The Stanford AI Index 2026, released in April, is among the most widely cited annual benchmarks of AI progress. Its 423 pages document a consistent pattern. China and the United States are converging on capability metrics. They are diverging on structural indicators that predict long-term leadership. For investors tracking China, the report separates signal from noise across five dimensions.

The model performance gap has effectively closed

As of March 2026, the top U.S. model leads the top Chinese model by just 2.7% on the Arena benchmark. Anthropic’s Claude Opus 4.6 scores 1,503. China’s Dola-Seed-2.0 Preview scores 1,464. The gap was far wider in 2023.

In February 2025, DeepSeek-R1 briefly closed within 0.4% of the leading U.S. model. Over the past year, the performance gap has fluctuated between near parity and low single digits.

This convergence has emerged from structurally different innovation environments, the report notes.

The U.S. still produced more top-tier models in 2025: 50 notable models to China’s 30. The absolute count matters less than performance at the frontier. The Arena leaderboard shows that China’s top models are competitive with all but the very best U.S. systems.

China dominates volume; the U.S. retains influence

The patent and publication data reveal a structural split likely to shape the next decade.

China holds 74.2% of global AI patents granted in 2024, up from under 20% a decade earlier. The United States accounts for 12.1%. China holds 97,990 AI patents, against 15,920 for the U.S.

On a per capita basis, South Korea leads at 14.3 patents per 100,000 people. China comes third at 7.0, ahead of the United States at 4.7.

In research publications, China accounted for 17.8% of AI papers in 2024, against 7.3% from the United States. Chinese AI publications also represented 20.6% of all AI citations globally, compared with 12.6% from the U.S. China’s share of the top 100 most-cited AI papers grew from 33 in 2021 to 41 in 2024.

The influence asymmetry cuts the other way. The U.S. accounts for more than 50% of all AI patent forward citations.Later inventions build more heavily on American foundational work than on Chinese patents. Chinese patents are frequently cited in U.S. filings. U.S. patents appear far less often in Chinese ones.

On the publication side, China’s government institutions contributed 25.1% of research output. In the U.S., industry leads at 24.6%. This reflects the different structural roles of the state in each country’s R&D pipeline.

Chinese AI research, while prolific, is not yet setting the reference standards that define technological direction globally.

The investment gap is large but understated

U.S. private AI investment reached $285.9B in 2025, more than 23 times China’s $12.4B in private funding. In generative AI specifically, U.S. investment exceeded the combined total of China and Europe.

The AI Index is explicit that this comparison understates China’s actual spending. Chinese government guidance funds have deployed an estimated $184B into AI firms between 2000 and 2023. These vehicles do not appear in private investment figures. The total AI capital deployment comparison between the two countries is therefore considerably closer than the private-funding gap suggests.

The structural difference matters. U.S. capital flows primarily through venture and growth equity, concentrating on frontier model companies and AI infrastructure. Chinese state capital operates through different vehicles with different timelines and incentives.

The U.S. created 1,953 newly funded AI companies in 2025, more than 10 times any other country. China’s entrepreneurial formation rate is not disclosed in the report. Organizational AI adoption data does show a different pattern. China and Europe posted the highest year-over-year increases in 2025 for generative AI use across business functions.

China’s robotics lead is widening

Industrial robotics is the dimension where China’s lead is most pronounced and accelerating fastest.

China accounted for 54% of global industrial robot installations in 2024, up from 51.1% in 2023. Global year-over-year growth was flat. The United States, Germany, and Italy all saw declines. China installed more industrial robots than the rest of the world combined. The margin is growing.

In humanoid robotics, Chinese vendors are scaling production at price points and unit volumes that resemble consumer hardware. Unitree’s R1 is priced from $4,900 and its G1 from $13,500.

AgiBot has manufactured approximately 10,000 units. Around 100 of its teleoperated humanoids run up to 17 hours per day. UBTECH’s Walker platforms integrate LLM-based planning with autonomous battery swapping for industrial deployment. Fourier Intelligence and DeepRobotics are also active in the segment.

By contrast, U.S. humanoid programs are operating in pilot mode. Figure AI’s Figure 02 spent 11 months on a BMW production line in South Carolina. It logged over 1,250 runtime hours and loaded more than 90,000 parts.

The scale and cost trajectory in China suggests Chinese hardware economics will define the next phase of industrial automation.

China surges in responsible AI research

China overtook the United States in responsible AI (RAI) research papers at major AI conferences in 2025.

China produced 812 accepted RAI papers at select conferences in 2025, compared with 394 from the United States. This reversed the 2024 picture, when the U.S. led with 788 papers to China’s 322. The AI Index notes that the reversal is consistent with China’s broader lead in AI publication volume. The scale and speed of the shift are nonetheless notable.

Over the full 2019-2025 cumulative period, the United States still holds the largest total of accepted RAI papers. China’s jump in 2025 is best understood as a structural shift in research focus. Chinese institutions are increasingly publishing on alignment, bias, interpretability, and governance.

What the data tells institutional investors

The AI Index 2026 maps a consistent picture across five dimensions.

China leads on volume: patents, publications, citations, and robot installations. The U.S. leads on influence: patent forward citations, notable model count, and private capital formation.

The frontier model performance gap is narrow but persistent. China is closing on capability, with model performance effectively at parity at the top of the leaderboard.

The U.S. retains a lead on talent attraction, though the number of AI researchers moving to the U.S. has dropped 89% since 2017. On governance, both countries remain outside formal international readiness assessments. Global trust in both trails trust in the EU.

For investors, the practical implications are sector-specific. In foundation models, the gap is narrow enough that China’s top-tier labs warrant tracking as direct competitors to U.S. frontier systems.

In industrial AI and robotics, China’s deployment scale and cost structure form a first-mover position hard to close. In responsible AI and governance, China’s increasing research presence signals intent to shape the emerging regulatory architecture. That has direct implications for standards-setting across Asian and developing-market jurisdictions.

The U.S.-China AI competition is no longer a story of one leader and one follower. It is a story of two systems strong on different dimensions, on different timelines, with different capital structures.

Share

The China Internet Network Information Center (CNNIC) publishes its Statistical Report on Internet Development twice a year.

The 57th edition, released in February 2026, covers data through December 2025. It is the closest Chinese equivalent to a Pew Research baseline on digital adoption — the authoritative census of who is online, what they use, and how Chinese internet infrastructure is evolving.

Its industrial AI and capital allocation data is drawn from IDC enterprise research and Ministry of Industry and Information Technology (MIIT) reporting. Most international coverage of the 57th report anchored on the consumer headline. The more consequential data sat two chapters deeper.

CIW Premium subscribers can download a more comprehensive report (April issue) here.

China’s generative AI user base: real scale, familiar shape

As of December 2025, 602M Chinese had used a generative AI product in the prior six months. That is up 353M from December 2024, a 141.7% increase. Population penetration reached 42.8%, a 25.2 percentage-point gain in twelve months.

The user count is real. It is also the wrong anchor for competitive analysis.

The St. Louis Federal Reserve’s November 2025 survey put US generative AI adoption at 54.6% of working-age adults, up roughly 10 percentage points YoY. China is behind on penetration, not ahead. GenAI adoption has outpaced both PC and internet diffusion curves globally — PC adoption sat at 19.7% three years after its mass-market launch; internet adoption at 30.1%. The speed is a worldwide phenomenon.

Use patterns are nearly identical across both markets. CNNIC reports 76% of Chinese GenAI users treat it primarily as a question-answering tool. 47.8% generate or process images and video. Just 10.8% generate code. OpenAI’s May 2025 analysis of 1.5M ChatGPT conversations found the same shape: 49% asking questions, 40% getting work done, 11% exploring ideas. Both markets remain in the shallow-tooling phase.

Consumer GenAI in China is additive tooling, not category-redefining behavior. The same is true in the West. The divergence begins one layer down.

Where China’s adoption curve actually diverged

Chinese industrial enterprise AI and agent penetration reached 47.5% in 2025, up from 9.6% in 2024. The figure originates from IDC enterprise research. NDRC cited it in its 29 August 2025 policy briefing. CNNIC reproduces it in the 57th report’s chapter on generative AI industry development.

A nearly fivefold increase in one year warrants scrutiny. McKinsey’s State of AI 2025 survey, based on 1,993 respondents across 105 countries, provides the benchmark: 88% of organizations now use AI in at least one function, up from 78% in 2024. But only 23% are scaling AI agents. Just 6% qualify as high performers attributing more than 5% of EBIT to AI.

The methodologies differ. McKinsey surveys enterprise respondents globally across functions. The IDC figure focuses on Chinese industrial enterprise deployment specifically. A clean comparison is not possible. The directional signal is clear. Global enterprise AI deployment remains stuck in the pilot loop McKinsey describes. Chinese industrial enterprises are moving from pilot to production at a materially faster pace.

The CNNIC report breaks down deployment across three functional buckets: research and design (33%), production manufacturing (24%), and management and operations (32%). The R&D-heavy distribution suggests Chinese industrial AI is concentrated in functions where productivity gains compound — not in the customer-facing use cases that dominate Western enterprise GenAI.

MIIT’s lighthouse factory program is the proof point. At designated lighthouse sites in 2025, AI penetrated more than 70% of business scenarios. Over 6,000 vertical-domain industrial models are deployed across the program. More than 1,700 smart-manufacturing equipment and software integrations have reached production scale.

If the penetration figure is directionally correct, Chinese manufacturing is compounding an AI productivity premium. Western industrial competitors cannot match it at current compute cost structures.

Read more

Two months after going public at a $6.5B valuation built on open-source credibility, MiniMax has locked its latest model behind a paywall. M2.7, released March 18, is the company’s first proprietary model.

It matches top US systems in software engineering at 1/50th the cost. And it helped build itself.

MiniMax is one of China’s “AI Tigers,” a cohort of foundation model startups racing to match US frontier labs. Yan Junjie, former vice president of SenseTime, founded the Shanghai-based company in 2022.

Its products span text, video, speech, and music AI, reaching 27.6M monthly active users across more than 200 countries. Over 70% of revenue comes from outside China, with the US accounting for roughly 20%.

MiniMax listed on the Hong Kong Stock Exchange on January 9, 2026, raising approximately $620M at HK$165 per share, according to Bloomberg. Shares doubled on debut, backed by cornerstone investors including Alibaba and Abu Dhabi’s sovereign wealth fund. Rival Zhipu AI had debuted one day earlier.

The company built its developer reputation on open-source models. M2, released in October 2025 under an MIT license, ranked first among all open-source systems on the Artificial Analysis Intelligence Index. M2.5 followed in February 2026, also open-source.

M2.7 breaks that pattern.

The shift carries implications for developers, investors, and the competitive landscape of Chinese AI.

Open-source to proprietary, a strategic U-turn

Before M2.7, MiniMax was a standard-bearer for open-source AI in China. M2 and M2.5 both shipped under permissive open-source licenses. Any developer could download, modify, and self-host the models.

Now, MiniMax has kept M2.7 closed. Developers can access the model through the MiniMax API and third-party model providers. They cannot download, fine-tune, or self-host it.

MiniMax is not the only Chinese lab making this shift. Zhipu AI (z.ai) released GLM-5 Turbo as a proprietary model in recent months. According to VentureBeat, Alibaba’s Qwen team is also reportedly moving toward proprietary development following the departure of senior leadership.

DeepSeek remains a notable holdout, keeping its V3 and R1 models open-source. The pivot is selective, not universal.

The pattern reflects a maturing business logic. Open-source was an acquisition strategy. When Chinese labs needed developer mindshare, freely available models attracted users who might otherwise default to OpenAI or Anthropic.

The strategy worked. MiniMax’s M2 became the top-ranked open-source model globally within weeks of release.

Proprietary is a monetization strategy. MiniMax is now publicly traded with a market capitalization of approximately HK$294B (roughly $38B). Shareholders expect defensible revenue streams. Giving away frontier models undercuts the API business that generates that revenue.

ZhenFund founding partner Huang Mingming invested in 6 consecutive MiniMax funding rounds. He framed the company’s appeal in terms of the “impossible triangle” of high performance, low cost, and commercial scalability, according to TechNode. Closing model access is one way to hold that triangle together while protecting margins.

At the IPO, Yan Junjie pledged to “ensure cutting-edge AI truly serves everyone.” M2.7 tests how far that mission stretches when shareholders expect returns.

For the past year, a useful shorthand described the global AI landscape. Chinese labs were open. US labs were closed.

The framing no longer holds. Global buyers now face proprietary options on both sides. Selection criteria increasingly center on capability, cost, ecosystem fit, and jurisdictional risk.

For enterprises that built production workflows on M2 and M2.5, the switch raises a practical question. Open-source models allowed self-hosting. M2.7 requires API dependency on Chinese infrastructure.

Teams that adopted MiniMax for its openness must now re-evaluate their vendor risk.

The model that helped build itself

M2.7’s most significant technical claim is not a benchmark score. It is the model’s role in its own creation.

MiniMax used earlier M2 versions to build an automated research system. The system manages data pipelines, training environments, and evaluation infrastructure. According to MiniMax, M2.7 then ran entirely autonomously through an iterative improvement loop.

The model analyzed its own failure patterns, planned code changes, and modified its training infrastructure. It ran evaluations, compared results, and decided whether to keep or revert each change. It executed more than 100 rounds of this cycle without human intervention.

During the process, M2.7 discovered optimizations that human engineers had not prioritized. These included parameter tuning, automated bug-pattern detection across files, and loop prevention in the training workflow.

The result, according to MiniMax, was a 30% performance improvement on internal evaluation sets. The company claims the system handled 30-50% of the operational work that would normally require human ML engineers.

The innovation is structural, not incremental. Most AI models are passive artifacts. Humans write the training code, tune the parameters, and fix the failures.

M2.7 inverts part of that process. The model actively manages portions of its own development pipeline. It identifies issues and applies fixes faster than a human team could cycle through them.

These claims deserve scrutiny. The 30-50% figure refers to operational tasks such as pipeline monitoring and evaluation reruns. It does not mean the model performed half of its own fundamental research.

Independent benchmarks also show mixed signals. On at least one coding test, according to VentureBeat, M2.7 actually scored lower than its predecessor M2.5. Self-evolution does not guarantee improvement on every axis.

MiniMax is not alone in exploring this approach. OpenAI recently described a similar process for GPT-5.3 Codex, according to The Decoder. The Codex team used early model versions to find bugs, manage deployment, and evaluate results during training.

Why it matters for enterprise leaders: If self-evolving training loops deliver consistent gains, the economics of AI research change. Labs with effective self-improvement systems could reduce the headcount needed per model generation. They could also compress development timelines.

The competitive moat would shift from model quality to iteration speed. For organizations evaluating AI vendors, the key question becomes whether a provider’s R&D engine can sustain compounding improvements.

Matching US frontier models at 1/50th the cost

The cost structure is the sharper competitive edge. M2.7 is priced at $0.30 per million input tokens and $1.20 per million output tokens. According to WaveSpeed AI, M2.7 costs roughly 1/50th as much as Claude Opus 4.6 on input tokens. On output, it costs roughly 1/60th as much.

In practical terms, an agent workflow that costs $100 to run on Opus would cost roughly $2 on M2.7. With automatic cache optimization, the effective blended cost drops to approximately $0.06 per million tokens.

M2.7’s efficient architecture activates only 10B parameters per query, keeping compute costs low while delivering high reasoning capability. The model runs at approximately 100 tokens per second, about 3 times faster than Opus.

The benchmark results support the pricing story. On SWE-Pro, a benchmark for real-world software engineering, M2.7 scored 56.22%, according to MiniMax. The score matches GPT-5.3-Codex.

On SWE-bench Verified, the model scored 78%, outperforming Claude Opus 4.6’s 55%.

For office tasks such as spreadsheet and document editing, M2.7 achieved the highest ELO score (1,495) on GDPval-AA among open-API models. In one demo, M2.7 independently read TSMC annual reports, built a sales forecast, and generated a presentation.

M2.7 integrates with major AI coding tools, including Cursor and Claude Code. Developers can configure existing environments to point at MiniMax’s API endpoint. The switching cost is low.

The ease of integration cuts both ways. It lowers the barrier to adoption. It also means developers can leave just as easily if a competitor matches the price.

The post-IPO valuation test

Shares opened at HK$235.40 on January 9, 2026, a 42.67% premium over the IPO price. They closed the first day at HK$345, up 109%, according to CNBC. The IPO was oversubscribed 1,837 times by retail investors.

As of early April 2026, the stock trades around HK$949.50, according to Investing.com consensus estimates. All 6 covering analysts rate it a buy, with an average 12-month target of HK$1,092.

The fundamentals tell a more complicated story. According to PitchBook, trailing 12-month revenue as of December 2025 was $79M. Revenue in the first 9 months of 2025 grew 170% YoY, with gross margins of 69.4%, according to TechNode.

Net losses in the same period reached $512M.

A $38B market cap on $79M in trailing revenue translates to a revenue multiple above 480 times. Even adjusting for growth, that pricing requires sustained rapid growth. No Chinese foundation model company has yet demonstrated a clear path to profitability.

The bull case rests on three pillars. First, MiniMax generates over 70% of revenue outside China, reducing single-market risk. Second, its API pricing undercuts every major competitor by an order of magnitude.

Third, self-evolving model development could compress R&D cycles and lower the cost of each successive model release.

The bear case is equally clear. The company allocated 90% of IPO proceeds to R&D, signaling years of heavy spending ahead. The AI model market is commoditizing rapidly. Multiple open-source alternatives exist at the same capability tier.

The Disney, Universal, and Warner Bros. Discovery copyright lawsuit against Hailuo AI, filed in September 2025, adds legal risk.

Geopolitical exposure adds a structural concern. Over 70% of MiniMax’s revenue comes from outside China. The US alone accounts for roughly 20%, according to eWeek.

With M2.7 now proprietary and API-only, enterprise customers have no option to self-host. If US export controls or data governance restrictions expand to cover Chinese model API access, MiniMax’s fastest-growing revenue stream faces regulatory risk.

For enterprises in regulated or government-facing industries, a proprietary Chinese API with no self-hosting fallback raises vendor risk. Expect risk committees to scrutinize these dependencies closely.

From model weights to training flywheels

M2.7 is a test case for whether AI labs can escape the commoditization trap by building models that improve themselves. If self-evolving training systems scale, the competitive moat shifts from model quality to training infrastructure.

For investors, the measurable proxy is release cadence. MiniMax shipped M2 in October 2025, M2.5 in February 2026, and M2.7 in March 2026. The gap between releases is shrinking.

Whether that pace holds will indicate whether self-evolution is a real structural advantage or a one-cycle narrative. A sustained acceleration would validate the thesis. A plateau would suggest the gains are front-loaded.

The question for strategists watching China’s AI sector has changed. The question is no longer who can build the best model. It is who can build the model that builds the next model.

China’s largest publicly listed tech companies released their Q4 and full year 2025 earnings between late February and mid March 2026. Baidu reported on February 26. Alibaba and Tencent both reported on March 18.

The results arrived at a moment when the global AI race is entering its most capital intensive phase. For institutional investors and corporate strategists, these earnings offer the sharpest comparison yet of three divergent AI strategies operating at scale in the same market.

The central question is not whether AI investment pays off. It is which architecture for investing pays off fastest: vertical integration across the full stack, incremental embedding into proven profit pools, or wholesale replacement of a legacy business.

PDD Holdings and Meituan are excluded. PDD posted no material AI narrative. Meituan reports on March 26.

Alibaba goes full stack on AI, absorbs the margin hit

No Chinese tech company is building across as many layers of the AI stack as Alibaba. Proprietary chips from T-Head. Foundation models through Qwen. Cloud infrastructure. Consumer applications with agentic transaction capabilities. The Q4 earnings show both the momentum this strategy is generating and the margin it is consuming.

Cloud Intelligence Group revenue reached RMB 43.3B, a 36% increase YoY. External customer revenue grew 35%, accelerating from prior quarters. AI related product revenue delivered triple digit YoY growth for the 10th consecutive quarter. Total company revenue was RMB 284.8B (+2% YoY, or +9% excluding the disposed Sun Art and Intime businesses).

On the consumer side, the Qwen app surpassed 300M monthly active users across all platforms. Alibaba integrated the app with Taobao, Tmall, Taobao Instant Commerce, Amap, Fliggy, and Alipay in January 2026.

Following a Chinese New Year promotional campaign launched on February 6, approximately 140M users had their first AI driven shopping experience through Qwen’s agentic features by the end of the month.

Alibaba’s foundation model suite, Qwen, surpassed 1B cumulative downloads on Hugging Face as of January 21, 2026. The company launched Qwen3.5 in February with improved multimodal reasoning, coding, and agentic task performance. Qwen is now the most widely used open source model family globally, according to Alibaba, based on download metrics.

T-Head, Alibaba’s chip design subsidiary, has brought its proprietary GPU into production at scale. The chip supports end to end AI workloads from training through inference. It is compatible with mainstream AI frameworks.

By pairing T-Head hardware with Qwen models and Alibaba Cloud infrastructure, Alibaba is assembling a vertically integrated AI offering that spans the entire value chain. That integration is the strategic moat. The cost is visible in the earnings.

Adjusted EBITA fell 57% YoY to RMB 23.4B. Net income dropped 66% to RMB 15.6B. Free cash flow fell 71% to RMB 11.3B.

The compression came from heavy investment in quick commerce, user experience, and AI infrastructure. The company has committed to spending at least RMB 380B on AI and cloud over 3 years. CEO Eddie Wu noted that figure may prove to be too small.

The question for investors is whether the margin pain is temporary or structural.

Tencent embeds AI into profit centers, avoids building a standalone AI business

Tencent took the opposite approach to Alibaba. Rather than building AI as a distinct business line, it wired AI into the products that already generate profit: advertising, gaming, cloud services, and the WeChat ecosystem. The result is the only company among the three that grew both revenue and profit in the period.

Full year 2025 revenue reached RMB 751.8B, up 14% YoY. Net profit rose 16% to RMB 224.8B. Free cash flow grew 18% to RMB 182.6B.

The AI impact shows up most clearly in two areas.

Business services revenue accelerated to 22% growth in Q4, driven by higher cloud services revenue including AI related workloads. Tencent Cloud achieved profitability at scale in 2025. The company attributed the acceleration to enterprise demand across domestic and international markets. Higher e-commerce technology fees tied to WeChat Mini Shops also contributed.

Marketing Services revenue reached a record RMB 145B for the full year, up 19% YoY. AI powered ad targeting, the automated campaign solution AIM+, and higher engagement across Video Accounts and Mini Programs contributed directly to this growth. AI also enhanced game engagement, content recommendation, and user retention across Tencent’s portfolio.

Capital expenditure hit a record RMB 79.2B. R&D spending reached RMB 85.8B. The company spent RMB 18B on AI products specifically in 2025, according to CNBC. It plans to double that in 2026.

President Martin Lau noted on the earnings call that 2025 capex came in below target due to GPU supply constraints. The company will expand computing power acquisition through both direct purchase and flexible leasing in 2026.

On the model front, Tencent is preparing to launch Hunyuan 3.0, its next generation large language model, in April 2026. The model is in internal testing.

The company also announced plans for an advanced AI agent within WeChat, branded as QClaw. The agent integrates as a mini program and handles tasks including file management, commerce, bookings, and PC control. This positions WeChat as a platform for agentic AI rather than just messaging.

Tencent also restructured its Hunyuan development teams into dedicated LLM and multimodal divisions, signaling a more focused approach to model development.

Tencent’s strategy treats AI as an operating lever rather than a separate bet. Chairman Pony Ma stated in the earnings release that AI capabilities improved ad targeting, supported game engagement, and delivered improving cloud revenue growth. He did not describe a standalone AI business. That framing is deliberate.

By embedding AI into existing profit pools, Tencent avoids the margin compression that Alibaba is absorbing. The trade off is that Tencent may be moving more slowly on frontier model capabilities and consumer AI products.

Baidu’s AI pivot hits an inflection point, but the legacy business keeps shrinking

Baidu faces the most structurally complex situation of the three. It is not supplementing an existing growth engine with AI. It is replacing one. The Q4 results show AI revenue growing fast enough to matter, while legacy search declines fast enough to offset the gains at the consolidated level.

Baidu’s AI powered business crossed RMB 11.3B in Q4, representing 43% of what the company now calls Baidu General Business (formerly Baidu Core) revenue. A year earlier, that share was 26%. For the full year, AI powered business revenue reached RMB 40B, up 48% YoY.

Within that category, the numbers tell different stories.

AI Cloud Infra revenue was RMB 5.8B in Q4, with subscription based revenue from AI accelerator infrastructure growing 143% YoY. Full year AI Cloud Infra revenue reached approximately RMB 20B, up 34% YoY. This segment is Baidu’s fastest growing and highest conviction bet.

AI native Marketing Services surged 110% YoY in Q4 to RMB 2.7B. Full year revenue of RMB 9.8B still represents a small share of total revenue. AI Applications contributed RMB 2.7B in Q4. ERNIE Assistant reached 202M MAU in December 2025.

Apollo Go, Baidu’s autonomous ride hailing service, delivered 3.4M fully driverless rides in Q4 with weekly rides peaking above 300,000. Total rides grew over 200% YoY. Cumulative rides surpassed 20M by February 2026.

The fleet has logged over 300M autonomous kilometers, including over 190M fully driverless kilometers. Apollo Go expanded to 26 cities globally, including partnerships with Uber and Lyft in London and Dubai. The international expansion builds regulatory and operational moats that are difficult for later entrants to replicate.

Baidu released an updated ERNIE 5.0 foundation model in January 2026. The company also announced the proposed spin off and separate listing of Kunlunxin, its AI chip subsidiary. Separately, Baidu unveiled a new $5B share repurchase program and its first ever dividend policy.

The challenge is on the other side of the ledger.

Total Baidu revenue fell 3% YoY for full year 2025 to RMB 129.1B. Legacy Business, which consists primarily of traditional search advertising, generated RMB 12.3B in Q4, declining as AI powered alternatives cannibalize the core. A RMB 16.2B impairment charge on a core asset group pushed full year operating results into a loss of RMB 5.8B. Excluding the impairment, operating income was RMB 10.4B.

Non-GAAP operating income for the full year was RMB 15B at a 12% margin. Operating cash flow turned positive in the second half of 2025, generating RMB 3.9B combined after negative cash flow in the first half.

The impairment charge underscores the shrinking value of the traditional business. Baidu’s ability to sustain AI investment depends on whether AI Cloud Infra and Apollo Go can scale revenue faster than legacy search contracts.

Three strategies, one shared problem

The Q4 earnings reveal three distinct approaches to AI among China’s tech leaders.

Alibaba is building full stack: models, cloud infrastructure, proprietary chips, and consumer applications. It accepts severe near term margin compression in exchange for the broadest possible AI moat.

The Qwen ecosystem and T-Head chip advantage give Alibaba vertical integration that no Chinese peer currently matches. But the steep declines across every profitability metric test investor patience.

Tencent is embedding AI into existing profit centers. It uses AI to improve advertising yield, game engagement, cloud monetization, and platform stickiness. Profitability and cash flow remain strong.

The risk is that this defensive approach leaves Tencent behind on frontier models and consumer AI products. The planned Hunyuan 3.0 launch and WeChat AI agent suggest the company recognizes that risk.

Baidu is replacing its core business with AI. Its AI powered revenue is growing faster than either peer. Apollo Go’s global expansion gives Baidu a unique asset in autonomous mobility. But the declining legacy business and significant impairment charge create a narrower margin for error.

One theme connects all three: AI investment is no longer discretionary. Alibaba plans to spend over RMB 380B on AI and cloud over 3 years. Tencent plans to double its AI spending in 2026. Baidu is channeling resources into AI Cloud Infra and autonomous driving as it phases out legacy search. The capex cycle is accelerating regardless of near term profit pressure.

For global investors and corporate strategists, the takeaway is structural. The AI transition for China’s tech giants is no longer a strategic option. It is the business.

The question is no longer whether these companies will spend on AI. It is which company’s AI revenue will be first to self-fund its own capex cycle.

Three metrics will tell the story over the next two quarters. For Alibaba: whether Cloud external revenue growth holds above 30% as capex continues to accelerate.

For Tencent: whether AI related revenue is broken out as a separate line item, signaling confidence that AI is a growth driver rather than an embedded feature. For Baidu: whether AI Cloud Infra growth can outpace the legacy search decline rate, narrowing the gap at the consolidated revenue level.

On February 26, Hefei-based Origin Quantum Computing Technology Co. released its quantum computer operating system, Origin Pilot, for public download. The announcement was reported by Xinhua and Global Times.

No other company has made a full quantum operating system available for local installation. The UK-based Riverlane released Deltaflow.OS in 2020 under more restrictive terms. Origin Pilot is the first quantum OS freely downloadable for local deployment.

A quantum operating system manages the relationship between software and quantum hardware. It schedules tasks, calibrates qubits, and provides the driver layer that lets developers interact with physical quantum chips.

In classical computing terms, it is the equivalent of Windows or Linux. In quantum computing, the challenge is greater. The processor’s basic units of computation are unstable, noisy, and require constant recalibration.

IBM, Google, and Microsoft all operate sophisticated internal quantum operating systems to power their cloud-accessible quantum hardware. None offers that underlying OS as a downloadable package. Origin Pilot occupies a space the Western quantum industry has left empty.

What Origin Pilot gives you

The strategic value of Origin Pilot is that it is hardware-agnostic, locally deployable, and free. It supports three major qubit types: superconducting, trapped ion, and neutral atom.

It integrates quantum processors with classical CPUs, GPUs, and AI acceleration in a unified workflow. And it ships in two tiers: a free Community Edition focused on noise mitigation and hybrid compilation, and an Enterprise Edition with post-quantum cryptography tools for industrial use, according to Quantum Computing Report.

The OS is now in its 4th major version.

First described in a 2021 research paper, it has undergone multiple upgrade cycles. According to Origin Quantum, it handles resource scheduling, parallel quantum task processing, automatic qubit calibration, and software-hardware coordination. The company calls it a “quantum-classical-intelligent” operating system.

For anyone building a quantum computer from components, this is the layer that makes the parts work together.No Western company currently offers that layer as a download.

Origin Wukong: the proof point

Origin Pilot is not theoretical software. It manages a live quantum computer that has processed real workloads from users across the globe.

The system powers Origin Wukong, China’s third-generation superconducting quantum computer. The Wukong chip contains 198 physical qubits: 72 working qubits and 126 coupler qubits that manage interactions between them. This tunable-coupler design is similar to approaches used by Google on its Sycamore and Willow processors.

The company reports that the Wukong cloud platform has attracted over 30M visits from users in more than 120 countries since January 2024, with over 339,000 quantum computing jobs executed across finance, biomedicine, and materials science. These are company-reported figures, according to Origin Quantum. Its 4th-generation control system, Tianji 4.0, deployed in 2025, supports over 500 qubits.

The company behind these systems is Origin Quantum, founded in 2017 as a spinout from the University of Science and Technology of China (USTC). Professor Guo Guoping and Academician Guo Guangcan lead the company.

According to PostQuantum, it is the only quantum computing company that builds the entire stack in-house: quantum processor chips, dilution refrigerators, control electronics, an operating system, a programming framework, and a cloud platform.

The company has raised approximately $150-165M in funding, entirely from domestic Chinese investors, according to PitchBook. Key backers include the China Internet Investment Fund (under the Cyberspace Administration of China), Guoxin Fund (State Council), Shenzhen Capital Group, and CITIC Securities. The deliberate exclusion of foreign capital reflects its positioning as a national security asset.

Why this is not just a framework update

The distinction between a quantum SDK and a quantum OS matters. IBM’s Qiskit and Google’s Cirq are powerful programming frameworks. They let developers write quantum circuits in Python and submit them to cloud-hosted hardware. They are not downloadable operating systems that manage the full hardware stack locally.

According to PostQuantum, the release connects directly to the Quantum Open Architecture movement. Institutions and governments worldwide are seeking to build sovereign quantum computing capabilities from modular, multi-vendor components. That effort requires an integration layer. It requires an OS.

The Western landscape offers few alternatives. A university or national lab assembling a quantum computer from parts has three choices: build an OS from scratch, contract with an emerging quantum systems integrator, or download Origin Pilot for free from China.

The strategic parallel to DeepSeek is direct. DeepSeek demonstrated that open-sourcing a capable AI model forces competitors into a response. It attracts a global developer community. It establishes conventions before others can.

But the quantum version of this play is more potent. In AI, DeepSeek entered a market crowded with open-source alternatives. In quantum operating systems, Origin Pilot enters a market with none.

Whoever provides the OS defines the driver standards, API conventions, and hardware abstraction patterns that others build against. Guo Guoping told Global Times that making Origin Pilot globally available marks a deliberate shift from “closed-door tech innovation” toward open-source ecosystem development. The goal is not just adoption. It is standard-setting.

The 15th Five-Year Plan puts quantum at the center

Origin Pilot’s release arrives at a moment of elevated policy support. China’s 15th Five-Year Plan (2026-2030), unveiled on March 5, identifies quantum technology as the first among seven “future industries” positioned to become new economic growth engines, according to China Briefing.

Read more

China’s silver economy describes the commercial ecosystem built around consumers aged 60 and above. The term entered mainstream business vocabulary in China around 2015, coinciding with the government’s first formal policy acknowledgment that population aging would reshape domestic consumption patterns over the following decades.

By H1 2025, the demographic reality is concrete. CNNIC‘s Digital Consumption Development Report for 2025 counts 117M consumers aged 60 and above actively participating in digital commerce.

That figure represents 52% internet penetration within the age group, a rate that has risen steadily. Mobile-first platforms have lowered the technical barrier to entry. The cohort itself has aged upward from a generation with higher baseline digital literacy.

The silver economy is frequently analyzed through a deficit lens: lower incomes, lower digital literacy, lower spend per transaction. The CNNIC data does not support this framing uniformly. It reveals a cohort with strong habitual purchase behavior, above-average healthcare spend, and documented vulnerability to specific commercial tactics.

Understanding all three dimensions simultaneously is the prerequisite for any credible analysis of this segment’s commercial value and regulatory risk.

72.7% habitual shoppers: the behavioral foundation

The first corrective to the low-value assumption is the habitual shopping rate. 72.7% of silver economy digital consumers are classified as habitual online shoppers. They make regular, recurring purchases through digital channels rather than treating online shopping as an occasional or experimental behavior.

This figure sits above the overall digital consumer average. It indicates that for the 117M who have adopted digital shopping, the behavior has become embedded in daily routine rather than remaining a supplementary channel. The implication for platforms is that silver economy users, once acquired, show strong retention characteristics.

59.4% read product reviews before purchasing. This is a further corrective to the assumption that older consumers are less discerning or more easily manipulated than younger ones. A cohort that systematically consults peer reviews before transacting is exhibiting informed purchase behavior. It is not the behavior of a passive or vulnerable consumer across the board.

51.2% prioritize quality within their budget. The dominant purchase logic among silver economy consumers is value-seeking rather than price-seeking. They are not buying the cheapest option available. They are buying the best option within a defined spending threshold. This distinction matters for brand strategy. Premium positioning with clear quality evidence performs better with this cohort than pure price competition.

Why it matters for platforms: The habitual purchase rate, review-reading behavior, and quality-within-budget logic collectively describe a consumer segment with above-average loyalty potential and below-average price sensitivity.

Platforms that have invested in trust signals, review infrastructure, and quality certification for products targeting older consumers are positioned to capture a disproportionate share of this segment’s spending.

Healthcare: the high-value spend concentration

The most commercially significant finding in the silver economy data is the healthcare spend profile. 26.1% of silver economy digital consumers spend RMB 300 or more per online healthcare transaction, exceeding the overall user average by 5.3 percentage points.

Online healthcare in China encompasses medical consultation platforms, prescription drug delivery services, health supplement e-commerce, and medical device purchasing. The category is growing rapidly as digital health infrastructure matures. The 60-and-above cohort is also developing confidence in online medical channels.

The 26.1% figure at RMB 300 or more per transaction is notable for two reasons. First, it contradicts the assumption that elderly consumers concentrate their digital spending in low-ticket categories. The healthcare spend concentration shows above-average transaction values in a high-frequency, high-need category.

Second, the figure implies significant total market value. 117M users with 26.1% spending RMB 300 or more per healthcare transaction represents a substantial addressable market even before accounting for transaction frequency.

Read more

China’s rural digital economy is one of the most consequential and least precisely analyzed segments in global consumer research. The standard framing treats rural China as a lagging market. Lower income, lower digital penetration, lower average spend. That framing was accurate in 2015. It is increasingly inadequate in 2025.

CNNIC‘s Digital Consumption Development Report for 2025 provides the most granular publicly available dataset on rural digital consumption behavior. It is based on a 30,000-respondent telephone survey covering all 31 provinces, with rural respondents defined as residents of townships and villages outside designated urban administrative boundaries.

The H1 2025 data describes a rural digital consumer base that has largely completed the adoption phase of online shopping and is entering the depth phase.

The gap metrics, the income data, and the forward-looking spending expectations all point in the same direction. Rural China is not catching up to urban consumption patterns. It is developing its own consumption profile. Its category priorities and platform preferences require separate analysis from the urban market.

83.2% penetration: the adoption phase is largely complete

The headline rural digital consumption figure is the online shopping penetration rate. 83.2% of rural netizens shopped online in H1 2025, according to CNNIC. The urban equivalent is 88.4%. The 5.2 percentage point gap has narrowed by 4.9 percentage points since the end of 2024.

At 83.2%, rural online shopping penetration is not a developing market metric. It is a near-mature market metric. The comparable figure for several developed economies in the same period sits below 80%.

Rural China’s e-commerce penetration rate is not a sign of catching up. It is a sign of structural completion.

The remaining 5.2 percentage point gap is not uniformly convertible. It is concentrated among the oldest rural residents, those with the lowest digital literacy, and those in remote areas with the least reliable logistics infrastructure.

Converting these remaining non-shoppers to online purchasers is a long-cycle project requiring infrastructure investment rather than platform marketing.

The strategic implication is the same one that applies to the overall digital consumer base documented in CIW’s analysis of China’s 46.5% digital consumption threshold: the rural market has shifted from an acquisition story to a spend-depth story.

The relevant competitive question is no longer how to get rural consumers online. It is how to capture a larger share of what they are already spending online.

Income growth: rural outpaces urban by 1.5 percentage points

The income data is the most structurally significant finding in the rural consumption section of the CNNIC report. Rural disposable income reached RMB 11,936 in H1 2025, growing at 6.2% YoY. Urban disposable income grew at 4.7% YoY over the same period.

The 1.5 percentage point rural income growth premium over urban is not a temporary distortion. It reflects several converging structural factors.

Rural labor migration patterns have shifted as urban manufacturing employment has consolidated. This has driven higher remittance flows back to rural households. Agricultural commodity prices have been supported by domestic food security policy. Rural producers have benefited directly. Government transfer payments to rural households have increased as part of the domestic consumption stimulus framework.

The income growth differential compounds over time. A rural household growing income at 6.2% YoY closes the absolute income gap by a measurable increment each year. An urban household growing at 4.7% YoY does not. The rural consumer in 2030 will have meaningfully higher purchasing power relative to urban peers than the rural consumer in 2025.

A consumer whose income is growing faster than average requires different positioning than one at a stable income level. The rural consumer is the marginal income growth story in China’s household sector.

The geographic dimension of the income data adds nuance. Income growth is not uniform across rural China. Eastern coastal rural areas, with stronger integration into manufacturing supply chains and proximity to urban labor markets, show higher growth rates than western interior rural areas.

Investment strategies calibrated to “rural China” as a monolith will systematically misallocate against the actual income distribution.

Trade-in subsidies: policy transmission into rural demand

The government’s trade-in subsidy programs, which drove RMB 2.9T in overall sales in H1 2025, showed strong rural penetration. 13.5% of rural netizens participated in subsidy programs, with 43.4% of rural subsidy participants buying appliances and 30.5% buying 3C products (consumer electronics).

The rural subsidy participation data matters beyond the aggregate number. Trade-in programs were structured to route purchases through digital channels.

For many rural consumers, a subsidized appliance purchase is their first high-ticket online transaction. The behavioral consequence is a first digital purchase in a high-ticket category.

First digital purchases of major categories tend to anchor subsequent behavior. A rural consumer who successfully purchased an appliance online through a subsidy program has demonstrated to themselves that the channel works for high-value transactions. The probability of their next appliance purchase also being digital is substantially higher than it was before the subsidy transaction.

The 43.4% appliance share among rural subsidy participants reflects genuine pent-up demand in the rural market. Rural appliance ownership rates for certain categories including dishwashers, dryers, and certain premium kitchen appliances remain significantly below urban rates. The subsidy programs have functioned as a demand-release mechanism for categories that rural consumers wanted but had not yet purchased.

30.5% buying 3C products through subsidy programs reflects a different dynamic. Consumer electronics in rural areas are increasingly driven by the same preference for AI-enabled smart terminals documented in CIW’s analysis of China’s GenAI consumer surge.

Rural Gen Z consumers show nearly identical smart product purchasing preferences to their urban counterparts. The price sensitivity differs. The product preference does not.

Platform preference: online leads offline decisively

The platform preference data in the rural subsidy context reveals an important structural shift. 66.3% of rural subsidy program participants preferred online platforms over offline stores (55.3%) for their subsidy-related purchases.

This figure is significant because it inverts the common assumption that rural consumers default to offline channels for major purchases due to familiarity and trust. For subsidy-mediated transactions, online platforms are the preferred channel.

The explanation lies in program architecture. Online platforms provided simpler subsidy verification, broader product selection, and faster fulfillment than local offline retailers in many rural areas.

The preference for online over offline in rural major-purchase contexts is a behavioral shift that will persist beyond the subsidy programs. Consumers who have completed a high-ticket online purchase successfully do not revert to offline channels for subsequent similar purchases.

The subsidy programs have functionally expanded the category range of rural online purchasing beyond the low-ticket goods that historically dominated rural e-commerce.

Rural platform preference among the broader shopping population shows Pinduoduo as the dominant operator, consistent with its performance in the silver economy segment.

The platform’s price-competitive positioning, social sharing mechanics, and group-buying infrastructure align with rural consumer behavior. Premium-positioned platforms do not replicate this alignment.

The services gap: where rural digital penetration lags most

Read more

Tencent Cloud engineers set up installation booths outside the company’s Shenzhen headquarters. The queue stretched across the north plaza. Attendees carried NAS devices, MacBooks, and mini PCs.

Some had flown in from Hangzhou the night before. A retired aviation engineer waited alongside a 9-year-old primary school student. They were all there to install OpenClaw.

The open-source AI agent framework, nicknamed “lobster” for its claw-shaped icon, has become the fastest-growing project in GitHub history. It surpassed 247,000 stars in just a few months, overtaking Linux and React. According to CGTN, nearly 1,000 people attended the Shenzhen event.

OpenClaw was created by Austrian developer Peter Steinberger in November 2025. It is an autonomous AI agent that connects large language models to messaging platforms. It can browse the web, execute shell commands, manage files, write code, and operate across applications.

Unlike chatbots that respond to prompts, OpenClaw runs tasks continuously and autonomously. For China’s cloud vendors, model companies, and device makers, this is not a community event. It is a land-grab.

Why China adopted OpenClaw faster than anywhere else

OpenClaw’s appeal in China runs deeper than novelty. It addresses a structural bottleneck that has stalled enterprise AI adoption for years.

China’s large enterprises average more than 150 independent IT systems, according to TMTPost. Roughly 60% are legacy platforms with no API documentation, no maintenance, and no open interfaces.

Traditional agent deployment requires mapping APIs, restructuring permissions, coordinating vendors, and rebuilding data pipelines. This “heavy integration” path is slow, expensive, and fragile.

OpenClaw bypasses it entirely. The framework operates at the screen level through visual recognition. It identifies buttons, text fields, and input boxes.

It then completes tasks through simulated mouse clicks and keyboard inputs. No API access required.

A late-2025 survey by Anthropic and Material found that nearly half (46%) of tech leaders cite system integration as the primary AI deployment barrier. A tool that skips integration altogether changes the calculus.

The tradeoffs are significant. OpenClaw takes 15 to 30 seconds per step, compared with 1 to 3 seconds for API-integrated agents.

A “log in and export a report” task might take an integrated agent 5 seconds. OpenClaw can take 3 to 5 minutes. But for enterprises trapped behind closed interfaces, a slow workaround beats no workaround.

Cloud vendors sell the OpenClaw shovels

China’s major cloud providers moved with striking speed.

5 providers launched one-click deployment within weeks. Tencent Cloud, Alibaba Cloud, JD Cloud, Volcano Engine (ByteDance), and Baidu Intelligent Cloud all rolled out OpenClaw services.

According to PANews, Alibaba Cloud created a dedicated “Coding Plan” API subscription for OpenClaw users. Tencent Cloud crossed 100,000 OpenClaw users on its Lighthouse lightweight servers.

A cottage industry emerged in parallel. On Taobao and Xianyu, paid installation services appeared overnight, charging 100 to 500 RMB per remote setup. The top-selling Taobao store logged over 1,000 orders, according to TMTPost.

The business logic is straightforward. Installation is free. Cloud server rental, bandwidth, and API calls are not.

Tencent’s official documentation states it plainly: “OpenClaw comes from the open-source community. The cloud app is free.” The next line adds: “Cloud servers and APIs are billed based on actual usage.”

The strategy mirrors China Telecom’s PHS playbook from the early 2000s, according to TMTPost. Attract users with low entry barriers. Retain them with recurring service fees.

PHS thrived not on product superiority. The carrier needed a user-acquisition vehicle.

Chinese model companies reap the OpenClaw windfall

OpenClaw is a “token black hole.” Each task involves sustained, multi-step API calls. Context windows balloon rapidly.

A single active session can exceed 200,000 tokens. Daily consumption for a power user can reach 50M tokens, according to TMTPost.

This created a sharp demand spike for model APIs. The winners were Chinese companies offering the best price-to-performance ratio.

On February 24, the top 10 models on OpenRouter consumed roughly 8.7T tokens. Chinese models accounted for 5.3T, or 61%, according to TMTPost. Four of the top 5 by call volume were Chinese: MiniMax M2.5, Kimi K2.5, Zhipu GLM-5, and DeepSeek V3.2.

By early March, Step 3.5 Flash from Stepfun had climbed to the global top spot.

The commercial impact was immediate. Moonshot AI’s Kimi K2.5 became the default free model for OpenClaw after version 1.30. OpenClaw subsidized compute costs.

Read more

China’s generative AI landscape reached a new phase in 2025. The term “generative AI” in China broadly encompasses large language model applications, AI-powered search and shopping tools, AI image and content generators, and the expanding category of AI-embedded smart terminals including phones, tablets, earbuds, and glasses.

CNNIC‘s Digital Consumption Development Report for 2025 tracks GenAI adoption as a discrete consumer behavior category for the first time. Its methodology captures users who have actively engaged with a generative AI product or feature in a given period, based on a 30,000-respondent telephone survey across 31 provinces.

The December 2024 baseline and H1 2025 figure together produce the sharpest adoption curve in the report. 515M users in six months is not a projection. It is a measured user count that places China’s GenAI consumer base larger than the population of most G20 economies.

For the macro picture of China’s digital consumption landscape, including the RMB 9.37T total and the services digitization thesis, see CIW’s companion analysis.

The doubling in context: why speed matters more than scale

Large user numbers in China require calibration. A country of 1.4B people with 1.123B internet users generates large denominators naturally. The 515M GenAI figure needs to be read against the pace of adoption, not just the absolute count.

The doubling from December 2024 to June 2025 occurred in a six-month window. For comparison, China’s mobile internet user base took several years to double at equivalent scale. Online food delivery, one of the fastest-adopted consumer categories in the CNNIC dataset, did not show comparable velocity at this stage of its adoption curve.

The speed signal matters for two reasons. First, it indicates that the friction threshold has been crossed. GenAI products in China are no longer in the phase where users require significant effort or technical literacy to engage. Second, the doubling rate implies that a substantial share of new adopters came from mid-tier cities and older cohorts, not just the urban youth demographic that drove early uptake.

Why it matters for investors: A user base that doubles in six months at 515M scale is structurally different from one that grows incrementally. Platforms with established AI interfaces are accumulating behavioral data at a pace that compounds their advantage. Latecomers face a widening gap across user count, model training quality, and personalization depth.

The hardware cycle: AI terminals accelerate faster than platforms

The consumer software story is well-covered. The hardware story is not. According to CNNIC, AI smart terminal sales on the Taotian platform (Alibaba’s combined Taobao and Tmall) reached RMB 14.22B in H1 2025, up 131% YoY. That growth rate on a multi-billion renminbi base is a product cycle signal.

AI smart terminals in this context include AI-enabled smartphones, wireless earbuds with real-time translation and ambient intelligence features, smart glasses, and AI-embedded home devices. The category is distinct from standard consumer electronics in that the AI capability is the primary purchase driver, not a feature add-on.

The policy dimension reinforces the hardware trajectory. The Chinese government has set explicit penetration targets: 70% smart terminal penetration by 2027, and 90% by 2030. These targets are backed by subsidy programs and procurement incentives that function as demand guarantors for the terminal hardware supply chain.

For semiconductor and component suppliers, the 131% YoY growth rate on AI terminal sales is the most actionable data point in the CNNIC report.

It maps directly to demand for edge AI chips, on-device inference memory, and low-power neural processing units. The supply chain implications extend from Chinese fabless designers through to international component makers with China exposure.

AI glasses represent the most promising sub-category for near-term growth. IDC projects the global AI glasses market at USD 300B by 2030.

Meta’s Ray-Ban smart glasses sold over 2M units through Q4 2024, validating mainstream consumer appetite for the form factor. Chinese manufacturers including Xiaomi, Huawei, and a cohort of emerging specialists are positioned to compete on price and AI integration depth.

Platform competition: how Alibaba and JD are competing for AI-mediated purchase intent

The most consequential near-term monetization arena is AI-powered product discovery. Both Alibaba and JD have deployed AI shopping assistants that intercept the consumer at the moment of purchase intent.

Taobao’s “AI Wan Neng Sou” (AI All-Capable Search) restructures the traditional search-and-browse model. Users can input natural language queries, image references, or cross-category descriptions and receive curated product recommendations rather than ranked listings.

The shift from keyword search to intent interpretation changes the competitive dynamic for merchants. Visibility is no longer determined by keyword bidding alone. It is determined by how well a product’s attributes match the AI’s interpretation of user intent.

JD’s “Jing Xiao Zhi” operates on a similar logic. It adds a business-facing layer serving both consumers and merchants, providing AI-generated product listing optimization, pricing recommendations, and demand forecasting. According toCNNIC, JD’s AI-optimized merchant listings achieved 52% higher payment conversion in overseas markets.

That conversion uplift, if sustained at scale, represents a structural margin advantage for merchants who adopt AI tooling early. The platform that owns the AI interface owns the moment between search and purchase. This is not a marginal efficiency gain. It is a restructuring of where value is captured in the e-commerce stack.

The 60,000 SME signal: AI adoption moves down-market

Large platform deployments are expected. The more telling adoption signal is what is happening among smaller operators.

According to CNNIC, 60,000 small and medium enterprises are using Alibaba International’s full-process AI for cross-border commerce operations. AI-published product listings across the platform have reached 7M.

This figure matters because SME adoption of AI tooling historically lags enterprise adoption by two to four years. An SME cohort of 60,000 actively using full-process AI in H1 2025 suggests the technology has crossed the accessibility and cost threshold that previously limited deployment to larger operators.

The cross-border context adds a further dimension. SMEs using AI for international commerce are deploying it for tasks including multilingual product description generation, localised marketing content, customs documentation, and demand signal interpretation across foreign markets.

These are not single-feature applications. They are workflow integrations that, once embedded, create switching costs and compound operational advantages.

The SME adoption curve is a leading indicator for the broader enterprise AI deployment wave. The 60,000 figure will grow. The question is which platform ecosystems capture that growth. The secondary question is what revenue share accrues from licensing fees versus total sales volume gains.

Gen Z as the proof of concept

The 34.8 percentage point gap between Gen Z GenAI usage (71.4%) and the overall average (36.5%) is the sharpest cohort divide in the CNNIC dataset. It is also the most strategically important one for platform operators planning a three-to-five year horizon.

Gen Z in China (born 1995-2009) entered digital consumption as mobile-native users. They are now entering it again as AI-native users. The difference matters.

Mobile-native behavior was shaped by interface constraints: tap, swipe, scroll. AI-native behavior is shaped by expectation of personalization, contextual response, and frictionless fulfilment.

A Gen Z consumer using Taobao’s AI search is not just running a more efficient query. They are developing an expectation that the platform understands their aesthetic, their budget range, their size history, and their social context. That expectation, once formed, is difficult to satisfy through a non-AI interface.

The Gen Z adoption rate of 71.4% represents a cohort that will carry AI-native consumption habits into higher income brackets over the next decade. The platforms that shape those habits now are building a switching-cost advantage at the most valuable point in the consumer lifecycle. A full analysis of Gen Z’s emotional spending, content payment behavior, and brand strategy implications is available in CIW’s companion piece on China’s AI-native consumer cohort.

Subscribe now

The monetization gap: users without revenue models

The risk embedded in the 515M figure is the gap between adoption and monetization. Chinese consumers have demonstrated consistent resistance to paying directly for AI features. The dominant monetization model to date has been indirect: AI as a conversion tool, AI as a retention mechanism, AI as an advertising targeting layer.

Direct subscription models for GenAI products remain limited in penetration. The consumer market has not yet produced a Chinese equivalent of ChatGPT Plus at meaningful scale. Enterprise and SME licensing is growing. It starts from a low base.

The platforms best positioned to close this gap are those that have embedded AI into transactional workflows rather than offering it as a standalone product. When AI generates a sale, the revenue model is clear. When AI provides a conversation, the revenue model requires construction.

The next twelve months will test whether China’s GenAI platforms can translate user scale into revenue density.The 515M user base is a structural asset. Its monetization trajectory is the variable that will separate durable platform value from inflated adoption metrics.

China’s GenAI consumer story in H1 2025 is real, fast, and commercially significant. The doubling of the user base reflects genuine friction reduction, not measurement artifacts.

The hardware cycle is accelerating on the back of policy support and consumer demand. The platform competition for AI-mediated purchase intent has begun in earnest.

The monetization models remain underdeveloped relative to the user base. That gap is where the next phase of competitive differentiation will be decided. The platforms that solve the monetization equation in the next two years will define the architecture of China’s AI consumer economy for the decade that follows.

Share

Read the rest of the series on China’s digital economies in 2026:

• Gen Z drives China’s premium digital spend via AI

• Services drive China’s next digital consumption wave

• How AI and Elderly Demographics Created China’s Most Profitable Micro-Drama Tier

• China’s elderly digital consumers defy the low-value assumption

• Rural China’s digital consumers narrow the urban gap

Alibaba’s instant retail bet signals AI agents’ next commerce frontier

During Spring Festival 2025, Alibaba’s AI model Qwen (Tongyi Qianwen) did not launch a red-packet campaign. It launched an infrastructure experiment.

Qwen deployed a RMB 3B subsidy program it called the “Spring Festival Treat Plan,” connecting its AI assistant for the first time to Alibaba’s full consumer ecosystem: fresh grocery platform Hema, travel booking app Fliggy, ticketing platform Damai, on-demand delivery service Taobao Instant Purchase, and Tmall Supermarket. The integration spanned food, drink, entertainment, and daily necessities.

The move signals a strategic calculation. Qwen’s architects selected instant retail, not traditional e-commerce search, as the primary deployment surface for consumer AI. That choice is not accidental. It reflects a structural argument about where AI agents deliver maximum operational impact.

Instant retail refers to the on-demand delivery of physical goods, typically within 30 minutes, sourced from local stores or forward warehouses within 1 to 3 kilometers of the consumer. The model has evolved from a convenience option into a daily necessity for tens of millions of Chinese households.

According to the Ministry of Commerce’s research institute, China’s instant retail market is projected to reach RMB 971.4B in 2025 and surpass RMB 1T in 2026. The sector now encompasses groceries, pharmaceuticals, household supplies, and prepared food, operating across major cities and smaller-tier markets alike.

Why Qwen chose instant retail, not search

Traditional e-commerce already has AI. Recommendation engines, personalized feeds, and AI-generated product descriptions are standard across Taobao, JD.com, and Pinduoduo. The marginal value of adding another AI layer to a search-and-scroll interface is limited.

Instant retail presents a different operating environment. The decision cycles are shorter. The product types are more standardized, the inventory is geographically dispersed, and the fulfillment window is measured in minutes rather than days. Each of these characteristics amplifies AI’s capability to add value.

According to retail e-commerce analyst Zhuang Shuai, founder of Bailian Consulting, AI agents can intervene most effectively when demand is spontaneous, supply is local, and fulfillment requires real-time optimization across multiple simultaneous variables. Instant retail satisfies all three conditions. Traditional e-commerce satisfies none.

The commercial implication is significant. Qwen’s Spring Festival experiment was not a marketing campaign. It was a test of whether an AI assistant could become the operating system for a consumer’s daily consumption, replacing keyword searches, manual cart additions, and price comparisons with conversational intent.

The four structural advantages of instant retail for AI

First: decision speed. Instant retail purchases are low-deliberation decisions. Consumers buying bottled water, fever medicine, or late-night snacks decide in under 10 seconds. They do not comparison-shop. They do not read reviews.

This makes AI intervention highly effective. A consumer says “I’m thirsty” or “my stomach hurts,” and an AI agent can match that intent to available inventory, confirm a substitution if the preferred product is out of stock, and complete the order.

According to Bailian Consulting’s analysis, intervention success rates are significantly higher in instant retail than in categories requiring extended deliberation, such as electronics or apparel.

Traditional e-commerce purchase cycles for consumer electronics or clothing can span hours or days. AI tools in that environment can suggest options. They cannot close the decision. In instant retail, AI closes the decision.

Second: product standardization.

The core product categories in instant retail are highly standardized. Bottled water, milk, diapers, and painkillers have fixed SKUs, uniform specifications, and predictable consumption rates. AI systems can recommend, substitute, and automate replenishment with minimal error.

This contrasts sharply with apparel, furniture, or electronics, where size variation, color deviation, and model differences introduce high uncertainty into AI recommendations. Standardization is the prerequisite for reliable AI-driven purchasing. Instant retail delivers it consistently.

Third: distributed inventory.

Instant retail inventory is held in thousands of local stores and forward warehouses scattered across urban neighborhoods. This creates a data management problem that human operators and legacy systems cannot solve efficiently: real-time monitoring of stock levels across thousands of locations, responding to demand shifts driven by weather, time of day, and local events.

AI solves this structurally. According to Zhuang Shuai’s framework, an AI agent can scan nearby inventory nodes in milliseconds after a consumer signals intent, select the closest stocked location, and route the order.

When a product is unavailable, AI can recommend a functional substitute and redirect fulfillment automatically. Beyond order-level decisions, AI can generate hourly replenishment forecasts for individual stores by integrating sales history, weather data, and local consumption patterns, reducing both stockouts and perishable waste.

Traditional e-commerce uses centralized regional warehouses with inventory measured in bulk quantities and replenishment cycles measured in weeks. The scale and stability of that model makes AI optimization useful but not essential. Distributed instant retail inventory, by contrast, cannot be managed efficiently without AI.

Fourth: minute-level fulfillment.

Read more

China’s Gen Z cohort was born between 1995 and 2009. They entered the internet as mobile-native users. They are now entering the AI era as the most digitally fluent demographic in the country’s consumer market.

At 261M active digital consumers, they represent 27.2% of China’s entire digital consumption base, according to CNNIC’s Digital Consumption Development Report for 2025.

The CNNIC report is based on a 30,000-respondent telephone survey across 31 provinces. It provides the most granular publicly available dataset on Chinese consumer digital behavior.

Its Gen Z findings are notable for their consistency across categories. Across GenAI adoption, emotional spending, smart product purchasing, and content payment rates, Gen Z systematically outperforms the overall average.

The pattern describes a consumer cohort that has upgraded not just the channels through which it spends, but the reasons and logic by which it spends.

The AI adoption gap that defines the cohort

The single most strategically significant number in the CNNIC Gen Z data is the GenAI adoption gap. 71.4% of Gen Z consumers use generative AI, against a 36.5% overall average. The 34.8 percentage point difference is the largest cohort divide in the report.

This gap is not simply a function of age or technology access. Older millennials in China have comparable smartphone penetration and internet access. The difference is behavioral.

Gen Z has adopted GenAI as a functional layer in daily decision-making, including product research, content discovery, and purchase decisions.

Before AI-powered search, product discovery on Chinese e-commerce platforms relied on keyword input, sponsored rankings, and social referrals. Gen Z consumers using Taobao’s AI Wan Neng Sou or Xiaohongshu’s AI recommendation layer interact with platforms through intent expression rather than keyword entry. They describe what they want in natural language, upload images for visual search, and receive curated outputs shaped by their behavioral history.

The implication for platform operators is direct. A user cohort that has adopted AI-mediated discovery at 71.4% penetration will find non-AI interfaces increasingly inadequate.

The barrier is not financial. It is experiential. A Gen Z consumer who has used an AI-powered interface for six months has already calibrated their expectations upward. Returning to keyword search feels like a regression.

Smart products and the hardware preference signal

Gen Z’s technology spending extends beyond software. 59.3% of Gen Z consumers bought smart products online in H1 2025, compared to 45% across all age groups. The 14.3 percentage point gap reflects a cohort that treats AI-enabled hardware as a default rather than a premium.

Read more

China’s digital economy is typically framed as a technology story. The CNNIC Digital Consumption Development Report for 2025 is based on a 30,000-respondent survey across 31 provinces. It reframes China’s internet economy as a structural consumption story.

CNNIC has tracked China’s internet economy since 1997. Its annual reports are among the most comprehensive primary datasets on Chinese digital behavior available to international researchers and investors. The H1 2025 edition is notable for one number above all others: 46.5%.

That is the share of total resident consumption now flowing through digital channels. At RMB 9.37T, it is not a rounding error or a pandemic-era distortion. It is a durable structural threshold that changes the baseline assumptions for China consumer analysis.

Read the rest of the series on China’s digital economies in 2026:

• China’s gen-AI user base doubled in 6 months

• Gen Z drives China’s premium digital spend via AI

• Services drive China’s next digital consumption wave

• How AI and Elderly Demographics Created China’s Most Profitable Micro-Drama Tier

• China’s elderly digital consumers defy the low-value assumption

• Rural China’s digital consumers narrow the urban gap

Subscribe now

Is 46.5% a floor or ceiling?

The 46.5% figure deserves decomposition before interpretation. The RMB 9.37T total breaks into three categories: physical goods bought online (66.3% of digital consumption), digital services (29.2%), and digital content (4.5%).

Physical goods online penetration is mature. The growth rate is decelerating toward its natural ceiling as offline retail stabilizes at its residual share. That category’s trajectory is largely set.

The services share is the more important signal. At 29.2% of digital consumption, digital services remain structurally underpenetrated relative to their physical goods counterpart. Food delivery, travel booking, healthcare, education, and home services are all digitizing at different speeds.

The policy architecture in 2025 is explicitly designed to accelerate this transition. The 46.5% threshold will rise. The increment will come almost entirely from services, not goods. Analysts modeling China consumer exposure using physical retail metrics are measuring the wrong variable.

Services crossed RMB 1T. Twice.

The clearest evidence of the services shift sits in two category-level numbers. Online food delivery reached RMB 1.09Tin consumption value, serving 569M users. Online travel and transportation reached RMB 1.08T, serving 514M users.

Both categories crossed the trillion-renminbi threshold in a single reporting period. Neither number was widely anticipated at the scale or speed at which it materialized.

Food delivery is the more structurally interesting case. The RMB 1.09T figure captures more than transaction volume. It reflects the digitization of daily household expenditure.

When food delivery crosses a trillion renminbi in semi-annual consumption, it signals that digital channels have captured a category previously considered offline by default.

Travel digitization at RMB 1.08T reflects a different dynamic: the recovery of deferred consumption combined with platform consolidation. Trip.com Group,

Meituan’s travel vertical, and Fliggy (Alibaba) have collectively built infrastructure that now intermediates the majority of domestic travel booking. The trillion-renminbi figure is as much a platform concentration story as a consumption story.

Together, these two categories account for over RMB 2T of digital services consumption in a single half-year. That is larger than the entire digital consumption base of most economies.

The user base is near-saturated. Spend depth is the new metric.

958M digital consumers is the headline user figure from the CNNIC report.

That is 68.1% of China’s population engaging in digital consumption. Total internet users reached 1.123B. The gap between internet penetration and active digital consumption participation stands at approximately 165M.

Those 165M are disproportionately elderly, rural, and lower-income. Converting them to digital consumers is both a policy objective and a commercial opportunity. It is a long-cycle project. The near-term growth story is not acquisition. It is spend depth.

Three metrics now drive platform revenue growth and consumer brand margins: average spend per digital consumer, category breadth per household, and services penetration among existing users. The 958M user base is mature enough that marginal user acquisition has diminishing returns in most categories.

Why it matters for investors: Consumer sector equity positions in China priced on user growth assumptions are measuring the wrong variable. The correct metrics are average revenue per user, category cross-sell rates, and services attach rates on existing physical goods transactions.

Policy as accelerant: trade-in programs and infrastructure build-out

The RMB 2.9T in sales driven by government trade-in subsidy programs in H1 2025 is the most direct evidence of policy transmission into digital consumption. The programs, covering appliances, consumer electronics (3C), and vehicles, were structured to route purchases through digital channels.

Read more

Exchange-style dining is a retail format that emerged across China in 2024. Restaurants in this category model their interiors on stock exchanges. Large circular screens display product names, prices, sales volumes, and real-time price movements.

Each item carries its own ticker code. Prices flash red on the way up and blue on the way down, borrowing the visual conventions of Chinese securities markets.

The format originated in the beer category. It has since spread to coffee, bakery products, ice cream, and hot pot. Some operators have woven exchange-style elements into broader dining concepts rather than building entire venues around them.

China’s food and beverage sector has seen a recurring pattern in recent years. Formats that prioritize visual novelty and social media virality over product depth tend to follow a predictable arc. Blind box dining, immersive themed restaurants, and interactive food installations have all followed this trajectory.

Exchange-style dining is the latest test case. Whether it follows the same path depends on one tension: the format’s visual appeal is its strongest acquisition tool. It is also the asset least likely to drive return visits.

For investors and operators evaluating China’s experiential dining market, the format offers an early-stage case study in how novelty economics play out in real time. The data is still early. The trajectory is already visible.

How exchange-style dining spread from Qingdao to 10+ cities

In May 2024, a beer exchange opened on Qingdao’s Tadong pedestrian street. The concept was direct.

A large screen displayed real-time “prices” and “transaction volumes” for each beer on offer. Customers poured their own drinks from smart taps arranged in a circular layout. The format was designed to feel like a trading floor, not a bar.

The response was immediate. According to Hongcan Wang (a Chinese food industry publication), daily footfall exceeded 10K visitors at peak periods. Single-day sales surpassed 2,400 cups, with daily transaction volume reaching 1,200 liters.

On Douyin, topics tagged “beer exchange” accumulated more than 40M views. The format moved from a local novelty to a nationally recognized concept within weeks.

Replication followed quickly. Exchange-style beer venues appeared in more than 10 Chinese cities within months. The spread ran from Harbin in the northeast to Sanya, Guangzhou, and Shenzhen in the south.

Most venues are independent, single-location shops. A small number have begun testing chain formats. Harbin’s Snow Bear Beer Exchange Center, Guangzhou’s Mengmeng Bear Smart Beer Exchange, and Foshan’s Gorilla Beer Exchange each operate 1-4 locations.

The format did not stay in beer. Coffee exchanges adopted the same ticker-screen model. Zhengzhou’s Coffee Exchange lists its products with alphanumeric codes.

Items like SPU001 and SPU002 are priced between 6.6 and 19.9 RMB per 500ml. Bakeries incorporated exchange-style aesthetics by displaying baking schedules on electronic ticker boards. Shanghai’s Wuwei Canba added a wall of self-pour alcohol dispensers with stock-style price displays to its Sichuan and Hunan dining concept.

The format spread because it produces shareable content. A screen full of real-time price fluctuations is inherently photogenic. For operators, exchange-style dining offered something rare in a crowded market: a reason for customers to film and post.

Earned social media visibility of this kind is difficult to replicate through conventional marketing spend. The speed of replication is itself a signal. The format requires no proprietary technology or specialized skill to copy.

Why the initial operator numbers looked compelling

The financial picture painted by early operators was encouraging. Foshan’s Beer Exchange reported daily sales between 1,000 and 3,000 cups, with stronger performance on evenings and weekends. Harbin’s Snow Bear Beer Exchange Center claimed to have broken even within its first month, according to Hongcan Wang.

These figures suggest attractive unit economics on paper. Labor intensity is low because customers serve themselves. Menu complexity is minimal, which keeps preparation costs down.

Marketing is largely organic, driven by social media sharing rather than paid acquisition channels. Customer acquisition costs are rising across China’s food and beverage sector. Self-service formats with built-in viral potential appeared to sidestep that trend entirely.

The proposition appeared straightforward. Combine the margins of a self-service bar with the visibility of a viral trend.

No celebrity chef required. No complex kitchen operation. A screen, smart taps, and a ticker display constituted the entire value stack.

The early numbers capture only the acquisition phase. The harder questions live in the pricing mechanics, the product constraints, and the durability of consumer interest over time.

Why the dynamic pricing is largely cosmetic

The stock-exchange aesthetic is the format’s central hook. The pricing system behind it is thinner than it appears.

Each business day, all products reset to a fixed “opening price.” The system runs a settlement cycle every 10 minutes. If more customers buy a given item within that window, its price adjusts upward for the next cycle.

Fewer purchases push it down. The color coding mirrors the red-and-blue conventions of Chinese stock displays.

The cap is where the mechanism loses credibility. According to Hongcan Wang, fluctuation limits vary by venue but remain narrow. Zhengzhou’s Coffee Exchange and Guangzhou’s Mengmeng Bear each cap price movements at 10%.

Qingdao’s Beer Exchange allows fluctuation of up to 20%. Foshan’s Beer Exchange permits the widest range at 30%. With base prices between 5 and 25 RMB, the actual price change on any item falls within 0-3 RMB.

The dynamic pricing is atmosphere, not economics. Consumer feedback is direct on this point. One visitor to Zhengzhou’s Coffee Exchange noted that the price movements were too small to matter.

The fluctuations create visual activity on the screen. They do not signal genuine demand shifts or serve an inventory-clearing function.

The 10-minute settlement window is too short to create meaningful price discovery. It is long enough to produce the visual impression of a functioning market.

If the pricing system does not meaningfully shape purchasing behavior, the exchange aesthetic offers no operational advantage over a standard fixed-price menu. The ticker becomes decoration.

How self-service caps product quality

Exchange-style dining is structurally tied to self-service delivery. Customers pour their own drinks or select from pre-prepared options.

This keeps staffing costs low and throughput high. It also sets a ceiling on product complexity.

Beer flows from taps. Coffee is predominantly cold brew or pre-mixed liquid. Juices rely on concentrate blending.

According to Hongcan Wang, consumer reviews consistently flag product quality as the format’s weakest point. One reviewer noted that coffee at a local exchange carried no aroma. Juice products had an obvious reconstituted taste.

Freshly ground, pour-over preparation is rare across most venues.

The equipment economics reinforce this constraint. According to a Beijing-based equipment supplier cited by Hongcan Wang, a 9-tap smart dispensing system costs approximately 35K RMB. A 30-tap circular setup with an automated price-fluctuation system runs 300K-400K RMB.

These are significant upfront costs for a single-location venue with no guarantee of sustained traffic. The payback calculation depends entirely on whether foot traffic holds beyond the novelty cycle. Most operators have no historical data to model that assumption.

Some operators have already responded by downgrading. One venue replaced its dynamic pricing screen with a standard illuminated sign, cutting total equipment costs to roughly 50K-60K RMB. The trade-off is decisive. Without the ticker, the venue loses its primary reason to exist as an exchange-style restaurant.

High-volume self-service also accelerates hardware wear. Dispensing systems built for tens of cups per day are being pushed to thousands. Breakdowns become frequent.

Maintenance costs compound over time, eroding the margins that made the format look attractive in early projections.

The repeat-visit problem and first signs of decay

Social media virality can launch a format. Sustaining one requires a different set of dynamics.

According to Hongcan Wang, a Nanjing-based consumer observed that the local beer exchange drew consistent queues during its first two months after opening. As cooler weather reduced seasonal foot traffic, the queues disappeared.

The venue continued to operate. The initial surge of interest did not repeat.

The sharpest data point is a rebranding case. Zhuhai’s “Three Bulls Hot Pot Exchange” styled its menu around beef cuts displayed as K-line charts on a large screen. Daily sales volumes and prices scrolled beneath.

The concept attracted initial attention. By the end of 2024, the restaurant had dropped “Exchange” from its name. It reverted to a conventional hot pot identity.

The exchange format had failed to sustain return visits.

Novelty decay is the format’s central structural risk. Exchange-style dining delivers a strong first-visit experience. The screens are visually striking.

The concept is easy to explain and share. Without a product layer that justifies a second visit, the format functions as a one-time attraction.

The pattern is consistent with previous experiential dining formats in China. Initial buzz drives a surge in first-time visitors. Without a durable product reason to return, the format settles into a niche or disappears.

China’s micro-drama industry entered 2025 facing a critical inflection point. The market reached 50.4B RMB according to DataEye. Premium productions escalated costs while platform guarantee payments dropped 60-80% starting September 2024. This forced mid-tier producers to confront whether high production values generate superior returns.

The answer challenges conventional assumptions.

The Profitability Paradox: Why Cheaper Content Makes More Money

Budget micro-drama producers discovered that 100K-300K RMB productions generate better risk-adjusted returns than 500K-1M premium content.

The math reveals the advantage. Premium tier operates as binary gamble. 500K-1M investment. Approximately 30% success rate. Total loss on failures.

Budget tier functions as portfolio strategy. 100K-300K per title. 70% achieve breakeven or profitability. 20-50% margins per successful title.

Dafang Hulian, a Zhengzhou-based production company with 100+ staff, experienced this directly. The company invested 600K RMB in a premium drama launched May 2025. Final return reached only 100K RMB, representing 83% loss.

Li Runzhe, commercial manager at Dafang Hulian, explained the shift. “We now invest in 10 dramas at 100k RMB each instead of gambling 500-600k RMB on one premium drama.”

The portfolio approach transforms economics. Ten budget productions totaling 1M RMB typically see seven titles return profit. Combined returns deliver 20-30% overall profitability in conservative scenarios. Returns reach 100% when multiple titles perform well. Single premium dramas offer higher ceiling but catastrophic floor.

Platform guarantee reductions accelerated this realization. Through 2024, platforms guaranteed up to 500K RMB for content. By September 2024, guarantees dropped dramatically.

“The maximum guarantee fell to around 200K RMB,” Li noted. “More risk transferred to producers.”

Investment psychology shifted accordingly. “Investors contributing 200-300k RMB for downmarket content view it differently than 500-600k RMB premium bets,” Li observed. Budget tier enables measured experimentation. Premium tier demands conviction.

The Audience Arbitrage: Why 310M Seniors Make Budget Content Profitable

The profitability paradox exists due to specific audience characteristics that eliminate the cost-to-quality correlation governing traditional media economics.

China’s elderly population reached 310M people aged 60+ by end of 2024, representing 22% of total populationaccording to National Health Commission data. This demographic will exceed 400M by 2035, adding 10M+ annually.

Three behavioral traits create structural profitability:

Time availability. Elderly viewers possess significantly higher daily viewing hours than younger demographics. Retirement and reduced work obligations translate to consumption capacity.

Ad tolerance. This audience demonstrates willingness to watch advertisements for free content. Younger viewers employ ad-blocking technology or pay for ad-free experiences. Elderly users complete ad views at substantially higher rates.

Piracy resistance. Lower technical literacy around VPN usage and torrent access means elderly viewers primarily consume content through official channels. Premium content loses substantial revenue to piracy. Budget content on official platforms captures full monetization.

Quality agnosticism. Most critically, elderly viewers prioritize complete narrative arcs and direct emotional payoffs over production sophistication. “They don’t require television or film-level narrative structure,” Zhang Lai, founder of Zhejiang-based Xinnengliang Fenghua Short Drama Dream Factory, explained. “Emotion delivered straightforwardly, story told completely, that meets standards.”

Blue-collar workers constitute the secondary audience segment. Security guards, delivery drivers, domestic workers, construction laborers share similar characteristics. Fragmented attention windows. Preference for accessible content. Limited willingness to pay premium prices.

Popular themes reflect audience preferences. Rural life dramas, mother-in-law conflicts, village suspense, middle-aged CEO stories dominate viewership. Content difficulty drops substantially from premium requirements.

This creates defensive moat. Premium producers cannot profitably serve these viewers. Their cost structure requires either $5+ CPM advertising rates or $0.50+ per episode direct payment. Elderly audiences in Tier 3-5 cities generate $2-3 CPMs and near-zero direct payment willingness.

Premium producers would need to cut costs 60% to compete. That reduction would eliminate their production quality advantage entirely. Budget producers already operate at this cost level. This makes their position structurally unassailable within this demographic.

The Market Mechanism: How Mid-Tier Collapse Forced Business Model Discovery

Three production companies arrived at identical conclusions within six months. This signals structural opportunity, not tactical luck.

Dafang Hulian shifted strategy in late 2024 after the premium drama loss. For 2026, the company plans allocating two-thirds of capital to downmarket productions focusing on original IP and copyright ownership.

Gangzi (pseudonym) co-founded Henan Yunye Tingfeng Film Industry in September 2024, launching directly into budget production. “Premium female-targeted content requires expensive cast. Premium male-targeted content demands high-quality script development over two to three months,” Gangzi noted. Both paths require capital beyond mid-tier capabilities.

Zhang Lai established Xinnengliang Fenghua Short Drama Dream Factory in Zhejiang in December 2024, also targeting budget tier exclusively.

Geographic distribution confirms broad-based movement. Production companies span Zhengzhou in Henan and various cities across Zhejiang. This isn’t isolated regional phenomenon.

Head consolidation drives mid-tier exodus. Tinghuadao captured 4 of top 10 positions on DataEye Red Fruit Hot Rankings for November. S-tier premium productions reached 1.5-3M RMB per title in 2025. Hit titles like “Shengxia Fendela” and “Jiali Jiawai 2” exceeded even that range.

Read more

China’s smartphone market contracted for another year in 2025. Total shipments fell to 284.6M units. The 0.6% year-over-year decline masks deeper strategic shifts. Brands navigated subsidy cliffs, surging component costs, and intensifying premium competition.

Huawei reclaimed the annual crown with 16.4% market share after years of US sanctions. Apple and vivo tied for second at 16.2% each. Their trajectories diverged sharply. Xiaomi and OPPO rounded out the top five. Xiaomi posted the strongest growth among major players.

The year split into two distinct phases. Government subsidies and Lunar New Year demand drove early momentum. But the stimulus proved short-lived as local funds depleted. Consumers pulled back. Fourth-quarter dynamics then scrambled the rankings. Apple seized a commanding lead on new product strength. Chinese brands retreated from low-margin segments.

IDC is a global market intelligence firm tracking technology shipments and trends. Their quarterly reports provide authoritative data on smartphone market dynamics. The firm surveys supply chain partners and retailers to compile shipment estimates.

Huawei’s return to dominance validates survival strategy

Huawei shipped 46.7M units in 2025. The company claimed 16.4% of China’s smartphone market according to IDC. The result marks a symbolic victory for the Shenzhen giant. It spent three years rebuilding its mobile business after losing access to Google services and advanced chip supplies.

The annual win came despite marginal shipment declines from 2024’s 47.6M units. Huawei’s 1.9% YoY drop actually outperformed the overall market’s 0.6% contraction. This suggests the brand maintained pricing power even as volumes softened.

Market share became the real battleground as total demand stagnated. Huawei’s 16.4% compared favorably to its 16.6% share in 2024. This narrow compression preserved its lead over Apple and vivo.

The company achieved this through premium positioning with its Mate and Pura series. It targeted buyers willing to pay for domestic alternatives to foreign flagships.

The strategy reflects a broader shift in China’s smartphone landscape. Volume growth has evaporated. Brands now compete on margins rather than scale. Huawei’s ability to defend share while maintaining premium pricing validates its focus on loyal customers over market expansion.

Geopolitical factors continue shaping Huawei’s competitive position. US export controls still limit access to cutting-edge chipmaking technology. This forces reliance on older-generation semiconductors. Yet the brand has turned constraints into differentiation. It markets devices as symbols of technological self-sufficiency amid rising nationalism.

Apple dominates Q4 with iPhone 17 surge

Apple shipped 16M units in Q4 2025. The company captured 21.1% of quarterly sales according to IDC. The performance marked a 21.5% YoY surge. The iPhone 17 series launched in September powered this growth.

The quarterly dominance contrasts sharply with Apple’s full-year second-place finish. Annual shipments reached 46.2M units for 16.2% market share. This represented a modest 4.0% gain from 2024’s 44.4M. The divergence underscores Apple’s dependence on new product cycles to drive China growth.

iPhone 17 differentiation proved critical in converting buyers. While specific feature details were not disclosed in IDC’s report, Apple’s ability to grow share during a market contraction suggests strong perceived value. Premium pricing held firm despite broader deflationary pressures in consumer electronics.

The Q4 surge also highlights seasonal dynamics in China’s smartphone calendar. Apple typically concentrates China demand in fall quarters around new launches. Share then compresses in spring when local brands refresh their lineups. This year’s 21.5% Q4 growth suggests the pattern intensified as competitors pulled back from aggressive holiday promotions.

Profitability concerns reshaped competitive behavior across the industry in Q4. With memory chip costs climbing, most Chinese brands reduced shipments of budget models to protect margins. Apple faced no such pressure given its premium positioning. This allowed it to capture volume that might otherwise have gone to sub-2000 RMB devices.

vivo and Xiaomi diverge on growth trajectories

vivo held third place with 46.1M units shipped in 2025. This matched Apple’s 16.2% market share on rounded figures. But the Dongguan brand contracted 6.6% YoY from 49.3M units in 2024. This marked the steepest decline among top-five players.

The drop reflects strategic missteps in China’s evolving premium segment. vivo built its reputation on selfie cameras and celebrity endorsements targeting younger buyers. As competition intensified at higher price points, the brand struggled to justify premium pricing without breakthrough differentiation. Share erosion accelerated in the second half as subsidies faded. Buyers became more selective.

Xiaomi posted 4.3% YoY growth. This was the fastest expansion among major brands despite ranking fourth overall. The Beijing company shipped 43.8M units for 15.4% market share. This compared to 42.0M and 14.7% respectively in 2024.

Xiaomi’s performance validates its dual-brand strategy. The main Xiaomi line targets value-conscious buyers with competitive specs at aggressive prices. The Redmi sub-brand defends ultra-budget segments where other players are retreating.

Meanwhile, the high-end Xiaomi 15 series competes directly with Apple and Huawei flagships. It does so at lower volumes.

OPPO rounded out the top five with 43.4M units and 15.2% share. This represented a marginal 2.1% gain from 2024. The Guangdong brand maintained stable positioning by balancing online and offline channels. It avoided the sharp corrections seen at vivo.

The divergent trajectories expose vulnerability in China’s smartphone mid-tier. Brands without clear premium differentiation or cost leadership face compression from above and below. vivo’s struggles suggest incremental product updates no longer suffice when buyers can either stretch for flagships or step down to capable budget devices.

Rising costs force industry-wide profitability focus

Memory chip prices surged throughout 2025. This squeezed margins across the smartphone supply chain. DRAM and NAND costs climbed as datacentre demand from AI applications tightened global supply. Phone makers absorbed the impact through Q3 before adjusting strategies in Q4.

Chinese brands reduced budget shipments in the year’s final quarter to protect profitability. IDC noted most local players decreased output of entry-level models. They prioritized margin over volume as component costs escalated. The shift marked a reversal from previous years when brands competed fiercely for market share in sub-1500 RMB segments.

The pullback created space for Apple’s Q4 surge. With Chinese competitors shipping fewer low-end devices, buyers looking for reliable alternatives faced a narrower range of options. Some traded up to iPhones rather than settling for aging Android models. This contributed to Apple’s 21.5% quarterly growth.

Industry observers expect further market contraction in 2026 as cost pressures persist. IDC warned that China smartphone shipments may decline more significantly if memory prices continue climbing. The projection reflects manufacturers’ limited ability to pass costs to consumers in a deflationary environment.

Storage configurations became casualties of the cost crunch. Several brands reduced availability of 512GB and 1TB options on mid-range models. This pushed buyers toward 256GB variants with better margin profiles. Flagship devices maintained higher storage tiers but at increased prices. This compressed demand from cost-sensitive buyers.

Subsidy cliff exposes demand fragility

Subsidies drove early-2025 growth before exhausting allocated funds. China’s “guobao” consumer stimulus program overlapped with Lunar New Year shopping season. This created a brief surge in smartphone demand during Q1 and early Q2. Local governments offered rebates on electronics purchases to stimulate consumption amid economic uncertainty.

The boost proved unsustainable as regional funding depleted faster than anticipated. According to IDC, multiple provinces exhausted subsidy budgets by mid-year. This removed a key demand driver just as spring product cycles matured. The timing left brands with elevated inventory entering the traditionally weaker summer months.

Market dynamics shifted sharply in H2 2025. Without subsidy support, YoY comparisons turned negative. The pulled-forward demand from H1 created a tougher baseline. Replacement cycles lengthened as buyers who upgraded early sat out subsequent product launches. This compounded the volume decline.

The subsidy experience revealed limitations of stimulus-driven smartphone demand. Unlike categories such as home appliances where subsidies unlock pent-up replacement needs, phone buyers often shift purchase timing rather than creating net new demand. The 2025 pattern mirrored earlier subsidy programs that produced short-term spikes followed by extended troughs.

Policy uncertainty complicates 2026 planning for manufacturers. With no major subsidy program announced for the coming year, brands must assume normalized demand conditions. This conservative posture reinforces the industry’s margin-over-volume pivot as companies prepare for potential further contraction.

Others category dominance masks market fragmentation

Smaller brands collectively shipped 58.4M units in 2025. This represented 20.5% of China’s market according to IDC. The “Others” category actually ranked first by share. It exceeded Huawei’s 16.4% and the other individual leaders. But the aggregate figure obscures intense fragmentation among second-tier players.

The Others segment contracted 3.3% YoY from 60.4M units in 2024. This marked a sharper decline than the overall market’s 0.6% drop. The divergence suggests smaller brands bore disproportionate pressure as component costs rose and subsidies evaporated. Many lacked the scale to negotiate favorable supplier terms. They also lacked brand equity to command premium pricing.

Subscribe now

Honor, Realme, and OnePlus likely comprise significant portions of the Others category. IDC did not break out individual sub-leaders. These brands compete primarily on value positioning. They offer flagship-adjacent specs at mid-tier prices. The strategy works during market expansions but becomes untenable when costs surge. Buyers polarize toward premium or ultra-budget extremes.

E-commerce specialists also populate the Others segment. Brands such as iQOO (vivo’s gaming sub-brand) and Redmi (Xiaomi’s value line, when reported separately) target online-first buyers. They use aggressive flash-sale tactics. Their performance depends heavily on platform promotions during shopping festivals. These softened in 2025 as e-commerce growth slowed.

The fragmented landscape creates consolidation pressure. With five brands commanding nearly 80% of shipments, remaining players fight over a shrinking pool. Profitability becomes nearly impossible for brands shipping fewer than 10M annual units. This suggests further exits or acquisitions ahead as the market matures.

2026 outlook: deeper contraction looms

China’s smartphone market faces intensifying headwinds entering 2026. IDC projects shipments may decline more significantly than 2025’s marginal 0.6% drop if current cost and demand trends persist. The warning reflects industry consensus that structural challenges now outweigh cyclical factors.

DRAM and NAND prices remain elevated despite corrections from 2023 peaks. AI infrastructure investments continue absorbing memory supply. This leaves consumer electronics as lower-priority customers. Phone makers cannot easily redesign products around cheaper components given storage and RAM requirements driven by operating system demands and user expectations.

Manufacturers confront difficult tradeoffs between volume and profitability. Maintaining shipments requires absorbing cost increases or passing them to consumers through price hikes. The former destroys margins. The latter suppresses demand in a deflationary economy where buyers have grown accustomed to declining electronics prices.

Replacement cycles continue extending as innovation plateaus. Flagship devices from 2022-2023 remain capable for most users. This reduces urgency to upgrade. The shift from 4G to 5G provided temporary stimulus. But no comparable technology transition appears imminent to drive mass device turnover.

Premium segment competition intensifies as brands chase profitability. With budget segments squeezed by cost pressures, manufacturers cluster around 3000-6000 RMB price points where margins improve. But this concentration creates overcrowding. It forces aggressive promotion to differentiate increasingly similar products.

The 2025 results showcase market maturation. China’s smartphone industry has transitioned from volume growth to margin focus. Success now requires defending share while maintaining pricing discipline. Component cost management becomes as critical as product innovation. Brands lacking scale advantages or premium positioning face existential pressure as consolidation accelerates through 2026.

Zhipu AI and MiniMax listed on the Hong Kong Stock Exchange on January 8 and 9, 2026, becoming the first foundation model companies to go public globally. The simultaneous debuts beat OpenAI and Anthropic to market by months. More importantly, their prospectus filings expose financial realities that Chinese AI startups have kept private for years.

The two companies represent opposite bets on AI monetization.

Zhipu generated 85% of H1 2025 revenue from selling on-premise model deployments to state-owned enterprises and financial institutions, according to its prospectus.

MiniMax derived 71% of revenue from consumer applications, with its Talkie AI companion app and Hailuo video generator driving growth in overseas markets.

Neither path shows clear signs of sustainability. But the data reveals which model comes closer to working.

Two divergent bets on AI monetization

Zhipu AI, officially Knowledge Atlas Technology, emerged from Tsinghua University’s computer science department in 2019. Co-founders Tang Jie and Li Juanzi built the company around their academic research in knowledge graphs and large language models.

The business model reflects its origins. Zhipu’s on-premise deployment revenue reached RMB 162M in H1 2025, representing 84.8% of total revenue of RMB 191M ($27M), according to prospectus data.

These deployments serve clients requiring strict data sovereignty—central state-owned enterprises, banks, and government agencies that cannot use cloud-based AI services.

The gross margin on this business is substantial. On-premise deployment achieved 59.1% gross margins in H1 2025, well above traditional enterprise software. The model packages AI capabilities as licensed software plus technical services, commanding premium pricing.

Zhipu’s cloud API business tells a different story. The segment generated just RMB 29M in H1 2025, representing 15.2% of revenue. More concerning, gross margins turned negative at -0.4% after operating at a thin 3.4% in 2024, according to industry analysis.

This reflects China’s brutal API pricing environment where DeepSeek offers inference at $0.14 per million tokens.

MiniMax chose consumer applications over enterprise services. Founded in December 2021 by former SenseTime vice president Yan Junjie, the company initially launched Glow, an AI character roleplay app, in October 2022. Four months later, the app had 5M users.

Regulatory pressure forced a pivot. MiniMax repackaged Glow as Talkie for international markets, targeting the US and Singapore.

The strategy worked. Talkie ranked 5th among free entertainment apps in the US by June 2024, surpassing Character.AI in downloads, according to company filings.

Consumer applications generated 71% of MiniMax’s 9-month 2025 revenue of $53.4M. Talkie contributed 35% of total revenue, while Hailuo AI video generator added another 33%. The remaining 29% came from B2B API access, according to prospectus data.

The geographic split matters. MiniMax derived 73.1% of revenue from overseas markets as of September 2025, with the US and Singapore as top markets. This international focus insulates the company from domestic price competition but exposes it to regulatory risk.

Disney, Universal, and Warner Bros. filed a $75M copyright lawsuit against MiniMax in September over Hailuo-generated content allegedly violating intellectual property.

The profitability mirage behind impressive margins

Both companies present gross margin figures that suggest viable unit economics. The reality is more complex.

Zhipu reported overall gross margins of 50% for H1 2025, driven primarily by its on-premise business. But revenue growth cannot keep pace with R&D spending. The company posted revenue of RMB 191M against R&D expenses of RMB 1.595B in H1 2025, producing an R&D expense ratio of 835%, according to financial analysis.

Approximately 70% of R&D spending goes directly to computing power costs. Zhipu burned RMB 1.145B on GPU clusters in H1 2025 alone.

The net loss reached RMB 2.36B ($330M) on revenue of RMB 191M ($27M), a loss ratio exceeding 12:1.

Annual trajectory shows the problem. Revenue grew from RMB 57M in 2022 to RMB 312M in 2024, a 130% CAGR. But 2024 R&D spending hit RMB 2.2B ($315M), representing 703% of revenue, according to prospectus disclosures.

For every yuan earned, Zhipu reinvested over eight yuan into model development.

MiniMax shows better trajectory but remains deeply unprofitable. The company achieved overall gross margins of 23.3% in the first nine months of 2025, up from 12.2% in 2024 and turning positive after -24.7% in 2023, according to financial filings.

The B2B segment drives margin improvement. API and enterprise services achieved 69.4% gross margins in 9M 2025, far exceeding industry averages.

This suggests MiniMax achieved significant technical advantages in inference cost control, likely through its Mixture of Experts architecture and custom attention mechanisms.

But losses remain staggering. MiniMax reported $512M in net losses on $53.4M revenue in 9M 2025, a loss ratio approaching 10:1. R&D spending reached $180M, representing 337% of revenue.

The company spent over $150M on cloud computing bills in 2025, with total infrastructure and R&D consuming approximately $250M annually, according to industry reports.

Revenue growth provides some optimism. MiniMax grew from $3.5M in 2023 to $30.5M in 2024, reaching $53.4M in 9M 2025, a 130% YoY growth rate.

Monthly active users expanded from 3.1M in 2023 to 27.6M by September 2025, demonstrating rapid global adoption, according to company disclosures.

The critical question is whether revenue can scale faster than compute costs. MiniMax reduced inference costs by 45% over the past year, but gross margins remain at 23% even after these improvements, according to financial data. At current burn rates, neither company approaches breakeven.

Why they went public now and what it reveals

The IPO timing reflects strategic urgency rather than financial readiness. Both companies raced to list before OpenAI and Anthropic begin their public offerings, expected in 2026.

OpenAI is projected to reach a $830B valuation if it completes a $100B fundraising round. Anthropic is valued at approximately $350B following major investments from Amazon and Google, according to market data. These mega-offerings will dominate global AI investment attention and liquidity.

Smaller companies face a closing window. Once US AI giants list publicly, Chinese startups will struggle to secure funding at favorable valuations. The market can absorb multiple billion-dollar IPOs but has limits on total capital allocation to unprofitable AI companies.

Zhipu raised HK$4.35B ($560M) at a HK$116.20 offer price, implying a market valuation of approximately $6.7B. Retail investors oversubscribed the offering 1,159 times, signaling strong Hong Kong appetite for AI exposure, according to exchange data. The stock closed 13.2% higher at HK$131.50 on debut.

MiniMax raised HK$4.8B ($620M) at the top of its HK$151-165 range, pricing at HK$165 per share. Institutional demand forced the company to halt bookbuilding a day early. The stock surged 109% on debut, closing at HK$345 and valuing the company at approximately $13.7B, according to market reports.

Market reception diverged significantly. MiniMax achieved 3x higher first-day gains than Zhipu despite Zhipu’s technically superior models. Zhipu’s GLM-4.7 currently ranks among the top global open-source models, ahead of MiniMax’s best offering, according to benchmark data.

The valuation gap reflects two factors. First, investors prefer consumer growth stories over enterprise deployment models in AI. Second, Zhipu faces regulatory headwinds that MiniMax avoids.

The US Commerce Department added Zhipu to its Entity List in January 2025, restricting access to American semiconductor technology and effectively banning the company from US markets, according to government filings.

The designation cites alleged support for China’s military modernization through research collaborations. MiniMax operates without similar restrictions and actively courts US developers through events in New York, Miami, and San Francisco.

Both companies plan similar capital deployment. Zhipu will allocate 70% of IPO proceeds to R&D for general-purpose AI models.

MiniMax will invest heavily in product development and infrastructure scaling, according to use-of-proceeds statements. Neither prioritizes near-term profitability. The capital provides runway to continue the compute arms race.

The DeepSeek disruption that reshaped everything

DeepSeek fundamentally altered Chinese AI economics in 2024, forcing both IPO companies to pivot their business models months before listing.

On May 6, 2024, DeepSeek released V2 with API pricing at $0.14 per million input tokens, representing a 5-10x reduction from prevailing market rates, according to market analysis. The launch triggered immediate price warfare.

Alibaba slashed its Qwen model pricing from $1.10 to $0.07 per million tokens within days. ByteDance followed with its Doubao model at $0.04 for the lite version. By January 2025, when DeepSeek’s R1 model gained global attention, average Chinese LLM prices had fallen 92% compared to May 2024 levels, according to pricing data.

The impact on API-focused revenue was severe. Companies relying on cloud inference margins saw profitability collapse overnight.

Zhipu’s cloud API gross margins fell from 3.4% in 2024 to -0.4% in H1 2025, making the business economically unviable at scale, according to financial disclosures.

The disruption forced strategic shifts. Zhipu accelerated its enterprise deployment focus, leveraging advantages in custom training and industry-specific implementations where DeepSeek’s standardized API could not compete effectively.

The company’s co-founder Zhang Peng stated publicly that Zhipu aims to increase API business to 50% of revenue, signaling recognition that on-premise deployment alone cannot sustain growth, according to executive statements.

MiniMax pursued differentiation through consumer applications. Rather than competing on API pricing, the company focused on packaged experiences—Talkie for social interaction, Hailuo for video generation—where users pay for outcomes rather than tokens.

This strategy insulated MiniMax from direct price competition while building international brand recognition.

Both companies also accelerated model performance improvements. As of July 2025, 41% of employees at the Six Tigers (China’s elite AI startups including Zhipu and MiniMax) listed job-hunting status, reflecting workforce reductions across the sector, according to employment data. But the survivors emerged with stronger technical foundations.

Zhipu’s GLM-4.5, released July 28, 2025, topped Hugging Face’s global trending list within 48 hours. By September, GLM-4.5 ranked in the top five globally on ChatBot Arena and WebDev Arena, prestigious benchmarks where user votes determine rankings, according to performance data.

The December release of GLM-4.7 achieved coding abilities comparable to Anthropic’s Claude Sonnet 4.5 on mainstream tests.

MiniMax’s M2 model, launched October 2025, claims performance rivaling Claude Sonnet 4.5 while operating at just 8% of the cost, according to company benchmarks.

Earlier, MiniMax-M1 set industry standards with 4M token context windows, positioning the company at the technical frontier.

DeepSeek’s impact extended beyond pricing. The company demonstrated that model training costs could be reduced by an order of magnitude through architectural innovations.

DeepSeek’s V3 model reportedly cost $6M to train, compared to $100M+ for comparable US models, using approximately one-tenth the computing power of Meta’s Llama 3.1, according to technical papers.

This efficiency breakthrough challenged assumptions about AI economics. If DeepSeek’s approach scales, massive capital expenditure may not be necessary to maintain frontier performance.

That proposition threatens the core valuation thesis for capital-intensive US AI giants while potentially leveling the playing field for Chinese startups operating under US chip restrictions.

Compute costs define the real business model

GPU spending dominates AI company economics in ways traditional software never experienced. Understanding compute costs is essential to evaluating whether these business models can work.

Zhipu spent RMB 1.145B ($160M) on computing power services in H1 2025 alone, representing 71.8% of total R&D spending of RMB 1.595B, according to expense breakdowns. This does not include personnel costs for the research teams designing and training models.

Annualized, Zhipu burns approximately $320M on compute while generating $54M in annual revenue at current run rates. The company must grow revenue 6x just to cover GPU costs, before accounting for salaries, infrastructure, or other operating expenses.

US export controls compound the challenge. The Commerce Department’s October 2023 restrictions limit Chinese companies’ access to Nvidia’s H100 and A100 chips, forcing reliance on less powerful H800 alternatives or Chinese-made alternatives from Huawei and domestic manufacturers, according to regulatory analysis.

Companies work around restrictions by renting capacity from cloud providers like Microsoft Azure and Amazon Web Services in Southeast Asian and Middle Eastern regions where controls apply less stringently.

But US lawmakers are pushing to close this cloud loophole, which could substantially increase costs, according to policy reporting.

MiniMax faces similar economics. The company spent over $150M on cloud infrastructure in 2025 while generating $53M in revenue, an infrastructure cost ratio of approximately 280%, according to financial data. Total R&D including personnel reached $180M, or 337% of revenue.

The critical question is whether inference costs decline faster than training costs rise. As models grow larger and more capable, training expenses increase exponentially. GPT-4’s training reportedly cost over $100M. Next-generation models may require $1B+ in compute, according to industry projections.

But inference costs per query have dropped significantly. MiniMax reduced inference costs by 45% year-over-yearthrough architectural optimizations, according to company statements.

DeepSeek’s technical papers describe custom GEMM routines using 8-bit floating point arithmetic that dramatically lower memory requirements and computational load, according to research documentation.

If inference efficiency continues improving while training costs plateau, unit economics may eventually work. Models could reach quality thresholds where incremental performance gains require less additional compute.

Current evidence does not support this optimistic scenario. Companies continue racing to train larger, more expensive models to maintain competitive positioning.

The infrastructure dependency creates strategic vulnerability. Unlike software companies that achieve high gross margins as revenue scales, AI companies face compute costs that scale linearly with usage. Every additional customer, query, or interaction requires proportional GPU capacity. This fundamentally limits margin expansion potential.

Zhipu’s 59% gross margins on on-premise deployment appear sustainable because customers provide their own infrastructure. The company licenses software and provides services, but does not bear ongoing inference costs. This model resembles traditional enterprise software economics.

MiniMax’s 69% gross margins on B2B API services suggest the company achieved architectural efficiencies that reduce inference costs below revenue per query.

But overall margins remain at 23% when consumer applications are included, indicating that infrastructure costs consume the majority of consumer subscription and advertising revenue, according tosegment analysis.

Neither model has proven it can support the $400M+ annual R&D budgets required to remain competitive while also achieving profitability. The IPO capital provides 12-18 months of runway at current burn rates, after which both companies must either raise additional capital or dramatically reduce R&D spending.

What $9B versus $500B valuations actually reveal

Market capitalizations tell a story about investor confidence in business model viability and market positioning.

Zhipu’s ~$9B valuation and MiniMax’s ~$14B valuation combine for approximately $23B in market capitalization, according to trading data.

Together, they represent less than 5% of OpenAI’s reported $500B valuation or 7% of Anthropic’s ~$350B valuation, according to comparative data.

Read more

Multi-agent AI systems represent teams of specialized AI agents working together to complete complex tasks. Unlike single large language models that handle all aspects of a query, multi-agent systems divide work among planning agents, execution agents, and validation agents. Each agent operates in its own computing environment with specific tools and expertise. This architectural approach emerged as scaling single models showed diminishing returns. Training costs grew exponentially while performance gains slowed.

The industry has spent years making individual models bigger and smarter. Now, the real performance gains come from making AI agents work together like specialized human teams.

Manus was developed by Butterfly Effect, founded in Beijing in 2022 by three Chinese entrepreneurs: CEO Xiao Hong, 33, a Huazhong University of Science and Technology graduate who previously built WeChat productivity tools; Chief Scientist Ji Yichao, 32, a high school dropout who created the Mammoth Browser and became Forbes China “30 Under 30”; and Product Director Zhang Tao.

The company initially operated dual headquarters in Beijing and Wuhan with backing from Beijing-based venture capital firm ZhenFund, Tencent, and HongShan (formerly Sequoia China).

Following Manus’s March 6, 2025 launch and intense international scrutiny over its Chinese origins, the company relocated its headquarters from Beijing to Singapore in mid-2025, laid off 80 Beijing employees, and discontinued services in mainland China—a move that enabled Silicon Valley’s Benchmark to lead a $75 million funding round and ultimately Meta’s $2+ billion acquisition, though it also triggered China’s export control investigation examining whether AI technologies developed while Beijing-based fall under national security regulations.

Multi-Agent AI Systems Unlock Performance Through Collaboration

For years, AI development followed a simple formula. Bigger models with more parameters would solve harder problems. GPT-4 reached 1.8 trillion parameters. Gemini Ultra added cross-modal fusion. Claude 3 pushed context windows to new limits.

The returns from this approach are diminishing. Training costs grow exponentially. Performance improvements slow to incremental gains.

Multi-agent systems take a different path. Instead of building omniscient models, they create specialized AI agents. These agents collaborate on complex tasks through clear division of labor.

According to Anthropic, multi-agent systems with Claude Opus 4 leading Claude Sonnet 4 subagents outperformed single-agent Claude Opus 4 by 90.2% on internal research evaluations. The performance gap becomes even larger for tasks requiring parallel exploration across multiple information sources.

Three factors explained 95% of this performance variance. Token usage accounted for 80%. Tool calling frequency and model selection explained the remainder. Multi-agent systems achieve superior results by intelligently distributing computational resources across specialized agents. Each operates within its own context window.

This architectural shift changes how AI systems scale. Single agents hit context limits. Multi-agent systems simply spawn additional agents with fresh context. Continuity is maintained through careful coordination.

Manus demonstrates this principle in production. According to Manus, since launching in March 2025, the platform has created over 80 million virtual computers. It has processed more than 147 trillion tokens. Each virtual machine functions as an isolated workspace. AI agents execute tasks autonomously, from market research to full-stack application development.

Manus Proves Commercial Viability at $100M ARR in 8 Months

Manus crossed $100 million in ARR just eight months after launch. This makes it the fastest startup to reach this milestone from zero. The company’s total revenue run rate now exceeds $125 million when including usage-based fees.

This growth trajectory validates a crucial hypothesis. When AI can reliably complete entire workflows rather than just assist with individual steps, its commercial value shifts. The shift moves from efficiency enhancement to labor replacement.

Pricing reflects this positioning. Manus charges $19-$199 monthly for subscription tiers. These tiers allow 2-5 concurrent tasks. The company operates a freemium model. This generated a waitlist exceeding 2 million users before monetization. Businesses pay for digital labor capacity, not just software features.

The architecture makes this possible. Manus maintains a lead orchestrator agent. This agent analyzes tasks, creates execution plans, and spawns specialized subagents. Subagents handle components in parallel. Planning agents act like project managers. Execution agents function as domain specialists. Validation agents ensure quality control.

State-of-the-Art Performance on GAIA Benchmark

In the GAIA benchmark designed to evaluate AI agents on real-world tasks, Manus achieved state-of-the-art performance across all three difficulty levels. According to multiple sources, it surpassed OpenAI Deep Research.

GAIA tests practical problem-solving. Tasks require multi-step reasoning, external tool usage, and long-horizon planning. Scoring at the top validates Manus’s ability to handle genuine business complexity.

The platform’s technology stack reveals sophisticated engineering beneath the surface. Each task spawns an isolated cloud virtual machine. The VM includes a complete development environment. Network isolation prevents data leakage across tasks. Sandboxes are destroyed immediately after completion to eliminate data residue.

Resource Optimization Drives Cost Efficiency

Resource optimization drives cost efficiency. Manus implements dynamic model selection. Simple queries route to lightweight open-source models like Llama 3. Complex reasoning reserves Claude 3.5 Sonnet. This tiered approach reportedly reduces token consumption to one-third of industry averages.

Virtual Machines and Token Economics Enable Production Scale

Multi-agent systems face unique engineering challenges. These don’t exist in single-agent architectures. Coordination complexity grows non-linearly. Early versions spawned excessive subagents. They searched endlessly for nonexistent sources. Communication bottlenecks emerged.

Anthropic’s production experience building Claude Research provides insight into these challenges. The team implemented explicit scaling rules. Simple fact-finding requires one agent with 3-10 tool calls. Complex research deploys 10+ subagents with clearly divided responsibilities.

Context management becomes critical at scale. Agents summarize completed work phases. They store essential information in external memory before proceeding to new tasks. This prevents context overflow while preserving conversation coherence.

Parallel Execution Delivers Dramatic Speedups

Parallel execution delivers dramatic speedups. Lead agents spawn 3-5 subagents simultaneously. Each subagent uses 3+ tools in parallel. According to Anthropic, these changes cut research time by up to 90% for complex queries.

The engineering challenge extends beyond algorithms. It reaches operational concerns. Multi-agent systems are stateful. They are non-deterministic. They are prone to emergent behaviors. Minor issues that might be manageable in traditional software can completely derail agent trajectories. This leads to unpredictable outcomes.

Production deployment requires rainbow deployments. These gradually shift traffic between versions. Both versions are maintained simultaneously. This prevents disruption to in-progress sessions. Robust error handling becomes essential. Detailed logging is critical. Comprehensive evaluation frameworks form necessary infrastructure.

Token Economics Matter Intensely

Token economics matter intensely. According to Anthropic, multi-agent systems use approximately 15 times more tokens than standard chat interactions. This makes them economically viable only for tasks where the value of the outcome outweighs the expense.

This cost structure shapes product strategy. Manus targets high-value business workflows. Automation delivers substantial ROI in market research, competitive analysis, application prototyping, and data pipeline creation. Tasks where hours of human labor can be replaced with minutes of AI execution justify the computational overhead.

Global Tech Giants Race to Build Multi-Agent Platforms

The strategic implications of Manus’s success haven’t escaped major technology companies. Multi-agent capabilities represent potential competitive advantage across three dimensions.

Multi-Agent Systems Redefine Competitive Advantage

Technologically, multi-agent systems amplify base model capabilities. A company with strong foundational models but weak application distribution can suddenly deliver end-to-end solutions. These compete with integrated platforms.

From a product perspective, multi-agent interfaces fundamentally change human-AI interaction patterns. Users transition from operating AI tools to delegating complete workflows to AI workers. This shift enables entirely new product categories.

Strategically, multi-agent platforms could become the next ecosystem control point. Just as iOS and Android became foundational to mobile applications, multi-agent orchestration platforms may become essential infrastructure. They will be critical for AI application development and deployment.

Meta Acquisition Triggers Geopolitical Response

Meta’s acquisition signals the stakes involved. Meta purchased Manus for over $2 billion in December 2025. This represents a 4-6x valuation increase from the company’s $75 million Series B at $500 million valuation in April 2025.

Read more

China’s supermarket sector faces unprecedented pressure. Foreign giants like Carrefour have shuttered operations while domestic chains post consecutive years of losses.

E-commerce platforms and membership warehouses like Sam’s Club capture market share with convenience and bulk pricing. Traditional hypermarkets lost ground as consumer preferences shifted and economic slowdown reduced spending.

Against this backdrop, a regional chain operating exclusively in Henan province has become the industry’s unlikely blueprint for survival. Pangdonglai, founded in 1995 in the fourth-tier city of Xuchang, now has executives from across China traveling to study its stores.

The profit paradox driving desperate imitation

The numbers tell a startling story.

Pangdonglai’s 13 stores generated RMB 23.5B ($3.24B) in revenue in 2025, representing 38.71% year-over-year growth, according to sales data published by Pangdonglai Group.

The 2024 revenue reached RMB 17B ($2.34B) with over RMB 800M ($110.2M) in profit, according to Henan’s Provincial Party Committee Publicity Department.

Meanwhile, Yonghui Superstores, China’s second-largest hypermarket chain, reported RMB 67.6B ($9.31B) in revenue but posted a net loss of RMB 1.47B ($201.7M) for the same period.

The per-store economics reveal the gap.

Each Pangdonglai location averages approximately RMB 1.3B ($179M) in annual revenue. Yonghui operates hundreds of stores nationwide yet generates far lower per-location returns.

The regional chain’s gross margins reached 30% in 2020, according to Southern Weekly, compared to Yonghui’s 21%.

The industry context makes Pangdonglai’s performance even more remarkable. The top 100 supermarkets in China posted sales of RMB 900B across 25,200 outlets in 2024, representing growth of just 0.3% year-over-year, according to the China Chain Store & Franchise Association.

Market share for the top 10 retailers declined by 2.4 percentage points as fragmentation accelerated.

Pangdonglai made the CCFA’s top 100 list for the first time in 2024, ranking 31st despite operating fewer stores than virtually every other entrant. While competitors chase scale, the Henan retailer demonstrates that superior unit economics can outperform geographic expansion.

Yonghui’s billion-dollar bet on transformation

Facing three consecutive years of losses, Yonghui made an extraordinary decision in May 2024. The struggling giant partnered with Pangdonglai founder Yu Donglai to completely overhaul its operations. Former Yonghui CEO Li Songfeng stated,

Pangdonglai has insisted on quality, serving employees and customers well, forming a positive cycle of trust. This is a path worth learning from for the retail industry.

The partnership intensified in September 2024 when Miniso acquired a 29.4% stake in Yonghui for RMB 6.3B ($870M), becoming its largest shareholder.

Miniso founder Ye Guofu, who took over as head of Yonghui’s reform committee and acting CEO, has repeatedly stated that “the Pangdonglai model is the only way out for Chinese supermarkets.”

Early results suggest the transformation is working. Yonghui’s revamped store in Zhengzhou reopened in June 2024, attracting nearly 13,000 customers on the first day—five times normal traffic, according to Caixin Global.

The location became a near-mirror image of Pangdonglai, adopting its customer-friendly service standards and stocking DL-branded products. Sales jumped tenfold in subsequent weeks.

The Xi’an Zhongmao store delivered even more dramatic improvements. Daily sales rose from RMB 200K to RMB 1.6M in the first two days after revamping, according to Yonghui, while customer flow increased from 3,000 to over 14,000.

Read more

• Geographic bifurcation emerging: Southeast Asia absorbs 55% of fintech users while US-Japan-Korea extract 60% of gaming revenue, forcing dual strategies for volume versus value.

• AI creates overtaking opportunity: Application layer leadership in companion apps, content localization, and agent platforms enables cost advantages even as model layer catches up to Western competitors.

• Regulatory walls rising faster than ecosystems build: Data localization requirements and compliance costs force strategic pivots from direct operations to partnership models in developed markets.

• Concentration risk disguised as dominance: Leading apps derive 60-70% of revenue from single markets, creating fragility as local competitors learn and regulatory hostility intensifies.

Six verticals show common pattern

Chinese internet companies command significant positions across gaming, social applications, content platforms, AI tools, online education, and financial technology, according to TMTPost.

Chinese self-developed games generated $18.6B in overseas revenue. Growth reached 13.4% annually. Short drama platforms claim 46% of global market share after 1,454% revenue expansion. And, AI companion apps posted 30x monthly active user growth over five years. Product counts surged 191.89% in the past year.

Business models China perfected domestically now define global markets.

Short drama platforms adopted the freemium structure Chinese apps pioneered. Free content hooks users. Paid unlocks generate revenue. This model had limited presence in Western markets before Chinese platforms popularized it. All top 10 global short drama applications in 2024 are Chinese-owned.

Read more

Chinese AI models have gained measurable traction in Silicon Valley’s startup ecosystem during 2025, challenging assumptions about the AI competitive landscape. The adoption reflects economic pressure on cash-constrained startups combined with improving technical capabilities of open-source Chinese alternatives. What began as cost-conscious experimentation has evolved into production deployment at notable companies and venture-backed startups.

The trend gained visibility in October 2025 when executives at major technology companies publicly endorsed Chinese models. The endorsements came as live trading competitions demonstrated Chinese models’ practical capabilities in high-stakes financial environments, providing empirical validation beyond benchmark scores.

For venture-backed startups operating under cash constraints, the economics are straightforward. API costs for closed-source models can consume 20% to 40% of monthly burn rates for AI-focused companies.

Chinese open-source alternatives reduce this expense to near-zero when self-hosted, extending runway by months without requiring additional fundraising.

Airbnb publicly endorses Qwen over ChatGPT

Brian Chesky revealed in October 2025 that Airbnb relies heavily on Alibaba’s Qwen models to power its AI-driven customer service agent, describing the model as very good, fast and cheap.

The CEO told Bloomberg and CNBC that OpenAI’s ChatGPT wasn’t quite ready for Airbnb’s needs, with the software development kit not robust enough for the company’s requirements.

The statement carried weight given Chesky’s friendship with OpenAI founder Sam Altman, representing a departure from typical Silicon Valley diplomacy around AI providers.

Airbnb uses 13 different AI models including those from OpenAI, Alibaba’s Qwen, Google and open-source providers, but Chesky’s specific praise for Qwen highlighted performance parity with cost advantages.

Chesky told Bloomberg the company uses OpenAI’s latest models but typically doesn’t use them much in productionbecause there are faster and cheaper models. For companies processing millions of queries monthly, open-source models eliminate per-token API fees, translating to cost reductions of 80% to 90% compared to GPT-4 or Claude pricing structures.

Airbnb’s AI system reduced average customer resolution time from nearly three hours to six seconds, cutting the need for human support by 15%. The operational improvements demonstrate that Chinese models meet enterprise-grade performance requirements for customer-facing applications at major technology companies.

Open-source startups show 80% adoption of Chinese models

Martin Casado, a partner at Andreessen Horowitz, revealed in The Economist that 80% of startups pitching to them that use open-source models are now using Chinese AI models.

Casado later clarified the statistic referred to 80% of the 20-30% of new applicants running open-source models, translating to 16-24% of all startups using Chinese open models.

Martin Casado is a general partner at a16z in charge of a $12.5 billion infrastructure fund, having invested in AI startups including World Labs, Cursor, Ideogram and Braintrust. His background includes co-founding network virtualization company Nicira, which sold to VMware for $1.2 billion in 2012.

The venture capital validation carries strategic weight. Firms like a16z conduct extensive technical diligence before investing, including architecture reviews and performance testing. Their willingness to back startups built on Chinese models indicates confidence in the technology’s scalability and long-term viability.

Nathan Lambert, a machine learning researcher, stated he has personally heard of many high-profile cases where the most valued American AI startups are training models on Qwen, Kimi, GLM or DeepSeek.

The adoption reflects changing requirements as AI applications move from proof-of-concept to production, where reliability and cost predictability matter more than cutting-edge capabilities.

Chinese models like Kimi K2 charge just 15 cents per million input tokens and $2.50 per million output tokens, while Alibaba’s Qwen3-Max costs as little as $0.459 per million input tokens.

For early-stage startups operating under cash constraints, this dramatically extends runway.

Qwen dominates Alpha Arena live trading competition

The first AI Crypto Trading Competition held on the AIpha Arena platform by US-based Nof1.ai laboratory ran from October 18 to November 3, gathering six AI models: DeepSeek Chat V3.1, Qwen3 Max from Alibaba, GPT-5 from OpenAI, Gemini 2.5 Pro from Google, Claude Sonnet 4.5 from Anthropic, and Grok 4 from X AI.

Each model received $10,000 and traded autonomously on Hyperliquid decentralized exchange with real capital, executing trades without human intervention. All models used identical prompts and market data, trading six major cryptocurrencies.

Qwen 3 Max won with 22.32% returns while DeepSeek secured 4.89%, and the four American models posted losses of 30% to 62%. The competitive gap demonstrated fundamentally different decision-making capabilities in high-volatility environments.

DeepSeek displayed traits of a quantitative fund manager, rapidly constructing a multi-coin, low-leverage diversified portfolio, strictly following buy-on-pullback discipline. This mirrored the quantitative hedge fund background of DeepSeek’s parent company.

Qwen’s strategy was aggressive, focusing on Bitcoin with high leverage and going all-in during the October 23 market rebound, pushing returns to 51%. The peak showdown occurred October 27 when Qwen shifted to ETH with 25x leverage, resulting in a $4,150 single-day loss as prices declined, while DeepSeek secured $7,463 profit.

The results challenge AI capability hierarchies. GPT-5 and Gemini rank highest on traditional benchmarks like MMLU and GPQA but exhibited poor judgment in financial markets. The disconnect between benchmark scores and real-world performance raises questions about existing evaluation frameworks.

Cost-performance gap reshapes competitive dynamics

In an analysis published by AllianceBernstein in February, DeepSeek’s pricing was estimated at a fraction of the cost of OpenAI’s.

American closed-source providers charge premium rates for custom model variants, often requiring enterprise contracts with minimum commitments exceeding $100,000 annually, while open-source Chinese models allow unlimited fine-tuning at no marginal cost beyond compute resources.

Developers such as Beijing-based Z.ai and Hangzhou-based DeepSeek reported using older-generation chips not subject to US export controls in relatively small quantities, dramatically reducing training and running costs.

To an average startup, what really matters is speed, quality and cost at scale, according to Aman Sharma, co-founder of Lamatic, a South Florida firm helping businesses build AI solutions. Chinese models consistently balance these three factors effectively.

Open-source models create strategic advantages beyond cost. Companies can inspect code, verify security properties and ensure data privacy in ways impossible with API-based services. For regulated industries or companies handling sensitive user data, these transparency benefits often outweigh pure performance considerations.

Margin pressure on American AI companies intensifies as Chinese competition commoditizes model capabilities. OpenAI and Anthropic built business models assuming customers would pay premium prices for superior intelligence. As Chinese alternatives demonstrate comparable capabilities at fractions of the cost, the closed-source value proposition erodes.

Social Capital validates execution advantage

Chamath Palihapitiya revealed in October 2025 that Social Capital migrated much of its work to Moonshot’s Kimi K2 as it was way more performant and a ton cheaper than models from OpenAI and Anthropic.

Palihapitiya made these comments while co-hosting the All-In podcast, providing public validation of Chinese model capabilities in production environments.

As a former Facebook executive, Palihapitiya helped scale the platform from 45 million to 700 million users. His venture capital success with Social Capital established reputation for identifying technology inflection points ahead of broader market recognition.

The investor perspective emphasizes deployment excellence. Chinese labs ship new models every few weeks with measurable gains, while American labs maintain semi-annual or annual release cycles. Shayne Longpre, an MIT researcher, emphasized the paradigm-shifting pace of Chinese model releases with weekly or biweekly iterations offering choices that American labs don’t match.

Data from Hugging Face platform compiled by the ATOM Project confirmed Chinese models overtook the US in cumulative downloads, with Qwen reaching 385.3 million compared to Llama’s 346.2 million by October 2025.

Derivative systems built on Qwen now account for more than 40% of new language models on Hugging Face, while Meta’s share fell to 15%.

Nathan Benaich, founder of Air Street Capital, noted that the biggest factor where Chinese model adoption matters is for government and high-stakes enterprise applications where security is paramount, with concerns over training data provenance. For many commercial applications, however, the cost-performance equation drives adoption decisions.

Chinese AI companies iterate faster, moving from research to production in months rather than years. They maintain closer connections between model development teams and end-user applications, enabling rapid feedback loops that create ecosystem advantages complementing technical capabilities.

Market displacement remains uncertain despite Chinese model gains

The adoption of Chinese AI models by US startups represents competitive pressure rather than market displacement.

Cost economics favor open-source alternatives when performance reaches acceptable thresholds, creating pricing ceilings that challenge American closed-source providers’ business models. American companies retain advantages in capital access, research talent and government partnerships, but these strengths matter less if markets value deployment efficiency and cost economics over capability ceilings.

What remains uncertain is whether this represents temporary dislocation or sustained dynamics. Bloomberg analyst Michael Deng stated he doesn’t expect Chinese models to take over the market anytime soon, but they will maintain competitive pressure and price ceilings on AI costs.

For government and high-security applications, data sovereignty concerns continue limiting Chinese model adoption regardless of cost advantages.

The pattern mirrors historical technology transitions where open-source alternatives gained traction in price-sensitive segments while proprietary solutions maintained positions in premium markets.

Whether OpenAI and Anthropic can evolve business models to compete with open-source economics will determine the next phase of AI market structure.

For now, Chinese models have established themselves as credible alternatives that startups and cost-conscious enterprises increasingly deploy in production.