At the 10th CSDI Summit in Shenzhen on October 16, SenseTime SenseCore vice president and CTO Xuan Shanming outlined how the company is rebuilding inference infrastructure for an agentic era in which multi-agent workflows and long unattended tasks drive token demand far beyond chatbots. The pitch is not how many accelerators a data center owns, but how steadily it can produce high-quality tokens under service-level and cost constraints.
SenseCore's answer centers on Prefill/Decode, or PD, separation inside 10,000-GPU-scale heterogeneous clusters at its Lingang AIDC. Prefill processes input tokens and is compute-heavy; Decode generates output tokens while repeatedly reading context cache and needs high memory bandwidth. In SenseCore's design, suited domestic accelerators concentrate on Prefill, high-bandwidth GPUs handle Decode, and the stages exchange state through KV Cache links that span mixed 800G, 400G and 200G fabrics in one cluster.
That split creates an operational risk: in some mixes, 30 to 50 mid- or lower-tier Prefill nodes may feed a single Decode node, so a Decode failure can strand a large Prefill pool. SenseCore says it keeps a small reserve of high-end cards in the Prefill pool for fast failover, and lets nodes switch Prefill and Decode roles dynamically under load. In a specific test configuration with the same count of flagship high-end GPUs as a baseline, the company reported about five times higher overall token throughput after heterogeneous PD scheduling. It is also exploring cross-domain PD, sending Prefill work to regional mid-tier capacity and Decode to remote high-bandwidth clusters as KV Cache for million-token contexts shrinks from near 100 GB toward about 1 GB on some models, plus Attention/FFN separation inside Decode.
SenseCore says its platform now serves about 6 trillion tokens per day and aims for 10 trillion tokens per day by the fourth quarter of 2026, with full-year token volume targeted to grow roughly 25 times. Those capacity and speedup figures are company statements. The talk framed domestic chips as taking a larger share of Prefill and mixed inference, not as a wholesale replacement for high-bandwidth GPUs on Decode.
SEE ALSO: Huawei Launches Grid-Interactive AIDC 1.0, Linking AI Data-Center Power, Cooling and Ascend Servers · Peng Cheng Cloud Brain III Puts 10,000P of Domestic AI Computing Into Trial Operation