Physix Frontier · News Briefing Card (Leiphone · Oct 23, 2026)

SiliconFlow Pitches Token Productivity Formula for SOE Inference Cost Cuts

KEY FACTS

SiliconFlow shared its large-model inference infrastructure and ROI practices at a CIO summit for central state-owned enterprises.
It proposed the Token Productivity Formula: capacity times utilization rate times delivery quality.
The company designed a four-layer decoupled architecture covering everything from model operations to compute scheduling.
It has adapted to nearly one hundred models and is working with partners such as Huawei Cloud to optimize inference on domestic chips.
In one central SOE case, the model lineup was consolidated to about ten, and cost per million tokens fell from over one hundred yuan to under ten yuan.

KEY DATA

1920 tokens/sSingle-card Decode throughput (DeepSeek-

4000 tokens/sSingle-card Prefill throughput (DeepSeek

1000 tokens/sSingle-card Decode throughput (DeepSeek-

from over one hundred yuan to under ten yuanCost per million tokens change

PHYSIX OBSERVATION

Large-model deployment at central state-owned enterprises is shifting from stacking compute to competing on inference efficiency. SiliconFlow's formula and case show that unified management and Serverless-ization can significantly cut costs, but the real value lies in opening up scenarios that previously could not run at all. The industry must guard against a disconnect between applications and infrastructure; Context optimization and DevOps collaboration will be the key to ROI in the next stage.

Source: Leiphone report