BlockBeats news, on July 19, the semiconductor and AI independent research firm SemiAnalysis stated that although approximately three-quarters of Kimi K3's network layers utilize KDA, reducing KV cache transmission bandwidth by up to 10 times compared to a full global attention model, this does not imply a significant contraction in the AI network switch market.
Kimi K3 has 2.8 trillion parameters and still requires approximately 1.5 TB of HBM bandwidth per forward pass, even with MXFP4. To achieve a profitable deployment while maintaining reasonable interaction speeds, a large number of chips must be connected via high-bandwidth networks such as GB300 NVL72, and WideEP scaling services must be leveraged.
WideEP distributes 896 expert models across multiple GPUs and performs token routing and result aggregation twice at each layer during every forward pass, requiring over 120 such operations per forward pass. In contrast, KV cache transfers between prefill and decoding occur only once per dialogue round; thus, the bandwidth savings from KDA are likely far smaller than the increased network demands introduced by large-scale expert models.
SemiAnalysis believes that more efficient attention mechanisms could also push context lengths beyond 5 million tokens, up from 1 million tokens. According to Jevons' Paradox, increased efficiency may expand AI usage, further increasing network demand.