AI Was Supposed to Kill the CPU. Instead, It Is Rewriting the Data Center

For years, the AI infrastructure story looked simple: CPUs run the system, GPUs run the intelligence.

That model is breaking.

AI is moving from answering questions to executing tasks. Agents can call APIs, query databases, retrieve information, execute code, enforce policies, maintain state and invoke models repeatedly. The GPU still performs the heavy tensor computation, but a growing share of the work around the model is general-purpose compute.

That is one reason the server CPU market suddenly looks much larger.

AMD now expects the server CPU market to exceed $120 billion by 2030, a major increase from its previous forecast. Citi has estimated roughly $132 billion, while Raymond James has published a much more aggressive $201 billion scenario. The estimates differ sharply, but they point toward the same architectural change: AI is creating more compute outside the GPU itself.

The GPU is not going away

The first generation of generative AI was naturally GPU-heavy.

A simplified architecture looked like this:

Application → CPU → GPU → Response

The CPU prepared workloads, handled operating-system services, I/O and orchestration. The GPU performed the computationally expensive model execution.

Nothing about the new CPU story changes that.

The GPU remains central to large-scale AI training and inference. The important change is what happens before, between and after model calls.

Then AI started doing things

A chatbot might need one model call to answer a question.

An agent can turn the same request into a workflow:

reason → retrieve → query → call a tool → execute code → call a model → validate → act

Some of those operations benefit from accelerators. Many do not.

They involve branch-heavy code, databases, APIs, security controls, data movement, scheduling and application logic. In other words, the CPU is increasingly running the system that makes the model useful.

That distinction matters because one user request can now generate many infrastructure operations.

The GPU computes the model. The CPU increasingly runs the world around the model.

Why the CPU market can expand so quickly

There is a difference between adding a faster CPU to a GPU server and creating a larger CPU layer around an AI environment.

AMD argues that agentic AI can create dedicated CPU-heavy infrastructure for orchestration, data processing, tool execution and other workloads rather than simply adding more CPU capacity inside accelerator servers. TrendForce similarly describes the CPU moving from an auxiliary role toward a core orchestrator as AI shifts toward inference and agentic workloads. Its research points to CPU-to-GPU ratios in inference infrastructure moving from roughly 1:8 toward 1:1, depending on workload.

That does not mean every AI server will have one CPU for every GPU. It means the CPU side of the architecture is becoming economically significant again.

And the industry's response is hard to miss.

NVIDIA is entering the CPU market for a reason

NVIDIA built Grace. Now it is building Vera.

NVIDIA describes Vera as a CPU for the agentic AI era, targeting workloads such as agentic AI, reinforcement learning and data processing. The strategic point is more important than the processor itself.

NVIDIA is trying to own the AI factory, not just the accelerator.

If AI infrastructure is becoming heterogeneous, controlling the GPU alone leaves part of the value chain outside the platform. CPU, GPU, networking, memory and software increasingly have to work together as one system.

Arm is following a different path toward the same destination. In 2026, Arm announced the Arm AGI CPU, extending its Neoverse strategy into production silicon for AI infrastructure. At the same time, hyperscalers continue to develop custom Arm-based CPUs optimized for their own workloads, economics and power constraints.

AMD has both sides of the equation through EPYC and Instinct. Intel is repositioning Xeon around orchestration, data movement and heterogeneous AI infrastructure.

This is no longer simply an x86-versus-Arm story.

It is a heterogeneous-compute story.

The new CPUs are not just faster. They are smarter.

The CPU story is not only about more cores. The latest generation of server processors from both Intel and AMD are embedding purpose-built accelerators directly into the silicon, offloading work that previously consumed CPU cycles or required separate hardware.

Intel Xeon 6 (6th Gen, Sierra Forest and Granite Rapids) ships with three built-in accelerators:

• Intel QAT (QuickAssist Technology) offloads encryption, compression and key exchange. In AI and cloud environments where data moves encrypted between hosts, QAT handles the crypto without touching the CPU cores. VMware Cloud Foundation 9.1 takes direct advantage of this: vSphere uses QAT to offload encrypted vMotion, meaning live VM migrations run at near-line-speed without the CPU performance hit that software encryption creates. For a deeper look at how Intel QAT integrates with VCF, see our complete hardware refresh guide.
• Intel DSA (Data Streaming Accelerator) handles memory copies, data fills and CRC operations. In virtualized environments with heavy data movement between VMs, containers and storage, DSA takes that burden off the CPU.
• Intel IAA (In-Memory Analytics Accelerator) accelerates compression and decompression for analytics and database workloads, reducing CPU overhead when processing large datasets.

These accelerators are not add-on cards. They are on the CPU die, available to every workload without extra hardware cost. That changes the economics: the CPU is not just orchestrating anymore, it is actively offloading specific work that would otherwise waste compute cycles.

AMD EPYC Turin (Zen 5 architecture) takes a different but equally aggressive approach. Turin scales up to 192 cores per socket, the highest core count in the x86 server market. For AI orchestration workloads that are thread-heavy but do not need GPU acceleration, that core density matters.

AMD also integrates SEV-SNP (Secure Encrypted Virtualization - Secure Nested Paging) directly into EPYC, providing hardware-level memory encryption and attestation for virtual machines. In AI environments where models and data carry significant IP value, SEV-SNP allows the infrastructure to enforce confidential computing at the hardware level. VMware vSphere supports SEV-SNP, enabling encrypted VMs that the hypervisor itself cannot inspect.

The competitive dynamic here is clear. Intel is embedding specialized accelerators to make each core do more. AMD is maximizing core density while adding security at the hardware level. Both approaches matter for AI infrastructure, and VCF supports both.

But there is an important catch

A growing amount of CPU work does not automatically mean CPU shipments will grow at the same rate.

Better software can reduce the compute required for each task. Higher CPU utilization can absorb more workloads on existing infrastructure. DPUs, SmartNICs, GPUs and specialized accelerators can also take work away from the CPU.

That is why the most aggressive 2030 market forecasts should be treated as scenarios rather than facts.

But this is exactly where the story becomes more interesting.

If the industry can extract more work from the same physical CPUs, GPUs and accelerators, then the strategic value moves upward in the stack.

The question stops being only:

How many processors will we buy?

It becomes:

How efficiently can we use the processors we already own?

The AI infrastructure problem is becoming a utilization problem

Modern AI infrastructure is not one type of machine.

A production environment can contain high-core-count x86 CPUs, Arm CPUs, GPUs, high-speed networking, DPUs, local NVMe, shared storage and different accelerator configurations. It may also run virtual machines, Kubernetes workloads and AI services simultaneously.

Different workloads want different resources at different times.

That makes infrastructure efficiency much harder.

A CPU-heavy agent may need capacity for one period. A GPU inference service may need the same server later. Another workload may need memory bandwidth or networking rather than raw compute. Static allocation creates stranded capacity. Overprovisioning solves the capacity problem by wasting money.

This is where virtualization stops being just a traditional server technology and becomes an infrastructure economics layer.

The old value of virtualization becomes more relevant in the AI era

Virtualization was built around a simple idea: abstract workloads from physical hardware and turn fragmented resources into a managed pool.

That principle becomes more valuable as the hardware pool becomes more heterogeneous and expensive.

Instead of forcing every application to understand exactly which physical server, CPU generation or accelerator it is running on, the infrastructure platform can manage placement, resource allocation, isolation, lifecycle and recovery centrally.

The objective is not merely to virtualize a GPU.

The objective is to make CPU, GPU, memory, networking and storage behave like an infrastructure platform rather than a collection of individual machines.

This is also why resource scheduling and utilization matter more than ever. If CPU and GPU utilization can be improved, the enterprise may delay buying additional hardware. If workloads can share infrastructure safely, the same expensive accelerator fleet can support more applications and teams. If workloads can be moved, resized or recovered without rebuilding the application stack, operations become less dependent on the underlying hardware configuration.

That is the economic layer sitting above the silicon race.

This is where VMware Cloud Foundation fits the story

VMware Cloud Foundation is no longer positioned simply as a virtualization stack. VCF is becoming a private-cloud platform for mixed enterprise and AI workloads.

VCF 9.1 supports mixed compute infrastructure across AMD, Intel and NVIDIA and is designed to run inference and agentic AI workloads alongside conventional applications. It also brings an API-first operating model, centralized operations and automation across large private-cloud environments.

That matters because the AI infrastructure problem is not only getting a model onto a GPU.

It is getting the entire application onto production infrastructure and then operating it.

An enterprise AI application may contain a model runtime, containers, a Kubernetes service, databases, APIs, storage, networking and security controls. VCF brings those environments under a common infrastructure and operations model rather than treating AI as a separate island.

The hardware integration runs deep. VCF 9.1 can leverage Intel QAT for encrypted vMotion offload, AMD SEV-SNP for confidential VMs, and NVIDIA GPUs for model execution, all managed through the same platform. That multi-vendor hardware abstraction is increasingly what enterprises need as they build AI infrastructure that spans multiple processor generations and vendors.

And the hardware abstraction is becoming more practical, not merely theoretical. In August 2026, vSphere 9.1 became an NVIDIA-Certified Hypervisor, with NVIDIA and Broadcom describing supported AI and HPC workloads as operating at near-bare-metal performance. Broadcom has also documented GPU resource management with reservations, DRS and vMotion in supported configurations.

The important point is not that virtualization suddenly makes GPUs free.

It is that the performance penalty and operational trade-offs of virtualization are becoming smaller while the benefits of sharing and managing expensive infrastructure become larger.

Private AI changes the equation again

The next step is not simply virtualizing AI infrastructure. It is operating AI as a private cloud service.

VMware Private AI Services are built into VCF and include capabilities such as Model Gallery, Model Runtime, Agent Builder, data indexing for RAG, API Gateway and MCP Tools Registry.

That changes the abstraction level again.

The enterprise does not have to think only in terms of servers, VMs and GPUs. It can consume AI capabilities through infrastructure services while keeping models, data and policies inside the private environment.

This becomes particularly important for agentic AI because agents are not passive workloads. They can access data, call tools and take actions. Security, identity, isolation, observability and governance therefore become part of the infrastructure problem, not just the application problem.

VCF's role is increasingly to provide that common operating environment.

And then comes AI Factory

Broadcom's 2026 introduction of VMware AI Factory makes the strategy even clearer.

AI Factory is described as the software-defined foundation of VMware Private AI Cloud, providing automation for deploying AI-ready infrastructure and supporting Day 2 operations, from infrastructure toward production inference.

This is an important distinction.

The AI infrastructure race started with:

Who has the fastest GPU?

It is moving toward:

Who can turn the entire infrastructure into a production AI platform?

That includes silicon, networking, storage, virtualization, Kubernetes, models, data, security, governance, automation and operations.

The value of the platform increases when hardware becomes more expensive and workloads become less predictable.

VCF 9.1.1 provides another example. Its multi-tenant model-sharing capability allows separate tenants or lines of business to share model infrastructure while maintaining data privacy, reducing the need to deploy redundant model copies and potentially avoiding unnecessary GPU allocation.

That is exactly the same economic principle behind the CPU story.

Use more of what you already bought.

The bigger AI infrastructure shift

The next data center is unlikely to be a room full of GPUs with a few CPUs attached.

It will be a heterogeneous compute system:

CPU + GPU + accelerator + networking + memory + storage

The GPU remains the engine for model computation. The CPU increasingly becomes the engine for orchestration and general-purpose work. And the infrastructure platform becomes the layer that decides where workloads run, how resources are shared, how systems are secured and how the whole environment operates.

That is why the CPU market matters.

But the deeper story is not really about CPUs.

AI is creating more kinds of compute, more expensive infrastructure and more dynamic workloads. The winner may not be the organization that buys the most silicon. It may be the one that extracts the most useful work from every piece of silicon it owns.

That is where the next infrastructure battle is moving.

---

Sources

• AMD, Agentic AI Changes the CPU/GPU Equation↗
• NVIDIA, Vera: The CPU for Agents↗
• Arm, Introducing the Arm AGI CPU↗
• Intel, Xeon and Agentic AI↗
• TrendForce, AI Inference Era: Server CPU Regains Data Center Centrality↗
• Broadcom, VMware Cloud Foundation 9.1↗
• Broadcom, VMware AI Factory↗
• VMware Cloud Foundation Blog, NVIDIA-Certified Hypervisor↗
• VMware Cloud Foundation Blog, VMware Private AI Cloud and AI Factory↗
• VMware Cloud Foundation Blog, Optimizing AI Deployments↗
• Intel Xeon 6 + VMware VCF 9 Hardware Refresh Guide
• Intel QAT, DSA and IAA Built-in Accelerators↗
• AMD EPYC Turin Processors↗