Kimi K3's 2.8T Open Weights Just Spooked the Chip Market

Moonshot AI released the full weights of Kimi K3 on July 27 — 2.88 trillion parameters, of which only 104B activate per token thanks to a Mixture-of-Experts design, plus a 1,048,576-token context window. It ships under a Modified MIT license, so commercial use is fair game. A new Kimi Delta Attention mechanism reportedly makes long-context inference up to 6x cheaper, and the model takes native text, image and video input through an OpenAI/Anthropic-compatible API.

The market read it as a threat. SK Hynix closed down 14.65% and Samsung Electronics lost over 13% in late-July trading, part of a broader memory rout that has dragged Micron and Sandisk down roughly 30%. Analysts pinned it on a stack of worries: a frontier-grade open model from Beijing undercutting the case for enormous proprietary training runs, HBM shortages easing as capacity comes online, and China's CXMT IPO funding yet more domestic supply.

Meanwhile Nvidia went the other direction on openness — launching the Open Secure AI Alliance with 36 partners including Microsoft, Cisco, Cloudflare, CrowdStrike, Hugging Face, IBM and the Linux Foundation, and open-sourcing NOOA, an Apache 2.0 framework for making agent behaviour testable, traceable and auditable. Its scope is the whole agent stack: identity, permissions, isolation, guardrails, logs, model scanning.

Why it matters: when the best open model is free and the memory it runs on is no longer scarce, the AI trade stops being about who can spend the most.