The field of Artificial Intelligence is no longer solely focused on training increasingly larger models. It is also beginning to heavily emphasize how to run AI efficiently in production. In this context, Intel and SambaNova have announced a new joint architecture designed for agentic AI workloads—an deployment type where models do more than just answer questions; they compile code, call tools, query databases, and coordinate complex workflows.
This proposal challenges a previously almost-unquestioned idea in the market: that the entire future of AI inference should be built on GPUs. Intel and SambaNova assert the opposite. Their approach starts from the understanding that new agentic workloads reveal the limitations of “GPU-only” stacks and require better workload distribution across different types of chips. The announced design combines GPUs for the prefill phase, SambaNova RDU’s for decode, and Intel Xeon 6 processors as both the host CPU and action CPU—responsible for executing tools, orchestrating tasks, and validating results.
The central idea: each inference phase on the appropriate chip
This announcement makes solid technical sense. In modern inference—and especially in code agents—not all tasks are equal. The prefill phase involves heavy parallel calculations as prompts are transformed into key-value caches. GPUs remain the natural choice here. But once the model enters the decode phase, what matters most is the rapid and sustained generation of tokens with low latency. For this, SambaNova aims to deploy their SN50 RDU, reconfigurable dataflow architecture accelerators.
The third component consists of Intel Xeon 6 processors. They aren’t just supporting actors here; Intel and SambaNova position them as the system’s control plane and as the layer handling what could be called “agent’s real work”: compiling, executing code, calling APIs, accessing tools, coordinating sandboxes, workload distribution, and overall system behavior. This role is much more ambitious than simply serving as a host CPU and aligns with Intel’s long-standing message: future AI systems will require more balanced architectures—not just more accelerators.
Intel aims to reposition the CPU at the center of the AI conversation
This move aligns with Intel’s broader strategy around Xeon 6. The company launched the full lineup in February 2025 and has since positioned it as the reference CPU for modern data centers, especially in scenarios where AI isn’t isolated but integrated with networks, storage, vector databases, and enterprise applications. During that presentation, Intel described Xeon 6 as the “foundational CPU” for AI systems and a processor that pairs especially well with GPUs in host nodes.
Now, the messaging escalates: it’s no longer just about accompanying GPUs but about regaining functional ground in the era of agentic AI. As Reuters noted on April 9, the rise of AI agents is boosting demand for general-purpose CPUs—because many of those workloads require heavy-duty tasks outside the pure model generation. Intel seeks to capitalize on this trend with two clear messages: data center software is still predominantly built on x86, and much of the production work still relies on the mature ecosystem running on Xeon.
SambaNova aims to differentiate itself in the most costly part of inference
For SambaNova, this move is equally strategic. The company has long argued that inference economics won’t be solved solely with GPUs, and that the decode phase needs specialized hardware if token costs are to be reduced and latencies kept competitive. Their announcement presents the SN50 RDU as a component designed to change the “tokenomics” of inference—that is, the balance between performance, cost, and scalability in real-world deployments of large models.
SambaNova also offers an interesting commercial argument: the joint architecture can be deployed in existing air-cooled data centers. This could appeal to companies and cloud providers wanting to scale agentic AI without radically redesigning their physical infrastructure. While this alone doesn’t guarantee widespread adoption, it provides a practical advantage over much more power- and cooling-intensive setups.
Much promise, but still a lot to prove
As with most announcements of this type, there is a mix of roadmaps and actual products. Intel and SambaNova state that this heterogeneous inference solution will be available to businesses, cloud platforms, and sovereign AI deployments in the second half of 2026. They also indicate that, under an existing agreement, SambaNova will standardize Xeon 6 as the host CPU alongside their RDU-based inference hardware for this architecture. This suggests a deeper relationship than mere marketing collaboration.
Nevertheless, many questions remain. SambaNova’s performance claims—such as over a 50% improvement in LLVM compilation compared to server Arm CPUs or up to a 70% performance boost in vector databases over available x86 competitors—are based on internal measurements and not widely published independent benchmarks. While this doesn’t invalidate the architecture, it means we should interpret this announcement as a forward-looking blueprint with strong technical principles, not a market victory already achieved.
This partnership clearly underscores a broader trend: agentic AI is pushing the industry toward more heterogeneous systems, where prefill, decode, orchestration, and tool execution may occur on different chips. If this idea takes hold, AI infrastructure debates will shift from “which GPU to buy” to “how to optimally distribute each phase of work.” Intel and SambaNova aim to position themselves right in that evolving discussion—and by 2026, that will be quite significant.
Frequently Asked Questions
What exactly have Intel and SambaNova announced?
They announced a heterogeneous architecture for agentic AI combining GPUs for prefill, SambaNova RDUs for decode, and Intel Xeon 6 processors for orchestration, tooling, and action execution.
What does it mean that Xeon 6 is both host CPU and “action CPU”?
It means Xeon 6 not only coordinates the system but also handles tasks like compiling and executing code, calling APIs, accessing tools, and validating results within agentic AI workflows.
When will this solution be available?
Intel and SambaNova expect it to be available to businesses, cloud providers, and sovereign AI deployments in the second half of 2026.
Why aren’t GPUs sufficient for some AI deployments anymore?
Because agentic AI involves different phases with distinct needs. GPUs remain useful for prefill, but decode, orchestration, tool execution, and other tasks benefit from CPUs and specialized accelerators.
via: sambanova.ai