-side talking point.

The broader list tells the same story. Three model-serving and inference platforms landed on Ramp’s trending list: Fireworks AI, fal AI, and DeepInfra. Vast.ai, a GPU cloud provider, also appeared. The picture extends beyond one Chinese model gaining traction. A cost-optimization infrastructure layer is forming around alternatives to premium-priced American providers.

The developer-side signal arrived earlier. OpenRouter data cited by Bloomberg in May showed 7 of the 10 most-used models on its developer routing platform were Chinese-built. Weekly token volume had reached 25 trillion, up from 5 trillion 6 months earlier. We examined the competitive dynamics and policy implications of that shift earlier this week.

OpenRouter showed where developers were routing tokens. Ramp shows which vendors are entering corporate payment systems. That difference matters for investors, and for Washington.

Read more

When Meituan reported Q1 2026 results on June 1, the headline figure dominated coverage: adjusted net loss narrowed to Rmb 5 billion from Rmb 15.1 billion the previous quarter. After a year in which the delivery war produced a Rmb 62 billion swing from profit to loss, coverage and early trading focused on stabilization: the adjusted net loss narrowed sharply from Q4, and the subsidy war appeared to be easing.

The more consequential signal was buried deeper in the filing. Starting this quarter, Meituan restructured how it reports revenue. The company merged commission income and advertising income into a single “merchant services” line. Simultaneously, it carved out “product sales” as a standalone category for the first time, separately disclosing revenue from Xiaoxiang Supermarket, Kuailv, and other self-operated retail categories such as medicine and alcohol, where both revenue and inventory cost are recorded on a gross basis.

Under the previous structure, investors could roughly isolate Meituan’s in-store and delivery business performance by comparing commission revenue growth against advertising revenue growth. The new merged category weakens that signal. The separate product-sales line makes a different signal visible: product sales reached Rmb 21 billion in Q1, up 46.6% year over year. Within core local commerce, product sales nearly doubled, rising 96.0%. Revenue disclosure changes of this kind often carry strategic intent. Self-operated retail becomes easier to track. Pressure in in-store services becomes harder to isolate. That gross-basis treatment makes the business look less like a marketplace and more like a retailer carrying its own economics.

The headline financials reinforce that reading. Total revenue rose 5.6% to Rmb 91 billion. Core local commerce, which includes food delivery, in-store services, and instant retail (Meituan Instashopping), grew just 0.1% to Rmb 64.1 billion. Meituan said core local GTV still benefited from higher annual purchase frequency and cross-selling, but revenue growth was held back by elevated incentives deducted from revenue. New initiatives, encompassing Xiaoxiang Supermarket (Meituan’s self-operated grocery platform) and overseas delivery brand Keeta, grew 21.3% to Rmb 27 billion. Operating loss narrowed to Rmb 6.5 billion from Rmb 16.1 billion in Q4 2025, helped by lower spending on promotion, advertising, and user incentives. Selling and marketing expenses fell to Rmb 23 billion from Rmb 31.7 billion.

Meituan’s Q1 results suggest a company using a narrow window of competitive relief to accelerate a fundamental shift: from an asset-light marketplace that earns commissions, toward a heavier operating model that controls procurement, manages inventory, and owns fulfillment. Product sales are becoming more visible. In-store pressure is becoming harder to isolate. The balance sheet already shows the financing pressure behind that shift: non-current borrowings more than doubled in a single quarter, rising to Rmb 42.7 billion from Rmb 18.8 billion at year-end 2025. Meituan is not just narrowing losses. It is borrowing to build.

The ceasefire is real. So is the borrowing. The question is whether Meituan can complete its shift before borrowing costs, renewed subsidies, or in-store competition narrow the window.

Read more

According to OpenRouter usage data cited by Bloomberg in May 2026, 7 of the 10 most-used AI models on the platform were built by Chinese companies. DeepSeek, Tencent, and Moonshot AI all appeared near the top of the rankings. Anthropic’s Claude Opus 4.7 was the highest-ranked American model in that snapshot.

OpenRouter, a platform that routes developer API calls across more than 400 AI models, has become a clearinghouse for the global inference market. It now processes 25 trillion tokens per week. Six months earlier, the figure was 5 trillion. On May 26, OpenRouter announced a $113 million Series B led by CapitalG, Alphabet’s investment arm, with NVentures, Nvidia’s venture arm, among the participants. TechCrunch, citing The New York Times, reported a roughly $1.3 billion post-money valuation.

The investor list is telling. The parent company of Google and the maker of the world’s dominant AI chips are both backing a platform whose core value proposition is helping developers find cheaper and more suitable alternatives across providers. That bet implies both expect the multi-model, cost-driven inference market to grow faster than any single provider’s share of it.

The volume shift carries its own weight. At 5 trillion tokens a week, OpenRouter was a niche tool for cost-conscious developers. At 25 trillion, it is infrastructure. OpenRouter CEO Alex Atallah has framed inference at scale as a multi-model routing problem. CapitalG, its lead investor, put the infrastructure analogy more directly: OpenRouter is trying to fill for AI inference the kind of infrastructure gap that Stripe filled for digital payments.

On a platform processing that volume, the models receiving the most traffic carry weight well beyond experimentation.

Washington has reached a different conclusion about what the data means.

On April 29, the House Select Committee on China and the House Committee on Homeland Security sent letters to 2 American companies: Airbnb and Anysphere, the maker of the AI coding tool Cursor. Both letters cited serious concerns about the national security and data-security implications of using AI models developed by Chinese firms.

The Airbnb inquiry was prompted in part by comments from CEO Brian Chesky, who said in an interview last October that the company preferred Alibaba’s Qwen, an open-weight large language model often described as open source, in certain situations because it was “fast and cheap.” In Anysphere’s case, the probe specifically targeted Composer 2, a model the company later disclosed was built on Kimi, developed by Beijing-based Moonshot AI.

The scope is now clear. The concern extends well beyond one company’s customer-service chatbot.

Airbnb’s customer-support AI bot, expanded over the past year, now handles about 40 percent of issueswithout escalating them to a human agent, up from about one-third earlier this year. In a series of Bloomberg interviews in May, Chesky pushed back on the probe. Airbnb uses “a variety of open-source models, including US open-source models,” he said. “We are not providing data to any Chinese companies. They don’t have access to any data.” The technical point, he added, is straightforward: “An open-source model does not have access to data. It doesn’t work that way. I think people need to understand how this stuff works.”

At Google I/O, Alphabet CEO Sundar Pichai shifted the frame from model origin to frontier capability. “I worry less about, ‘Are we adopting open-source models from China?’” he said, “and more, ‘Are we doing enough in the US to make sure we are staying at the frontier?’”

Two prominent technology CEOs are pushing back on the Congressional framing from different positions. The one deploying Chinese models says the security concern misreads how the technology works. The one competing against them says the strategic priority should be American frontier capability.

Their responses converge on a shared observation: the investigation risks treating model origin as a proxy for data transfer. In many open-weight deployments, that proxy does not match the technical architecture.

There is an important distinction between the cases. OpenRouter captures routed inference traffic, where prompts move through an API gateway and model providers. Airbnb’s defense rests on a different architecture: open-weight models deployed inside its own controlled infrastructure. The market signal is similar. Companies are choosing cheaper, capable, and more flexible models. The data-flow risk is not the same.

The competitive dynamic Congress is responding to, meanwhile, keeps accelerating.

In May, DeepSeek announced that a 75 percent discount on its flagship V4-Pro model would become permanent after the end of the month. Cached input pricing starts at RMB 0.025 per million tokens. Uncached input costs RMB 3. Output runs RMB 6. Even against Claude Sonnet 4.6’s discounted prompt-caching economics, DeepSeek’s cached-input price is still around 1 percent of Anthropic’s comparable input cost.

Making a promotional discount permanent does not prove durable profitability. It does show that DeepSeek is willing to set the market around that price. Whether the rate reflects durable unit economics, temporary subsidy, or a capacity strategy remains opaque. The pricing floor is now visible.

The pressure on the other side of the equation is equally specific. Uber president and chief operating officer Andrew Macdonald recently acknowledged that the company had reportedly exhausted its annual AI budgetjust four months into 2026, as Claude Code usage and token consumption rose faster than the company could clearly connect to shipped consumer features.

As enterprises deploy agent workflows that chain dozens of inference calls per task, the cumulative cost of premium-priced tokens scales from a line item into a constraint on how aggressively a company can deploy AI at all. When budgets evaporate that fast, model origin becomes harder to prioritize. Cost becomes the first screen.

Three months ago, we documented the moment Chinese models first surpassed American ones in weekly token consumption on OpenRouter. The pattern was emerging. The data now shows something more consolidated: dominant market share, 5-fold volume growth, and permanent pricing set far below American competitors.

The question Congress is investigating, whether American companies should be using Chinese AI models, may already trail the question the market is answering every day.

The open-weight distinction Chesky drew has specific consequences for what policy can and cannot do. Behind it lies a cost structure that permanent pricing has now made visible, and a set of enforcement constraints that neither hearing nor letter can easily resolve.

Read more

From Principle to Price War

In late May, Fuli Luo, who leads Xiaomi’s MiMo large language model team, posted a technical explanation for why MiMo’s API prices had just dropped by up to 99 percent. A hierarchical KV cache optimization had increased cached token capacity by 5 times, she wrote, cutting caching costs by roughly 80 percent; the 99 percent headline reduction applied specifically to cache-hit input tokens, with broader input and output prices dropping 60 to 80 percent. The optimization lowered unit costs enough for MiMo to price its new API rates at near break-even, Luo argued.

In April, Luo had publicly argued that the AI industry’s rush to slash token prices was structurally unsustainable. Xiaomi’s own Token Plan, launched in April, was built on the opposite premise: a credit-based system that billed developers for tokens consumed rather than offering unlimited access at a flat rate.

Between those two positions, MiMo distributed 100 trillion free tokens to developers worldwide and briefly topped OpenRouter’s Hermes agent model rankings. By late May, MiMo-V2-Pro had fallen to 16th on the same ranking. Then came the permanent price cut, aligning MiMo’s rates with those of DeepSeek, the Chinese AI lab whose open-source models have set the industry’s pricing floor. The shift from structural pricing critique to competitive price-matching happened within the same reporting season.

Xiaomi’s Q1 2026 earnings, released the same week, supply the financial context. Revenue fell 10.9 percent year-on-year to RMB 99.1 billion. Adjusted net profit dropped 43.1 percent to RMB 6.1 billion. After reporting RMB 0.9 billion in operating income for full-year 2025, the segment housing both electric vehicles and AI posted a RMB 3.1 billion operating loss in Q1 2026.

MiMo adds one more cost-sensitive front to a company already trying to protect margin across phones, cars, and IoT. The surface reads as broad-based weakness. The composition points to a more selective set of trade-offs. Xiaomi’s Q1 numbers read more clearly as four different businesses carrying four different burdens: smartphones preserve the user base, IoT supplies the margin buffer, EVs sustain the growth narrative, and AI absorbs the capital and R&D spending. Q1 reveals which of the four is bearing the most weight, and it is not the one that usually dominates the market narrative.

Shrinking on Purpose

A significant part of the revenue decline reflected deliberate trade-offs, though not every segment fits that pattern.

Smartphone shipments fell 19.2 percent year-on-year to 33.8 million units. Management described this as proactive control of mid-to-low-end inventory, a response to the memory cost supercycle that has been driving up component prices since mid-2025. Average selling price rose 8.2 percent to a record RMB 1,310, with premium models reaching 23.5 percent of mainland China sales. Xiaomi halted first-generation SU7 sales for two months to prepare the new-generation launch in March. Q1 EV deliveries of 80,856 units reflected the transition gap, not demand failure: the new SU7 accumulated over 80,000 configuration-locked orderswithin 48 days.

IoT was different. Its revenue fell 23.7 percent year-on-year to RMB 24.7 billion, pulled down by a subsidy-inflated prior-year quarter in China, though overseas IoT revenue hit a record high. The common thread across segments was not that every decline was chosen, but that Xiaomi kept prioritizing margin structure over headline revenue. Whether that margin structure could absorb the costs piling up elsewhere depended on the least visible of Xiaomi’s major hardware businesses.

The Q1 numbers reveal a company making controlled trade-offs across four business lines. The division carrying the most financial weight is the one that tends to receive the least attention. What follows starts with a margin comparison that reframes the entire story and asks how long the current structure can hold.

Read more

Dozens of Products, One Thin Foundation

Between March and May 2026, Tencent launched an unusually broad wave of AI agent products across office productivity, coding, desktop control, data analytics, design, and operating-system-level assistance. WorkBuddy for office productivity. CodeBuddy for software development. QClaw for personal desktops. DataBuddy for data analytics. Ardot for UI/UX design. Marvis, an operating-system-level assistant released in late May, uses a six-agent system led by a main agent that delegates tasks to File, Computer, App, Browser, and Search agents. DataBuddy targets data pipelines. Ardot targets UI/UX design and code handoff. These are not general-purpose chatbots. Tencent is pushing agents into work categories where task completion, not conversation quality, determines value. Alongside these new products, Tencent upgraded Yuanbao, QQ Browser, Sogou Input, and Tencent Meeting with agent capabilities and began preparing to convert WeChat Mini Programs into AI-callable skills, meaning services that agents can invoke directly rather than pages users open manually.

Chairman Pony Ma personally promoted the product lineup on WeChat Moments, listing over a dozen agent variants and noting that more were coming. Tang Daosheng, Tencent’s Senior Executive Vice President and CEO of the Cloud and Smart Industries Group (CSIG), declared that AI’s application paradigm was shifting from chatbots to agents.

The product wave landed against a less flattering set of recent numbers. Yuanbao, Tencent’s consumer AI app, reported 57.35 million monthly active users. ByteDance’s Doubao reported 345 million. Alibaba’s Qwen app reported 166 million. In China’s public cloud IaaS market, Tencent Cloud ranked 5th at 8% share. In Model-as-a-Service (MaaS), the segment where token-based AI services are growing fastest, ByteDance’s Volcano Engine, Alibaba Cloud, and Baidu AI Cloud held nearly 90% combined. Tencent Cloud did not appear among the named players. New AI products implied roughly Rmb 8.8 billion ($1.2 billion) in quarterly non-IFRS operating-profit drag.

That contrast matters beyond Tencent: in the agent era, can the company that occupies the most user touchpoints first build a durable advantage, even when the model underneath is still catching up and the cloud infrastructure is still undersized?

In my March analysis, I described Tencent’s strategic logic as “interface layer beats model layer.” In my May analysis, I described the company as a cash machine funding a loss-making AI venture. Both frameworks hold. What has changed is the volume of execution data available to stress-test them.

WeChat won the messaging market through ubiquity: be everywhere, optimize later. Applied to agents, the strategy translates into a deliberate bet: plant a flag on every surface where users might interact with AI, then improve the products from inside the ecosystem. The bet assumes that agents will reward early occupancy the way messaging once did. Three months of execution data offer a first evaluation. The early evidence is mixed.

Tencent’s agent proliferation, model rebuild, and early user feedback reveal a set of tensions that may determine whether the company’s distribution advantage translates to AI-era market share. Below, I trace the evidence across each layer of the strategy.

Read more

The Mirror Filing

In March, I dissected Unitree Robotics’ prospectus and found a number that reframed the entire humanoid robot sector: industrial applications accounted for just 9% of humanoid revenue. The world’s largest humanoid shipper by volume was, at its core, a research hardware company selling to labs, developers, and AI startups.

On May 18, the Shanghai Stock Exchange accepted a second IPO application from a Hangzhou robotics company in the “embodied intelligence” sector, the Chinese industry term for AI systems housed in physical robots. DEEP Robotics (云深处) filed to raise RMB 2.5 billion ($345 million) on the STAR board. Its prospectus tells the mirror story. Industry applications generated 79.33% of main business revenue in 2025. Power grids, fire departments, police and security operators, industrial facilities, infrastructure managers, and logistics companies purchased its robots for scheduled deployment. Same city. Same exchange. Same “embodied intelligence” label. A completely different business underneath.

According to Sullivan data cited in the filing, DEEP Robotics ranked first globally in industrial quadruped application revenue in 2025, second in overall quadruped revenue, and fourth in embodied-AI robot revenue.

The contrast between these two filings gives investors a clearer map of China’s embodied-AI industry than product demos or fundraising announcements provide. The sector operates as multiple parallel markets, differentiated by form factor and by who pays and why.

DEEP Robotics’ filing reveals what the robot business looks like when machines perform actual work in defined field environments, where terrain, safety, and reliability matter more than general-purpose intelligence. The economics differ from the research-hardware model in several important dimensions: smaller revenue base, slower growth curve, heavier deployment burden, longer sales cycles, and a value proposition that does not depend on a technological breakthrough that has not yet arrived.

A Different Growth Curve

DEEP Robotics’ financial trajectory looks like a hardtech company crossing from R&D into field revenue after years of investment.

Revenue grew from RMB 50.1 million in 2023 to RMB 337.5 million in 2025, a compound annual growth rate of 159.51%. The company turned profitable for the first time in 2025, reporting net income of RMB 28.68 million after losses of RMB 25.85 million and RMB 13.29 million in the preceding two years. Operating cash flow turned positive at RMB 63.75 million. Gross margins climbed from 33.48% to 52.83% over the same period.

These numbers describe a credible technology commercialization arc. They are not the explosive growth profile that Unitree’s filing revealed: RMB 1.71 billion in 2025 revenue, 60% gross margins, RMB 672 million in operating cash flow. DEEP Robotics generated roughly one-fifth the revenue and one-tenth the operating cash flow of its crosstown competitor.

The margin expansion reflects manufacturing scale: unit production costs for the Jueying X series fell from RMB 230,300 in 2023 to RMB 131,200 in 2025 as procurement volumes grew, while revenue mix shifted toward higher-value models.

The scale gap points toward a structural difference in customer economics, not a difference in execution quality. Research hardware serves a broad, fragmented buyer base purchasing standardized platforms at accessible price points. Unitree sold over 3,500 humanoid units in the first nine months of 2025 at an average price of RMB 167,600. DEEP Robotics sold 681 Jueying X-series industrial quadrupeds at an average price of RMB 287,500 each. Higher prices, fewer buyers, longer sales cycles. The revenue per unit is larger. The path to scale is narrower.

The Safety Premium

The most consequential disclosure in DEEP Robotics’ filing is the breakdown of revenue by downstream application.

In 2025, RMB 267.3 million of main business revenue came from industry applications: power-grid inspection, emergency firefighting, police and security operations, industrial monitoring, infrastructure patrol, and logistics. The company has cumulatively sold more than 300 robots in power inspection, more than 300 in police and security, more than 150 in fire rescue, and approximately 300 in industrial settings. These are production deployments with cumulative volumes that suggest ongoing procurement cycles. The classification is based on the company’s disclosed understanding of customer use cases rather than independently audited end-use tracking.

The value proposition bypasses what I have previously described as the binding constraint on industrial robot economics: the payback arithmetic of displacing low-cost labor. DEEP Robotics’ customers are not calculating whether a robot is more efficient than a human inspector. They are calculating whether a robot can enter environments where human presence carries unacceptable risk.

Utility tunnels with toxic gas accumulation where human entry requires evacuation protocols. Substation bays where high-voltage contact is lethal. These are environments defined not by labor cost but by the price of a single incident. A substation failure or a firefighter casualty creates losses that dwarf the RMB 287,500 average selling price of a Jueying X-series industrial quadruped.

This distinction matters for understanding the sector’s trajectory. DEEP Robotics’ 2025 numbers suggest that safety substitution can already support a profitable robot business within defined industrial domains. The machines do not need general-purpose cognition to walk a predetermined patrol route through a cable tunnel and report thermal anomalies. They need reliable locomotion, environmental hardening, and sensor integration. Its flagship industrial quadrupeds carry IP67 protection ratings and operate across temperature ranges of -20 to 55 degrees Celsius. These are industrial reliability specifications, not research prototype benchmarks.

This analysis continues below for Hello China Tech Premium subscribers. The free section above establishes what DEEP Robotics sells and who buys it. The paid section examines the more important question: why the company’s quadruped, wheeled-legged, and humanoid product lines appear to sit on three different commercialization timelines, and what that says about the future of China’s embodied-AI industry.

Read more

At this week’s Alibaba Cloud Summit, the company launched a new product website called Qianwen Cloud. Open it in a browser and you see nothing recognizable as a cloud service. No catalog. No console entry point. No navigation structure. A single command-line instruction fills the page: npx skills add QianWen-AI/qianwen-ai. That line is readable by an AI agent, meaningless to a casual visitor.

In 17 years of operation, Alibaba Cloud had never launched a standalone product website outside its main cloud site. The fact that its first one is designed to be read by software rather than people encapsulates what happened at this week’s summit: a major cloud provider publicly restructured its entire technology stack around the premise that AI agents, not human developers, will become the dominant consumers of cloud services.

The summit announcements spanned every layer. A new AI chip. A 128-card server cluster. A flagship model that ran autonomously for 35 hours. A platform retooled so every cloud product can be called by agents as a discrete function. The interface change matters because it changes who initiates consumption. A human developer buys capacity in planned blocks. An agent can trigger token, compute, storage, and tool usage repeatedly within a single task, without a human initiating each call. Taken individually, these are product launches. Taken together, they amount to a thesis: for the fastest-growing AI workloads, the unit of cloud revenue is migrating from compute hours to tokens, and the infrastructure must follow.

This thesis already has financial backing. One week before the summit, Alibaba’s March quarter earningsdisclosed that AI-related products now account for 30% of Cloud Intelligence Group’s external revenue, annualizing above Rmb 35.8 billion. Model-and-application-services ARR, which includes API calls on Bailian, Alibaba Cloud’s Model Studio platform, and AI software subscriptions, exceeded Rmb 8 billion. Management expects this model-and-app ARR to surpass Rmb 10 billion in the June quarter and Rmb 30 billion by year-end. Wu told analysts that this revenue stream carries higher gross margins than traditional infrastructure-as-a-service. If the trajectory holds, agent-driven model-and-app revenue appears positioned to overtake Alibaba Cloud’s legacy compute product line as the largest revenue contributor.

That revenue shift points toward a broader structural question: if the fastest-growing workloads on a cloud platform are generated by software rather than people, the assumptions baked into two decades of cloud product design may need revisiting.

In my March analysis, I argued that Alibaba’s most probable failure mode was not any single assumption proving wrong but multiple timelines slipping simultaneously. Last week’s earnings confirmed the company chose growth over near-term profit, burning Rmb 17.3 billion in free cash flow. The summit a week later revealed what that cash is building. Below, I trace the architecture Alibaba is constructing for an agent-first cloud, where the vertical integration could create margin leverage, and what the next two quarters need to prove before investors treat this as more than a product launch.

Read more

Three Chinese semiconductor companies reported financial results in the same week of May 2026. Taken individually, each tells a company-specific story.

SMIC, the country’s largest contract chipmaker, posted $2.5 billion in Q1 revenue and guided Q2 growth of 14 to 16 percent sequentially. Hua Hong Semiconductor, a smaller foundry specializing in power and analog chips, saw net profit attributable to shareholders rise 458.1 percent year-on-year. And CXMT, China’s sole significant DRAM manufacturer, updated its IPO prospectus to reveal Q1 net income attributable to shareholders of Rmb 24.8 billion, enough in a single quarter to exceed the Rmb 23.5 billion it lost cumulatively in 2023 and 2024 combined.

Taken together, these three sets of numbers describe a shared structural driver. The global AI buildout, which concentrates the most advanced manufacturing capacity on GPU and HBM production, is squeezing supply for everything else. That squeeze is redirecting orders toward the geography with the most aggressive mature-node expansion underway. China’s chip sector appears to be benefiting, in part, from the very supply dynamics that trade restrictions helped set in motion.

SMIC’s co-CEO Zhao Haijun used a specific phrase on the company’s earnings call to describe this dynamic: the “AI siphon effect”. The term is worth unpacking. In Zhao’s usage, it describes AI demand pulling capacity and orders away from traditional chip categories, while giving Chinese fabs with mature-node supply more room to raise prices. It captures a transmission mechanism that trade-restriction frameworks tend to miss.

Zhao used the phrase for SMIC’s business specifically. But the same gravitational logic applies more broadly. In foundry, the siphon appears as order reshoring and mature-node tightness. In specialty processes, it appears as rising demand for power, MCU, and analog platforms. In memory, it appears as global DRAM allocation shifting toward data center customers, leaving Chinese domestic demand underserved.

Five channels, one direction

Zhao laid out five pathways through which AI demand translates into orders for SMIC’s fabs at 28nm and larger nodes. They fall into three tiers of causation.

The first tier is direct pull. AI servers require large quantities of supporting chips, particularly power management ICs and data-transmission components, manufactured on mature process nodes. SMIC’s BCD platform, which produces power management chips for AI server boards, is capacity-constrained. At the same time, TSMC, Samsung, and memory IDMs are reallocating fab capacity toward AI-related production, exiting or de-prioritizing niche products like NOR Flash and SLC NAND. Those orphaned product lines seek alternative foundries.

The second tier is displacement. Global customers whose capacity allocation at leading foundries has been reduced are placing orders with Chinese fabs. Domestic chip designers serving AI-adjacent markets (electric vehicles, robotics, edge inference) are ramping volume, pushed by supply uncertainty abroad. A separate channel is policy-driven: localization preferences are bringing domestic logic, networking, and communications customers onto SMIC’s platforms even where the immediate trigger is not AI capacity pressure.

The third tier is behavioral amplification. Customers are building inventory preemptively, concerned that supply constraints will worsen. This channel is the most fragile: it depends on sentiment rather than structural demand, and it can reverse rapidly if supply conditions ease.

The distinction matters for durability. Tier-one demand persists as long as AI infrastructure investment grows. Tier-two demand persists as long as leading foundries prioritize advanced nodes over mature ones. Tier-three demand could unwind in a single quarter if customers perceive a shift in supply balance. Together, they explain why SMIC can guide 14 to 16 percent sequential revenue growth in Q2 while simultaneously raising prices on supply-constrained product categories.

SMIC’s Q1 data aligns with this three-tier reading. China-region revenue share rose to 88.9 percent, from 84.3 percent a year earlier, consistent with order reshoring. Blended average selling price increased 2.5 percent sequentially. Eight-inch wafer revenue grew 6 percent quarter-on-quarter, driven primarily by pricing power rather than volume. The industrial and automotive segment expanded from 9.6 percent of wafer revenue a year ago to 14 percent. The utilization rate fell to 93.1 percent from 95.7 percent the prior quarter, but management attributed the decline partly to new start-up capacity entering the denominator and partly to smartphone order adjustments tied to earlier memory supply concerns. The overall pattern points less to an underutilization problem than to a business reallocating capacity toward higher-demand categories. Management also noted that SMIC is signing long-term agreements with customers to lock in future demand, suggesting an effort to convert displacement orders into stickier relationships.

What connects a foundry raising prices, a specialty fab posting 458.1 percent profit growth, and a memory maker generating more net income in one quarter than its cumulative losses over two years? The analysis below examines how the siphon effect operates at each layer, what the consolidation moves by SMIC and Hua Hong reveal about the sector’s next phase, and where the tailwind meets its limits.

Read more

Three Signals Behind One Metric

On May 13, Robin Li took the stage at Baidu’s annual developer conference and proposed a new unit of measurement for the AI era. He called it DAA: Daily Active Agents. The number of AI agents completing work and delivering results each day.

The framing was deliberate. Token consumption, Li argued, measures input. DAA measures output: how many agents are completing work and delivering results. Li predicted global DAA could eventually exceed 10 billion.

A new metric from a conference keynote would normally register as marketing. Three structural moves underneath it suggest otherwise.

First, Baidu changed how it reports revenue. Starting in late 2025, the company stopped splitting its business into “advertising” and “non-advertising.” The new structure breaks out Baidu Core AI-powered Business, which includes AI cloud infrastructure, AI applications, and AI-native marketing services. That category generated Rmb 40 billion ($5.7 billion) in 2025, growing 48% year over year. This is a company telling investors: judge us as an AI infrastructure firm, not a search advertising firm.

Second, Baidu has repositioned ERNIE from a standalone consumer story toward an infrastructure engine for the agent layer. Wenxin Assistant remains sizable: Baidu reported 202 million monthly active users for ERNIE Assistant in its Q4 2025 earnings. But the company’s strategic narrative at Create 2026 was increasingly centered on agents, cloud, and full-stack AI infrastructure. ERNIE’s role is increasingly to power the agent layer above it, rather than to win a pure consumer-attention battle against ByteDance’s Doubao, Alibaba’s Qwen, or Tencent’s Yuanbao.

Third, the product announcements at Create 2026 filled in what the new game looks like. DuMate (百度搭子), a general-purpose work agent. A code-generation platform (秒哒) that, according to Baidu, has produced over 1 million applications for more than 10 million users. A decision-optimization agent (伐谋) deployed in logistics routing, manufacturing scheduling, and financial risk management. A multi-agent digital human platform (百度一镜) for e-commerce and content production. These are not primarily chat interfaces. They are execution-oriented products, and their workloads can consume far more tokens than ordinary conversational AI.

These three signals encode the same structural judgment: in a market approaching billions of agents, the race to build agents matters less than the race to build what agents run on. Baidu has oriented its entire stack toward that objective: chips (Kunlun P800, with 256-card super-node systems that Baidu says will become available in June), cloud (a redesigned AI cloud optimized for agent workloads), model serving (KV Cache pooling, P/D separation, global inference scheduling), and agent orchestration (sandboxes, long-context management, sub-agent coordination, skill libraries).

The judgment is partly earned, partly forced. Baidu holds assets that compound in an infrastructure game: proprietary chips, a mature cloud business, two decades of search-scale data engineering. It lacks the content ecosystem or transaction network that would give a consumer agent natural distribution. But the same thesis is visible in companies that face no such constraint, which suggests something broader is at work.

Payment Rails, Device OS, National Standards

Baidu’s positioning becomes legible alongside what other Chinese companies are building. Each occupies a different layer of the same stack.

Ant Group, Alibaba’s fintech affiliate, is building the payment rails. Alipay AI Pay surpassed 100 million usersin February 2026 and processed over 120 million transactions in a single week. More structurally, Ant launched AI Collect (支付宝 AI 收) in April, enabling agents built on frameworks like OpenClaw to charge users per-task through Alipay. Ant researchers frame AI payments in three evolutionary levels: Level 1, where AI triggers payment but the user confirms; Level 2, where AI pays autonomously within preset limits; Level 3, full autonomous payment, which remains in standard-setting discussions. Ant’s security teams are also framing the problem around KYA (Know Your Agent): agent identity verification, transaction intent validation, permission boundaries, and audit trails. This is a governance framework under development, not a single shipped product.

This matters because execution without settlement remains a demonstration, not a product. An agent that books a hotel but cannot pay for it has not completed a task. Ant is solving the last step: how agents transact, who authorizes them, and who bears liability when they err. The contrast with the US is not that America lacks payments. It is that China’s most visible agent-commerce experiments are already tying payment to authorization levels, agent identity, and transaction audit trails, making agent payment a new trust layer rather than a new checkout flow.

Xiaomi is building the device-level embedding. Its MiMo-V2-Pro model, with over 1 trillion total parameters and 1-million-token context, powers miclaw, a system-level AI agent integrated into Xiaomi’s HyperOS. Xiaomi began a limited closed beta for miclaw on March 6, with a broader agent ecosystem platform moving into public testing in April. CEO Lei Jun committed at least Rmb 60 billion ($8.7 billion) to AI over 3 years. Xiaomi ships phones, cars, and more than 800 million connected IoT devices, based on prior company disclosures. If system-level agents are eventually embedded across that fleet, they gain native access to device context that app-layer agents cannot obtain. Xiaomi controls the hardware and OS surface where such deployment can happen without permission from any super-app.

As I examined in my analysis of how OpenClaw conquered China, system-embedded agents from device makers may matter more than open-protocol frameworks over the long term. They occupy a position that tools requiring manual deployment structurally cannot reach.

The Chinese government completed the picture on May 8 by issuing two documents simultaneously. Three ministries published implementation guidelines defining agents as “intelligent systems with autonomous perception, memory, decision-making, interaction, and execution capabilities,” identifying 19 application scenarios, and establishing a tiered governance framework. The same day, a separate multi-ministry group released a national standard (GB/Z 177–2026)creating a grading framework for AI terminal intelligence across categories such as phones, PCs, TVs, glasses, vehicle cockpits, speakers, and earbuds.

Few major economies have attempted to standardize both the software entity and its hardware carrier simultaneously at the national level. The US leaves these definitions to market competition. The EU regulates by risk category without defining product form. China chose to build the coordinate system itself, defining what an agent is and grading what it runs on in a single policy gesture.

The pattern across Baidu, Ant, Xiaomi, and the regulatory apparatus answers a “what” question: China is constructing the infrastructure layer for an agent economy. The harder questions follow. Why does China’s approach diverge from America’s? What does “Harness Engineering” mean for competitive positioning? And if China’s agent economy does not monetize through SaaS subscriptions, what does its business model look like?

Read more

The most important number in the H200 story is zero.

The license is large enough to matter. Reuters says selected Chinese buyers, including Alibaba, Tencent, ByteDance and JD.com, can each purchase up to 75,000 H200 chips. Lenovo and Foxconn also received distributor approvals. The commercial result, so far, is still zero shipments.

That is the new power balance. Washington can open a gate. Nvidia can prepare the product. Top Chinese buyers can qualify. Beijing can still decide whether access becomes deployment.

The Beijing trip did not change that. US Trade Representative Jamieson Greer said chip export controls were not a major topic in the latest talks with Chinese officials. Trump said China had not approved H200 delivery because it wanted to develop its own chips. The signal is not a breakthrough. It is a stalemate with clearer rules.

The headline version looks bullish for Nvidia. The investor version is more awkward. China has moved from pure denial toward managed access for one high-end but non-frontier Nvidia chip. The power now sits not only in Washington’s export license, but also in Beijing’s ability to decide when Chinese demand becomes real.

H200 remains a serious product. Nvidia says it carries 141GB of HBM3E memory and 4.8TB/s of memory bandwidth. It is designed for large language model inference and training workloads that still stretch domestic Chinese accelerators. For Alibaba Cloud, Tencent Cloud or ByteDance, H200 remains useful for frontier model work, benchmark-sensitive projects and high-end research.

That is why the zero deliveries matter. Beijing is not hesitating because H200 lacks value. It is hesitating because the chip is valuable enough to recreate dependency.

Nvidia’s China problem is large enough to distort the whole story. The company once held about 95 per cent of China’s advanced AI chip market, and China previously contributed 13 per cent of its revenue. Jensen Huang has put the country’s AI opportunity at $50bn this year. That is why zero shipments carry more information than the approval list.

H200 now passes through 3 filters: US licensing, Chinese permission and corporate deployment. The first decides whether Nvidia may sell. The second decides whether Chinese buyers should buy. The third decides whether the chip enters frontier experiments or long-lived infrastructure. Only the first filter has moved in Nvidia’s favor.

This is demand governance, not merely supply security. Beijing is shaping which foreign chips Chinese companies are allowed to want at scale. Beijing’s recent supply-chain and counter-extraterritorial rules do not name Nvidia or H200. They still give Beijing a broader vocabulary for reviewing critical technology dependence.

Some Chinese AI firms still have workload-level reasons to want H200: frontier training, benchmark races and elite lab work. They also need predictable supply for multi-year inference clusters, regulated data centers and cloud infrastructure. A chip whose shipment depends on Washington, Beijing and Nvidia’s inventory is powerful. It is also a fragile foundation.

Read more

On May 13, Alibaba and Tencent reported quarterly earnings within hours of each other. Both missed revenue estimates. Both saw their US-traded shares rise.

Alibaba posted its first operating loss in five years. Free cash flow swung from a Rmb 3.7 billion inflow a year earlier to a Rmb 17.3 billion outflow. Its US-listed shares gained 8.2%. Tencent reported its slowest revenue growth in six quarters, at 9% year over year. Its US stock climbed nearly 5%.

The market’s willingness to reward deteriorating headline numbers points to a shared question that now dominates both investment cases: when does AI spending convert into AI revenue? Both companies have been spending aggressively. Neither has fully answered the question. But on the same day, they offered strikingly different frameworks for how the conversion will happen.

Alibaba chose to sell first. CEO Eddie Wu disclosed two AI metrics for the first time. AI-related cloud product revenue now annualizes above Rmb 35.8 billion, accounting for 30% of the Cloud Intelligence Group’s external revenue. Separately, the narrower model-and-application-services (MaaS) ARR had already exceeded Rmb 8 billion, is expected to surpass Rmb 10 billion in the June quarter, and is on track to exceed Rmb 30 billion by year-end. Cloud external revenue grew 40%. Wu told analysts that not a single AI card on Alibaba’s servers sits idle. The company is monetizing its compute as fast as it can deploy it.

Tencent chose to build first. Chief Strategy Officer James Mitchell told analysts that the company had “consciously” delayed monetizing AI through Tencent Cloud because it prioritized a “multiplicity of internal services” over external sales. GPU capacity went to Hunyuan model training, advertising optimization, agent development for WeChat, and productivity tools like WorkBuddy and CodeBuddy. Tencent Cloud’s AI agent solutions are already generating initial token monetization, but external cloud has not been the primary allocation priority. The difference is sequencing: which company exposes scarce compute to external revenue first.

Chip supply does not explain everything, but it determines the sequence each company can afford. Both face the same binding constraint: the number of AI chips they can source shapes what they can do, and in what order. Alibaba has proprietary chips from T-Head and an external cloud business already at scale; it chose to maximize external revenue per chip. Tencent does not operate a comparable in-house AI chip business but has more internal surfaces where AI can improve ads, games, and productivity tools; it chose to maximize internal product value per chip before supply loosens in the second half of 2026.

This quarter gave investors more AI economics than either company had previously disclosed. Below, I trace how each path works, where the financial evidence supports management’s framing, and which signals will show whether May 13 marked a real inflection or simply a better story.

Read more

In mid-January 2026, ByteDance promoted its AI study feature, Doubao Aixue (豆包爱学), to a core entry point inside one of China’s most popular chatbot apps. Almost simultaneously, Alibaba’s Qwen chatbot added a function that lets students search and download exam papers from top schools across the country. Within days of each other, two of China’s largest tech companies signaled that Chinese students’ after-school hours were once again worth competing for.

Mobile analytics data from late 2025 shows AI education apps in China crossing 120 million monthly active users, a 340% jump year over year. That number makes AI-powered learning one of the fastest-growing consumer app categories in the country. And the companies flooding in carry very different product logic, cost structures, and paths to monetization. Understanding who wins requires understanding what each side brings to the table.

Two Product Architectures, One Unsolved Problem

The companies racing into this space carry different DNA, and that DNA shapes what their products can and cannot do.

ByteDance and Alibaba entered through their existing chatbot apps. Doubao Aixue and Qwen Zhixue (千问智学) are AI conversation interfaces with an education tab bolted on. A student types or speaks a question; the underlying large language model reasons through it in real time. ByteDance’s version uses voice interaction that mimics a phone call with a patient tutor. The advantage is distribution: neither company needs to acquire education users from scratch. They already sit inside products with enormous reach.

Zuoyebang (作业帮, one of China’s largest homework-help platforms) and Yuanfudao (猿辅导, a former tutoring company that pivoted to AI) started from the opposite end. Their apps are built around structured learning workflows: scan a math problem, match it against a proprietary question bank curated by human teachers, then layer AI explanation on top. Yuanfudao’s version simulates a one-on-one lesson with a virtual blackboard and interactive pauses. Here, AI sits on top of an existing curriculum architecture rather than replacing it.

The gap between these two approaches shows up in a specific place: what happens when the model gets the answer wrong. For a general-purpose LLM, a hallucinated math solution is a known failure mode with no backstop. For an app built on a verified question bank with decade-old pedagogical scaffolding, the structured content layer offers at least a partial guardrail. In an industry where a single wrong answer can erode parental trust, that structural difference matters.

Yet neither camp has found a way to charge for what it offers. The flagship products competing for users today still offer core AI features for free, with some labeling the access as “limited time.” The tech giants subsidize education as a means of driving engagement across their platforms. The education companies absorb the cost to retain users they fear losing to those platforms. For now, monetization is a problem both camps are willing to postpone.

The traffic-first camp treats education as a retention feature. The content-first camp treats it as a product line. Which model monetizes first will likely depend on which side can demonstrate measurable learning outcomes to paying parents.

Why the Demand Is Structural

The user growth is not accidental. China’s household education spending is structurally inelastic. A 2023 Stanford study found that families spend an average of 17.1% of household income on education. Among the poorest quartile, that figure reaches 56.8%. Education spending in China functions as a necessity across income levels, not a discretionary expense.

The 2021 regulatory reset of the for-profit tutoring sector removed the most visible commercial layer serving that demand. For 4 years, supplementary learning lacked a scalable delivery mechanism. Large language models provided one. AI study tools do not employ human teachers. They do not operate physical classrooms. They sit inside apps that families already use. And because they deliver supplementary learning through software rather than human instruction, they occupy a different product category from the classroom tutoring businesses that were restructured in 2021.

What both platform giants and education incumbents have already accomplished is a structural fact: supplementary learning is back inside the daily routines of Chinese families, delivered through mobile apps at near-zero marginal cost. The platform companies compete for traffic and engagement. The education companies compete on content depth and parental trust. Together, they have rebuilt the delivery layer for after-school learning in a new form.

Beijing Moves to Shape the Market

Beijing is not standing aside. On April 10, 2026, the Ministry of Education released a five-ministry AI+Education Action Plan (人工智能+教育行动计划) extending through 2030. It calls for AI literacy courses at every educational level, mandates the development of “intelligent learning companions” (智能学伴) for students, and envisions centralized data systems for student performance and resource allocation. It is the clearest national framework yet for AI’s role in education.

For companies in this space, the plan creates both opportunity and uncertainty. On the opportunity side, it opens a government procurement channel for AI teaching tools. On the uncertainty side, it introduces compliance and standards requirements that have not yet been specified. The plan is heavy on infrastructure ambition: compute platforms, model infrastructure, pilot bases, and future classrooms. What it leaves less defined is who absorbs the recurring cost of adoption inside individual schools: devices, software, maintenance, training, and subscriptions.

That matters because Chinese education policy has long faced a familiar implementation risk, one that officials themselves describe as “emphasizing construction while neglecting application” (重建设、轻应用). A 2024 review found a shortfall of 8.5 million computers against the government’s own baseline of 1 device per 15 students. Commercial classroom AI systems add another layer of cost: a 2022 procurement record shows iFlytek (科大讯飞), one of China’s largest education technology providers, charging a single school roughly Rmb 1.74 million for its smart-classroom software suite.

For AI education startups and platform companies alike, the B2G (business-to-government) channel carries real revenue potential. The question is timeline and predictability. Until the Action Plan’s standards and procurement rules take concrete shape, the near-term paying market remains concentrated in urban households that already have devices and the willingness to pay.

What to Watch

The products competing for students today optimize for one thing above all: how quickly they can deliver an answer. Speed is the selling point. Every major AI tutoring app now offers instant step-by-step solutions. Few offer a reason to pause and think before requesting one. Whether that trade-off matters for learning outcomes is an open pedagogical question. Whether it matters for monetization is a more immediate business question: parents will pay for tools that visibly improve test scores. The link between AI-assisted homework and exam performance has not yet been demonstrated at scale.

More than 120 million monthly active users are already using AI education apps. The companies serving them have yet to find a sustainable business model. The next phase of competition will not be about who acquires users fastest. It will be about who converts free usage into revenue before the subsidy window closes.

User growth is already here. Whether AI changes learning outcomes, or simply speeds up the race, is a much harder question.

On May 6, 2026, Samsung Electronics announced it would stop selling televisions, monitors and home appliances in mainland China. The list covers refrigerators, washing machines, air conditioners, audio products, projectors, vacuum cleaners and air purifiers. Production at its Suzhou home-appliance factory is not affected. Smartphones and chips remain on sale in China. But after 34 years, Samsung’s TV, display and home-appliance presence on mainland Chinese shelves is effectively over.

Samsung says existing customers will continue to receive after-sales support under Chinese consumer-protection rules, and the company is reviewing support measures for business partners.

The same day, Samsung’s stock surged 13%. Its market capitalization crossed $1 trillion, making it the second Asian company after TSMC to reach that threshold. The semiconductor division had just reported quarterly operating profit equivalent to roughly 94% of the entire group’s earnings.

Investors did not treat this as a defeat. They treated it as overdue housekeeping.

The juxtaposition is revealing. A company can lose a major consumer market and still become more valuable when investors believe that market no longer matters to its profit base. That paradox points to a deeper structural shift in who captures value in the electronics supply chain, and what “winning” means at different layers of the stack.

Samsung’s TV and home-appliance division recorded its first-ever operating loss in 2025. WSJ put the deficit at 200 billion won, or about $138 million. In China specifically, the numbers had long been worse. Market research firm AVC Revo put Samsung’s offline share at 3.62% for color TVs, 0.41% for refrigerators and 0.38% for washing machines as of early April. Separately, full-year 2025 AVC data cited by Jiemian showed its online TV share at just 0.67%.

The last meaningful presence was a niche following for Samsung’s “The Frame” lifestyle television on social media. The exit even triggered a final burst of demand. Jiemian reported that some stores received a wave of customer inquiries, while JD.com said Samsung TV transaction value rose more than 200% year over year after the announcement. But that was residual brand equity, not a national business. Samsung was no longer a mass-market player. It had a following too small to sustain a national sales and service operation.

From Peak to Footnote in a Decade

Samsung entered China’s appliance market in 1992. Its broader China business peaked in 2013, when total China revenue hit $25.7 billion across categories including smartphones and large-screen TVs, with market-leading positions in both.

The competitive story was structural, but Samsung also weakened its own position through a series of local missteps. Three moments help explain the decline.

The exit was not abrupt. Jiemian reported that Samsung had largely stopped supplying the mainland market after the Lunar New Year holiday. One Samsung store employee estimated total remaining inventory at roughly 30,000 units.

The first was the Galaxy Note 7 battery crisis in September 2016. Samsung initially excluded China from its global recall, publicly claiming Chinese units were safe. When Chinese units also caught fire, the company’s local team delayed acknowledgment until regulators intervened. The episode attached a “double standard” label to Samsung’s brand in China. The trust damage bled from smartphones into appliances and televisions. Smartphone revenue in China dropped 38% that year and never recovered.

The second was a series of factory closures that hollowed out Samsung’s local manufacturing base. The Shenzhen and Tianjin phone factories closed in 2018. The last Chinese phone production line, in Huizhou, shut in 2019. The sole remaining television factory in Tianjin ceased operations in 2020. Each closure removed a layer of local supply chain integration.

The third was a failed outsourcing experiment. In 2021, Samsung handed its white-goods business, including refrigerators, washing machines and air conditioners, to two Chinese distributors, one focused on premium retail and the other on small-town shops. Within two years, the arrangement collapsed. The premium distributor lacked volume to sustain a meaningful retail presence; the small-town partner sold Samsung hardware through channels that undercut its positioning. Pricing discipline broke down and after-sales service became disorganized. By the time Samsung reclaimed direct sales in 2024, the retail infrastructure it needed had already consolidated around domestic brands.

The later failures, especially channel outsourcing and product adaptation, point to the same organizational constraint. Samsung’s China operation functioned as an execution arm, not a decision-making unit. Product definition, pricing authority, and channel strategy all required sign-off from Seoul. This architecture might work in stable markets where global product-market fit translates across geographies. China’s appliance market after 2016 was not stable. Competitors iterated monthly. Samsung’s approval cycles ran quarterly to annually. The mismatch compounded over time, and by the time it showed up in market-share data, recovery had become unlikely.

Category-level data sharpen the picture. The appliance business peaked slightly later than Samsung’s overall China revenue. Samsung sold roughly $3 billion of TVs and $1 billion of home appliances annually in China during 2014–2015. By its 2026 operating plan, TV sales targets had fallen to about 5% of that peak. White-goods targets stood at about 1%.

Samsung’s exit completes a pattern that began with Japanese brands a decade earlier. But the more interesting question is not why foreign appliance makers lost China. It is what Samsung’s financial results reveal about where value has migrated in the electronics supply chain, and why the market rewarded a retreat. The answer involves three mechanisms: panel self-sufficiency, Mini LED cost compression, and memory pricing power.

Read more

In the span of a single week this April, three figures from three different corners of the entertainment industry arrived at essentially the same conclusion.

Cristóbal Valenzuela, CEO of AI video startup Runway, told an audience at the Semafor World Economy summit that studios should take $100 million and spread it across 50 films instead of spending it on one. “Same quality. Same amount of output, visually. But you make way more content,” he said. “It’s a quantity problem.”

In Cannes, Mathieu Kassovitz, the French director behind the 1995 film La Haine, predicted that “in two years from now nobody will care” whether characters on screen are human or generated by AI. He announced plans for an AI film studio in Paris and noted that AI had cut his anticipated visual effects budget from $50 to $60 million down to $25 million. Kassovitz was speaking at the World AI Film Festival, a separate event from Cannes’ official competition, which has signalled strong resistance to AI-generated films.

And in Beijing, Gong Yu, the founder and CEO of iQIYI, often described as China’s Netflix, laid out what he called the “112 law”: AI would reduce unit content costs by one order of magnitude, increase the number of creators by one order of magnitude, and expand total output by two orders of magnitude. He unveiled Nadou Pro, an AI production toolkit designed to cover nearly every stage of filmmaking from scriptwriting to final rendering, and declared the shift “once in a decade.”

Three continents. Three distinct markets. One shared thesis: AI compresses production costs, which means more content can be produced, which means the odds of producing something great improve through sheer volume.

The logic is intuitive. It has also already been tested. Short dramas are not feature films. But China’s AI short drama market offers the clearest industrial-scale test so far of the logic now spreading across film and television: lower costs should lead to more experiments, and more experiments should improve the odds of producing something people care about.

For readers unfamiliar with the format: AI short dramas are serialised, mobile-first shows with episodes lasting 2 to 5 minutes, distributed through algorithm-driven platforms and monetised through advertising and micropayments. Think of them as soap operas rebuilt for smartphone attention spans. In January 2026, Chinese platforms launched more than 14,600 new AI-generated titles. That works out to roughly 470 per day. By February, 127,800 were in active circulation.

The production economics behind those numbers have shifted just as dramatically. A live-action short drama that cost upward of Rmb 1 million ($137,000) to produce in 2024 can now be generated for Rmb 50,000 to 100,000 ($7,000 to $14,000) using AI tools. At the cheapest tier, outsourced workshops quote Rmb 30,000 to 40,000 per complete series. Entry barriers have compressed dramatically: a single operator working with current video generation tools can produce 40 minutes of distribution-ready content per day.

This is, by any measure, the quantity thesis in action. The question is whether it has produced the results that Valenzuela, Kassovitz, and Gong are promising.

More Shots, Fewer Goals

The quantity thesis rests on a probabilistic assumption: if a hit is partly a matter of chance, then producing more content improves the absolute number of hits. More shots on goal, more goals.

China’s AI drama market offers the clearest available test of this logic. Of the 127,800 AI-generated dramas in circulation by February 2026, the proportion that crossed the 100-million-view threshold stood at 0.117%. In 2025, Douyin’s ecosystem launched 60,000 AI dramas. 96 reached the benchmark. And the breakout rate has been declining as production volume climbs.

The performance ceiling tells a parallel story. The highest-viewed AI drama in circulation has reached approximately 1 billion views. The most successful live-action short drama accumulated 4.4 billion. AI content accounted for roughly 30% of total short drama viewership during the 2026 Spring Festival holiday. Yet that share was carried by sheer volume rather than audience loyalty. No AI-generated title has yet entered the broader cultural conversation in the way that popular live-action hits routinely do, generating fan communities, launching performer careers, or driving merchandise.

The quantity thesis assumes that hit probability is independent of production method. China’s market, the largest test case available, suggests otherwise. What the data reveals about audience behaviour, structural dynamics, and the real bottleneck in entertainment carries implications well beyond short dramas and well beyond China.

Read more

Editor’s Note: FlashPoint is our premium quick-strike column on market-moving China tech events. Today: DeepSeek’s talks with China’s Big Fund, and what a possible $45 billion valuation is really pricing. In April, I argued that DeepSeek’s $300 million fundraising story was less about cash than about talent, options, and industrial scale. Weeks later, the buyer changed. So did the meaning of the round.

Earlier reporting placed DeepSeek in talks to raise at least $300 million at a valuation above $10 billion. At that stage, according to The Financial Times, Liang Wenfeng initially wanted to raise a relatively nominal sum to assign a market value to employee stock options, partly to counter competitors poaching researchers with packages 2 to 3 times their current pay. That effort addressed a human-resources problem with a financial instrument.

A potential Big Fund-led round would address something else. The China Integrated Circuit Industry Investment Fund has not publicly backed any other Chinese large-language-model company. Its mandate covers semiconductor equipment and materials. Its portfolio reads like a supply chain map of China’s chip-independence ambitions: SMIC for fabrication, YMTC for memory, and the surrounding equipment ecosystem. Tencent is also in talks for a stake, and the final investor line-up has not been finalized. But if the Big Fund leads, DeepSeek’s entry into that orbit would suggest Beijing now places frontier AI models in the same category as chip foundries: national infrastructure requiring assured domestic supply.

For investors applying venture frameworks to Chinese AI labs, the shift creates a valuation problem. Moonshot AI, the company behind Kimi, can still be read through a more familiar pricing logic: revenue growth, paid usage, and a possible public-market exit. DeepSeek cannot. For subscribers, the question is what kind of asset DeepSeek is now being priced as: a startup with a path to revenue, a chip-stack coordination node, or a piece of national infrastructure.

Read more

In January 2026, OpenAI began rolling out ChatGPT Go worldwide. The $8 per month tier, first tested in India in August 2025, expanded access at a lower price point while OpenAI prepared to test advertising in the free and Go tiers. According to projections reported by The Information, OpenAI reportedly expects Go to reach 112 million subscribers, even as its $20 per month Plus base contracts from 44 million to 9 million. The direction is clear: downmarket.

Four months later, on May 4, ByteDance’s Doubao moved the other way. After nearly two years of free operation, China’s largest AI-native consumer app updated its App Store listing with a subscription-service notice, disclosing tested tiers at RMB 68, RMB 200, and RMB 500 per month. The direction is equally clear: upmarket.

ByteDance stressed that free Doubao will remain. The company described the paid tiers as value-added services still under testing, with full details to come through official channels at formal launch. The target is not basic chat. Reported paid features center on compute-heavy productivity tasks: PPT generation, data analysis, and video production. Daily conversation, copywriting, and information lookup remain free.

Both companies face the same structural pressure: inference costs that scale with usage, and free tiers that bleed cash at volume. Their opposite responses reflect different market maturities, different user psychology, and different assumptions about where the money sits.

The thesis: the world’s two largest AI consumer markets are running parallel, opposing experiments on the same question. The results will shape whether consumer AI monetizes through subscriptions, advertising, ecosystem commerce, or some hybrid that has not yet stabilized.

Why Doubao Is China’s Pricing Referendum

Doubao holds 345 million monthly active users as of March 2026, more than the second and third-place Chinese AI apps combined. During the 2026 Lunar New Year, its daily active users reportedly peaked near 145 million. Its daily token consumption reached 120 trillion, a 1,000x increase from its May 2024 launch. Few other consumer AI products operate at comparable token throughput. ChatGPT and Gemini are the most visible.

This scale makes Doubao’s pricing decision an industry referendum rather than a single company’s business choice. China’s other major AI apps have been watching. Baidu’s Ernie Bot went fully free in April 2025 after more than a year of paid membership experiments, a reminder that Chinese consumer AI subscriptions had not yet become a settled model. Alibaba’s Qwen and Tencent’s Yuanbao still compete on free access and subsidies. If Doubao, with the largest user base by a wide margin, cannot convert a meaningful fraction to paid subscribers, the implication extends beyond ByteDance. It would suggest that standalone AI subscriptions face structural resistance in the Chinese consumer market.

ByteDance’s own financial position sharpens the stakes. The company’s 2025 operating profit margin came under pressure. An executive acknowledged that slowing e-commerce growth and rising AI infrastructure investment weighed on second-half margins, though he disputed earlier reports characterizing the decline as severe. Doubao’s token consumption doubled in three months between January and March 2026. ByteDance is reportedly investing at an intensity that approaches North American cloud providers, while operating under tighter chip constraints. That configuration likely makes inference economics harder, though the exact cost gap with OpenAI is not public.

The pricing itself carries a signal. Doubao’s RMB 68 entry point sits above the RMB 48.3 monthly averagethat surveyed Chinese users say they would accept for AI subscriptions, according to a Citi Innovation Lab study of 1,800 respondents in March 2026. ByteDance appears to be pricing for value rather than volume, testing whether a smaller cohort of heavy users will pay a premium rather than chasing mass conversion at a lower price.

The intensity of the Chinese public reaction is not purely about price. It reflects a long consumer memory of internet products that launched free, degraded the free experience over time, and nudged users toward paid upgrades. The immediate question among Chinese users was not whether the paid tier is worth RMB 68. It was whether the free tier will quietly deteriorate.

Doubao and OpenAI are running opposite experiments on the same question: who pays for consumer AI, how much, and what happens when subscriptions are not enough? The structural forces shaping their answers, and what the results suggest for every company seeking a sustainable AI business model, are where the real analysis begins.

Read more