Oracle taps AMD CPUs and networking chips for its next cloud compute generation

Oracle’s next AMD-based cloud compute generation will address both application processing and the infrastructure work competing for it. At Oracle AI World on October 27, 2026, the companies detailed upcoming E7 instances using sixth-generation AMD EPYC processors, alongside Pensando chips that handle networking, storage and security tasks for AMD-powered Oracle cloud instances.

Giving agents room to do the surrounding work

The collaboration update extends AMD’s multigeneration relationship with Oracle Cloud Infrastructure, or OCI. Its E7 component focuses on enterprise applications and the CPU services surrounding AI models, rather than just the computing used to generate a model’s answers.

Oracle’s companion announcement describes that work in concrete terms: CPUs assemble context, route actions, retrieve knowledge, query databases, parse documents and execute tools. These steps affect application performance alongside model inference—the process of running a trained model to produce a response.

E7 is intended to support several layers of that application stack, not a single agent function:

• Agent orchestration and concurrent agent sandboxes: coordinating agents and providing environments where multiple agents can run.
• Retrieval services and in-memory caches, supporting access to information used by applications.
• Web, API and enterprise application tiers—the services through which applications receive requests and carry out business tasks.

More CPU resources, with infrastructure work moved aside

Oracle identifies E7’s processors as sixth-generation EPYC “Venice” CPUs from AMD’s EPYC 9006 series. It lists processor capabilities of up to 256 cores and 512 threads per socket, plus 16 memory channels. Those are processor specifications, not a disclosed allocation for every E7 virtual machine.

E7 Acceleron Standard virtual machines will offer flexible CPU and memory configurations. Oracle says customers can match those resources to their services, while the processors’ standards-based x86 foundation supports continuity with existing enterprise software stacks.

The networking side tackles a different demand on the host processor. Oracle Acceleron network interface cards for AMD compute and accelerator instances incorporate Pensando data processing units, or DPUs. These dedicated chips accelerate networking, storage and security functions in hardware. AMD says that preserves host CPU capacity for customer applications.

Oracle says the co-engineered cards combine host networking and cloud-control functions in one architecture while keeping customer and cloud-managed functions isolated. Their programmable design lets OCI change networking, storage, security and observability services as workloads evolve, rather than treating the hardware’s infrastructure role as fixed.

The next step is sizing the application, not just the model

Oracle’s guidance puts workload testing ahead of a universal hardware prescription. It recommends evaluating CPU capacity, the number of tasks running at once and response-time targets alongside GPU requirements. Flexible sizing provides a way to adjust resources; testing determines the balance an application actually needs.

The company also points to cost per token—the pieces of text a model processes or produces—or completed workloads as practical measures of value. Its proposed assessment is application-specific: judge the resources against the performance a service needs, rather than treating a processor’s headline specifications as the final result.