Moonshot's Kimi K3 may be more about memory than compute

1:00 AM · July 20, 2026

Moonshot's Kimi K3 may matter less for what it says about computing power and more for what it says about memory demand, according to Bloomberg. The model's scale, 2.8 trillion parameters and a context window of roughly 1 million tokens, requires far more memory capacity to run than earlier generations of large language models, a distinction Bloomberg draws against DeepSeek's earlier breakthrough, which centered mainly on making models cheaper to train and run rather than on raw scale. That memory intensive design points toward sustained, and potentially growing, demand for memory chips from a market currently dominated by a small handful of suppliers, chiefly South Korea's SK Hynix and Samsung Electronics, rather than easing pressure on that supply chain the way a smaller, more efficient model might have. The framing complicates the initial market reaction to Kimi K3's release, in which chip and AI infrastructure stocks broadly sold off on fears that a cheaper, more efficient Chinese rival could reduce demand for the hardware underpinning the leading US labs; if Kimi K3's actual bottleneck is memory rather than raw compute, the read through for memory chipmakers specifically may be more positive than the initial selloff suggested. The distinction matters for investors trying to parse exactly which parts of the AI hardware supply chain benefit or suffer as increasingly large, open weight Chinese models continue to close the gap with the leading closed labs.