Estimates of K3's GPU demands and expense versus DeepSeek V4.1 Flash

One commenter argues that at EP64 and a 77,000-token average context length reported by DeepInfer, both K3 and DeepSeek V4.1 Flash can reach compute-bound batch sizes, despite K3's claimed 16-fold larger KV cache. They estimate K3 would need 21,600 GB200 GPUs to serve the same traffic as 2,400 for DeepSeek, citing sixfold greater activations and no CED/YOCO. Another commenter estimates K3 is 25 times as expensive as V4.1F and highlights GPU availability.