US Semiconductors: Kimi K3: More Models, More Tokens, More Silicon

Research evidence excerpt

US Semiconductors: Kimi K3: More Models, More Tokens, More Silicon

Accessible version

US Semiconductors

Kimi K3: More Models, More Tokens, More

Silicon

Industry Overview

AI labs must increase compute to defend vs. Chinese alts. 17 July 2026

Moonshot yesterday unveiled Kimi K3, a 2.8T-parameter open-weight MoE model with a Equity

1M-token context window, again raising the post-DeepSeek debate: do better Chinese United States

models reduce semiconductor demand or expand it? Note the timing as media reports Semiconductors

also suggest Google’s Gemini 3.5 Pro is months behind schedule, with coding Vivek Arya

performance short of internal goals. Thus, frontier leadership may become increasingly Research Analyst

harder to defend. We believe the likely response from U.S. frontier labs is not less BofASvivek.arya@bofa.com

compute, but more. If open-weight Chinese models close the performance gap, OpenAI, Duksan Jang

Anthropic and Google must preserve differentiation through larger training runs, more Research Analyst

BofAS

RL/synthetic-data loops, heavier test-time reasoning and faster release cadence. The duksan.jang@bofa.com

cost of staying ahead rises as the gap narrows. Michael Mani

Research Analyst

LLM leadership is not the ultimate end goal BofASmichael.mani@bofa.com

Investors should not confuse the leaderboard with the business model. Model leadership Liam Pharr

can rotate quickly across benchmarks and use cases. Enterprises ultimately buy reliable Research Analyst

outcomes, not the “best LLM” in isolation. The durable moat is increasingly the ability to liam.pharr@bofa.com

deliver accurate, low-latency, high-uptime AI at the lowest cost per useful output.

MoE makes infrastructure (and silicon) matter more Glossary on page 5