China’s AI deployment is no longer measured in research papers or benchmark scores. According to Digital in Asia, citing data presented by China’s National Data Administration at the China Development Forum, the country’s daily AI token consumption exceeded 140 trillion in March 2026. At the start of 2024, that figure stood at roughly 100 billion. The implied growth of more than a thousandfold in two years reflects not a research breakthrough but a deployment wave: millions of applications running models at scale, across industries, around the clock.
The scale of this activity is partly visible in third-party platform data. Digital in Asia reported that Chinese models accounted for 61 percent of total token consumption among the top ten models on OpenRouter, the AI inference routing platform, during February 2026. In the week of February 16 to 22, Chinese models processed an estimated 5.16 trillion tokens on the platform, compared with 2.7 trillion for U.S. models. “These aren’t projections,” wrote Digital in Asia author Tom Simpson. “They’re usage metrics published by China’s National Data Administration at the China Development Forum in March 2026, and confirmed by third-party platform data.”
A Deployment Scoreboard, Not a Benchmark
The significance of these numbers lies in what they are, and what they are not. Token consumption is a measure of inference demand: the number of words, characters, and reasoning steps processed through deployed AI systems. It is not a measure of model quality, economic productivity, or research capability. OpenRouter is a large and analytically serious platform. Its December 2025 State of AI report drew on more than 100 trillion tokens of anonymized real-world usage, but it represents a specific slice of the global market, weighted toward developers who use model-routing services. Platform token share and total global AI market share are not the same thing.
What the data does indicate, with appropriate caveats, is that Chinese open and low-cost models have gained substantial developer traction at the inference layer. Open-weight availability, aggressive pricing, and a mature domestic infrastructure for deployment have together made Chinese models a default choice for cost-sensitive applications. Digital in Asia links the token surge to open-weight model proliferation, falling prices, and rapid expansion of domestic data center capacity.
Infrastructure Demand Follows Usage
If inference volumes continue growing at anywhere near the rate implied by these figures, the strategic implications extend into hardware and energy. Every trillion tokens processed requires compute, electricity, and networking capacity. China’s reported build-out of Huawei Ascend-based clusters (with approximately 600,000 Ascend 910C chips planned for 2026 and the Atlas 950 SuperPoD architecture linking 8,192 chips per cluster, according to Digital in Asia) is consistent with an AI economy orienting itself around inference at industrial scale rather than only frontier model training.
The OpenRouter State of AI report also identifies agentic inference as a rising usage pattern: multi-step, tool-calling workflows that sustain longer active sessions and consume far more tokens per user interaction than simple chatbot exchanges. If agentic workloads scale in China as Baidu and others are projecting, the current token numbers may understate the compute demand that lies ahead.
Caveats and Context
The methodological limits deserve attention alongside the headline figures. OpenRouter’s dataset is built on anonymized request-level metadata, covering timing, model and provider identifiers, token usage, and system performance metrics, but the platform is not a representative sample of all AI usage globally. It skews toward technically sophisticated users and developers who route requests through an aggregation layer rather than using proprietary APIs directly. Enterprise deployments at large companies, government systems, and embedded industrial applications may not appear in OpenRouter usage at all.
The 140 trillion daily token figure sourced to China’s National Data Administration carries its own caveats: official government statistics on AI activity are not independently audited, and methodologies for counting tokens consumed across an entire national economy have not been publicly disclosed in detail.
What the convergence of these datasets suggests, held with appropriate uncertainty, is directional rather than precise: Chinese AI deployment has reached a scale that was not anticipated when chip export controls were designed, and the inference layer, not only frontier training, is becoming the contested terrain. The next policy question is whether that matters more for the competitive outcome than the benchmark gap that controls were designed to preserve.
The broader implication of the token data is that the AI competition is running at multiple layers simultaneously. At the frontier, U.S. labs retain benchmark advantages in the most capable closed models. At the inference layer, Chinese models are competing aggressively on cost and availability. At the industrial deployment layer, Chinese firms are embedding AI into manufacturing and logistics at scale. Each layer has different competitive dynamics, different policy levers, and different timelines. Token consumption data captures one of those layers with some fidelity. It does not reduce the complexity of the others.