AI models are increasingly being accepted in practice. This raises the question of how to reconcile the results of these models with the results of traditional models. The article presents the application of AI models – Transformer to calculations related to the risk of the Bitcoin cryptocurrency. The results are presented to solve the problem of changing the risk assessment methodology, even in the case of Bitcoin portfolios, from traditional GARCH-based models to algorithmic models based on AI tools. The authors show that the presented algorithmic models have long-term memory properties. Thanks to this, the use of Transformer models allows to overcome the main limitation of heteroscedastic models – the lack of possibility of forecasting within a reasonable time horizon.