semiconductor-manufacturing
(5)
The Slow Beam Learned to Move in Parallel
Electron-beam lithography was always wonderfully precise and painfully slow. Multibeam's bet was that the beam did not need to become faster - it needed colleagues.
The Synchrotron That Fit in a Room - Until the Bet Got Bigger
Lyncean proved a national-lab machine could become a laboratory appliance. Then a far larger gamma-ray contract exposed the unforgiving economics of frontier hardware.
EthonAI
EthonAI (now branded Ethon) is a Zurich-based industrial AI company, spun out of ETH Zurich in 2021, that builds a Manufacturing Analytics System for factories. Its no-code platform layers on top of existing factory data sources to detect visual defects, monitor processes and run autonomous root cause analysis, using causal AI to explain what drives production variation. Customers including Siemens, Lindt & Sprüngli, Bosch and Roche use it to cut quality losses by up to 50% and raise productivity while keeping engineers in the decision loop.
Halo Industries, Inc.
Halo Industries is a Santa Clara deep-tech company that uses lasers instead of diamond wire saws to slice silicon carbide crystal boules into wafers, the foundational substrate for high-voltage power electronics. Spun out of Stanford in 2014, its proprietary light-based process aims to eliminate kerf loss, cut defects, and slash water and energy use, giving more usable wafers from every expensive crystal. Backed by an $80M Series B, Halo is scaling SiC production for electric vehicles, grid infrastructure, renewables, and aerospace, and extending the same approach to silicon, sapphire, and diamond.
Averroes.ai
Averroes.ai is a San Mateo-based startup building a no-code AI visual inspection platform that helps manufacturers automatically detect, classify, segment and monitor defects on their production lines. Founded in 2021 by Tareq Aljaber and Omar Altamimi, the company targets semiconductor, electronics, pharmaceutical, food and beverage, and oil and gas producers, promising 99%+ inspection accuracy, deployment in under five days, and models trained on as few as 20-40 images per defect class without writing code or buying new hardware.