Enterprise AI projects often run into a hard infrastructure constraint: the most valuable prompts, model weights and intermediate data may be too sensitive to expose beyond tightly controlled systems. Advanced Micro Devices, Inc. (NASDAQ: AMD) said its 5th Gen EPYC processors now provide the host infrastructure for Google Cloud, part of Alphabet Inc. (NASDAQ: GOOGL), to support TPU 8tx systems with hardware-enforced confidential computing.
AMD disclosed the expanded collaboration Nov. 9. The companies said the configuration combines EPYC host processing, AMD Infinity Guard security capabilities and x86 software compatibility with Google Cloud TPU acceleration for inference, fine-tuning and reinforcement-learning workloads. Their stated objective is to let organizations run sensitive AI work at cloud scale while preserving data protection.
The change matters because the host CPU does more than supply general compute in an accelerated system. AMD said EPYC processors prepare data, coordinate workloads and manage high-throughput communication with the TPU accelerators. That position lets the host participate in the controls that govern how data enters, moves through and is processed by an AI system.
In the TPU 8tx configuration, AMD said its Secure Encrypted Virtualization capabilities work with the TPU to extend confidential-computing protections from the CPU to connected accelerators. The company said the design verifies an accelerator’s identity and security configuration before granting direct access to protected memory. It is intended to keep customer prompts, model weights and intermediate data encrypted in CPU memory and in transit between the host and accelerators, with hardware isolation during accelerator execution.
For IT leaders, the practical benefit is not simply a stronger security feature to add to a model deployment. It is a path to place workloads involving proprietary data into cloud AI infrastructure with a defined hardware protection model. That can matter for reasoning and agentic applications, where prompts and intermediate results may contain business information that enterprises cannot treat as ordinary test data.
AMD said EPYC CPUs already support Google Cloud Confidential VMs, Confidential GKE, Confidential Space, Confidential Managed Service for Apache Spark and Confidential Dataflow. The companies also cited previous EPYC deployments across Compute Engine C4D, N4D, H4D, Z4D and G4 virtual machines. TPU 8tx extends that relationship to AI acceleration.
The company also said EPYC processors serve as host CPUs across more than 10 hyperscalers and frontier AI labs, covering systems built around GPUs and custom accelerator silicon. That breadth gives enterprises a familiar x86 foundation across a range of accelerator choices, although the announced configuration is specific to Google Cloud’s TPU 8tx system.
The companies have not described this as a replacement for enterprise access controls, data classification or workload governance. Those remain necessary around any production AI program. But the system architecture adds a hardware-rooted layer at the point where organizations increasingly need to connect sensitive data with specialized accelerators.
Relative to earlier confidential-computing deployments centered on CPU-backed services, AMD’s stated TPU 8tx design pushes the protection boundary into an accelerator-based AI configuration. That changes the competitive baseline for cloud AI infrastructure: enterprises can assess whether a platform protects only the virtualized host or also provides an attested, hardware-enforced path for the data moving between host processors and AI accelerators.