Same model, multiple clouds: cross-cloud hosting of frontier models and model diversity

The Cloud Security Alliance (CSA) published a research note stating that, after 2026, when OpenAI models became available on Amazon Bedrock and Claude became available on AWS, Google Cloud and Microsoft Foundry, multi-cloud architecture no longer equals model diversity, and cross-cloud deployment only distributes infrastructure risk rather than model-level risk. Model weights, safety training, usage policies and regulatory exposure are the same on every hosting cloud, and switching clouds cannot hedge against model takedown, policy tightening, jailbreaks or government suspension. The note cited as an example the June 2026 suspension of Anthropic's Fable 5 and Mythos 5 under a U.S. export control directive, in which both models went offline simultaneously on their own platform and on three clouds; according to Anthropic, the directive was related to a report by Amazon researchers finding a method to bypass Fable 5's safety guardrails, and access was restored only after a new safety classifier was deployed. It recommended that enterprises map dependencies along three axes—hosting channel, model lineage and jurisdiction—treat different cloud paths to the same model as having no model-level redundancy, and test scenarios in which a model is taken down entirely.