Reports have emerged that some of Amazon's key artificial intelligence startup customers are finding its custom-designed AI chips, Trainium and Inferentia, to be underperforming compared to GPUs from Nvidia. Startups such as Cohere and Stability AI have reportedly cited issues with speed and cost-effectiveness, with one document suggesting that access to the newer Trainium 2 chip was "extremely limited" and subject to frequent disruptions. This customer feedback presents a challenge to Amazon's strategy of reducing its reliance on Nvidia for the computationally intensive tasks of training and running AI models. While Amazon Web Services (AWS) has stated it is working to address the performance issues and values the customer feedback to improve its chips, the dissatisfaction could impact the adoption rate of its custom silicon. The company maintains that its in-house chips are a multi-billion dollar business and that the Trainium 2 chips are "fully subscribed."