Topic

Moonshot AI's Kimi K3 (2.8T params, largest open-weight model ever, released July 16) triggered a real market reaction this week: an independent Arena.AI benchmark ranked it #1 overall in blind Frontend Code testing, ahead of Claude Fable 5. TSMC fell 7% on July 17 despite reporting a 77% jump in quarterly operating profit, echoing the Nvidia -$590B single-day crash from DeepSeek R1 (January 2025). The twist: full weights are not even public yet, they land July 27. The market repriced on anticipation, not proof. Same week, China's National Data Administration confirmed daily AI token consumption hit 140 trillion in March 2026, up 40%+ from December 2025's 100T and ~1,000x from 100B in early 2024. ByteDance's Doubao alone accounts for ~120T tokens/day.

This is a direct sequel to Run 43 (Kimi K3 beats Opus 4.8 on GDPval-AA, drafted 2026-07-16 but never posted and never engaged, per the Chrome-unavailable note on that page). Today's angle is different and fresher: the market panic itself, and the buried detail that it happened before the model was even fully out.

Key facts used

• Kimi K3: 2.8T parameter MoE, largest open-weight model ever shipped, Modified MIT license, full weights land July 27
• Arena.AI independent benchmark: K3 ranked #1 overall in blind developer testing, and specifically #1 in Frontend Code Arena at 1,679 points, ahead of Claude Fable 5
• Company's own benchmarks still show K3 trailing Fable 5 and GPT-5.6 Sol on overall performance, but ahead of everything else tested
• TSMC stock fell 7% on Friday July 17, the SAME DAY it reported a 77% jump in quarterly operating profit — the drop was pure sentiment, not fundamentals
• Echo of DeepSeek R1 (Jan 2025): Nvidia lost ~$590B in market cap in a single session on that release
• China's National Data Administration (Liu Liehong): daily token consumption exceeded 140 trillion in March 2026, up 40%+ vs end-2025 (100T), ~1,000x vs early 2024 (100B/day)
• ByteDance's Doubao model alone consumes 120T+ tokens/day, the majority of China's total
• Buried detail (the angle nobody else is using): the semiconductor selloff happened BEFORE Kimi K3's weights are even public. The market priced in a benchmark claim and a July 27 promise, not a downloadable model.

Pillar

Market-share / infra-choice + contrarian-reframe. Direct continuation of Run 33/36 (China open-weight vs US frontier) and Run 43 (Kimi K3 vs Opus 4.8 benchmark). New wrinkle this run: the STORY is now the market's reaction, not just the model's specs. Also ties the token-consumption stat in as a second, independent data point supporting the same thesis (China running away with inference volume regardless of who's "on top").

Post (LinkedIn)

Nvidia lost $590 billion in a single day when DeepSeek dropped R1.

Eighteen months later, it just happened again. Before the model was even out.

Moonshot AI's Kimi K3 shipped July 16. 2.8 trillion parameters, the largest open-weight model ever. An independent Arena.AI benchmark ranked it #1 overall, ahead of Claude Fable 5, in blind developer testing on frontend code.

Full weights aren't public yet. They land July 27.

Didn't matter. TSMC fell 7% on Friday, the same day it reported a 77% jump in quarterly profit. The market didn't wait for proof. It repriced on the announcement alone.