Embodied AI investor: scaling law holds in pretraining, data—not compute—is the bottleneck
创业邦 · wechat · 2026-10-25
In a DEMO CHINA interview, Hui Fund partner Xiao Li argues embodied scaling law is preliminarily validated in pretraining (citing PI's π0.5), but high-quality data—not compute—remains the real bottleneck: top embodied-brain models run at B-to-tens-of-B parameters on hundreds-to-1000 GPUs (tens of millions RMB/year rent). He covers full-stack investing across 10 startups, joint-module integration trends, 10,000-hour industrial reliability demands, sim2real gaps, and why model iteration still outweighs single-scene ROI.
More from Venture
- AI Killed Video Courses? Matt Pocock Says It's His Best Year, With Courses at 98% of Income — mattpocockuk · 2026-10-25
- ByteDance, Alibaba and Tencent order 3M+ Nvidia RTX PRO 5500 cards worth $35B+ — Miles_Brundage · 2026-10-25
- $399 robot duck rakes in $2.6M in preorders within 24 hours — CodeByPoonam · 2026-10-25
- OpenRouter analysis finds evidence of spoofing-style price manipulation by inference providers — AccBalanced · 2026-10-25
- OpenAI offers up to 90% discounts to poach Anthropic's biggest customers — The Decoder · 2026-10-25
- Ambient assistant Ambi exits beta with $26M backed, lands on Apple devices — FinanceYF5 · 2026-10-25