The new Kimi K3 model, with 2.8T parameters and a 1M token context window, is the largest open-weight model ever released, despite its efficient MoE architecture activating only 50B active parameters per token. Its design requires all 2.8T weights to reside in memory at all times, making memory capacity, memory bandwidth, and interconnect the primary serving bottlenecks. Hyperscalers' combined 2026 capex guidance is around $725B, up 77% year-over-year, indicating continued hardware investment despite model efficiency gains.
The release of Kimi K3 will drive immediate demand for HBM, optical memory pooling, and coherent DCI, shifting hardware investment focus from raw FLOPs to memory and interconnect.