The AI That Actually Builds Your Chips: Inside the World’s Smartest Factories
Minds.ai CEO It Gilboa discusses deploying AI for semiconductor fab optimization, addressing challenges like dynamic factory conditions, data quality issues, and organizational resistance. The company's Maestro platform enables multi-KPI optimization across entire fabs, delivering hundreds of millions in value per factory annually.
Summary
This podcast features It Gilboa, CEO of Minds.ai, discussing AI implementation in semiconductor manufacturing. Gilboa, a materials scientist with extensive fab operations experience, explains how modern semiconductor fabs are too complex for human optimization alone, requiring AI to handle hundreds of thousands of variables simultaneously. The company's DeepSim platform serves as a neural network training foundation, while Maestro provides semiconductor-specific workflows including forecasting, scheduling, and reinforcement learning.
The discussion reveals significant deployment challenges, particularly around data quality issues like inconsistent labeling, corrupted data, and poor data hygiene across fab systems. Gilboa emphasizes that organizational factors often present bigger barriers than technology, with customers treating factory data as crown jewels and being reluctant to share control. The company addresses these concerns through customer tenant deployment, data anonymization, and ensuring models remain under customer custody.
Regarding ROI, tool-level optimizations show results in days to weeks, while factory-level improvements take months due to the 6-9 month wafer cycle times. The company reports 1-3% factory output improvements, translating to hundreds of millions in annual value for large fabs. Gilboa stresses the importance of human-AI collaboration, noting that many fabs face knowledge transfer challenges as experienced operators retire, making AI augmentation valuable for consistency and expertise preservation.
Key Insights
- Gilboa argues that modern fabs are too complex and dynamic for humans to optimize based on simple heuristic models or intuition, requiring AI to handle hundreds of thousands of variables simultaneously
- The speaker claims that small percentage improvements deliver massive value because a 1% output improvement on a $30 billion factory equals $300 million annually
- Gilboa explains that most deployment failures start with data ingestion problems, including poor data hygiene with inconsistent labels, corrupted data, and inadequate saving frequencies
- The CEO reveals that organizational and human political factors often present bigger implementation barriers than technology, with customers treating factory data as crown jewels and being reluctant to give up control
- Gilboa describes how fab operators frequently tell him they don't need AI to be better than their best human, just as good as the best one but consistent day in and day out
Topics
Full transcript available for MurmurCast members
Sign Up to Access