Everything About AI
NextFin News - China's artificial-intelligence industry is shifting its center of gravity from building ever-larger language models to deploying autonomous AI agents, a transition driven by a price war that has pushed Chinese model tokens to roughly one-tenth the cost of U.S. rivals and turned the model layer into a commodity battleground. The pivot marks the moment the race for bigger models gave way to the race for usable systems: with frontier capability now purchasable for pennies per million tokens, the competitive prize has moved up the stack to the platforms, workflows, and data that turn models into work.
The shift is measurable. Active enterprise AI agents in China are expected to jump from nearly 2 million in 2025 to an estimated 5 million in 2026, according to International Data Corporation, as companies move from pilot projects to embedding intelligence directly into business processes. An industry research report released in March by the China Business Industry Research Institute projected China's AI agent industry chain would exceed 100 billion yuan in 2026, with the global AI agent market reaching $17.5 billion. Meanwhile, the model layer that powers those agents has been collapsing in price: DeepSeek made a 75% cut on its flagship V4-Pro model permanent in May, and Xiaomi slashed API costs by as much as 99% days later.
Here is the tension the market has not fully priced: the companies that won the model race may not be the ones that capture the agent economy. When the core input becomes cheap enough, the bottleneck — and the margin — migrates to orchestration, enterprise data, governance, and the workflows that sit between the model and the outcome. The model builders are fighting a war of attrition; the platform builders are collecting the tolls.
The Model Layer Is Becoming a Commodity
The first question to ask is why a price war erupted at all. The answer is that Chinese labs had little choice. U.S. export controls on advanced AI chips have constrained access to the compute needed to scale brute-force model training, forcing Chinese developers to compete on efficiency rather than sheer parameter count. A report by the U.S.-China Economic and Security Review Commission noted that Chinese AI labs have made models more accessible than their U.S. counterparts, creating shared technical commons that lift the overall capabilities of China's AI ecosystem even as frontier performance lags by several months. That constraint became a strategy.
The numbers show how far costs have fallen. DeepSeek cut V4-Pro API prices to between 0.025 and 6 yuan per million tokens — approximately $0.0035 to $0.83 — from a previous range of 0.1 to 24 yuan, according to the company's statement. Xiaomi followed with cuts of up to 99% for its MiMo-V2.5 series, which third-party platforms list at roughly $0.435 per million input tokens and $0.87 per million output tokens. On Vercel's AI Gateway, which routes production traffic for deployed applications, Chinese open-weight models processed 29% of all tokens in June 2026, up from roughly one-ninth of total volume in April, while consuming less than 4% of total spending because they were priced at about one-tenth the average rate of U.S. frontier systems.
When a task doesn't need the best model, teams are beginning to route it to the cheapest one that's good enough, and the recent wave of models coming out of China is winning that trade.
Harpreet Arora, Vercel's head of agentic infrastructure, said in interviews. DeepSeek alone accounted for 22.6% of token volume on the gateway in June, ranking third overall behind Anthropic and Google and within two percentage points of Google's 24% share. ByteDance's Seedance captured nearly half of all video-related spending on the platform despite producing roughly one-third of video output, a sign that Chinese models are winning not just on price but on higher-value workloads.
The strategic logic is that falling inference costs are not promotional discounts but structural efficiency gains. Sanchit Vir Gogia, chief analyst and CEO at Greyhound Research, noted that V4-Pro was engineered to run at roughly a quarter of the single-token compute and a tenth of the memory footprint of its predecessor at very long contexts. "It is not a discount," he said. "It is an efficiency gain being passed through."
That distinction matters. A promotional discount can be reversed when margins come under pressure — and DeepSeek has already signaled that a floor exists, announcing it would double peak-hour prices for its V4 models once the full version goes live. Output tokens for the flagship model would rise from 6 yuan to 12 yuan per million during peak periods. But the underlying trajectory of cost per unit of capability is down, and that is what makes the model layer a commodity: when the marginal cost of intelligence approaches zero, pricing power evaporates.
Agents Are the New Value-Capture Point
If models are becoming the electricity, agents are the appliances — and the platform that connects them to the enterprise is the wiring. IDC's August report, "New Enterprise AI Platforms Are Emerging," found that companies are no longer satisfied with isolated AI applications and are instead building unified enterprise-grade AI platforms. The value proposition of AI has shifted from "technical toys" to "operational assets," and its application depth has moved from "single-point question answering" to "embedded in core business flows." The role of employees is evolving from "system operators" to "intelligent commanders."
The report describes a six-layer full-stack architecture: the AI foundation layer, data layer, intelligent agent development layer, process and business layer, application entry layer, and governance and security layer. Two construction paths dominate. The first, intelligent process automation, embeds AI into the business-process-management lifecycle, treating the agent as a "digital employee node" that performs initial reviews, fills forms, and flags risks. The second, data-AI integration, centers on data infrastructure — lakehouse integration, governance, model training and inference, and agent development — for industries with large data volumes and complex rules.
The scale of the shift is visible in the deployment numbers. IDC expects active enterprise AI agents in China to more than double from 2025 to 2026, and the China Business Industry Research Institute projects the market for enterprise-grade AI agent solutions to grow from 5.6 billion yuan in 2024 to 59.1 billion yuan in 2029, a compound annual growth rate of 60.2%. Digital human agents — one of the most commercially mature forms — are forecast to reach 36 billion yuan in 2026, up from 20 billion yuan in 2024.
Progress from pilot validation to AI-native operations will increasingly depend on governance and risk-management capabilities.
Lu Yanxia, IDC's research director for AI and big data, said. That is the quiet hinge of the whole transition: the bottleneck is no longer whether the model can answer a question, but whether an autonomous system can be trusted to act inside a live business process without breaking compliance, leaking data, or making an unrecoverable mistake.
This is why the competitive set is changing. The report identifies three types of players now standing side by side: full-stack ecosystem providers, vertical specialists, and ecosystem enablers. Vendors such as Inspur, Primeton, and Yanhuang Yingdong are building platforms that wrap models in process orchestration, low-code tooling, and governance — the layers where switching costs are high and margins can persist even as model prices fall. Primeton positions itself as an "enterprise-level digital intelligence foundation," with model-connector services that integrate legacy systems into AI workflows; Inspur encapsulates complex processes into callable business-skill packages so that models can understand enterprise data.
The Infrastructure Layer Beneath the Pivot
The agent economy does not float above the ground; it rests on computing power, cloud services, and domestic chips. The same industry research report projected China's AI computing-acceleration chip market to reach 381.39 billion yuan in 2026, with a competitive landscape of "multiple strong players coexisting" as domestic suppliers substitute for restricted U.S. hardware. China's cloud computing market, the carrier for agent services, is expected to reach 1,398.6 billion yuan in 2026, while the AI large-model market itself is forecast to grow from 68 billion yuan in 2026 to 325 billion yuan by 2030.
Those figures expose the second-order structure of the pivot. The model layer — the 68 billion yuan segment — is being squeezed by the price war. The infrastructure underneath it — the 381 billion yuan chip market and the 1.4 trillion yuan cloud market — is where the capex is flowing and where pricing power is more defensible. The agent layer in the middle, growing at 60% annually, is the battleground where value will either be captured or passed through.
Policy is accelerating the shift. China's State Council issued opinions on deepening the implementation of the "Artificial Intelligence+" initiative, framing AI as a core pillar of economic and technological leadership and directing deployment across manufacturing, public services, finance, and smart cities. That directive channels procurement toward domestic platforms and gives the agent build-out a customer with deep pockets: the state itself.
The Second-Order Consequence: Who Wins the Agent Stack
The first-order read of this story is simple: cheaper models help everyone. The second-order read is sharper and less comfortable for the model builders themselves. When the core input is commoditized, value migrates to the scarce complements: proprietary enterprise data, workflow integration, distribution, and trust. A model lab that competes only on price is selling into a market where the buyer's switching cost is near zero; an enterprise platform that owns the workflow owns the customer.
The market has already begun to price this intuition. Chinese AI and semiconductor stocks have rallied since the breakout of DeepSeek's R1 model in January 2025, with the Hang Seng Tech Index rising 23% in 2025. Zhipu AI, billed as the world's first publicly listed large AI model company, debuted in Hong Kong in January 2026 at HK$116.2 per share, implying a market capitalization above HK$51.1 billion, and traded as high as HK$182.3 in early call-auction trading — valuing the company at more than HK$80 billion. But the rally has been uneven: large Chinese tech groups have largely missed the AI stock-market frenzy, while smaller model startups and chip designers have captured the premium.
That dispersion is the market's way of asking which layer will hold value. The base case is that platform and infrastructure providers — the companies selling the six-layer stack, the governance tooling, and the domestic compute that substitutes for restricted chips — will retain pricing power longer than pure model vendors. The model vendors that survive will be those that can move up the stack themselves, bundling agents and workflows with their APIs rather than selling raw tokens.
The Counter-Thesis: Model Capability Still Matters
The strongest case against this pivot is that it confuses cost with capability. American frontier models still maintain a clear lead in overall performance across benchmarks, from math and reasoning to code generation and long-horizon agentic tasks, according to the U.S.-China Economic and Security Review Commission. Agentic workflows are exactly where capability matters most: a system that autonomously decomposes tasks, invokes tools, and executes over long horizons amplifies every weakness in the underlying model. A hypothetical 10% error rate in a single-turn chatbot is an annoyance; the same error rate in an autonomous agent that can take actions is a liability.
There is also evidence that the price floor is rising, not falling. DeepSeek's decision to double peak-hour prices for its V4 models suggests that even the cheapest provider recognizes that inference capacity is not infinite — the chips underneath are scarce, and demand is growing faster than domestic supply. If agentic workloads prove more compute-intensive than chat workloads, as the higher per-token complexity of back-office agents suggests, the cost advantage could compress faster than the adoption curve expects.
This counter-thesis is real but incomplete. It correctly identifies the ceiling on how far commoditization can go, but it mistakes the near-term battleground. For the vast majority of enterprise use cases — document processing, customer service, internal workflow automation — the quality bar is "good enough," and the purchasing decision is a budget line, not a technical benchmark. Kyle Chan, a fellow at the Brookings Institution's John L. Thornton China Center, noted that rising token prices from leading U.S. labs have made model selection a budget-line decision rather than a purely technical one. The model race is not over; it has simply moved to a different field, where efficiency and deployment matter as much as raw capability.
What to Watch: The Next Phase of the Pivot
The outlook splits by time horizon. In the short term, the agent build-out is a sentiment and capex story: Chinese tech and chip names that can credibly claim agent-platform exposure will continue to attract momentum capital, but the rally is vulnerable to profit-taking. Exchange filings show that 73 listed semiconductor companies completed share-reduction plans in 2026, a sign that insiders see elevated valuations. In the medium term, the winners will be determined by enterprise traction: the companies that can show agents embedded in core business flows, with governance that satisfies compliance, will separate from the vendors selling demos. In the long term, the structural question is whether the agent layer can defend margins against the same commoditization pressure that just hit the model layer — or whether the value simply migrates again, this time to the chip and cloud infrastructure underneath.
Three signals will tell the story. First, whether the 2-million-to-5-million agent deployment figure for 2026 is met — a miss would suggest the pivot is ahead of actual enterprise readiness. Second, whether model pricing stabilizes or resumes falling; a renewed price cut from DeepSeek or Alibaba would confirm the commodity thesis, while sustained price increases would signal scarcity returning. Third, and most important, whether any Chinese agent platform can demonstrate durable gross margins above the model layer — the proof that value has truly migrated up the stack.
The falsifying signal for the structural-shift thesis is specific: if, by the end of 2027, model vendors still capture more than half of total AI-related spending in China while agent-platform gross margins remain below 40%, the pivot has not transferred value — the model layer remains the tollbooth, and the agent economy is just a cheaper way to use it.
China's AI industry has spent the past two years racing to build the best model. The next two will decide whether the prize was the model at all — or the system that makes it work. The price war settled the first question. The agent race will settle the second.
Explore more exclusive insights at nextfin.ai.
