An enterprise agent may retrieve proprietary data, invoke a model, call an application, execute code and apply security policies — often multiple times before completing a task. That makes agentic AI not just a model challenge, but an infrastructure challenge.
New IDC research commissioned by Intel shows how quickly that shift is taking shape. Eighty-nine percent of IT decision makers expect agentic AI to be important to their IT responsibilities within the next 12 to 24 months. Yet agentic AI infrastructure was the least-funded category tested, with only 56% of organizations saying it is well or adequately funded.
The opportunity is not simply to add more AI infrastructure. It is to define an infrastructure strategy that fits the way enterprises actually operate: hybrid in where it runs, heterogeneous in how it computes, and increasingly dynamic in matching resources to the work being done.
This also means looking beyond any one component to how the system works as a whole.
Enterprise AI Will Follow Enterprise Data
The IDC survey found that 55% of respondents take a balanced approach between private IT and public cloud, while 38% primarily or exclusively prefer private IT. At the same time, organizations expect spending to grow across core data centers, edge environments, colocation, hosting and public cloud in 2026.
By The Numbers
- 55% take a balanced approach between private IT and public cloud
- 38% primarily or exclusively prefer private IT
- 45% cite modernizing or consolidating data center infrastructure as a top IT priority
The reason is straightforward: enterprise applications and data already live across these environments. Agentic AI makes that reality even more important.
Some workloads will benefit from the scale and services of public cloud. Others will need to run closer to enterprise data because of latency, cost, security, sovereignty, governance or integration requirements.
The strategy is to run AI where it makes the most sense. That is why hybrid AI is emerging as the natural architecture for enterprise AI.
It is also why modernization matters. The IDC survey found that modernizing or consolidating data center infrastructure is the top IT priority, cited by 45% of respondents, with security, compliance and resilience also ranking near the top. Agentic AI is not separate from that modernization agenda. It is increasingly one of the reasons to accelerate it.
Agentic AI Turns Architecture Into a Systems Question
The research also shows how buyers are thinking about compute. For agentic AI, 27% of respondents see GPUs as the dominant performance factor and 24% see CPUs as most important. But the bigger signal is in the middle: about half see performance as context-dependent, with both CPUs and GPUs playing a role.
Agentic AI is not simply a matter of acquiring a cluster of GPUs. It requires a heterogeneous mix of compute designed to maximize token value — not only the quantity of tokens delivered per dollar, but the quality and usefulness of the work produced.
The future of enterprise AI is therefore not CPU versus GPU. At a minimum, it is CPU plus GPU — and increasingly CPU plus GPU plus other accelerators.
In other words, it is a systems problem.
Survey deployment preferences reinforce that point. Approximately forty percent of buyers prefer standard servers for agentic AI solutions, including CPU-based servers with specialized vector or matrix acceleration and systems that can support optional PCIe accelerators.
That is precisely the kind of environment Xeon was engineered to support. Agentic workflows require strong local inference, embeddings and vector database performance alongside the throughput needed for high-concurrency tool routing, microservices, and API calls. Xeon combines built-in matrix and vector acceleration on Performance-core platforms with high-density Efficient-core platforms designed to scale orchestration efficiently — adding specialized compute without forcing customers to re-architect their data centers.
Agentic workflows move repeatedly among model inference, retrieval, application logic, code execution, data movement, networking, and security. No single processor does all of that work. The infrastructure has to coordinate different compute engines and keep them connected, utilized, and matched to the task.
That is where the systems story becomes important. Systems is not about owning every component. It is about making the components work better together.
As agentic AI scales, performance and economics will increasingly be determined not by one chip in isolation, but by how efficiently the full system completes useful work.
x86 Gives Enterprises a Practical Path to AI
Enterprises already operate a vast installed base of general-purpose infrastructure. Their applications, databases, security systems, operational tools — and much of their data — already run there.
Agentic AI does not mean creating a separate AI island. AI can come to the infrastructure that already runs the enterprise.
The survey shows x86 remains well positioned for that transition. Buyers rated x86 and Arm nearly identically in suitability for agentic AI — 87.5% versus 87.2%. But once the discussion moves beyond theoretical suitability and into running AI at scale, clearer differences emerge. x86 led on nine of the 12 attributes tested, including software compatibility (88% versus 80%), scalability (89% versus 81%), reliability and resilience (88% versus 82%), current adoption (88% versus 77%) and in-depth knowledge (92% versus 83%). Arm scored marginally better on price, power efficiency and satisfaction, but x86's advantages aligned closely with the operational realities enterprises face when moving agentic AI from experimentation into production.
That incumbency matters because enterprises do not deploy AI into a clean room. They bring decades of applications, data, governance processes, and operational expertise with them. The Xeon installed base gives organizations a familiar platform for bringing agentic capabilities closer to proprietary data and existing workflows rather than isolating AI in a separate silo.
Software is what helps turn that hardware foundation into an agile AI environment. Intel contributes optimizations upstream to open frameworks including PyTorch, vLLM and OpenVINO so models, orchestration frameworks and specialized accelerators can integrate with existing enterprise software stacks.
The hardware and software stories are inseparable. Open software gives enterprises greater flexibility to deploy models across Xeon CPUs and complementary accelerators without tying their infrastructure strategy to a single proprietary architecture.
Intel’s Role: Make AI Work With How the Enterprise Already Runs
At Intel, we believe the next phase of AI will be defined by how effectively organizations can put intelligence to work across the enterprise.
That requires infrastructure that uses the assets enterprises already have while adding AI capability over time.
Xeon remains foundational to the applications, data and general-purpose computing that power enterprise environments. As AI becomes more agentic, Xeon also plays a critical role in orchestration, application execution, data retrieval, security, networking and tool use around accelerated inference. GPUs and other accelerators provide specialized compute where workloads demand it.
The opportunity is to make those resources work better together.
That is what a systems approach means for Intel: not owning every component, but designing our technology and software so CPUs, accelerators, memory, networking and the broader software stack operate as a more efficient whole.
For enterprise customers, the goal is not to maximize any single piece of silicon. It is to maximize the useful work the system can deliver across the infrastructure they already operate and the AI capabilities they are adding.
The right intelligence. The right compute. In the right place for the workload.
That means infrastructure that is hybrid in deployment, heterogeneous in compute, open in software and grounded in the enterprise environments customers already operate.
The AI industry has spent the last several years proving what models can do. The next phase is about putting that intelligence to work across real enterprise data, applications and workflows.
Ultimately, that is how agentic AI moves from impressive technology to useful work in the real world.
For detailed survey findings and architectural insights, explore the full IDC research paper commissioned by Intel.
Intel does not control or audit third party data. You should consult other sources to determine accuracy.