Etched, an artificial intelligence chip startup recently valued at $21 billion after raising $700 million, is facing fresh scrutiny over the technical claims behind its AI inference hardware even as it says shipments have begun and orders exceed $1 billion. The company has promoted a low-voltage inference approach for sparse mixture-of-experts models, saying its architecture can push model flop utilization above 80% of theoretical peak. Critics including the tiny corp AI computing team founded by George Hotz and chip designer Wesley Yue argue that investor backing, customer orders and hardware photos are not substitutes for third-party performance data, and that high utilization alone does not prove stronger real-world performance if a chip's peak compute is lower than rivals'. Etched has not publicly released a full set of compute, power-consumption and independent benchmark results, saying only that early customer tests reached leading levels and that more detailed figures will be published later. There is no evidence Etched fabricated its claims, and the company has not publicly responded to the latest criticism.