Kimi K3 Shakes Wall Street: End and Rebuilding of Compute Belief
创业邦 · wechat · 2026-07-20
The recent release of Chinese large model Kimi K3 triggered a collective decline in global AI stocks. At first glance, the model's unexpected capability caused the shock, but in essence, it undermined the "compute expansion" financial narrative that Wall Street had firmly believed for the past three years.
The Shaken Compute Belief
Previously, the market assumed "stronger model = larger parameters = more GPU demand = more valuable Nvidia." But DeepSeek and Kimi K3 twice proved that Chinese labs, constrained by compute resources, were forced to achieve extreme efficiency in synthetic data, post-training, and reinforcement learning. When latecomers can produce top-tier performance at 40% or less cost, the massive GPU clusters of US giants become cost burdens.
Demand Doesn't Disappear, It Explodes
Just as Wall Street panicked, Kimi suspended new subscriptions for C-end users due to demand far exceeding capacity, confirming Jevons paradox: efficiency gains lower barriers to entry, leading to demand explosion. Compute demand won't vanish but will shift from low-frequency, centralized training to high-frequency, massive inference (e.g., 24/7 Agent services).
Old Maps vs New Continent
This sell-off is essentially a shakeout of crowded trades. AI competition is moving from Phase 1 (compute reserve: whoever has more GPUs wins) to Phase 2 (engineering efficiency: whoever uses GPUs best and has lowest inference cost wins). Kimi didn't change AI itself, but the way capital understands AI.
Related event: Moonshot AI Plans HK IPO as Kimi K3 Shakes Wall Street's Compute Faith(10 posts)→
More from AGI Musings
- The Evolution of LLM Business Models: Selling Outcomes Over Tokens — yacineMTB · 2026-07-22
- Bindu Reddy says GPT-6 is coming soon, with Alibaba, DeepSeek and Kimi close behind — bindureddy · 2026-07-22
- Bindu Reddy says the industry still lacks a way to train 20T models and scale post-training RL — bindureddy · 2026-07-22
- Advanced AI Models Are Becoming Impossible to Plug and Play — emollick · 2026-07-22
- AI suggested a better composition, and that made one user uneasy — Sydde · 2026-07-22
- The Thimble and the Waterfall: AI's Data Bottleneck and Feedback Loops — dyamins · 2026-07-22