AMD says its 5th Gen EPYC chips host Google's confidential TPU 8tx

AMD said on 9 November 2026, from Santa Clara, California, that its 5th Gen EPYC processors run the host side of Google Cloud's TPU 8tx, which AMD calls the first Tensor Processing Unit system in the cloud to offer hardware-enforced confidential computing.

AMD says the EPYC hosts coordinate workloads, prepare data and carry fast traffic to and from the TPU accelerators. Google's own post the same day describes TPU 8tx as paired with 5th Gen AMD EPYC hosts. Google says the system is tuned for secure, confidential inference, also suits small-scale training, and is aimed at regulated enterprises and governments that run frontier or in-house models and agents on their most sensitive data.

AMD says the Secure Encrypted Virtualization (SEV) feature in EPYC, working with the TPU, carries confidential computing protection past the CPU to the attached accelerators. According to AMD, each accelerator has its identity and security settings checked before it may reach protected memory directly, so customer prompts, the weights of a model and working data stay encrypted in host memory and in transit between host and accelerators. AMD adds that the hardware keeps the data isolated while an accelerator is running.

Mark Lohmeyer, Google Cloud's vice president and general manager for AI and Computing Infrastructure, said in AMD's release that with this pairing Google's cloud TPUs offer "hardware-enforced confidential computing" for the first time. Google says TPU 8tx is built on TPU 8t's silicon and adopts the low-latency Boardfly network, and claims latency up to two times lower than TPU 8t with almost no loss in performance. Google says teams can adapt models on their private data, compress large models into more compact ones, and train agents by reinforcement learning on one system, without shifting sensitive data from one cluster to another.

AMD also says EPYC is the host CPU for more than 10 large cloud operators and frontier AI labs, across systems built on GPUs and on custom accelerator chips. It says it and Google Cloud have worked together across several EPYC generations, and that Google security experts have reviewed AMD's security processor firmware in depth.