成立九年，中科类脑把积累装进Token工厂
Ranking
No observed public metrics; popularity remains neutral/archived.
Merged summary
TL;DR - Zhongke Leinao launched an integrated “electricity-compute-token” platform for operating AI infrastructure as a “Token factory,” jointly optimizing heterogeneous accelerators, model inference, workloads, and power consumption. The approach matters because AI infrastructure economics increasingly depend on useful tokens produced per unit of energy rather than raw compute capacity.
- Its BitaHub platform reportedly manages more than 3,000 compute nodes totaling over 5,000P and serves more than 80,000 enterprise and research users.
- A “1+3” architecture combines a central decision engine with compute, Token, and power subsystems, using workload, electricity-price, renewable-output, and hardware-efficiency forecasts for scheduling.
- The system can separate inference prefill and decode across suitable chips, migrate KV caches, and shift flexible batch jobs across locations or low-cost power periods while preserving latency-sensitive workloads.
- A three-city scheduling test reported control responses within 200 seconds, 100% cross-region migration success, and 98% energy-consumption forecasting accuracy.
Sources (1)
成立九年，中科类脑把积累装进Token工厂
TL;DR - Zhongke Leinao launched an integrated “electricity-compute-token” platform for operating AI infrastructure as a “Token factory,” jointly optimizing heterogeneous accelerators, model inference, workloads, and power consumption. The approach matters because AI infrastructure economics increasingly depend on useful tokens produced per unit of energy rather than raw compute capacity.
- Its BitaHub platform reportedly manages more than 3,000 compute nodes totaling over 5,000P and serves more than 80,000 enterprise and research users.
- A “1+3” architecture combines a central decision engine with compute, Token, and power subsystems, using workload, electricity-price, renewable-output, and hardware-efficiency forecasts for scheduling.
- The system can separate inference prefill and decode across suitable chips, migrate KV caches, and shift flexible batch jobs across locations or low-cost power periods while preserving latency-sensitive workloads.
- A three-city scheduling test reported control responses within 200 seconds, 100% cross-region migration success, and 98% energy-consumption forecasting accuracy.