NextFin

How AI Is Transforming China's Economy: The Factory-Floor Bet Behind the Boom

Summarized by NextFin AI
  • China's core AI industry exceeded 1.2 trillion yuan in 2025, with over 30 percent of large manufacturers adopting AI, as Beijing treats AI as industrial policy rather than a consumer craze.
  • Chinese AI models run at one-sixth to one-quarter the cost of comparable U.S. systems, with DeepSeek's API at roughly $0.028 per million tokens, driving adoption across emerging markets.
  • Huawei planned around 750,000 Ascend 950PR units in 2026, delivering 2.8 times the FP4 performance of Nvidia's China-compliant H20, as domestic chip substitution advances under U.S. export controls.
  • The bull case requires two conditions: domestic compute substitutes scale on schedule and AI spending converts into measured productivity, while the World Bank warns weak market competition may limit productivity gains.

NextFin News - China's core artificial-intelligence industry crossed 1.2 trillion yuan ($174 billion) in 2025, and more than 30 percent of the country's large manufacturers had adopted AI by the end of the year, the industry minister said in March. The question now facing investors is whether that buildout can do what Beijing needs it to do: lift an economy stuck in a property-led slowdown without firing off a debt bubble of its own.

The answer is not a simple boom story. China is pursuing AI not as a consumer-facing craze but as an industrial policy - a state-directed push to retool the world's largest manufacturing base, substitute domestic silicon for restricted imports, and export cheap intelligence to emerging markets. The 15th Five-Year Plan, approved in March 2026, mentions AI more than 50 times and puts it at the center of the digital-economy agenda. The World Bank kept its 2026 China growth forecast at 4.4 percent, citing high-tech investment - up 4.5 percent year over year between January and May - as a key support. But the same institutions warn that the productivity payoff is still unproven, the chip bottleneck is real, and the property drag has not gone away.

This is a structural shift with a cyclical overlay. The structural leg - an industrialized model-production pipeline, a cost structure a fraction of the American one, and a policy mandate running to 2030 - will not revert on its own. The cyclical leg - export-control shocks, capex timing, and a K-shaped domestic recovery - can and will fluctuate. Getting the two apart matters, because it decides whether China's AI push is a durable growth engine or a well-funded detour.

The Factory Floor Comes First

Unlike the United States, where AI's most visible impact runs through consumer software and cloud hyperscalers, China's AI story is anchored in the factory. By the end of 2025, more than 30 percent of manufacturing enterprises with annual main business turnover of at least 20 million yuan had adopted AI technologies, according to Li Lecheng, minister of industry and information technology, speaking on the sidelines of the National People's Congress. The core AI industry was valued at more than 1.2 trillion yuan in 2025, with the number of AI companies surpassing 6,200. China also ranked first worldwide in downloads of open-source AI models over the past year, and Chinese companies had released more than 300 humanoid-robot products.

The transmission mechanism runs through productivity at the margin of a vast, low-margin manufacturing base. A 0.3 percentage-point lift to activity from AI-driven productivity gains - the IMF's own estimate for what could materialize if the technology diffuses - would be meaningful for an economy growing at roughly 4.5 percent. The channel is not glamorous: predictive maintenance, visual inspection, supply-chain optimization, and demand forecasting. But in an economy where manufacturing accounts for nearly a quarter of GDP and where China represented 54 percent of global industrial-robot deployments in 2024, according to the International Federation of Robotics, small percentage improvements compound across hundreds of thousands of plants.

State policy is the accelerant. The 15th Five-Year Plan (2026-2030) calls for AI to be embedded throughout the economy, launches new infrastructure projects for hyper-scale intelligent computing clusters, and encourages large-scale commercial application in key sectors. The plan is not a production quota in the Soviet sense - Beijing abandoned that model in the 1990s - but it is a high-level signal that moves credit, land, and procurement. Local governments compete to build computing-power hubs; state-owned enterprises get marching orders to adopt; and the capital markets are being steered toward AI and semiconductor listings. GPU players Biren, Moore Threads, Illuvatar, and Enflame, along with model developers Zhipu, Minimax, and Moonshot, have been tapping exchanges in Shanghai and Hong Kong.

The hook here is the contrast: while Western debate fixates on chatbots and consumer subscriptions, China is treating AI as a capital-good upgrade cycle - the kind that shows up in fixed-asset investment before it shows up in earnings.

The Cost Revolution: Open Models and the Repricing of Intelligence

The second structural leg is that China has industrialized the production of near-frontier models at a cost the West is not matching. In roughly eight weeks between June and August 2026, five Chinese developers shipped six frontier-adjacent releases - Kimi K3 from Moonshot, Alibaba's Qwen3.8-Max, DeepSeek's V4-Flash and V4 Pro, Z.ai's GLM-5.2, and ByteDance's Seedance video model. State Street's assessment captures the shift:

This is more than a rerun of the DeepSeek shock, but not yet proof that Chinese and US frontier models are on equal footing.

The gap has narrowed, not closed - but the system now looks built to produce repeatedly rather than to surprise once.

Cost is the weapon. Chinese AI models run at roughly one-sixth to one-quarter the cost of comparable American systems, according to a RAND report published in early 2026. DeepSeek's API pricing sits at approximately $0.028 per million tokens - roughly one-180th of equivalent GPT pricing. That is not a promotional price war; it is a structural cost advantage built on three factors: cheaper engineering labor, architectural efficiency under compute constraints, and an open-source feedback loop that distributes improvement across a global developer base. The result is a Jevons-paradox dynamic: as the price of a unit of intelligence falls, consumption rises faster than efficiency gains, and the surplus accrues to the deployers - the enterprises and software vendors - rather than to the model developers themselves.

The adoption data backs this up. An Andreessen Horowitz partner estimated that roughly 80 percent of US startups built on open source now run on Chinese base models such as Qwen rather than OpenAI or Anthropic. Chinese models' weekly token consumption on OpenRouter surpassed US models in February 2026, and the gap has widened since. In Southeast Asia, Singapore's OCBC bank runs more than 30 internal tools on DeepSeek and Qwen. The implication for China's economy is indirect but powerful: if Chinese models become the default infrastructure for emerging-market digitization, the renminbi-denominated revenue may lag, but the standards, the data flows, and the dependency do not.

But there is a catch for the Chinese model developers: fast-rotating leadership means a thinner moat. When Kimi K3 paused sign-ups because GPUs ran out, or when GLM-5.2 briefly held the top open-model spot before being displaced, it showed capability is real but durable pricing power is not. The economics of AI, in Jensen Huang's "AI cake" analogy, concentrate profit at the compute base and at the deployment layer - not in the contested middle where Chinese labs are fighting.

The Choke Point: Chips, Sanctions, and the Race to Substitute

The third leg is the one that could break the thesis: compute. Washington's export controls remain the binding constraint, and China's response - domestic substitution - is advancing faster than many expected but is not yet proven at scale. Huawei planned around 750,000 Ascend 950PR units in 2026, with mass production beginning in April and full-scale shipments in the second half of the year, according to reporting in April. The chip delivers 1.56 PFLOPS at FP4 precision with proprietary high-bandwidth memory, achieving roughly 2.8 times the FP4 performance of Nvidia's China-compliant H20. GLM-5.2, a 744-billion-parameter model, was trained entirely on Huawei Ascend chips. DeepSeek's V4 is optimized for Ascend.

This is the mechanism that makes the sanctions story double-edged. Restrictions raise the cost of frontier training in the short run - Nvidia took a $5.5 billion charge after the United States limited exports of its H20 chip to China - but they also force a domestic supply chain into existence that would not otherwise have customers at scale. A Council on Foreign Relations analysis argues the opposite: if the United States were to export three million H200 chips to China in 2026, it would hand Chinese firms more AI computing power than they could produce domestically until 2028 or 2029 at the earliest. On that reading, keeping the lid on is the point - and loosening it would be a strategic error.

The near-term signal to watch is whether the 750,000-unit Ascend plan ships on schedule through the second half of 2026. If it does, the substitution thesis moves from story to data. If it slips, the model wave runs into a wall.

The Counter-Thesis: A K-Shaped Economy and a Productivity Lag

The strongest case against the transformation narrative is that AI investment is the bright spot in a K-shaped economy, not the engine of a broad recovery. Citi's second-half 2026 outlook frames AI-related activity as "the main anchor for growth, particularly through supply-side strength," while maintaining a 4.7 percent GDP forecast and warning that property, consumption, and fiscal deployment remain the swing factors. On that view, AI is doing what infrastructure investment has always done in China: filling a demand gap while the household side stays weak.

There is a deeper structural objection, and it comes from the World Bank itself. In its January 2026 Global Economic Prospects, the bank argued that the productivity benefit of AI at the firm or economy level is limited so far because weak market competition inhibits the reallocation of resources from less productive to more productive firms - which in turn lowers technology adoption. In plain terms: if state-directed credit keeps zombie firms alive, AI tools land in the wrong hands, and the aggregate productivity number disappoints. This is the single most important falsifying condition for the bullish case. The IMF's +0.3 percent activity lift is conditional on productivity gains materializing. If they do not, the investment shows up in GDP as capex today but not as total-factor-productivity growth tomorrow.

The valuation risk runs the other way. The IMF has warned that a moderate valuation correction in AI assets paired with tighter financial conditions could cut global growth by about 0.4 percent, with larger losses if investment falls harder. China's AI rally has been carried by sentiment and policy as much as by cash flow. If the global AI capex cycle rolls over, China's exporters of components and its domestic chip capex cycle would both feel it - even if the long-run structural story is intact.

So the adversarial verdict is this: the bull case requires two things to be true simultaneously - that domestic compute substitutes scale on schedule, and that AI spending converts into measured productivity rather than just measured investment. The bear case needs only one of those to fail.

What Comes Next: Beneficiaries, Exposure, and the Signals That Matter

Cash in the mechanism, and the picture sorts itself by time horizon.

In the short term - sentiment and liquidity - the beneficiaries are the listed proxies: domestic GPU designers, the Ascend supply chain, and the model developers queuing for Hong Kong and Shanghai listings. These trade on policy headlines and on whether the Ascend shipment numbers land. The exposed are the import-dependent cloud providers still waiting on licenses and any company whose valuation assumes uninterrupted access to leading-edge foreign silicon.

In the medium term - fundamentals - the winners are the enterprises that can turn cheap intelligence into measurable cost savings: exporters under margin pressure, logistics firms, and consumer-electronics manufacturers adding AI to their product mix. The IMF's baseline scenario has AI lifting activity by about 0.3 percent relative to baseline if productivity gains materialize; the downside scenario has a valuation correction and tighter financial conditions cutting global growth by about 0.4 percent. Those two numbers define the dispersion around the base case.

In the long term - structural - the question is whether China has built a self-sustaining AI stack that can keep producing frontier-adjacent capability under sanctions. If the 750,000-unit Ascend plan ships, if the core AI industry stays on a trajectory toward the 10 trillion yuan-plus scale officials have projected for 2030, and if manufacturing adoption climbs well beyond the current 30 percent, then the shift is durable and the West's leverage erodes. If adoption stalls and domestic chips miss, the 15th Five-Year Plan becomes another ambitious document that met reality.

The falsifying signals are concrete. First, watch the Ascend shipment cadence in the second half of 2026 - a material miss against the roughly 750,000-unit plan would break the substitution thesis. Second, watch whether AI adoption among large manufacturers climbs beyond 30 percent over the next two years; a flat read would support the World Bank's reallocation critique. Third, watch total-factor-productivity growth: if AI-driven productivity does not show through in the data by 2027, the investment-led story has not become a transformation story.

China is not trying to win the AI race the way the United States is - with frontier models and consumer platforms. It is trying to win it the way it won solar panels and batteries: by making the technology cheap enough, embedded deep enough in industry, and exported widely enough that the question of who leads at the frontier matters less than who owns the factory floor. That is a slower path to headlines, but a harder one to reverse.

Explore more exclusive insights at nextfin.ai.

Insights

What is China's core AI industry value?

How many factories adopted AI by 2025?

Why focus on factories not consumers?

What is the 15th Five-Year Plan goal?

How cheap are Chinese AI models?

What limits China's AI chip progress?

Can Ascend chips challenge Nvidia?

What risks face China's AI investment?

Is AI productivity gain proven yet?

How do US chip controls hurt China?

What signals prove AI transformation?

Who benefits from cheap AI models?

Why is property drag a risk for AI?

How does open source aid Chinese AI?

What is the AI Jevons paradox effect?

Can China substitute domestic chips?

What is IMF AI growth estimate?

How does China export AI globally?

What defines K-shaped economy risk?

Will AI fix China's economic slowdown?

Search, ask or analyze anything...