The CPU Era of AI Agents: x86 vs ARM

```

The main thread of AI hardware is expanding from "Is the GPU sufficient" to "Where in the system could there be bottlenecks." In the training era, GPUs and custom accelerators consumed most incremental budgets; moving into inference, Agentic AI, and enterprise AI phases, tasks are no longer just a single model call, but include planning, retrieval, tool invocation, API interaction, state management, database access, and multiple cycles. CPUs are re-entering the pricing framework—but the more crucial question is: for this round of growth, will x86 or ARM take the lead?

According to Hard AI, Bank of America's analyst Vivek Arya’s core judgment in their research report is: "By 2030, the server CPU TAM will increase from about $43 billion in 2026 to $125 billion, with a 31% compound annual growth rate; the rise of CPU’s role in Agentic AI is an addition to the current data center TAM, not a replacement for components like accelerator racks."

In this calculation, the change in market share is more striking than total expansion: by 2030, Intel and AMD will each account for about 28% of server CPU value share, ARM commercial CPUs about 7%, ARM custom/ASIC about 37%. In other words, x86 remains, but more incremental value flows to ARM, especially to CPUs developed or customized by cloud vendors. AMD remains the stronger side in x86, and Nvidia enters the CPU narrative due to its full-stack architecture and Vera CPU rack.

But this is not a "CPU replacing GPU" story. Even if server CPUs reach $125 billion in 2030, they only account for about 6% of the data center system TAM; AI accelerators remain the main player, with an estimated scale of about $1.17 trillion in 2030. More accurately: Agentic AI turns CPUs from secondary roles into one of the system bottlenecks, and x86 and ARM are contesting different parts of this new role.

Agentic AI pushes CPU from "feeding the GPU" to the control plane

In the training phase, CPUs had clear tasks: decompressing, tokenizing, batching, scheduling, and feeding data to GPUs. The real large-scale matrix computations are carried out by GPUs or AI accelerators, so from 2022 to 2025, the AI accelerator market saw a 139% compound annual growth rate, while server CPUs only saw 4%.

By 2025, AI accelerators account for 88% of data center computing expenditures, CPUs only 12%. This ratio explains why the market has focused almost exclusively on GPUs in recent years.

In the inference phase, the CPU's role expands. LLM inference can be broken down into loading, prefill, and decode stages: loading relies on disk I/O and CPU speed; prefill is mainly influenced by GPUs, network, and large matrix computation, but CPUs still participate in tokenizing, routing, batching, memory setup; in the decode phase, KV cache reuse, memory management, token-by-token control flow, sampling, guardrails, logits processing, all increase the CPU's presence.

Agentic AI takes it further. A user request may break down into multiple subtasks: context retrieval, tool invocation, database access, cross-model routing, intermediate result evaluation, then deciding the next step. GPU handles model calculations, CPU schedules these steps. Thus, the CPU is no longer just a host processor but is part of the AI inference control plane—and x86 and ARM have different architectural paths for this control plane.

What’s new isn’t just a host CPU, but a row of CPU-only racks

AI CPU opportunities split into two types.

The first type is the host CPU in GPU/accelerator racks. In modern accelerator racks, CPUs usually exist in a configuration of roughly one CPU per two GPUs, expanding as GPU deployments expand. This is where x86 still holds its existing advantage.

The second type is new CPU-only Agentic racks, used for RAG pipelines, tool execution, small and medium model inference, data processing, scheduling, vector databases, memory services, and enterprise workflows. This is the core of incremental TAM—and also the area where ARM custom solutions are most competitive.

The Nvidia Vera CPU-only rack is an example. The Vera CPU rack is scheduled to launch with the Vera Rubin platform in the second half of 2026, with each high-density liquid-cooled rack integrating up to 256 Vera CPUs, used to test, execute, and verify results from Vera Rubin NVL72 compute racks and LPX low-latency racks.

Looking at a 40-rack pod cluster, a Vera Rubin POD contains 1,152 Rubin GPUs, with 576 Vera CPUs already in compute racks; if you add 2 independent CPU racks, that's 512 more Vera CPUs. Totaling 1,088 CPUs to 1,152 GPUs, nearly 1:1. This isn’t CPUs defeating GPUs, but Agentic AI systems requiring more "control, scheduling, memory, and I/O"—and ARM architecture is taking a large share of these new needs.

How is $125 billion broken down: x86 is still large, custom ARM is the biggest variable

The breakdown of the $125 billion server CPU TAM in 2030 is as follows:

Intel: about $35.1 billion, roughly 28% value shareAMD: about $35.2 billion, roughly 28% value shareARM commercial CPUs: about $8.4 billion, about 7% shareARM custom/ASIC: about $46.5 billion, about 37% share

x86 totals about 56%, ARM about 44%. x86 still leads in overall volume, but growth dynamics have completely inverted—custom ARM is the fastest-growing and most rapidly increasing segment.

From 2026 to 2030, server CPU total shipments are expected to rise from 39.7 million to 95 million, a 24.4% compound annual growth rate; average selling price from $1,075 to $1,317, a 5.2% compound annual growth rate. Both volume and price are rising, but volume contributes more, and ARM is absorbing most of the incremental volume.

The $46.5 billion for custom ARM CPUs comes from a cross-validation: ARM Royalty IP business grows at about 20% CAGR from FY26 to FY31; by 2030, infrastructure business will account for about 50% of Royalty sales, server CPUs for about 80% of infrastructure, i.e., server CPUs account for about 40% of Royalty sales; as high-core CPUs and CSS products increase in share, comprehensive royalty rate for server CPUs rises from 3-4% now to 5-6%. Back-calculating with this royalty rate, custom ARM server CPU TAM is about $46.5 billion.

Another, more aggressive scenario not used as a base: if custom ARM CPUs are priced like commercial CPUs, rather than at about half the cost, the server CPU TAM in 2030 would become about $172 billion, with custom ARM’s share exceeding 50%. But this does not reflect the actual costs paid by AWS, Google, or other cloud vendors.

AMD wins x86, ARM wins increment, Intel’s defense is harder

This is the core conclusion of this competitive landscape, with three distinct paths.

Intel: By 2030, server CPU value share drops from about 54% in 2025 to about 28%. Enterprise customers remain a strength, but cloud will continue to be diverted to ARM and AMD. The defensive line is narrowing, and pressure mounts from both sides—AMD grabs share from x86 internally, ARM substitutes at the architectural level.

AMD: The path is more subtle, but it is the winner in the x86 camp. In cloud and enterprise markets for x86, AMD continues to gain share, expected to reach a peak of about 38% server CPU value share in 2026; afterward, ARM products scale up, AMD’s overall share drops to about 28% by 2030. This isn't AMD getting weaker, but the market structure changing: within x86, AMD is stronger, but ARM grows faster in the overall CPU market. AMD is x86’s final gatekeeper but guards a shrinking territory.

ARM: Has the most opportunity, but its value is more fragmented. Commercial ARM CPUs include new products like ARM AGI CPUs, expected to ramp up in 2027-2028, reaching about 7% in 2030. The larger share comes from custom ARM CPUs, including AWS Graviton 5, Google Axion, Microsoft Cobalt, etc. ARM's win comes from the cloud vendors’ push for autonomy and custom economics, not from single products’ competitiveness.

Nvidia: Is placed in a more advantageous position, not because of individual CPUs, but because it can put CPUs, GPUs, networks, storage, and memory racks in the same full-stack system. If Vera CPU expands with the Rubin platform, Nvidia gets not just CPU ASP, but the entrance to system architecture—standing above the x86 and ARM race.

CPUs become more important, but money in data centers still flows mainly to accelerators

By 2030, data center system TAM is estimated at about $2.1 trillion, server CPUs about $125 billion, only about 6%. Compared to about 5% at the 2025 low point, it’s a comeback but not dominance.

AI data center system TAM is expected to reach more than $1.7 trillion by 2030, with AI servers about $1.3 trillion, about 75%; AI network about $316 billion, about 20%; AI storage about $82 billion, about 5%.

Within AI servers, AI accelerators are still the absolute main player: estimated TAM is about $1.17 trillion in 2030. HBM is expected to reach $168 billion, about 14% of accelerator spending. AI CPUs grow from $8.8 billion in 2025 to $95.7 billion in 2030, a 61% compound annual growth rate, but still much smaller in scale than accelerators.

So the most easily misunderstood point: CPUs are repricing, and the x86 vs ARM pattern is being reshaped, but this doesn’t mean the GPU logic is over. Agentic AI lengthens the inference chain, increases the number of model calls, and increases the load for scheduling, retrieval, memory, network, and tool execution. GPU demand remains, CPUs merely reclaim some system value—but who owns this value is now the core of the x86 vs ARM competition.

Short-term data is already shifting, PC side not as strong

In Q1 2026, server CPU data already reflects the directional shift: overall server CPU shipments grew 6% quarter-on-quarter, 19% year-on-year; revenue grew 9% quarter-on-quarter, 45% year-on-year.

The change in competitive landscape is also starting to show: AMD server CPU unit share climbs to 27.4%, up 230 basis points quarter-on-quarter; value share rises to 40.2%, up 380 basis points quarter-on-quarter. Intel server CPU value share at 46.8%, down 500 basis points quarter-on-quarter. ARM server CPU value share reaches 13.0%, up 120 basis points quarter-on-quarter—ARM’s climb is now visible in the data, and Intel’s rate of loss is equally clear.

PC side is weaker. In Q1 2026, PC CPU shipments fell 13% quarter-on-quarter, revenue fell 7% quarter-on-quarter. Meanwhile, PC MPU shipments were 67.5 million units, while IDC-calculated PC sales were 59.1 million units, a difference of 8.4 million units, likely reflecting increased ODM or customer inventory.

Overall, Bank of America leans towards AMD and Nvidia: AMD benefits from continued expansion of x86 share, making it the most certain direction within the x86 camp; Nvidia benefits from Vera CPU and full-stack systems, standing above the x86 vs ARM competition, earning the system architecture premium. ARM grows fastest but its opportunities are spread across various cloud vendor custom projects, and its royalty model means its gains differ from the other two; Intel faces continued share loss pressure in cloud and enterprise, losing ground on both fronts, making its defense the hardest in this competition.

The x86 vs ARM competition is essentially a long tug-of-war between cloud vendors’ push for autonomy and the inertia of legacy architectures. The landscape of 2030 won’t arrive overnight, but the direction is already visible in quarterly data.

This article is from WeChat public account "Hard AI". For more frontier AI news, move here.

Risk warning and disclaimerThe market is risky, investment must be cautious. This article does not constitute personal investment advice and does not take into account individual users’ specific investment goals, financial situation or needs. Users should consider whether any opinions, viewpoints, or conclusions in this article fit their particular situation. Investing based on this article is at your own risk. ```