Cambricon Technologies, China’s most visible listed domestic AI-chip company, reported first-quarter revenue of Rmb2.88bn, up 160 per cent year on year. Net profit reached Rmb1.01bn, a 185 per cent increase. Operating cash flow came in at Rmb834mn, compared with a negative Rmb1.4bn a year earlier.

The scale of the acceleration is worth pausing on. That single quarter of revenue roughly matches what Cambricon booked in the entire first half of 2025. The stock responded accordingly, hitting its daily price limit at Rmb1,700 on April 30, with turnover exceeding Rmb26bn in a single session. Market capitalisation crossed Rmb710bn. On a 2025 earnings base, the stock was trading at roughly 348 times earnings. Even after incorporating Q1, the valuation still implies a multiple above 250 times trailing profit.

Eight months ago, the central question around Cambricon was whether it could make money. That question has been answered. The company turned its first-ever annual profit in 2025 and has kept the momentum going, with Q1 net income more than doubling the Rmb455mn earned in Q4 2025. Gross margins held above 54 per cent. Net profit excluding non-recurring items grew even faster, at 239 per cent, suggesting recurring operations are driving an increasing share of the profit.

At a 2025-earnings multiple of that scale, investors are no longer paying for current profitability. They are pricing an assumption about duration: that China’s AI compute buildout will run long enough, and that Cambricon will capture enough of it, to justify the premium.

Read more

In April 2026, Tencent Research Institute published findings from a survey of 2,903 Chinese internet users aged 18 to 40. One data point stood out above all others.

When respondents were asked who they would turn to with thoughts they found too shameful or painful to say aloud, 56% chose AI. The share who chose another human being was 14.4%.

This is not a question about convenience or productivity. The survey isolated one of the most emotionally vulnerable moments in a person’s social life and suggests that many users now prefer a machine to another person in that moment. The ratio was nearly four to one.

The finding arrived two weeks after five Chinese government agencies jointly published the Interim Measures for the Management of Anthropomorphic AI Interactive Services. The regulation, effective July 15, 2026, is one of the earliest dedicated national frameworks China has published for this category. It does not merely regulate content or data. It attempts to set terms for how a machine may occupy the role of confidant, companion, and emotional anchor.

The usual framing for Chinese tech regulation emphasizes state preemption: Beijing sees a risk on the horizon and moves to contain it before it fully materializes. That framing does not entirely fit here. In the surveyed population, the behavioral shift preceded the rule. What the regulation governs is a relationship that users had already begun to form.

From Trial to Routine in Under Three Years

The Tencent survey measured more than a single preference. It mapped the broader architecture of AI’s integration into Chinese youth social life.

Among respondents, 94.4% had heard of at least one AI social product. Of those aware, 98.8% had tried at least one. The average usage rate across the four product categories surveyed was 50%. The near-total conversion from awareness to trial suggests that AI social tools have moved from novelty into mainstream exposure with unusual speed.

Usage frequency told a parallel story. 15% of respondents reported interacting with AI socially every single day. Nearly 40% used these services at least once per week. For a substantial portion of the sample, AI social products had crossed from novelty into routine.

The emotional dimension was equally striking. 78.4% of users reported at least one moment of feeling that “AI understands me.” Among those, 23.4% described the feeling as frequent or habitual. Only 21.5% maintained a purely instrumental relationship with these tools. The rest had crossed into territory where the interaction carried emotional weight.

A separate survey published in March 2026 by the China Youth Research Center extended the picture downward in age. Researchers surveyed 8,563 primary and secondary school students across 7 provinces. Over 60% had used AI. Among those users, 35.2% reported chatting with AI or confiding personal feelings to it. And 20.5% said they preferred relying on AI to think rather than thinking independently. Rural students showed higher rates of dependency than urban ones, suggesting meaningful regional differences in how AI is being used and relied upon.

Read together, the two surveys suggest that emotionally meaningful AI use now spans a very wide age range. The shift is not hypothetical. It has measurable depth and width across both youth and adult populations.

ARK Invest has projected that the global AI companionship market could reach $70 billion in its base case and $150 billion in its bull case by 2030, up from roughly $30 million in annualized revenue in 2023. The exact trajectory remains uncertain, but the scale implied by that forecast helps explain why regulators are moving early.

The data confirms that the shift has happened. But what exactly did people shift toward, and what does the new regulation reveal about how Beijing reads the risks? The answers lie in the structure of the law itself, in what it chose to prohibit, what it chose to permit, and the enforcement gaps it left open.

Read more

Since GitHub Copilot launched in mid-2022, subscribers have paid a flat monthly fee and used AI coding assistance without thinking about what each interaction cost. That arrangement ends on June 1.

GitHub announced this week that every Copilot plan will transition to usage-based billing. A new virtual currency called GitHub AI Credits replaces premium request units. Each credit is worth $0.01. Copilot Pro at $10 per month includes $10 in credits. Pro+ at $39 includes $39. How quickly those credits deplete depends on which model a developer selects and how large each session’s context grows.

The subscription prices look unchanged. What they buy has not. Under the old model, a quick chat question and an agent session running for hours could cost a user the same amount. Under the new model, the gap between those two interactions shows up on the bill. Annual subscribers face a sharper adjustment: cost multipliers for premium models jump on June 1. Opus 4.7, currently at a 7.5x multiplier per request, rises to 27x.

Code completions and Next Edit Suggestions remain included at no credit cost, which matters for developers whose primary use stays within the editor. But anyone running Copilot’s chat, agent mode, or code review features now faces a visible price tag tied to the model and context each session consumes. To ease the transition, GitHub is launching a preview bill in early May that projects costs before the new system takes effect. Business and Enterprise customers receive promotional credit pools for June through August, with usage pooled across the organization rather than locked to individual seats.

GitHub’s chief product officer framed the shift in structural terms. Copilot “has evolved from an in-editor assistant into an agentic platform.” The premium request, a billing unit built for the chat era, could no longer represent the cost of what Copilot had become. In effect, GitHub acknowledged that its AI coding tool was not a software subscription. It was a compute service with a subscription wrapper, and the wrapper had to go.

When Agents Broke the Subscription

GitHub reached this point alongside most of the industry.

Anthropic cut off third-party agent frameworks from routing API calls through its consumer subscriptions earlier this month. OpenAI debuted a $100 subscription tier to boost usage of its Codex model. Cursor moved from fixed request quotas to a credit pool weighted by model and task complexity. Google tightened usage policies for Gemini CLI. Each arrived at its own version of the same conclusion: flat-rate pricing absorbs light, predictable workloads well enough. It falters with agent-scale computation, where a single task can spawn sub-agents, traverse entire repositories, and consume millions of tokens over hours.

The OpenClaw surge made the hidden subsidy harder to ignore. When the open-source agent framework went viral earlier this year, developers began running continuous coding sessions through subscription plans designed for intermittent use. The usage distribution that made flat rates viable shifted hard toward the heavy end. The Register’s Thomas Claburn compared the situation to Red Lobster’s Endless Shrimp promotion, which famously helped push the restaurant chain into bankruptcy.

The structural pressure runs deeper. Gartner estimates that cumulative capital investment in AI data centers will approach $6.3 trillion between 2024 and 2029. To clear even a minimal return on that build-out, model providers would need to generate close to $7 trillion in cumulative AI-driven revenue across the same period. Room for subsidized inference is narrowing, and the companies absorbing the heaviest losses from agent workloads are moving first.

GitHub’s response: expose the token cost and let users manage it.

Three Layers, One Market

Over the same month, Chinese AI platforms confronted the same pressure from agent workloads. Their responses look different from GitHub’s single billing switch, and from each other. But the split is telling: application platforms are tightening access, cloud providers are repackaging usage into consumer-friendly bundles, telecom operators are distributing token packages through mobile channels, and infrastructure players such as DeepSeek are still cutting prices to pull workloads toward their stack.

The first layer is a clampdown. Alibaba Cloud stopped accepting new sign-ups for its entry-level coding plan in March and ended renewals in April, leaving only a ¥200-per-month (about $28) Pro plan available through a daily limited-quantity drop at 9:30 AM. According to developer reports, slots typically vanish within minutes. Zhipu, the company behind the GLM model family, is shutting down auto-renewal for its older coding subscriptions without weekly caps on April 30. Replacement plans introduce tighter periodic quotas. The logic parallels GitHub’s: low-price tiers cannot absorb the inference demands that agent frameworks generate.

The second layer is packaging. Where GitHub chose to surface token costs directly, several Chinese platforms chose to embed usage-based economics inside a consumer-friendly shell. Tencent Cloud’s Hy Token Planis the clearest example. Explicitly positioned for agent workloads, it packages access to Tencent’s latest models with tiered monthly credit pools compatible with OpenClaw, Claude Code, Cursor, and other popular frameworks. Users pick Lite, Standard, Pro, or Max. They do not see a per-token invoice. Alibaba introduced a parallel team-tier Token Plan that unifies text and image model consumption under a single credits system.Moonshot folded Kimi Code into its broader membership system alongside document generation, slides, and agent capabilities, starting at ¥39 per month (about $5), with no token pricing visible to the user. China Mobile’s provincial branches in Jiangsu, Beijing, and Guangdong began offering AI token packages this month, with one regional plan providing 20 million tokens for coding use and another starting at ¥5.99 (under $1) per session pack. Tokens are migrating from developer platforms toward telecom distribution channels, extending a pattern already visible in China’s broader token economy. The underlying math across these products resembles GitHub’s AI Credits. The user-facing design looks closer to a mobile phone plan: pick a tier, use until the quota resets.

The third layer runs in the opposite direction. While application-layer platforms tightened quotas or raised effective prices, DeepSeek cut. On April 26, it reduced cache-hit input pricing across its entire model lineup to one-tenth of launch rates and extended a 75% discount on V4-Pro through May 31. Lower per-token costs make long-running, cache-heavy agent sessions cheaper to operate, pulling developer workloads toward DeepSeek’s infrastructure regardless of which frontend tool sits on top. China is not moving in one direction. Platforms that sell to developers are tightening. The infrastructure layer underneath them is still competing on price.

The Meter vs. The Bundle

Beneath the surface-level packaging difference lies a structural question: who absorbs the cost uncertainty that agent workloads create?

GitHub’s model shifts that uncertainty onto the developer. A $10 monthly credit pool buys a variable amount of work depending on model choice, context length, and how many turns an agent needs to complete a task. Developers noticed quickly. “Your ‘usage-based billing’ will make it significantly harder for you to predict your usage,” one commenter wrote in the community thread. Another asked a question that several echoed: “If I have to pay per token, where’s the advantage compared to using the API of my favorite models directly?” Under this framing, Copilot’s remaining competitive edge compresses to a few specific areas: the VS Code integration, the GitHub workflow features, and the frictionless procurement path for enterprise teams already operating inside the Microsoft ecosystem.

The bet GitHub is placing: visible costs will push both users and framework developers to reduce wasted tokens, improving efficiency across the ecosystem over time.

Chinese platforms have largely chosen to keep that uncertainty off the invoice and move it into product constraints. Users encounter monthly caps, five-hour rate windows, and weekly quotas rather than per-token bills. Users feel less exposed to price risk. But cost pressure still has to go somewhere. In practice, it tends to surface as slower inference during peak hours, lower-quality model outputs under heavy load, and hard quota resets rather than as a surprising invoice. The wrapper changes how the friction feels. The underlying economics remain.

GitHub also quietly altered the value exchange in another direction during the same week. On April 24, it updated its data policy: interaction data from Free, Pro, and Pro+ users, including inputs, outputs, and code context, will feed GitHub’s model training by default unless users opt out. Business and Enterprise customers are not affected. For individual subscribers, this adds a separate layer to the transaction. They pay for tokens. Unless they opt out, their interaction data can also improve the models behind those tokens.

What Neither Path Has Proven

GitHub and Chinese platforms agree on the diagnosis. Agent workloads make AI coding subscriptions economically unstable. Per-session costs are high, variable, and growing. The subsidy that made flat-rate plans viable for chat-era usage cannot stretch to cover agentic computation.

They disagree on the response. GitHub bet on transparency, exposing token costs and trusting market pressure to drive efficiency. Chinese platforms broadly bet on packaging, translating cost variance into familiar consumer formats and managing the strain through rate limits and quota structures rather than user-facing bills. Agentic coding is turning software subscriptions into something closer to compute markets, but China is trying to preserve the consumer psychology of subscriptions even as the underlying economics shift.

Neither has demonstrated what both sides need: a billing model that makes agentic coding sustainable for the platform while remaining affordable and predictable enough that developers use it daily without hesitation.

China’s experiment may be the more revealing one. Cloud providers, model companies, telecom operators, and consumer app platforms are running distinct pricing strategies simultaneously, while DeepSeek’s aggressive discounting applies competitive pressure from the infrastructure layer. The pricing question GitHub just resolved with a single billing switch, Chinese platforms are answering in at least five different ways at once. How those experiments play out over the coming months may tell us more about the long-term economics of AI coding tools than GitHub’s billing switch alone.

When I wrote about Moore Threads in September 2025, the IPO was the signal. China was using capital markets to buy time for a domestic GPU champion.

That was the point of the listing. The STAR Market gave Moore Threads public capital, legitimacy, and a policy-aligned financing channel at a moment when US export controls were forcing Chinese customers to take domestic AI chips more seriously.

Now the question has changed.

Moore Threads is no longer being judged only by prospectus language, order pipelines, or the symbolism of becoming China’s first listed domestic GPU champion. Its latest annual report and first-quarter results move the company into a more concrete test: whether a state-backed capital-market story is beginning to turn into a real chip business.

The headline numbers are strong. Moore Threads reported 2025 revenue of RMB 1.51 billion, up 243.37% from 2024. Its first-quarter 2026 revenue reached RMB 738 million, up 155.35% year-on-year. The company also reported a small quarterly net profit attributable to shareholders, at RMB 29.36 million.

That is the new story. The old question was whether China could fund a domestic GPU challenger. The new question is whether Moore Threads can convert substitution demand into repeatable chip economics.

From IPO signal to revenue proof

The most important number in the new disclosure is revenue.

Moore Threads generated RMB 1.51 billion in 2025 revenue, compared with RMB 438 million in 2024. Gross profit reached RMB 987 million, with a reported gross margin of 65.57%.

This matters because the September IPO story still carried a large conversion risk. The prospectus and market narrative pointed to a large order pipeline, but many of those orders still needed to move from intention, testing, or phased evaluation into actual revenue. The old article identified this as one of the central execution risks.

The new report suggests that part of that pipeline is now entering the income statement.

The revenue mix also points in the right strategic direction. In 2025, cloud products contributed RMB 1.46 billion of revenue, while edge and terminal products contributed only RMB 25.51 million. Moore Threads is no longer mainly a consumer GPU or graphics-card story. It is increasingly an AI compute infrastructure story.

That is important for investors and strategists. A Chinese domestic GPU company cannot justify a sovereignty premium through gaming cards. It has to become relevant to AI training, inference, cloud infrastructure, and large-scale computing clusters.

The March 2026 RMB 660 million KUAE intelligent-computing cluster order reinforces that direction. It suggests Moore Threads is selling larger systems, not only discrete chips or cards.

There is still a limit to what the revenue proves. RMB 1.51 billion is meaningful for a young Chinese GPU company. It is still small relative to the valuation now attached to the company and small relative to the infrastructure ambition implied by the Nvidia comparison.

Revenue proof has arrived. Platform proof has not.

The rest of this analysis covers the profit quality behind the headline, what a 200x revenue multiple actually prices, and where Moore Threads stands in building a software ecosystem that matters.

Read more

Most model releases are written for developers, customers, and benchmark watchers. DeepSeek’s V4 report has another reader in mind: investors pricing China’s AI hardware gap.

The surface story is straightforward. DeepSeek has released a new model family with a one-million-token context window. In its V4 technical report, the company says V4-Pro uses only 27% of the single-token inference FLOPs and 10% of the KV cache of DeepSeek-V3.2 in a one-million-token context setting. V4-Flash goes further, at 10% of the FLOPs and 7% of the KV cache.

For investors following China AI, the point is not simply whether DeepSeek has another strong model. It is whether Chinese AI labs can extract more economic value from limited compute.

In China, compute is more than an expense line. It is a strategic constraint shaped by export controls, chip supply, cloud capacity, domestic hardware readiness, and inference economics. A million-token context window sounds like a product feature. For a model company, it is also a memory, bandwidth, latency, and power problem.

If the cost curve remains steep, long-context agents remain impressive demos. If the cost curve bends down, they become usable infrastructure.

This is why DeepSeek-V4 should be read as a compute-shaping document, not only as a benchmark report. DeepSeek spends significant space on fused kernels, expert parallelism, KV cache management, FP4 quantization, and software-hardware co-design. One section even offers proposals to hardware vendors.

That is the real signal. DeepSeek is turning compute scarcity into a set of design specifications.

Read more

Six months ago, Zhipu AI still modeled its strategy on OpenAI, building general-purpose models across every modality. Today it calls itself the Chinese Anthropic, focused almost entirely on coding tools. MiniMax spent its first three years deliberately avoiding enterprise customers, convinced that China’s B2B market would trap it in low-margin customization. It recently hired a former Huawei Cloud executive to lead an enterprise push. DeepSeek, the company that defined itself by refusing outside money, is in discussions to raise at least $300 million. Moonshot AI, which launched as a consumer chatbot, just released a model designed to coordinate up to a thousand sub-agents in parallel.

These shifts did not happen in sequence. They happened within the same few months of early 2026, and they share an underlying cause. The AI model market in China has matured enough to reveal its structural fault lines. The widening gap between independent model companies and the platform giants entering the same lanes is becoming the most consequential competitive divide in China’s AI industry.

Platforms Get the Whole Menu

By early 2026, three commercial pathways had emerged for AI model companies globally. General-purpose AI assistants, led by ChatGPT, demonstrated that consumer subscriptions could reach significant scale. Coding and agent tools, led by Anthropic’s Claude Code, became one of the highest-value product categories in the industry. And video generation, among the largest consumers of inference compute, established a third axis of monetization.

Chinese companies recognized these lanes early. The divergence lies in how different types of players can approach them.

ByteDance shows what the platform advantage looks like in practice. Volcano Engine said in early April that Doubao’s daily token usage had surpassed 120 trillion. The company is pushing on multiple fronts at once, from video generation to agent products to cloud model services. Chinese media reports say Volcano Engine’s MaaS revenue target for 2026 has been set above RMB 10 billion, up from around RMB 2 billion in 2025. Even if the exact number is not officially disclosed, the direction is clear: ByteDance is trying to turn token demand into a cloud business, not just an app feature.

Alibaba tells a similar story from the organizational side. In March, the company separated its AI business from the cloud arm and created a new AI group led directly by CEO Eddie Wu. The new structure consolidates the company’s AI research lab, model-as-a-service platform, and consumer and enterprise AI products under a single organization. The point is not that Alibaba has already won. It is that platform companies can enter several monetization lanes at once because they already control infrastructure, distribution, and cash flow from other businesses.

That combination of advantages has three dimensions. First, distribution: ByteDance feeds inference demand through its own consumer apps, its cloud API, and embedded AI across the product portfolio. Independent model companies must acquire each user and each enterprise customer from scratch. Second, infrastructure: platform companies own or control their compute at scale, while most independents still procure it as a service. Third, cross-subsidy: a platform can treat AI model development as a strategic cost center for years, absorbing losses that would bankrupt a standalone company. A loss at that scale is manageable inside a platform company’s consolidated business. For an independent model company, it is existential.

Four Pivots, One Pressure

The identity shifts underway at China’s AI model companies reveal how each has tried to resolve the tension between ambition and constraint.

Zhipu made the most deliberate pivot. Through 2025, the Tsinghua-backed company spread its resources across model types and product categories, broadly tracking OpenAI’s full-stack approach. The shift came when it concentrated on coding. Its GLM-5.1 model and Coding Plan subscription sold out within minutes of each limited-quantity release. Developers wrote automated scripts to secure access.

The survival thesis behind this focus is that model quality creates pricing power. Zhipu’s first post-IPO earnings offered early evidence: API prices rose 83%, coding subscriptions went up 30%, and demand still exceeded supply. But the pivot also exposed how thin the operational margin is. Outages and degraded response times have become more frequent as usage scales, and the company has priced its models above most Chinese competitors on third-party aggregators, steering traffic to its own infrastructure rather than competing for platform volume. The coding niche is working. The infrastructure to sustain it is still catching up.

MiniMax made the pivot it had spent years trying to avoid. The company was founded on a deliberate rejection of China’s enterprise AI market, where bespoke projects and thin margins had defined the previous generation. It built consumer social products for global users, and its open-source distribution flywheel pulled in developers at near-zero marketing cost, with models among the most called on major aggregator platforms, priced at roughly one-twentieth of Claude Opus.

But consumer revenue hit growth ceilings earlier than expected, and margins stayed thin. The company recently brought on a former Huawei Cloud regional vice president and began appearing at government and enterprise industry events. It is investing in its own data center infrastructure and partnering with major streaming platforms and state media on AI video production. Each step marks a departure from the capital-light consumer model MiniMax was designed to be. The distribution advantage is real. Whether it can be converted into enterprise revenue without falling into the customization trap the founders originally fled is the open question.

DeepSeek faces a transformation that is partly chosen and partly imposed. The company that defined itself by turning away outside money is now in discussions to raise at least $300 million. The immediate driver is talent retention: competitors have reportedly offered two to three times what DeepSeek can match, and without external funding, the company’s equity has no market-validated price.

The less visible pressure is hardware. Beyond building V4, the team has invested substantial engineering hours adapting the model to run on Huawei’s latest Ascend processors, work that people familiar with the effort say has materially extended the development timeline. Fifteen months without a major model release is a long gap in a market where competitors ship updates in weeks. DeepSeek remains a research lab being asked to simultaneously perform industrial-scale hardware migration, and the budget for both is stretching thinner.

Moonshot AI placed the most technically ambitious bet. Kimi K2.6, released in April, can orchestrate hundreds of sub-agents in parallel, scaling up to a thousand according to the company, with demonstrations showing autonomous operation over multiple days. Rather than competing on model benchmarks, Moonshot is wagering that agent orchestration architecture will matter more than any individual model’s performance. The commercial path from impressive demo to repeatable revenue remains unclear for this category globally. But the bet is that getting the architecture right early creates structural advantages that a better model alone cannot replicate.

What Listing Didn’t Fix

MiniMax and Zhipu both completed Hong Kong IPOs in early 2026. The listings resolved the immediate liquidity crisis. When both companies filed in December 2025, each had narrowed its runway to months. MiniMax ended 2025 with $1.05 billion in reserves, before counting $614 million in IPO proceeds. Zhipu raised substantial capital as well.

The IPO fixed liquidity. It did not fix industrial position.

Zhipu reported RMB 724 million ($104.8 million) in 2025 revenue, with R&D spending running at 4.4 times revenue and net losses widening to RMB 4.72 billion. MiniMax generated $79 million in 2025 revenue against an adjusted net loss of $250 million. Both companies showed improving trajectories, with revenue growing faster than R&D spending for the first time at each firm. But the absolute gap remains vast, and neither has yet shown that the improvement can survive the next competitive shock.

The comparison with platform companies underlines the scale difference. According to Chinese media reports, ByteDance’s cloud division has set a MaaS revenue target above RMB 10 billion for 2026. Alibaba has set a goal of reaching $100 billion in annual cloud and AI revenue within five years. These figures suggest a different kind of contest. Independent model companies compete for customers. Platform companies compete for market architecture.

A Temporary Window in Coding

One pattern cuts across the Chinese AI landscape. Coding, agent capabilities, and multi-step reasoning have become the focal point of both product strategy and commercial investment.

The convergence follows economic logic. AI coding tools represent one of the clearest paths to sustained, high-value inference demand. Token consumption per coding session can run ten to a hundred times above a typical chat interaction. Enterprise willingness to pay for developer productivity has proven more resilient than for general-purpose assistants.

Model companies and platform companies are converging on this market from opposite directions. Independent model companies enter from the model layer: Zhipu’s coding subscriptions, MiniMax’s aggressive API pricing, Moonshot’s agent architecture. Platform companies enter from the application and distribution layer: ByteDance has begun sponsoring infrastructure for the OpenClaw ecosystem’s Chinese developer community, signaling a play for ecosystem control. Alibaba is launching coding and agent products through its newly reorganized AI group.

For now, coding may offer focused model companies a temporary window before the platforms fully commit. Industry assessments suggest that platform companies have not yet concentrated their resources on coding and agent products to the same degree, and Zhipu’s coding tools have reportedly surpassed several domestic Big Tech competitors in developer evaluations, though the gap between Chinese offerings and Anthropic’s Claude remains meaningful. This window exists in part because platform companies initially treated AI coding as a feature of their cloud business rather than a standalone product category. Model companies that build user bases and developer habits before the platforms arrive in force may establish enough switching cost to defend their position. But the window shows signs of narrowing. ByteDance’s recent shift from aggressive API price-cutting to margin-conscious pricing suggests a company moving past the land-grab phase.

Too Big to Be Startups, Too Small to Be Platforms

China’s independent AI model companies occupy a position with few precedents. They are too large, too well-capitalized, and too technically capable to be called startups. Zhipu and MiniMax are publicly listed. DeepSeek may soon carry a valuation above $10 billion. Their models compete credibly on global benchmarks.

Yet they lack what defines platform companies: owned distribution channels, diversified revenue, compute infrastructure at scale, and the financial capacity to sustain multiple business lines simultaneously. They must function as research laboratories, product companies, and infrastructure builders at the same time, on budgets that strain to cover any one of those roles.

Each company’s strategic choice looks different on the surface. Beneath that difference sits the same budget constraint. And each lane leads toward territory where platform companies are arriving from the other direction, carrying resources that independents cannot come close to matching.

The deeper pattern is not that these companies lack strategy. They do not. The pattern is that their strategic freedom is being compressed from two sides at once: by the capital intensity of staying competitive in model development, and by the speed at which platform companies are entering the same product categories. Whether the depth of focus that comes from having to choose one bet can outweigh the breadth of resources that comes from not having to choose at all is the question that defines the next chapter of China’s AI industry.

On April 19, a humanoid robot built by Honor, the smartphone brand spun out of Huawei, crossed the finish line of the Beijing E-Town Half Marathon in 50 minutes and 26 seconds. The time beat the human world record. It finished ahead of every pure-play robotics entrant on the course.

Honor swept the top three positions, all running autonomously. According to race-day reporting, Honor-series robots filled all six top overall positions after the weighted calculation penalized remote-controlled entries. The team had existed for roughly one year. Its machines navigated 22 turns and multiple terrain changes across 21 kilometers without human intervention.

The companies that dominated last year’s inaugural race, including the Beijing Humanoid Robot Innovation Center’s Tiangong Ultra and Songyan Dynamics, were displaced from the podium. Unitree, the sector’s top revenue generator by disclosed financials, has a $610 million IPO filing pending on the Shanghai STAR board. But its humanoid revenue still comes overwhelmingly from research buyers rather than industrial labor substitution. Even Unitree could not break through Honor’s formation despite demonstrating sprint speeds that would have shattered the human 1500-meter record.

A year ago, the same race looked like a blooper reel. Most robots could not finish. The champion needed 2 hours and 40 minutes for a distance that elite human runners cover in under an hour. This year, 47 of 102 teams completed the course, compared to six last year. The performance gap between 2025 and 2026 compressed years of visible progress into one season of iteration.

The speed improvement will generate headlines. The identity of the winner tells a more consequential story.

Why a Phone Company

Honor’s engineering lead told reporters after the race that the winning robot achieved full-stack in-house development, from hardware to algorithms, within a single year. The machine relied on dual lidar sensors and satellite positioning for autonomous navigation. When it collided with another robot mid-race, an internal dynamic balancing algorithm pulled it back to stable posture without human intervention.

The thermal management system shows where the structural advantage sits. Running generates heat that degrades motor performance and eventually triggers system failure. Honor addressed this by adapting liquid cooling technology from its smartphone product line. The cooling supplier, Huake Lengxin, confirmed that it designed a miniature pump one-quarter the size of competing products, capable of dissipating over 2,000 watts. The system went through five design iterations and accounts for less than 2% of the robot’s total bill of materials.

Battery energy density. Motor reliability. Thermal management at miniature scale. Precision component manufacturing with consistent quality across thousands of units. China’s consumer electronics supply chain has refined these capabilities over two decades of shipping billions of devices. An executive at Shoucheng Holdings, the Hong Kong-listed firm managing Beijing’s RMB 10 billion robotics fund, attributed the result to this lineage: companies with access to mature supply chains could guarantee the kind of battery performance and component-level consistency that smaller startups, with limited procurement leverage, struggle to match.

The pattern extends beyond smartphones. CITIC Securities has estimated more than 60% hardware overlap between smart vehicle and humanoid robot supply chains. XPeng, the electric vehicle maker, claims 70% software reuse between its car AI and its robot AI. The marathon results offer the clearest public demonstration yet that the hardware bridge connecting consumer electronics to robotics carries real competitive weight.

But there is a catch embedded in that confirmation.

This analysis continues below for Hello China Tech Premium subscribers.

Read more

For 2 years, one of DeepSeek’s most distinctive features had nothing to do with its models. The company consistently turned away outside money.

Liang Wenfeng, the founder, made this clear from the start. He had built DeepSeek out of profits from High-Flyer, his quantitative hedge fund, which posted a 56.6% annual return and over Rmb 5 billion in revenue in 2025. He reportedly turned down partnership approaches from Tencent and Alibaba, wary that outside involvement of any kind could compromise his decision-making authority. “VCs manage money for LPs. They all need returns,” he said in one conversation. “So we can never see eye to eye.”

The strategy appeared vindicated in January 2025, when DeepSeek released R1. Trained for roughly $5.6 million, R1 matched the performance of systems that cost orders of magnitude more. Here was evidence that a lean, independent team could compete at the frontier without billion-dollar compute budgets.

15 months later, that story needs updating.

According to The Information, DeepSeek is now in discussions to raise at least $300 million at a valuation above $10 billion. This would be the company’s first external funding round. In the same period, at least 5 core researchers have left for competitors. The next-generation V4 model has reportedly slipped from around the Lunar New Year to February, then March, and now late April. And DeepSeek has begun recruiting data center operations and delivery staff in Ulanqab, a major computing hub in Inner Mongolia, suggesting a move toward heavier physical infrastructure.

Each of these signals tells a story on its own. Together, they point toward a common structural problem. DeepSeek is trying to operate simultaneously as a frontier research lab, a company diversifying its chip supply chain, and a startup that needs to retain expensive talent in an overheated market. All 3 roles compete for the same finite pool of engineering hours, management attention, and money. The tension among them offers a compelling explanation for DeepSeek’s unusually long release gap.

The deeper question is why all 3 pressures intensified at the same time, and whether $300 million can resolve tensions that originate in the company’s identity rather than its bank account. Full analysis for subscribers below.

Read more

Editor’s Note: This is FlashPoint, Hello China Tech’s premium quick-strike column on one market-moving China tech event and why it matters.

Today’s column examines CATL’s Q1 earnings, which beat analyst consensus by 40% on revenue and delivered Rmb20.7bn in net profit. When 12 brokerages miss by that margin, the surprise is not just in the numbers. It is in what the models were failing to capture.

CATL, the Chinese battery group that commands roughly 40% of the global EV battery market, posted Q1 revenue of Rmb129.1bn ($18.9bn), up 52%, with net profit rising 48.5% to Rmb20.7bn. The revenue surprise, 40% above the median forecast of 12 brokerages, suggests the sell-side was still modelling a battery supplier. The results read more like those of an energy conglomerate in formation.

Two fronts drove the beat. Domestic EV battery share crossed 50% for the first time in 5 years. Energy storage, where margins run about 3 percentage points above those on EV cells, rose to roughly a quarter of Q1 cell shipments; April production schedules show storage climbing to 41.3% of cell output, up from less than 20% a year earlier. The Hungary plant, a €7.3bn investmentserving Mercedes-Benz, BMW, Stellantis, and Volkswagen, began mass production. About a third of revenue now comes from outside China.

Read more

After Sora’s retreat in late March 2026, Seedance 2.0 looked like the model to beat in China’s AI video market. It topped benchmarks, powered a viral content wave across Douyin, and attracted enterprise clients willing to build production pipelines around it. What happened next surprised the market.

Rather than consolidating its lead through ecosystem expansion, Bytedance raised prices. Repeatedly. Within a single month of Sora’s shutdown, Jimeng, Bytedance’s consumer-facing AI creation platform, implemented at least three price increases. The premium tier’s annual subscription jumped from Rmb 2,599 to Rmb 3,099. Monthly credit allocations were cut by a third to more than half, depending on the subscription level. The cost of generating raw materials for a 2-minute AI comic drama climbed from roughly Rmb 7 to Rmb 80. Access to Seedance 2.0’s realistic human video mode required a separate enterprise agreement with a Rmb 5 million ($690,000) spending floor.

Premium subscribers still faced queue times exceeding 3 hours. Ordinary users sometimes waited behind 80,000 others. Only large production companies with partnership agreements could access the full-capacity, no-queue version of Seedance.

The frustration among paying customers was tangible. Content studios that had built their entire production pipelines around Seedance found their cost structures upended between one billing cycle and the next. Small and mid-sized teams, many of them producing AI comic dramas at slim margins, were particularly exposed. For a segment that depends on predictable per-unit costs, the rapid repricing felt less like a market adjustment and more like a penalty for early adoption. A secondary market emerged on Xianyu (Alibaba’s resale platform) and Xiaohongshu, where legacy Jimeng membership cards with grandfathered pricing were being resold for as much as Rmb 20,000.

These pricing moves carried a signal that mattered more than benchmark scores. They revealed the compute economics that AI video’s subsidy period had obscured.

Consider the contrast. Sora shut down because its costs could not be justified against roughly $2.1 million in total revenue, a 30-day user retention rate of 1%, and an estimated investment-to-revenue ratio of 2,500 to 1. Bytedance avoided Sora’s fate by building genuine demand. Its consumer-facing strategy, built around viral templates such as pets fighting movie monsters, pushed AI video into mainstream creator workflows. But demand built on subsidized compute remains a cost center until pricing catches up. The viral wave that made Seedance famous also consumed compute at a rate that the original pricing could not sustain.

The broader lesson extends to any company running AI video at commercial scale. Foundation model inference is resource-intensive. Video generation consumes tokens at rates that dwarf text or image workloads: a single 15-second clip burns through roughly 300,000 tokens. When millions of recreational users generate clips with minimal revenue attached, the arithmetic deteriorates quickly. Bytedance’s official API pricing, announced in March at Rmb 28 per million tokens with video input and Rmb 46 without, translates to roughly Rmb 15 ($2) per 15-second clip, or about Rmb 1 per second. At that rate, even modest daily usage across a large creator base adds up to significant compute costs.

Bytedance’s price hikes were a correction, not a strategy. They signaled the gap between viral growth and sustainable unit economics. And they opened a window.

The AI video market in China had been a two-player contest between Bytedance and Kuaishou, the short-video company behind Kling AI. Kling had quietly built an annualized revenue run rate exceeding $300 million by January 2026, with fourth-quarter 2025 revenue reaching Rmb 340 million ($47 million). But a third entrant was preparing to change the geometry of competition.

The market expected consolidation after Sora’s exit. What arrived instead was a challenger, timed to exploit the distance between Bytedance’s pricing ambitions and its customers’ patience. The next phase of AI video competition looks less like a technology race and more like a contest over infrastructure.

Good Enough at the Right Moment

Read more

China’s Big Six state banks spend $18 billion on their own AI transformation and lend $3.2 trillion to the technology sectors producing it. Nothing comparable appears among the major U.S. banks examined here.

On 6 April 2026, JPMorgan Chase released its Annual Report 2025. The senior executive letter to shareholders described AI as a source of “measurable value” and “tangible financial results” across fraud detection, credit scoring, and personalisation. The bank disclosed a technology budget of roughly $19.8 billion for 2026, with production AI use cases doubling over the previous year. Management has cautioned that efficiency gains from AI will ultimately be competed away to customers rather than retained as lasting margin improvement.

The same week, across the Pacific, Industrial and Commercial Bank of China (ICBC), the world’s largest bank by assets, made a different kind of announcement. ICBC upgraded its four-year-old digital strategy from “D-ICBC” to “AI-ICBC,” deploying large language models across more than 500 business scenarios, up from roughly 200 a year earlier. The technology spend reached approximately RMB 28.6 billion ($3.9 billion).

That was the smaller number. ICBC’s technology lending balance crossed RMB 6 trillion ($820 billion) in 2025. Through its investment subsidiary, it has established 48 equity funds with committed capital exceeding RMB 108 billion ($14.8 billion), backing technology companies in AI, semiconductors, biotech, and commercial aerospace across 18 pilot cities.

JPMorgan uses AI. ICBC uses AI and bankrolls the companies producing it. This dual role, where the same institution both deploys the technology and finances the industries creating it, does not appear among the major U.S. banks surveyed in this analysis. Understanding it helps explain a dimension of China’s AI expansion that remains largely invisible in English-language coverage: the state banking system functions simultaneously as an AI adopter and a technology funding engine, and the two roles reinforce each other.

Read more

On April 8, Alibaba CEO Eddie Wu sent another company-wide letter. He has been writing more of them lately: the previous one, announcing a sweeping AI business group called Alibaba Token Hub, landed just 23 days earlier. This letter announced a Technical Committee for AI, upgraded the company’s model research laboratory into a full business unit, and reshuffled three of its most senior technology leaders. One appointment stood out: Zhou Jingren, previously CTO of Alibaba Cloud, received the title of Chief AI Architect.

The word carries intention. An architect designs systems for consistency and scale. The role implies that Alibaba’s AI challenge has moved beyond pushing scientific boundaries and into building machinery that produces reliably, at volume, around the clock. In the space of three weeks, Alibaba has executed three organizational moves that together amount to one of the clearest attempts this year to turn AI research into industrial production. Each addressed a different layer of the same problem: how to convert AI research into a commercially viable industrial operation.

Three Moves in Three Weeks

The sequence began on March 16, when Wu announced the formation of Alibaba Token Hub (ATH), a new first-tier business group that unified five previously separate AI units under a single commercial framework organized around the production, distribution, and monetization of tokens. As I analyzed in my earlier reporting, ATH formalized the thesis that Wu had been articulating since late 2025: that the token is the foundational unit of AI commerce, the way the kilowatt-hour measures electricity.

The second move arrived on April 8. Wu established a four-person Technical Committee at the group level, heading the committee himself. Zhou Jingren, Li Feifei (a database systems expert who spent years building Alibaba Cloud’s data infrastructure, and a different person from Stanford’s Fei-Fei Li), and Wu Zeming (Group CTO) fill the remaining seats. On the same day, Tongyi Lab, which houses the Qwen model family, was upgraded into the Tongyi Large Model Business Unit under Zhou Jingren, a structural elevation that formalizes the lab’s central role in the company’s model development strategy. Zhou now leads both the business unit and serves as the committee’s Chief AI Architect.

The third move was the quietest and possibly the most consequential. Between late March and early April, Alibaba released three new AI models in rapid succession: the multimodal Qwen3.5-Omni, the image generation model Wan2.7-Image, and the flagship Qwen3.6-Plus. Across these releases, Alibaba kept its most commercially important new capabilities inside hosted distribution rather than making them freely downloadable as open weights. Qwen3.6-Plus, the most significant of the group, is available through Alibaba Cloud’s API, with smaller variants in the Qwen3.6 family expected to be open-sourced later. Qwen3.6-Plus reached the top of OpenRouter’s usage rankings shortly after launch.

Each move addressed a distinct layer. ATH defined the commercial logic: how tokens get sold. The Technical Committee and the new Tongyi Business Unit defined the production logic: how tokens get made, transported, and delivered. The closed-source pivot defined the product logic: the best capabilities are no longer free. Together, the three steps completed a transformation from a research institution pursuing scientific ambition to a production operation pursuing commercial scale.

An Assembly Line with Three Foremen

The Technical Committee’s composition maps precisely onto the token supply chain that ATH was designed to monetize.

The most revealing appointment may be Li Feifei’s. He replaces Zhou Jingren as CTO of Alibaba Cloud, but he is a fundamentally different kind of technologist. Li is a database and storage systems specialist who joined Alibaba in 2018, an ACM Fellow and IEEE Fellow whose career was built on making large-scale data infrastructure perform reliably under load. Putting him in the Cloud CTO seat signals that Alibaba now treats its AI infrastructure layer as an engineering discipline, where uptime, cost efficiency, and throughput matter more than algorithmic novelty.

Zhou, freed from the Cloud CTO role, takes ownership of the asset that everything else depends on: model development. As I reported in my earlier column on Qwen’s restructuring, the departure of Qwen’s technical lead Lin Junyang in early March left a vacuum at the top of model research. Zhou stepping into a dedicated role fills that gap with a veteran whose career has spanned cloud infrastructure, search algorithms, and AI research across two decades. Every token in the system begins as output from a foundation model. Zhou now owns that origin point as his primary mandate.

The third seat belongs to Wu Zeming, Group CTO, who takes responsibility for the inference platform and serves as the committee’s Convener. Inference is where models process requests and generate output at production volume. In a system where every commercial transaction is an inference call, this is the bottleneck that determines whether the token economy can operate profitably. Wu’s convener role also makes him the coordination point between the other two members.

Read together, the three roles form a supply chain rather than a research consortium. One produces the core asset. One runs the machinery underneath. One manages the delivery layer in between. The CEO sits above all three. The committee exists to keep the line synchronized.

Continue reading with a paid subscription.

Read more

Last week, Anthropic cut off third-party tools from routing API calls through Claude Pro and Max subscriptions. The decision triggered immediate backlash from developers who had built workflows around that access. Most commentary framed it as a platform defending revenue.

Luo Fuli read the move differently. Luo, who leads Xiaomi’s MiMo large model team and previously worked at DeepSeek, posted a detailed thread that rapidly circulated across Chinese tech circles. The crackdown was “not surprising,” she wrote. What followed was a structural indictment of how the entire industry prices AI coding, and a warning that selling tokens dirt cheap while leaving the door wide open to third-party harnesses “looks nice to users, but it’s a trap.”

Her timing was deliberate. Across China, AI coding subscriptions have become among the most sought-after products in tech. Across major Chinese cloud platforms, limited-quantity coding plans often disappear within minutes when new slots open. Developers set morning alarms and write auto-purchase scripts to secure monthly access.

The demand is real. The underlying economics point in the opposite direction. Subscription pricing, built for an era when each AI interaction consumed a few hundred tokens, is now absorbing agent workloads that consume 10 to 100 times more per task. China’s coding plan market may be the clearest visible test of what happens when that collision goes unaddressed.

When Every User Becomes a Power User

Luo’s analysis begins with the mechanics of flat-rate pricing. A fixed monthly subscription assumes a distribution of usage. Light users subsidize heavy ones. Averages hold, and the business survives. Gyms, streaming services, and cellular data plans all depend on this balance.

Anthropic built Claude Code’s subscription on that logic. Its native framework manages context carefully, maximizes prompt cache reuse, and consolidates tool calls to keep per-query costs within bounds. Within that controlled environment, the cross-subsidy between light and heavy users can work.

Third-party agent frameworks shattered the distribution from both ends. “I’ve looked at OpenClaw’s context management up close,” Luo wrote. “It’s bad.” Within a single user query, the harness fires off rounds of low-value tool calls as separate API requests, each carrying a long context window, often exceeding 100,000 tokens. The actual request count per query runs several times higher than what Claude Code’s own framework would generate. Converted to API rates, the real cost reaches tens of times the subscription price. In Luo’s assessment: “That’s not a gap. That’s a crater.”

A second problem compounded the damage. Claude’s caching system depends on consistent prefixes in conversation context. When the same token sequence opens successive requests, the system reuses prior computation and skips redundant processing. As Luo noted, “many third-party harnesses compress tool responses every 3 steps when approaching the context limit, leading to very low cache hit rates.” Each compression rewrites the prefix, and the model is forced to reprocess the full context window from scratch. Cached work becomes wasted work.

The combined effect pushed every query toward maximum cost. Users of third-party harnesses generated requests with the cost profile of power users, regardless of their individual usage intensity. The statistical distribution that made subscriptions viable simply collapsed. Anthropic absorbed the gap.

One striking case illustrates the scale. A Claude Max subscriber paying $100 per month generated over $5,600 in equivalent API costs in a single billing cycle. A subsidy ratio of 25 to 1. That sits at the extreme end, but the structural pattern held broadly: pricing designed for chat-era consumption funded agent-scale computation.

Luo framed the crackdown as the only rational outcome. Third-party tools retain access to Claude through standard API billing. They lost the ability to ride the subscription. Short term, she acknowledged, costs for agent users will jump easily by tens of times. “But that pressure is exactly what pushes these harnesses to improve context management, maximize prompt cache hit rates to reuse processed context, cut wasteful token burn. Pain eventually converts to engineering discipline.”

China’s coding plan market is deep in the same trap. Alibaba’s Pro tier sells out by 9:30 AM each morning. Tencent’s promotional slots show as perpetually unavailable. Beneath the scarcity, throttled speeds, degraded models, and quota walls suggest a product category under severe cost pressure. What follows examines why a Xiaomi executive’s proposed fix drew fierce resistance, and what the shift from token price to token efficiency means for China’s AI coding market.

[Continue reading with a paid subscription]

Read more

In January 2026, Chinese platforms launched more than 14,600 AI-generated short dramas. That works out to roughly 470 new titles each day. By February, 127,800 were in active circulation across the country’s streaming and social media ecosystem.

These figures require a moment of calibration for readers who have not encountered the format. Short dramas are serialised, mobile-first shows with episodes of two to five minutes, distributed through algorithm-driven platforms and monetised through advertising and micropayments. Think of them as soap operas rebuilt for smartphone attention spans: melodramatic, formulaic, and staggeringly popular. The audience for AI short dramas alone is projected to reach 280 million in 2026, up from 120 million the year before.

Before AI entered the picture, this market was already producing content at a volume no other national entertainment industry could match. Generative video models compressed the production cycle further and collapsed costs so sharply that the industry’s economics, its labour structure, and its relationship with the real human faces on screen have all been remade in under 18 months.

What followed is worth examining closely. The forces at work, collapsing production costs, displacement of human performers, industrial-scale identity appropriation, and belated regulatory response, are not unique to China.

Everyone Is a Studio Now

The production economics of AI short dramas have changed so fast that the industry’s own vocabulary has struggled to keep pace.

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. Per-minute costs have fallen from RMB 3,000 to 5,000 in early 2024 to a typical range of RMB 500 to 1,000 today, with some outsourced teams quoting as little as RMB 200 ($27).

The collapse in cost has created production enterprises that would have been difficult to imagine two years ago. Jiangyou Culture, a Hangzhou-based company backed by China Literature (the Tencent-affiliated publishing group), has grown to roughly 1,000 employees and generates annual revenue of approximately RMB 1 billion ($137 million) with net profit of RMB 200 to 300 million. A separate production house, Judian, has scaled to a comparable size and produces roughly 100 AI photorealistic dramas per month alongside 1,000 to 2,000 AI voiceover dramas. Yet another company scaled from zero to 200 employees in under a year, with a ten-person team completing a 30-episode series in 20 days.

At the individual level, the figures are equally striking. At an AI entertainment forum earlier this year, a startup founder demonstrated that a single operator working with ByteDance’s latest video generation tools could produce 40 minutes of distribution-ready content per day at a profit margin of 45 per cent. The operators who run these tools, known colloquially as “card-pulling technicians” for the lottery-like process of generating and selecting usable clips, earn between RMB 2,000 and 3,000 ($275 to $410) per month. That is barely above China’s urban minimum wage.

The underlying technology has improved at a pace that caught even its creators off guard. In January 2025, fully AI-generated titles accounted for just four of Douyin’s top 5,000 short dramas. By November, the number had reached 217. When ByteDance released Seedance 2.0 in early 2026, the founder of Game Science, the studio behind the global hit Black Myth: Wukong, said he was “deeply shocked” by the model’s capabilities. Complex physical actions, multi-character scenes, sustained character consistency across shots: problems that had constrained AI video for years appeared to have been meaningfully resolved.

The commercial trajectories of the leading models have diverged. Kling AI, the video generation platform built by Kuaishou (TikTok rival Kwai’s parent company), had crossed a meaningful commercial threshold by early 2026. Official disclosures showed an annualised revenue run rate of approximately $240 million as of December 2025; market reporting indicated the figure had exceeded $300 million by January 2026. Kuaishou has disclosed adoption across professional creative sectors including film, television, short dramas, marketing and e-commerce, but has not published a revenue breakdown by category. OpenAI shut down Sora on 24 March as it redirected compute and talent toward coding tools, enterprise products and its broader AGI ambitions. Reporting also suggests the product’s compute demands had become a meaningful internal constraint. I examined what drove those two trajectories apart, and what they reveal about where AI revenue actually accumulates, in a separate analysis.

Six Days of Work in March

The industry’s expansion has come with a human cost that is visible, immediate, and concentrated among the people who least expected it.

Li Wenhao, a short drama actor based in Chongqing, was busy enough after entering the industry in 2023 to shoot for 50 consecutive days. In March 2026, he worked six. Casting calls that once arrived at a rate of more than 20 per day now come once every two or three days. Of roughly ten short drama companies he knew in his city, only two or three still hire human performers.

His trajectory reflects an industry-wide contraction. Actors who had ridden the short drama boom to unusual earnings, with daily rates reaching RMB 50,000 to 60,000 ($7,000 to $8,200) for breakout performers, report that available scripts have roughly halved. One leading actor who previously fielded more than 30 offers per month now receives seven or eight.

Some companies have made the transition with startling directness. Chengdu Zhongdu, a mid-sized studio, announced a complete exit from live-action in March, converting all staff to AI workflows. Yaoke Media, a company behind several well-known Chinese television dramas, signed two AI-generated performers to formal contracts. Public reaction was sharply negative, partly because the digital performers resembled real celebrities, and partly because of the symbolism: a major production house officially filling roles with software.

Hao Lei, one of China’s most respected dramatic actresses, delivered an assessment on a nationally broadcast variety show that travelled widely: AI will replace 90 per cent of actors. She added that in certain roles, AI already outperforms its human equivalent. The remark landed hard because of its source. Hao Lei has decades of experience on screen. She was describing what she sees in her own industry, not promoting a product.

The speed of displacement has few parallels in entertainment history. Sound took years to replace silent film. CGI eliminated specific crafts while creating others over two decades. In China’s short drama market, where productions carried no union protections and cycles were already measured in weeks, AI compressed the adjustment into months. Actors who were fielding 30 script offers a month at the start of winter found themselves with seven by spring.

Whose Face Is That?

The displacement of professional actors was, in retrospect, predictable. A less anticipated category of harm emerged alongside it: the mass, unauthorised appropriation of real people’s faces.

In early 2026, a 72-episode AI period drama appeared on Hongguo, ByteDance’s dedicated short drama app, and accumulated a popularity score exceeding 40 millionbefore anyone noticed the faces. A fashion blogger who creates content about traditional Chinese clothing discovered that one of the drama’s characters bore her likeness. The character had been written as greedy and promiscuous. A second blogger found her face used for an antagonist depicted as violent and cruel. Neither woman had been contacted, compensated, or informed.

The technique has become a common method in AI drama production. Producers download photographs from social media and feed them into AI models as reference material, a process the industry calls “reference imaging” (垫图). The output is adjusted enough to appear distinct while retaining the structural features of the original face. Producers claim independent creation. Anyone who knows the source can see the resemblance. The financial logic is straightforward: building a cast from scraped photographs eliminates the single largest variable cost in short drama production.

The practice reached high-profile targets quickly. Multiple AI dramas featured characters whose faces closely matched those of Yi Yangqianxi (known internationally as Jackson Yee), one of China’s most prominent young actors. Two such titles had accumulated combined popularity scores exceeding 100 million before being flagged and removed. His management commissioned lawyers and issued a formal statement. Other major Chinese entertainers, including Xiao Zhan, Yang Zi, and Dilraba Dilmurat, had been targeted through similar methods.

The celebrity cases drew headlines. The cases involving ordinary people may carry larger implications. A well-known performer who is misused commands legal resources and public sympathy. A social media user whose face appears, without her knowledge, in a production she never consented to, playing a character designed to be despised, has almost none. She may never discover the appropriation. The producer may operate through layered outsourcing that obscures accountability.

This is not a concern confined to Chinese regulatory debates. The necessary ingredients, publicly available photographs, generative models capable of photorealistic output, and financial incentives to skip licensing, exist in every market where people post images of themselves online. Any jurisdiction where AI video production reaches industrial scale is likely to face a version of the same collision between generative capability and the right to control one’s own face. China surfaced the problem first because it produces at a speed that compresses years of gradual erosion into months of visible crisis.

Seven Days in April

China’s regulatory system responded with a velocity that surprised even seasoned observers of Beijing’s technology governance.

On 1 April 2026, AI dramas formally entered the national content registration system, requiring producers to register titles and undergo classification review before distribution. On 2 April, the actors’ committee of the China Broadcasting Association prohibited the unauthorised use of performers’ likenesses and voice prints, explicitly noting that labels such as “non-commercial” or “public interest sharing” do not constitute legal justification.

On 3 April, two actions landed simultaneously. The Cyberspace Administration of China published draft rules titled “Digital Virtual Person Information Service Management Measures,” governing the creation and commercial use of AI-generated human characters, with a public comment period running through 6 May. The same day, Douyin Group, ByteDance’s domestic entity, announced a RMB 200 million ($27.5 million) fund to support live-action short drama production, widely read as an acknowledgement that the AI transition had overshot.

By 6 April, Hongguo disclosed the results of a comprehensive audit: 15,000 works reviewed, 670 penalised for violations including unauthorised use of real performers’ likenesses, trademarked character designs, and copyrighted brand imagery. Penalties ranged from content removal to permanent account termination and legal referral. Updated platform-level content review standards, including a new layer of “thematic orientation review,” took effect on 7 April.

The pace reflects a recurring pattern in Chinese technology governance that Western analysts frequently underestimate. Beijing tolerates rapid commercial experimentation with minimal pre-emptive constraint. When the social costs become visible to ordinary citizens, the response comes as a dense cluster of platform actions, industry association statements, and draft national rules within days. Not all of these carry the same legal force: the CAC’s virtual-person regulation entered public consultation, while the platform review standards took effect immediately. But no comparably concentrated response has yet emerged across major Western jurisdictions.

The contrast with other governments is instructive. In March 2026, the United States Supreme Court declined to hear a case involving AI-generated visual art. The ruling leaves in place a clearer standard for fully AI-generated works lacking human authorship, which cannot receive copyright under existing law. Broader questions around AI-assisted creation remain contested. In the United Kingdom, the government abandoned a proposal allowing AI companies to train on copyrighted works under an opt-out model, retreating under pressure from artists including Sir Elton John and Dua Lipa, yet implementing no alternative framework.

This regulatory asymmetry matters beyond the specifics of short dramas. Jurisdictions with slower governance cycles face longer windows in which harmful practices can scale without formal constraint. China’s experience offers an uncomfortable data point: the costs of waiting for consensus are borne by the people whose faces and livelihoods remain unprotected while the deliberation proceeds.

Plenty of Content, Nothing to Watch

The production explosion carries a paradox the industry has begun to confront openly.

Of the 127,800 AI dramas in circulation by February 2026, the proportion that crossed the 100-million-view threshold stood at 0.117 per cent. In 2025, Douyin’s ecosystem launched 60,000 AI dramas. Ninety-six reached the benchmark. That breakout rate of 0.16 per cent 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, by comparison, accumulated 4.4 billion. AI content accounted for roughly 30 per cent of total short drama viewership during the 2026 Spring Festival holiday. The share is growing, but on the strength of sheer volume rather than audience loyalty. No AI drama has entered the cultural conversation the way live-action hits routinely do, generating fan communities, launching performer careers, spawning merchandise.

Audience surveys in early 2026 found that photorealistic AI dramas registered the lowest willingness to pay of any AI content format. Viewers detect the synthetic quality. The uncanny valley, where visual fidelity approaches realism without quite arriving, appears to suppress the emotional engagement that converts sampling into spending.

The industry has responded by leaning into distribution economics. AI drama production increasingly runs on what Chinese producers call traffic arbitrage: manufacture content cheaply, spend aggressively on platform ads to generate views, and profit from the margin. In March 2026, daily ad spend on AI dramas across Douyin broke RMB 70 million ($9.6 million), surpassing live-action short dramas for the first time. Survival depends on media buying efficiency, not storytelling.

One AI entertainment executive characterised the resulting structure with unusual candour: “model companies sell compute, platforms sell traffic, capital sells stories.” Each participant in the chain profits from the system’s activity regardless of whether the output finds a real audience. The closed loop can sustain itself financially without producing a single work that resonates culturally. Whether that constitutes an entertainment industry or merely a revenue mechanism wearing the costume of one is a question the market has not yet been forced to answer.

From Bollywood to Hollywood

China’s experience would command less attention if it appeared to be unique. The evidence from other markets suggests the pattern is broader.

India’s film industry, the most prolific by output in the world, has begun adopting AI production tools along a trajectory that broadly tracks China’s. A leading Bollywood talent agency has built an AI studio in Bengaluru to engineer digital actors inspired by Hindu mythology. AI-assisted production costs have reportedly fallen to one fifth of traditional budgets, with timelines compressed to one quarter. One Indian media company re-released a 2013 Bollywood hit with an AI-altered happy ending. The modified version sold 35 per cent of tickets during its release month, 12 percentage points above the cinema chain’s 2025 average. The company is now reviewing its back catalogue of 3,000 titles for additional AI adaptation opportunities.

India faces far fewer formal labour constraints than Hollywood, even as audience acceptance remains uneven. The country’s entertainment industry operates without equivalents to SAG-AFTRA or the Directors Guild of America, and studios have moved quickly to exploit this flexibility. Yet India’s quality challenge mirrors China’s with striking precision: an AI-generated mythological television series aired on a major Indian streaming platform to 26.5 million views and a 1.4 out of 10 audience score. The willingness to sample AI entertainment and the reluctance to invest emotional commitment in it appear consistent across cultures.

In Hollywood, the friction takes a different form. Union agreements negotiated after the 2023 strikes require consent and compensation for digital replicas and maintain minimum creative staffing. These constraints have slowed visible AI adoption at the production level. Below that level, AI use percolates through junior staff informally, without institutional guidance.

On the copyright front, the MPA and major studios including Disney, Warner Bros. Discovery, Paramount, Netflix, and Sony sent cease-and-desist letters to ByteDance over Seedance 2.0 in February. Separate lawsuits and complaints against other AI companies over image generation, character rights and training data are proliferating. The legal architecture governing AI-generated entertainment remains unsettled in every major jurisdiction. The commercial incentive to produce it has not paused to wait for resolution.

Market coverage citing Goldman Sachs projects the global AI video market at approximately $29 billion by 2030, up from $3 billion in 2025. China’s experience suggests the supply side of that projection is already plausible.

What 470 a Day Tells Us

China’s AI drama market has produced the world’s most detailed case study of what happens when the cost of creating entertainment content collapses toward zero.

Quantity arrives first, in overwhelming abundance. Production enterprises scale from nothing to a thousand employees within months. Anyone with access to a generative model and a platform account can operate as a studio. The barriers that once shaped the entertainment industry, specialised equipment, trained performers, physical production infrastructure, compress toward irrelevance.

The consequences follow, and they arrive faster than the institutions designed to manage them. Actors who were fielding 30 script offers a month find themselves with seven. Ordinary people find their likenesses embedded in productions they never authorised. Audiences sample the output in vast numbers, but willingness to pay remains weak. Regulators, confronted with a market moving at machine speed, sprint to impose structure on terrain that was open land the month before.

Every element of this sequence is portable. The short drama format, optimised for mobile distribution and algorithmic amplification, is not unique to China. TikTok, YouTube Shorts, and Instagram Reels have built equivalent distribution infrastructure worldwide. India’s entertainment industry is already travelling the same path without the friction of union protections. Hollywood’s contractual defences are real, and they are also the product of specific historical negotiations that most other countries’ creative industries have never replicated.

China has settled the first question for every entertainment market: AI production can reach industrial scale. The question that remains is who absorbs the cost when it does. The performers whose livelihoods dissolve. The individuals whose faces are harvested. The audiences asked to watch content that nobody invested enough to make compelling. Or the institutions that arrive to govern a transformation already well under way.

This article examined what happens on the ground when AI entertainment reaches industrial scale. A companion analysis looks at the other side of the equation: who is making money, and why. It covers the commercial divergence between Sora, Seedance, and Kling in detail, Meitu’s $530 million application-layer export model, and what the Manus acquisition tells us about how Chinese AI companies can and cannot go global.

On March 30, Baidu’s Apollo Go began commercial driverless service in Dubai. Apollo Go is China’s largest robotaxi fleet, with over 20 million cumulative rides across 26 cities. The company signed an exclusive partnership with Dubai Taxi Company, the emirate’s largest state-owned taxi operator, and became the first platform in Dubai to offer fully autonomous rides through its own app. For Baidu, the internet search giant that has poured more than a decade of investment into autonomous driving, the Dubai launch extended a global footprint that already covers Hong Kong and Abu Dhabi.

The next evening, nearly a hundred Apollo Go vehicles stopped simultaneously on elevated highways across Wuhan, a central Chinese city of 13 million. Passengers found themselves trapped in stationary cars on busy motorways. SOS buttons did not respond. One rider reported waiting thirty minutes before reaching customer service by phone. Others were trapped for up to ninety minutes. Some passengers forced doors open and walked along highway shoulders to find their way out. Police reported the mass shutdown compressed three traffic lanes into one across the city’s Second and Third Ring Roads. No injuries were recorded. The fleet resumed operations the following day.

Apollo Go was not alone in making global moves that quarter. WeRide, a Guangzhou-based autonomous driving company that received a $100 million strategic investment from Uber in 2025, launched fully driverless commercial operations in Dubai. Grab, Southeast Asia’s dominant ride-hailing platform, began offering WeRide-powered robotaxis in Singapore as the region’s first autonomous ride service. Pony.ai, a Nasdaq-listed Chinese autonomous driving company, partnered with Uber and a startup called Verne to bring what they called Europe’s first commercial robotaxi service to Zagreb.

Three companies. Three regions. One quarter.

The expansion reflects a clear competitive calculation. Waymo, Alphabet’s autonomous driving subsidiary, now provides about 500,000 paid rides a week across 10 U.S. cities and says it is on track to reach one million a week by the end of 2026. Waymo has stayed within the United States. Chinese competitors see a window to secure partnerships and regulatory footholds in markets where autonomous driving policy is still being drafted.

The ambition behind the global push is understandable. The Wuhan incident complicates it. China’s robotaxi industry has entered a phase where the speed of its international expansion has outpaced the maturity of its operations. The vehicles can navigate foreign roads. The cloud infrastructure, remote monitoring systems, and emergency response chains that support those vehicles remain fragile under real-world stress. The financial models meant to underpin everything still depend on a handful of showcase cities and earnings that dissolve under scrutiny.

The pace of the global rollout is the strategy. Whether that pace is also the vulnerability is the question this industry has yet to answer.

Read more

On March 22, at the China Development Forum in Beijing, Liu Liehong, director of China’s National Data Administration, gave tokens an official Chinese name: 词元, pronounced cíyuán, a compound meaning “word-unit.” He then went further. Tokens, Liu declared, are “the value anchor of the intelligent era” and “the settlement unit connecting technology supply with commercial demand.”

Two days later, at a press conference held by the State Council Information Office, Liu disclosed the latest figure. China’s daily token consumption had reached 140 trillion. In early 2024, the number was 100 billion. Growth over two years exceeded a thousand-fold. During the week of March 16, Chinese AI models consumed 7.36 trillion tokens on OpenRouter alone, accounting for 36 percent of the platform’s global volume and surpassing American models for the third consecutive week.

No other government has taken this step. Tokens are a concept from machine learning engineering. They measure how language models process information, much the way kilowatt-hours measure electrical consumption. Every prompt sent and every response generated breaks into tokens for computation. Cloud providers bill by the token. Benchmarks rank models by tokens processed per second. The term has lived in technical documentation for years without attracting political attention.

What China did in March was elevate this engineering concept into an economic indicator. The National Data Administration now tracks daily token volume and reports it at major policy forums. An eight-ministry directive issued in January 2026 set a target of 1,000 industrial AI agents by 2027. The most recent five-year plan formally adopted “AI+” as a national action framework. Industry forecasts project that China’s AI-related revenue could exceed 10 trillion yuan by the end of the decade. JPMorgan estimates that China’s inference token consumption will grow roughly 370-fold between 2025 and 2030.

When a government assigns a measurement unit, defines its significance, and builds reporting infrastructure around it, the unit acquires weight beyond its technical function. China is constructing a national accounting framework for AI activity with the token at its center. The question worth examining is what that framework captures and what it leaves out.

The Hundred-Trillion Club

According to the most widely cited industry estimates, only three companies have crossed the 100-trillion threshold in daily cloud-based model inference: OpenAI, Google, and ByteDance.

ByteDance crossed it in early 2026. By April, daily token consumption through its Volcano Engine platform had reached 120 trillion, doubling in three months and growing a thousand-fold from the 120 billion daily level reported at launch in May 2024. Token consumption from individual users has grown approximately sixteen-fold over the past month, driven largely by OpenClaw and similar agent frameworks that execute autonomous multi-step tasks rather than single-turn conversations. Only two American companies operate at the same scale, and both serve significantly larger international user bases. ByteDance’s consumption remains concentrated in China, though overseas volumes are expanding.

The consumption surge is rewriting cloud economics. In 2024, China’s entire Model-as-a-Service market was worth 710 million yuan, according to IDC. That figure represented less than a quarter of one percent of a public cloud market exceeding 300 billion yuan. Eighteen months later, a senior Alibaba Cloud executive projected that MaaS revenue could reach 30 percent or more of total cloud income. Volcano Engine initially set a MaaS revenue target above 10 billion yuan for 2026, then raised it after growth from new model launches and the OpenClaw surge exceeded projections. The previous year, the company had raised its MaaS target twice for the same reason.

For ByteDance, tokens are the entire competitive strategy. Volcano Engine entered cloud computing years after Alibaba, Tencent, and Huawei had established their positions. Closing the gap on traditional infrastructure services, databases, and enterprise software was structurally difficult. MaaS offered a different entry point. Starting in 2024, Volcano Engine reorganized its sales incentives so that selling model services at a given revenue level generates higher commissions than selling equivalent traditional cloud products. The message to the salesforce was explicit: lead with tokens.

The feedback loop is deliberate. Models deployed into real business environments through MaaS generate usage data, failure cases, and performance benchmarks that feed directly into model improvement. Enterprise customers revealed a significant pattern early on: heavy demand for agentic coding workflows, where AI models plan, write, test, and debug software autonomously over extended sessions. That signal shaped the direction of ByteDance’s own model development.

Alibaba and Tencent have each responded with strategic pivots reflecting different theories of how the token economy will consolidate. As I examined last month, Alibaba CEO Eddie Wu created a first-tier business group called Alibaba Token Hub and restructured the company’s AI operations around token production and distribution. Tencent took a different path. On March 27, at a cloud summit in Shanghai, the company rebranded its MaaS platform as TokenHub and launched a unified billing system called Token Plan. Developers can now access Tencent’s own Hunyuan model alongside DeepSeek, MiniMax, and others through a single API with consolidated billing. Dowson Tong（汤道生）, Tencent’s senior executive vice president, framed the reasoning: the capability gap between leading models is narrowing, and what matters now is who can engineer models into reliable production systems.

Three strategies, one destination. Alibaba integrates vertically: model lab, cloud platform, enterprise product, and proprietary chips under one management structure. ByteDance attacks from the MaaS layer, using token economics to offset its late start in traditional cloud. Tencent positions as the neutral aggregator, betting that model commoditization will shift value toward platform engineering and workflow orchestration. All three have reorganized significant portions of their businesses around the token as the central unit of account.

The thesis behind these moves draws on a pattern I reported in my coverage of OpenClaw’s rapid adoption. Agent workloads consume tokens at ten to a hundred times the rate of conversational chat. If agents become the primary interface between software and AI models, whoever controls the most efficient token delivery infrastructure holds a structurally advantaged position.

Token volume has become the metric everyone tracks. But metrics can illuminate or obscure. Understanding whether China’s 140 trillion daily tokens represent genuine economic activity requires examining who is paying, at what price, and for what outcome.

[Continue reading with a paid subscription]

Read more

Zhipu AI opened its first earnings report with a formula.

“AGI Commercial Value = Intelligence Upper Bound x Token Consumption Scale.”

No public AI company has ever reduced its entire business logic to a single equation. OpenAI publishes blog posts about safety. Anthropic talks about responsible scaling. Google buries its AI economics inside a $350 billion conglomerate. Zhipu, the Tsinghua University-backed startup that became the world’s first pure-play AI model company to go public, chose a different path. It told investors exactly what it believes and staked its future on that belief being correct.

The formula carries a specific claim. Model quality determines pricing power. Pricing power multiplied by usage scale determines commercial value. Every dollar Zhipu spends on research, every price increase it pushes through, every customer it serves flows from this logic. The “intelligence upper bound” is the variable the company believes it can control. Token consumption is the one it believes will follow.

The first year of numbers offers an early test. Revenue reached RMB 724 million ($104.8 million) for 2025, a 132% increase. Total losses widened to RMB 4.72 billion. Gross margins compressed from 56.3% to 41%. The stock, which had surged roughly sixfold from its January IPO price to briefly push market capitalization past HK$300 billion, fell 5.5% ahead of the results. Bloomberg consensus had expected RMB 756 million in revenue.

The headline story is a company growing fast but burning faster. A full year of revenue barely covers three months of R&D spending.

But the headline obscures the structural story. During the industry-wide price war that swept China’s AI model market in late 2024 and early 2025, Zhipu had cut its flagship GLM-4-Plus by as much as 90%. By early 2026, the dynamic had inverted. The company raised API prices by 83% and its coding subscription by 30%. Demand still exceeded supply. CEO Zhang Peng said the bottleneck was compute capacity, not customers.

That reversal, from price war to pricing power in a single year, is the most significant signal in the report. And the most uncertain.

In December, I argued that Zhipu and MiniMax, the two pure-play AI companies that rushed to list in Hong Kong, were both caught in the same structural trap: profitable at the unit level, insolvent at the business level, with competition dictating the pace of spending. Three months later, MiniMax’s earnings suggested one escape route through open-source distribution and global consumer reach. The gap between R&D growth and revenue growth closed for the first time.

Viewed through the lens of the last two earnings reports, the divergence between the two companies is now sharp. Zhipu’s earnings test a fundamentally different hypothesis. Where MiniMax bet on distribution, Zhipu bet on quality. Where MiniMax went global and consumer-facing, Zhipu stayed domestic and enterprise-focused. Where MiniMax let open-source adoption drive volume, Zhipu raised prices and let scarcity do the work.

The question the formula poses is whether an AI company can achieve durable pricing power through model quality alone.

[🔒 Keep reading with a paid subscription]

Read more

In 2024, Meituan, China’s dominant food delivery and local services platform, earned an adjusted net profit of Rmb 43.8 billion. It was one of the most reliable cash generators in Chinese tech. Twelve months later, the company reported an adjusted net loss of Rmb 18.6 billion for 2025. The swing exceeded Rmb 62 billion in a single year.

The cause was a subsidy war of extraordinary scale. When JD.com, China’s largest direct e-commerce retailer, launched its own food delivery service in early 2025, it triggered a three-way battle with Meituan and Alibaba’s Ele.me for control of on-demand delivery. Over roughly 6 months, the three companies collectively spent an estimated Rmb 80 billion to Rmb 100 billion in direct consumer subsidies. Total war-related spending ran far higher: Meituan’s promotion and advertising expenses alone nearly doubled, from Rmb 39.1 billion in 2024 to Rmb 74.5 billion.

All of it to compete for orders averaging Rmb 30. About $4.

From Merchant Margins to the CPI

The damage extended well beyond balance sheets.

A Fudan University survey of over 40,000 merchants found that after the war escalated in mid-2025, average daily order volumes rose 7% while average daily revenue fell 4%. Platforms subsidized orders at price points that undercut restaurants’ own pricing, generating traffic that diluted margins rather than building them.

The macro footprint followed. Food and dining carry roughly 30% weight in China’s consumer price index. During Q2 and Q3 of 2025, core CPI (excluding food and energy) was actually rising. The drag came specifically from food prices, pushed down by platform subsidies that returned restaurant meal prices to levels not seen in a decade. Goldman Sachs estimated the combined profit reduction across the 3 platforms at approximately Rmb 70 billion, with associated tax revenue losses exceeding Rmb 10 billion.

By late 2025, the war was showing signs of exhaustion on all sides. Meituan began narrowing its per-order losses in Q4. Alibaba merged its Taobao Flash Purchase unit with Ele.me to reduce redundancy. JD scaled back its delivery team and pivoted toward niche categories.

Regulatory attention reinforced the shift. In February 2026, China’s State Administration for Market Regulation convened meetings with executives from the major platforms on anti-unfair-competition practices. On March 25, Economic Daily, a State Council-affiliated newspaper, published a commentary titled “The Delivery War Should End.” The regulator reposted it the same day. Investors read the signal clearly: Meituan surged 13%, Alibaba rose 4.6%, JD gained nearly 5%.

The delivery war appears to be winding down. But the question of who actually benefited, and whether the cost was worth it, demands a closer reading of Meituan’s annual report. What follows is an analysis of the economics that made this war self-limiting, the structural shifts it produced, and the divergent strategies the 3 platforms are now pursuing.

Read more

On 25 March, OpenAI shut down Sora, the text-to-video model that had captivated the AI world since its debut two years earlier. The reasons were familiar: the product appears to have consumed more in compute than it could justify commercially, and never found a sustainable business model.

The same day, Kuaishou Technology, the Chinese short-video platform behind TikTok rival Kwai, published its annual earnings. Tucked inside the results was a striking figure. Kling AI, the company’s video generation platform, had reached an annualised revenue run rate of $300 million as of January 2026. Fourth-quarter revenue came in at RMB 340 million ($47 million). Management projected that Kling’s 2026 revenue would more than double.

One product collapsed under the weight of its own compute bill. The other crossed a commercial threshold rare among standalone AI video tools worldwide. The contrast frames a question that extends well beyond video generation: in the emerging AI industry, where does the revenue actually accumulate?

Western discourse about Chinese AI tends to cluster around two narratives. The first centres on model-layer breakthroughs: DeepSeek’s ascent, its open-source challenge to OpenAI. The second follows founder migration: Manus relocating to Singapore, HeyGen moving to Los Angeles, a generation of Chinese AI engineers voting with their feet. Both stories are real. Neither captures the most commercially significant development in Chinese AI today.

That development sits in the application layer. A cohort of Chinese companies that build AI-powered products for specific use cases, from image editing to video production to e-commerce design, are generating measurable, recurring global revenue. They treat foundation models as inputs rather than outputs. They solve problems granular enough that general-purpose AI cannot easily replicate the solutions. This application-layer advantage may prove more durable than model superiority or talent arbitrage. The evidence is accumulating faster than most Western observers appreciate.

The Scoreboard Everybody Misreads

Andreessen Horowitz published the sixth edition of its Top 100 generative AI consumer applications ranking in March 2026. The headline, as usual, belonged to ChatGPT: 900 million weekly active users, web traffic 2.7 times that of Google’s Gemini.

The more revealing signal sat further down the list. Chinese-background products occupied a significantly larger share than in any prior edition. DeepSeek ranked fourth globally on the web. Kling AI, Manus, Kimi (a well-funded Chinese AI chatbot), and Alibaba’s Qwen all featured. On mobile, ByteDance’s Doubao (China’s most popular ChatGPT equivalent) and Baidu’s AI search product entered the rankings.

English-language commentary latched onto DeepSeek. The model-layer narrative proved irresistible. Yet the majority of Chinese entrants across both web and mobile were application-layer products: image editors, video generators, creative tools. Meitu, the Chinese photo and design software company, placed five separate apps in the global mobile Top 50. All five serve primarily overseas users.

A structural pattern emerges. In the previous edition of the same ranking, 22 Chinese-origin applications featured in the global mobile Top 50. Nineteen of them targeted overseas audiences. The sixth edition suggests that ratio has held or widened. These products compete on functionality, design, and pricing in markets where Chinese brand recognition carries no inherent advantage. They succeed or fail on product merit.

When a16z mapped the global AI consumer landscape, three distinct regional ecosystems appeared: a Western bloc anchored by ChatGPT, Gemini, and Claude; a Chinese domestic bloc led by Doubao and DeepSeek; and an emerging Russian cluster around Yandex. The Chinese companies that show up in both domestic and global rankings are almost exclusively application-layer players. Model companies serve their home markets. Application companies cross borders.

Meitu: Selling Where Giants See No Market

Read more

Xiaomi’s 2025 annual results, released on March 24, read like a victory lap. Full-year revenue reached RMB 457.3 billion, crossing the RMB 400 billion mark for the first time, up 25 percent year-on-year. Adjusted net profit hit RMB 39.2 billion, a 43.8 percent increase and an all-time high. The smart electric vehicle business alone generated RMB 103.3 billion in revenue. The broader segment that houses it, which also includes AI research, robotics, and chip development, recorded positive income from operations for the first time.

By almost any conventional measure, this was Xiaomi’s best year. But the most important story in this earnings report has nothing to do with records.

One year ago, Xiaomi’s smartphone and AIoT segment generated 91 percent of total group revenue. In 2025, that share fell to 76.8 percent. A 14-percentage-point shift in revenue composition within a single year is not incremental adjustment. It is the financial signature of a company whose economic center of gravity is migrating from smartphones toward electric vehicles and artificial intelligence.

The shift is undeniable. What remains far less clear is how to read the financial results of a company in transit. The profit numbers the market is celebrating have become increasingly difficult to decompose, blending the economics of a maturing car business with the costs of an AI program whose commercial model remains unproven.

The RMB 0.9 Billion Fog

The segment Xiaomi calls “Smart EV, AI and Other New Initiatives” reported income from operations of RMB 0.9 billion for full-year 2025, the first time this division has finished a year in positive territory. Vehicle deliveries tripled to 411,000 units. Average selling prices rose 7.1 percent to RMB 251,000. By any operational reading, the car business has achieved genuine scale in its second year of sales.

The RMB 0.9 billion figure, however, requires careful unpacking. This segment does not measure the car business in isolation. CFO Alain Lam was direct on the earnings call: “You cannot just look at this segment as just an auto segment. It includes other new businesses.” Those businesses include Xiaomi’s trillion-parameter MiMo foundation model, a robotics program that recently demonstrated three hours of autonomous factory operation, in-house chip development, and the Xiaomi miclaw AI agent platform now in beta testing.

The segment generated RMB 25.8 billion in gross profit on RMB 106.1 billion in revenue, a 24.3 percent margin. Operating expenses reached RMB 24.8 billion, an 87.7 percent year-on-year increase, consuming nearly all of that gross profit. The RMB 0.9 billion is what remained. The question the earnings report leaves unanswered: how much of the RMB 24.8 billion relates to building and selling cars, and how much funds AI, robotics, and chip R&D?

Read more