SiliconFlow's Tang Anbo shares practices for central state-owned enterprises' large model inference infrastructure

On October 23, 2026, SiliconFlow's Tang Anbo gave a keynote sharing titled "From 'Compute' to 'Token Productivity': Inference Infrastructure and ROI Practices for Scaled Deployment of Large Models in Central State-Owned Enterprises" at the Central State-Owned Enterprise CIO Summit. According to a same-day report by Leiphone, in his sharing he proposed that Token productivity = capacity × utilization × delivery quality, and introduced a four-layer decoupled architecture—model gateway, model service platform, high-performance inference framework, and compute scheduling—to bridge the conversion layer from compute to tokens. The report also mentioned that SiliconFlow and Huawei Cloud launched DeepSeek R1/V3 Token services based on domestic chips, with DeepSeek-R1 single-card Decode throughput exceeding 1,920 tokens/s; after a central state-owned enterprise group consolidated forty to fifty model instances on about 50 machines into about ten and switched to Serverless calls, the cost per million tokens for the Qwen model dropped from over one hundred yuan to under ten yuan. Currently this sharing is a single-source report, with no follow-up developments yet.

AI-generated from public reports

1. Oct 23· 1 reports

   SiliconFlow Tang Anbo shares practices for central state-owned enterprises' large model inference infrastructure

   Source:English text is being prepared.