AMD used Oracle AI World 2026 to update its multigeneration work with Oracle, and the news centers on processors and network chips rather than a new GPU. Oracle Cloud Infrastructure (OCI) plans to offer E7 compute instances that run on 6th Gen AMD EPYC chips. Meanwhile, Oracle’s Acceleron network cards will carry AMD Pensando data processing units.
The focus makes sense once you consider what an AI system actually does all day. Accelerators get most of the attention, yet CPUs coordinate agents, retrieve data and run the enterprise applications around them. Networking gear then moves data between machines and takes over background chores that would otherwise eat into compute capacity.
E7 instances target that surrounding work. AMD points to agent orchestration, sandboxes for agents running side by side, retrieval services and in-memory caches, plus web, API and enterprise application tiers. The E7 Acceleron Standard virtual machines add flexible mixes of compute and memory, so customers can size each workload separately.
The Pensando DPUs sit on the network cards. They speed up networking, storage and security tasks by handling them in hardware, so the host CPU can spend its time on customer applications. AMD also says the chips are programmable, so OCI can rework its networking, storage, security and monitoring services when workloads change. That room to adjust matters, since AI training, inference and agent workloads all put new strain on the network. The cards also serve systems that pair AMD CPUs and GPUs on OCI.
However, the details so far leave gaps. The announcement carries no release dates, regional availability, pricing or benchmark results for the new instances. Therefore, customers cannot yet compare E7 against their current setups on cost or speed. Likewise, claims about preserved CPU capacity need real workload numbers before anyone treats them as proven.
Overall, the pairing says something about how Oracle and AMD see AI infrastructure. They treat the CPUs and network chips around the accelerators as a core part of the AI build, and buyers will soon test whether that holds up under real workloads.
