Money laundering not only facilitates the perpetration of dangerous and illegal activities it also damages the credibility and integrity of the global financial system and the financial institutions through whom money is laundered. Despite most financial institutions adhering to prevailing laws and regulations designed to prevent the practice of money laundering, it has been difficult to stop illicit activity using conventional methods. Hence, to combat money laundering, financial institutions are increasingly focused on the adoption of new technologies involving the use of artificial intelligence (AI) and machine learning (ML). One barrier to adoption of these new techniques for anti-money laundering (AML), however, is the need to maintain the confidentiality of the massive quantities of data required to train AI models, a financial data is the subject of regulatory controls and a target for cyber threat actors. In response to these challenges, this paper presents a secure and scalable architecture for AI implementation that uses confidential computing technology to provide complete end-to-end protection of sensitive financial data and the intellectual property of AML algorithm developers. Generative adversarial networks (GANs) are demonstrated using cloud infrastructure secured using Intel® Software Guard Extensions (Intel® SGX). The reported solution architecture can be adapted to support federated machine learning (FML), at scale, between mutually distrusting institutions, with independent control of data security at rest, in transit, and in use by individual data owners.