On November 3, the AI Industry Reference published its 2026 31st issue. Latest data shows that China’s AI inference data volume has exceeded its training data volume for the first time, with average daily token calls topping 175 trillion, marking the transition of the AI industry from ‘feeding the model’ to ‘using the model’ at scale.
Inference volume surpassing training volume means that many large models have entered production and are being called continuously by real workloads, rather than staying in training and evaluation. The 175-trillion daily token count corresponds to massive generation, code completion, multimodal understanding and real-time inference requests on edge devices; this structural shift moves the center of compute demand from training clusters to inference and edge deployment, putting cost, latency and privacy at the forefront.
From a competitive-structure perspective, inference is where AI value is realized. Offloading high-frequency, low-latency, privacy-sensitive tasks to the edge and keeping complex tasks in the cloud is becoming the mainstream architecture. For the compute industry, inference clusters, edge-inference boxes and on-device NPUs will see more durable growth, and the turning point between training and inference is reshaping the landscape of chips, servers and edge devices.
Inference moving to the edge is exactly the opportunity for edge hardware. Junyi Microelectronics’ edge AI hosts AI-BOX-3588 and 5G-EDGE-3399Pro and K-series edge motherboards are built for on-device inference, running lightweight models locally to cut backhaul bandwidth and latency; combined with touch all-in-ones, they let industrial and commercial sites deploy ‘model-plus-interaction’ applications right where the data originates.
Source: http://topic.scol.com.cn/2026/26aicyck/qs/202611/83349900.html (2026-11-03)