Intel and Oracle bet on the CPU as the engine of agentic AI on OCI: a multi-year plan with Diamond Rapids

By Intel · October 27, 2026.

At Oracle AI World (Las Vegas), Intel and Oracle announced a multi-year collaboration to expand Intel Xeon processor-based computing within Oracle Cloud Infrastructure (OCI). According to Intel's press release, the agreement covers current and future Xeon platforms, including the planned OCI X14 offerings built on Diamond Rapids, the codename for Intel's next generation of server processors.

The two companies say they will coordinate their roadmaps and processor capacity planning to support OCI's expansion. The release does not include investment figures, chip volumes or availability dates.

On the technical side, Intel highlights two new features planned for Diamond Rapids. The first is an enhanced version of AMX (Advanced Matrix Extensions, matrix compute units built into the CPU for inference), with support for the FP8, FP16 and BF16 numeric formats. The second is APX (Advanced Performance Extensions), which Intel says reduces instruction overhead in optimized applications. The company ties them to tasks such as generating embeddings (numerical representations of text), reranking (reordering search results by relevance) and inference with small models.

The entire announcement is worded conditionally: "planned," "expected," "intended." The release itself includes legal disclaimers stating that it is not a commitment to deliver and should not be used for purchasing decisions. It is best read as a statement of direction.

Our take: the underlying message is that agentic AI does not run on accelerators alone. An agent spends much of its time running applications, moving data, calling tools and querying enterprise systems. All of that is CPU work. Intel and Oracle emphasizing this is not neutral, since Intel needs a story of its own against GPU dominance in AI, but the technical argument makes sense for modest workloads closely tied to enterprise software.

The most interesting part is the economics. Oracle and Intel propose scaling application compute separately from the accelerators that serve large models, to match cost to each stage of the workflow. If it works, it makes it cheaper to bring agents to existing systems without rebuilding infrastructure. It is a company promise, still without public performance or pricing data to back it up.

Winners and losers: Oracle wins, strengthening its enterprise cloud offering with a well-known silicon supplier, and Intel wins if it manages to make Xeon part of the AI story and not just classic cloud. Anyone wanting a rigorous comparison loses, because without independent benchmarks it is impossible to know how much of this is a real advantage over alternatives.

More broadly, the enterprise AI market is shifting from training models to running and orchestrating them at scale, and there cost efficiency per task matters more than peak performance. If that trend holds, CPUs that are efficient at lightweight inference will be a less flashy but necessary piece of the computational abundance that will make AI affordable for everyone. For now, we need to wait for the silicon and independent measurements before believing the promises.