NuAxiSemiconductor trade intelligence

[News] AI-Designed CPUs Match Decade-Old Smartphone Chips, Cambricon Says; AI Also Reshapes Materials, Equipment

What happened

Cambricon Technologies co-founder Chen Yunji said, according to a Shanghai Observer report, at Saturday's 2026 World Laureates Forum in Shanghai that AI-designed chips have already reached performance levels comparable to human-designed processors from more than a decade ago. His team developed an iterative system that automatically verifies and corrects AI-generated designs hundreds of times per second; the approach produced Qimeng-1 in 2021, with performance comparable to Intel's 486 processor from the 1990s, and by 2025 its successor, Qimeng-2, had reportedly reached performance levels similar to mainstream smartphone CPUs from more than a decade earlier. According to Shanghai Observer, the team also analyzed more than 100 AI algorithms and identified a common set of underlying computational operations, paving the way for a dedicated AI instruction set architecture (ISA); Chen noted that Huawei subsequently adopted the team's processor architecture for its first-generation AI chips, paying approximately CNY 200 million in licensing fees. Separately, per China National Radio, Nobel laureate Konstantin Novoselov highlighted AI's potential to modify the atomic structures of 2D materials and predict properties while cautioning that real-world materials are more complex than idealized AI models, and Northwestern University professor Jeffrey Snyder cited the Apple Watch's proprietary inverse-design alloy as an example. Per YiCai Global, citing Piotech Chairman Lu Guangquan, equipment makers are increasingly adopting AI-driven virtual manufacturing: Lam Research uses physics-based digital twins to accelerate certain etch-modeling steps by up to 30 times, Applied Materials reportedly employs AI for real-time process analysis handling more than one million 3D wafer measurements per hour, and Piotech uses bidirectional data mapping between physical and virtual process chambers for real-time process prediction, predictive maintenance and wafer transport optimization; Lu also noted AI agents can automate failure analysis, cutting workloads that once required hundreds of labor-hours to under 5.5 hours. The source does not state the cause.