2026 Sustainable Global Leaders Summit was held from October 20 to 22 at the Huangpu Expo Park in Shanghai.
Yang Huiqian, General Manager of Intelligent Computing Business at Sensetime-W (HK: 00020) Large Device, attended and delivered a speech.
Yang Huiqian stated that the current idleness or vacancy of domestic compute is a structural problem, which stems from both historical reasons and is closely related to the phased migration of compute application scenarios.
He recalled that in earlier years, compute application scenarios were primarily focused on large model training as the main track, and investment and applications generally concentrated on compute card type selection, networking architecture, and other aspects.
However, as token volume surged massively from 2025 to 2026, the track is shifting from large model training to model inference and agent applications.
Subdivided scenarios such as text-to-video are also gradually becoming increasingly mainstream, and new research-oriented inference demands like AI4S are simultaneously emerging.
In Yang Huiqian's view, the structural migration of compute application demand from training to inference is an important reason for the structural mismatch at the underlying compute level.
Whether it is chip selection, cluster networking, architecture design, or regional adjustment, structural imbalances exist.
Yang Hui pointed out that currently, whether domestic compute or high-end compute, there is still room for further releasing application value.
On one hand, facing new inference demands, previously idle compute can be reapplied through promoting new technology explorations such as PD separation and AF separation.
On the other hand, domestic compute itself is moving from "usable" to "easy to use," and Sensetime-W (HK: 00020) has already achieved relatively good economic benefits in inference applications on domestic chips.
Yang Huiqian believes that these structural problems currently faced by domestic chips can be gradually improved through technological architecture innovation, task decomposition, and resource reallocation, thereby alleviating the mismatch between existing compute and application demands and unlocking the potential value of existing compute assets.
Comments