Kimi K3가 왜 AI의 병목을 계산에서 메모리로 옮겼는지 조사해 봤습니다.

I dissected Kimi K3, the 2.8-trillion-parameter model released by China's Moonshot, along four axes. A sparse MoE that calls only 16 of 896 experts, Kimi Delta Attention that compresses a long past into a single state, Attention Residuals that pull out only the past you need, and even MXFP4 low-precision training and Mooncake disaggregated inference. I trace how all of this design shifted AI's bottleneck from 'compute' to 'memory, communication, and cache,' and why the more efficient a model becomes, the heavier the memory, network, and power burden on the data center actually grows. ※ This video is a structural analysis based on publicly available technical materials and third-party evaluations, and is not a recommendation to buy or sell any particular security. ⏱️ Timestamps 00:00 Intro — the question hidden behind the noise 01:54 Status check — what was released and what is noise 04:41 The real meaning of 2.8 trillion — a vast complex, 16 call bells 07:12 The physics of memory — at least 1.5 tera, and