AI CPU

The AI CPU concept covers the host and server CPUs that orchestrate AI accelerators and run the non-matrix work of AI: scheduling, tool calls, retrieval, and data movement. As agentic AI turns one prompt into a thousand-step task, the CPU’s role grows, from NVIDIA’s Vera and Grace to AMD’s EPYC, Intel’s Xeon, and Arm’s architecture.

CategoryThematic / Semiconductors — CPU & agentic compute
Representative companies4
Related ETFKODEX US AI Semiconductor TOP3 Plus ETF (0151S0)
Last updated2026-06-15
Key takeaways
  • The AI CPU concept covers the host and server CPUs that orchestrate AI accelerators and run the non-matrix work of AI, plus the architecture those CPUs are built on.
  • Agentic AI changes the workload mix. One prompt can launch a thousand-step chain of reasoning, retrieval, and tool use, much of it CPU work such as Python runtimes, sandboxed code, and orchestration, so the CPU's role in an AI system is rising.
  • NVIDIA now pairs every GPU generation with a CPU. The GB200 NVL72 rack uses 36 Grace CPUs to 72 Blackwell GPUs, and NVIDIA's new 88-core Vera CPU, which it calls 'the CPU for agents,' anchors the Vera Rubin platform shipping in 2026.
  • Arm architecture is the common thread across AI host CPUs (NVIDIA Grace, AWS Graviton, Google Axion, Microsoft Cobalt); Arm's data-center royalties more than doubled year over year in fiscal Q4 2026.
  • x86 still anchors the general server: AMD's EPYC data-center revenue grew 57% to $5.8 billion, and Intel's Xeon remains the incumbent host CPU, with Intel also the only US-headquartered leading-edge foundry.

What is the AI CPU concept?

It is the layer of AI compute that sits next to the accelerator, not inside it. Every AI server needs a host CPU to boot the system, schedule jobs, feed data to the GPUs, run the parts of a workload that are not matrix math, and increasingly to drive agents through their reasoning loops. The concept holds the leaders of that layer. Arm licenses the architecture inside most AI host CPUs. AMD and Intel supply the x86 server CPUs, EPYC and Xeon. And NVIDIA, long the GPU leader, is now a CPU vendor too, shipping the Arm-based Grace and the custom Vera processor with its accelerators.

How does agentic AI change the role of the CPU?

Because an agent is not a single forward pass. NVIDIA describes agentic AI as a new kind of workload in which one prompt launches a long journey of reasoning, retrieval, tool use, and response generation, much of which runs on the CPU: Python runtimes, sandboxed code execution, orchestration logic, and analytics. That is why NVIDIA built a CPU specifically for agents, the 88-core Vera, claiming 1.8 times faster task completion than x86 and, at the rack level, 10 times the agent throughput of the prior Grace Blackwell generation (NVIDIA, May 31, 2026; NVIDIA, May 31, 2026). NVIDIA founder and CEO Jensen Huang put the bet plainly:

“AI agents will be the largest users of computing. Vera is the first CPU designed for that future—built to run agentic AI at hyperscale with extraordinary performance, efficiency and programmability.”

— Jensen Huang, founder and CEO, NVIDIA (NVIDIA, May 31, 2026)

What is the CPU-to-GPU ratio in an AI rack?

It is a deliberate design choice, and NVIDIA sets it per system. The GB200 NVL72 rack connects 36 Grace CPUs to 72 Blackwell GPUs, a 1:2 CPU-to-GPU ratio, in 18 compute trays of two CPUs and four GPUs each (NVIDIA GB200 NVL72). The Vera Rubin platform that follows it, unveiled at CES in January 2026 and ramping into full production in 2026, keeps the CPU at the center of the rack and links Vera to Rubin over a 1.8 TB/s NVLink-C2C interconnect (NVIDIA, May 31, 2026). The number of GPUs still dominates, but as agent workloads add orchestration and code execution, the value of a fast host CPU rises, which is why NVIDIA, AMD, Intel, and Arm all compete for that socket.

Who should consider AI CPU exposure?

It suits an investor who wants the AI build-out at a steadier, lower-beta layer than GPUs, and who understands the CPU is the smaller line item in a rack. The revenue is real and dated: Arm posted record fiscal Q4 2026 revenue of $1.49 billion, up 20%, with data-center royalties more than doubling year over year (Arm, May 6, 2026), and AMD’s data-center revenue grew 57% to $5.8 billion (AMD 8-K, May 5, 2026).

The caveat is that the architectures compete with each other inside the same theme. Hyperscalers increasingly design their own Arm-based CPUs, which lifts Arm’s royalty stream but can pressure merchant x86 from AMD and Intel. For an individual investor that argues for treating AI CPU as a satellite beside a broader chip allocation; for an institution it is a way to own the AI build-out with less single-name GPU concentration, while budgeting Intel’s turnaround and political risk from its 9.9% US government stake (Intel 8-K, Apr 23, 2026).

Which companies represent AI CPU?

CompanySectorWhat it does
Arm Holdings (ARM) Information Technology · Semiconductors (IP) Arm licenses the CPU architecture inside most AI host CPUs: NVIDIA's Grace, AWS Graviton, Google Axion, and Microsoft Cobalt are all Arm-based. Fiscal Q4 2026 revenue rose 20% to a record $1.49 billion, and data-center royalties more than doubled year over year as hyperscalers standardize on Arm hosts.
AMD (AMD) Information Technology · Semiconductors AMD's EPYC is the leading x86 server CPU and the head-node processor paired with Instinct accelerators in AI systems. Data-center revenue grew 57% to $5.8 billion in Q1 2026 on EPYC CPUs and Instinct GPUs, making AMD the credible second source for AI compute.
NVIDIA (NVDA) Information Technology · Semiconductors Beyond GPUs, NVIDIA now ships a host CPU with every accelerator generation: the Arm-based Grace, and the custom 88-core Vera CPU it calls 'the CPU for agents' in the Vera Rubin platform. The GB200 NVL72 rack pairs 36 Grace CPUs with 72 Blackwell GPUs.
Intel (INTC) Information Technology · Semiconductors & Foundry Intel's Xeon is the x86 incumbent host CPU in most data centers, and Intel is the only US-headquartered leading-edge logic foundry. Q1 2026 revenue was $13.6 billion, above its own guidance for a sixth straight quarter, as a state-backed turnaround runs.

What are the risks of AI CPU?

The layer is durable, but it is not the highest-beta way to own AI.

  • Capped upside per system. The CPU is a smaller share of an AI rack’s cost than the GPUs, so revenue per system is bounded by the accelerator.
  • Architecture competition within the theme. Hyperscaler custom Arm CPUs lift Arm royalties but can erode merchant x86 unit share at AMD and Intel.
  • Execution and policy risk at Intel. The foundry turnaround is unproven at scale, and the 9.9% US government stake injects political risk into a commercial story (Intel 8-K, Apr 23, 2026).
  • Cyclicality. Server CPU demand still moves with the broader data-center and enterprise capex cycle, even as AI lifts the secular trend.

Related concepts, securities & terms

Related indices & ETFs

These references describe index-tracking relationships as a matter of fact and are not a recommendation to buy any product. Akros, as the index provider, may receive licensing fees from product sponsors. Review the product's prospectus before investing.

Frequently asked questions

What is the AI CPU concept?

It is the host and server CPUs that orchestrate AI accelerators and run the work AI cannot push onto a GPU: scheduling, tool calls, retrieval, data movement, and code execution, plus the architecture those CPUs use. The basket spans Arm (ARM), the architecture under most AI host CPUs, AMD's EPYC (AMD) and Intel's Xeon (Intel) on x86, and NVIDIA's own Grace and Vera CPUs (NVIDIA).

How does agentic AI change the role of the CPU?

Agentic AI runs many sequential steps per task. NVIDIA describes agent work as a chain of reasoning, retrieval, tool use, and response generation, much of which is CPU work such as Python runtimes, sandboxed code execution, and orchestration. NVIDIA built its 88-core Vera CPU specifically for this, claiming 1.8x faster task completion than x86 and 10x agent throughput at the rack level versus the prior Grace Blackwell generation (NVIDIA, May 31, 2026).

What is NVIDIA's Vera CPU and the Vera Rubin platform?

Vera is NVIDIA's custom 88-core CPU with an LPDDR5X memory subsystem delivering up to 1.2 TB/s, paired to the Rubin GPU over a 1.8 TB/s NVLink-C2C link. NVIDIA calls it 'the CPU for agents' and made it the host processor of the Vera Rubin platform, unveiled at CES in January 2026 and ramping into full production in 2026 (NVIDIA, May 31, 2026; NVIDIA, May 31, 2026).

Is the CPU-to-GPU ratio in AI systems changing?

The ratio is a design choice that NVIDIA sets per rack, and it reflects how much non-GPU work a workload needs. The GB200 NVL72 rack pairs 36 Grace CPUs with 72 Blackwell GPUs, a 1:2 CPU-to-GPU ratio (NVIDIA GB200 NVL72). As agentic workloads add orchestration, retrieval, and code execution, the value of fast host CPUs rises even where the GPU count dominates, which is why NVIDIA, AMD, Intel, and Arm all compete for the host-CPU socket.

Which companies represent the AI CPU concept?

Four names: Arm (ARM), whose architecture sits inside NVIDIA Grace, AWS Graviton, Google Axion, and Microsoft Cobalt; AMD (AMD), whose EPYC is the leading x86 server CPU; Intel (INTC), whose Xeon is the x86 incumbent and which runs the only US leading-edge foundry; and NVIDIA (NVDA), now a host-CPU vendor with Grace and Vera. Arm's data-center royalties more than doubled year over year in fiscal Q4 2026 (Arm, May 6, 2026).

What are the main risks of the AI CPU theme?

The CPU is the smaller line item in an AI rack, so this is a lower-beta way to play AI than GPUs, with the trade-off that the upside per system is capped by the GPU. Competition is intense across architectures: hyperscalers design their own Arm CPUs, which lifts Arm royalties but can pressure merchant x86 from AMD and Intel. Intel's turnaround also carries execution and political risk from its 9.9% US government stake (Intel 8-K, Apr 23, 2026).

Is the AI CPU theme suitable for an individual investor?

It suits an investor who wants AI exposure with a different risk profile than a pure GPU bet, since CPUs are a steadier, lower-beta layer of the same build-out, sized as a satellite. The cash flows are dated: AMD's data-center revenue grew 57% to $5.8 billion (AMD 8-K, May 5, 2026) and Arm posted record fiscal Q4 2026 revenue of $1.49 billion, up 20% (Arm, May 6, 2026). Always check fees, holdings, and risk before investing.

Is there an ETF that captures the AI CPU concept?

There is no CPU-only ETF, but the KODEX US AI Semiconductor TOP3 Plus ETF (0151S0) tracks the Akros US AI Semiconductor index and holds the CPU-relevant names NVIDIA, AMD, Intel, and Arm alongside GPUs and memory (Samsung Asset Management). It is a broad AI-chip basket rather than a pure CPU play. Always check fees, holdings, and risk before investing.

Sources & references

  1. NVIDIA Unveils Vera, the CPU for Agents · NVIDIA Corporation (Newsroom), 2026-05-31
  2. NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide · NVIDIA Corporation (Newsroom), 2026-05-31
  3. NVIDIA GB200 NVL72 (36 Grace CPUs, 72 Blackwell GPUs) · NVIDIA Corporation, 2026-01-05
  4. Arm Holdings plc Reports Results for the Fourth Quarter and Fiscal Year Ended 2026 · Arm Holdings plc, 2026-05-06
  5. AMD Reports First Quarter 2026 Financial Results (SEC 8-K, Exhibit 99.1) · Advanced Micro Devices, Inc. / SEC EDGAR, 2026-05-05
  6. Intel Reports First-Quarter 2026 Financial Results (SEC 8-K, Exhibit 99.1) · Intel Corporation / SEC EDGAR, 2026-04-23
  7. NVIDIA Announces Financial Results for First Quarter Fiscal 2027 · NVIDIA Corporation (Newsroom), 2026-05-20
  8. KODEX 미국AI반도체TOP3플러스 ETF (0151S0) — product page · Samsung Asset Management, 2026-01-13