NextFin News - China’s AI campaign is turning into an economic policy, not just a technology theme. The country’s official data chief says daily token consumption has surged to 140 trillion as of March 2026, up from 100 billion at the start of 2024, while Beijing has paired that usage boom with a two-track policy push: an industrial plan that aims to secure core AI technologies by 2027 and a July action plan that widens AI deployment across data, computing power, ecosystems, talent, standards, and governance.
Those numbers matter because they point to a shift in how Beijing wants growth to happen. The state is no longer treating AI as a niche software layer sitting above the economy. It is treating AI as part of the production function itself, something that can be embedded in factories, logistics chains, industrial software, and public systems. That is a different ambition from a consumer-internet cycle. It is an attempt to rebuild productivity in an economy still dealing with property weakness, slower labor-force growth, and pressure on margins.
The question for investors and policymakers is not whether China has embraced AI. It clearly has. The real question is whether this becomes a durable productivity regime or another state-backed capital cycle that looks bigger than its economic payoff. The answer is likely to differ by horizon. In the short run, the policy push will keep creating bursts of spending and sentiment. In the medium and long run, the more important effect is whether AI changes the way Chinese firms allocate labor, capital, and time.
The official language already leans toward the structural view. In January, China said it aims to achieve secure and reliable supply of key core AI technologies by 2027, while also targeting three to five general-purpose large AI models in manufacturing, 100 high-quality industrial datasets, and 500 typical application scenarios. In July, the government added an action plan for AI cooperation and development that calls for deeper AI application across industries, joint cultivation of digital talent, and stronger development of rules, standards, security, and governance.
That combination is telling. Beijing is not merely subsidizing model training. It is trying to build an AI stack that runs from data collection to compute, from industrial deployment to standards-setting. The strategy is broader than a chip race, but chips remain the pressure point because no AI diffusion plan can outrun hardware constraints forever. If the supply of advanced semiconductors remains tight, the country will have to lean harder on efficiency, software optimization, and selective deployment in industries where the return on each unit of compute is clearest.
The latest token figure gives a sense of how fast the diffusion is moving. Liu Liehong, head of the National Data Administration, said daily token consumption had risen more than 1,000-fold in a little over two years. On one reading, that is evidence of broad adoption. On another, it is evidence of feverish experimentation, with firms and local governments trying to avoid being left behind. The distinction matters because token usage is not the same thing as profit, productivity, or durable cash flow. It is a usage proxy, not a full measure of economic value.
There is also a subtle economic difference between a token surge and a growth surge. Token usage can rise because more enterprises are moving from pilot projects to real workflows, but it can also rise because the same firms are repeatedly testing prompts, retraining models, or chasing efficiency gains that never fully arrive. The policy rhetoric makes the first explanation sound inevitable. The balance-sheet reality often makes the second one more common. That tension is why the market should be careful not to confuse activity with output.
What the Token Surge Really Says About China’s Growth Model
The deeper story is not that China has more AI activity. It is that China needs AI activity to do a job that old growth engines no longer can. Property investment is no longer the reliable multiplier it once was. Demographics are less favorable. Local-government balance sheets are tighter. In that environment, AI offers a new way to lift output without relying on the same construction-heavy or credit-heavy channels that dominated earlier cycles.
That is why the industrial plan is framed the way it is. The January document does not just ask for more models; it calls for secure supply, industrial datasets, manufacturing integration, and core-process deployment. This is a productivity play. It aims to reduce friction inside firms: planning delays, inventory mismatches, inspection errors, maintenance outages, and the lag between data generation and decision-making.
The wording matters because it shows where Beijing sees the bottlenecks. The plan says the country will promote the coordinated development of AI chips’ hardware and software, support innovations in model training and inference methods, foster key industry-specific large models, and deeply embed large model technology into core production processes. That is not a consumer-app strategy. It is a supply-chain and factory-floor strategy.
Those gains are often modest at the level of a single plant, but they compound across a national industrial base. A 1% improvement in throughput, a small reduction in scrap, or a faster procurement cycle can matter more for aggregate growth than a handful of headline-grabbing model launches. That is the second-order effect the market often misses. The first-order story is that AI spending rises. The second-order story is that firms with better data and more disciplined workflow design can outcompete peers that merely buy the same tools and hope for better results.
China’s own official framing points in that direction. The July action plan says the goal is to strengthen international AI collaboration and bridge the digital divide while also promoting deeper AI application across industries and supporting rules and standards. That sounds diplomatic, but economically it means the state wants diffusion without chaos: broad adoption, but with guardrails around data, security, and governance.
That framing also makes the 2027 targets easier to interpret. Three to five general-purpose large AI models in manufacturing is not a vanity metric. It implies an effort to move from one-off demonstrations to reusable industrial platforms. One hundred high-quality industrial datasets means the country is trying to build the raw material for reliable training and deployment. Five hundred typical application scenarios means policymakers want AI embedded in repeatable cases that can be copied across firms and regions rather than trapped in isolated pilots.
“The document outlines actions in eight areas covering data, computing power, ecosystems, industrial empowerment, talent development, rules and standards, governance and AI ethics,” the government said in its July 17 action plan.
That list is the important part. It shows the state understands AI as an ecosystem problem rather than a single-product race. Models are only one layer. The other layers are data quality, industrial know-how, compute access, training pipelines, and the rules that decide how quickly firms can deploy. In that sense, China is trying to compress an entire technology stack into its industrial policy apparatus.
The structural argument is strongest where AI can be embedded into repetitive, measurable processes. Manufacturing inspection, scheduling, logistics routing, energy management, customer service, code generation, and procurement are all areas where incremental gains can be tracked. If those gains show up, AI becomes a capital-efficiency story as much as a growth story. If they do not, the buildout remains largely a capital cycle.
That is why the token surge should be read carefully. Token volume can rise because models are genuinely doing more useful work, but it can also rise because firms are testing, retraining, benchmarking, and using AI in ways that are not yet deeply monetized. The usage curve is real. The earnings curve is still the harder proof.
There is a second-order implication that is easy to miss. If AI really improves industrial efficiency inside China, the winner is not just the AI vendor. It is also the exporter that can quote lower prices because its plant runs better, the logistics operator that can turn inventory faster, and the utility that can feed data centers and factories more reliably. In other words, AI can act like a margin multiplier across the supply chain even where the headline model business itself is small. That is why policy makers keep returning to manufacturing rather than consumer services: the spillovers are bigger and more measurable.
The Bottleneck Is Not Enthusiasm. It Is Translation Into Output.
The strongest reason to stay cautious is that policy momentum does not guarantee productivity payoff. China has plenty of experience with strategic sectors that attracted investment before the economics fully worked. If too many firms chase the same AI narrative too quickly, the outcome can be overcapacity in infrastructure, thin utilization, and a lagging return on capital. In that scenario, the economy gets the capex, but not the productivity.
That is the most serious counter-thesis, and it deserves weight. A state-directed rollout can look impressive in announcements while actual business adoption remains uneven. Local governments may fund projects because they fit policy priorities. Companies may announce deployments because they signal alignment. Neither automatically proves the tools are changing output. If AI spending rises sharply while industrial margins, software penetration, and workflow efficiency remain flat, then the narrative is too optimistic.
But the skeptical case has a limit. It assumes AI is just another investment cycle. The problem is that the official plans are aimed at altering how firms work, not just how they spend. The January plan’s targets for industrial datasets, manufacturing models, and production-process integration are designed to create repeatable usage, not one-off pilot projects. That suggests the intended mechanism is structural even if the near-term spending pattern is cyclical.
In other words, the cycle is in the rollout; the structure is in the objective. That is why the most useful frame is time-horizon specific. In the short term, AI can still behave like a theme that attracts capital, lifts supplier names, and fuels policy headlines. In the medium term, the winners will be the firms that convert AI into lower costs and better throughput. In the long term, the broader economy either absorbs AI into its production system or proves unable to translate token volume into productivity.
The difference between those outcomes is not abstract. It will show up in measurable signs: industrial profit margins, manufacturing value added, cloud and data-center utilization, enterprise software adoption, and the pace at which AI moves from trial deployments into core processes. Those are the numbers that matter. Without them, token growth is impressive but incomplete.
One useful way to read the policy mix is as a search for a new industrial operating system. China wants compute, data, and industrial process knowledge to interact the way capital, labor, and logistics once did in earlier growth phases. That makes AI less like a consumer gadget and more like a coordination technology. Coordination technologies matter because they lower the cost of getting many actors to do the right thing at the same time. That is where large productivity gains often come from.
Still, coordination only works if the system has enough reliable hardware, enough trained workers, and enough incentives to deploy the tools inside real workflows. The government can accelerate all three, but it cannot eliminate the trade-offs. Scarce chips must be allocated. Talent must be trained. Data must be cleaned and secured. Governance must avoid both stifling adoption and letting risk run ahead of control. These frictions do not kill the thesis. They slow it down and make the payoff uneven.
The broader consequence is that AI becomes a filter for organizational quality. Firms with clean data, fast decision-making, and high process discipline can turn the same model into better margins and faster throughput. Firms with messy inputs and weak management cannot. That means the AI push may widen differences inside sectors as much as it lifts sectors as a whole. The macro story becomes partly a micro sorting story.
What Would Prove The Skeptics Right?
The best counter-argument is that AI in China may be over-read as a structural growth engine when it is still mostly a policy-led capital cycle. That view is supported by history. Many strategic campaigns begin with strong official support, rapid capex, and fast headline adoption, but then run into utilization problems and weaker-than-expected returns. The same could happen here if firms deploy AI because they feel pressure to do so rather than because the tools are clearly improving output.
The signal that would prove that skeptical view right is quantifiable: if, over the next four quarters, AI-linked investment and token usage keep rising while industrial margins, enterprise adoption depth, and measured productivity gains do not improve, then the structural thesis fails. At that point, AI would look like a spending wave rather than a new growth engine.
For now, though, the official evidence points the other way. China has tied AI to manufacturing, datasets, ecosystems, and governance, which means it is building for durability rather than a short-lived theme. The country’s AI push is not just about generating more tokens. It is about changing the architecture of growth in an economy that can no longer rely on its old formula.
The short-term outcome is likely to remain noisy. Policy announcements, conference speeches, and infrastructure spending can keep sentiment hot. The medium-term outcome will be decided by whether firms use AI to cut costs and raise productivity. The long-term outcome is more consequential: whether AI becomes part of China’s industrial base in the same way rail, telecom, and payments once did.
The base case is uneven but real productivity improvement. The upside case is that AI helps offset slower labor-force growth and weak legacy sectors. The downside case is a capex cycle that leaves behind more hardware than output. The line between those scenarios will be drawn by the operating data, not the slogans.
China is trying to turn AI into a substitute for lost growth engines. If it succeeds, the winners will be the firms that convert data into output. If it fails, the country will still have built a lot of infrastructure — but not enough new growth.
As of Aug. 8, 2026, China’s AI push is best understood as a structural ambition carried by a cyclical rollout.
Explore more exclusive insights at nextfin.ai.
