The AI infrastructure landscape is undergoing a profound transformation, shifting from a GPU-dominated build-out to a more balanced CPU-GPU ecosystem. It is driven by the rise of agentic AI workloads that demand far more sequential, CPU-intensive orchestration than traditional model training ever did.
The purpose of this article is to examine the emerging CPU supercycle, explore how the competitive landscape is evolving with new merchant entrants such as Nvidia, Arm, and Qualcomm challenging incumbents Intel and AMD, and assess what a claimed $200bn+ CPU market opportunity by 2030 means for the broader AI infrastructure landscape.
In recent years, AI infrastructure spending has been dominated by GPUs, whose parallel architectures (16,000+ CUDA cores per Nvidia H100 and 20,000+ per Nvidia B200) are suited to training large language models (LLMs). CPUs, built around fewer and faster cores optimised for sequential logic, played a supporting role: running the operating system, scheduling tasks, and feeding data to the accelerators. That balance is changing as the workload mix shifts from training toward inference and increasingly agentic AI. Various reports expect AI inference workloads to take over AI training workloads by 2030.
An agentic system breaks a goal into steps and works through them iteratively: querying databases, calling external tools and APIs, executing code, calling the GPU cluster for reasoning, verifying outputs and deciding whether to loop again. Most of the steps in that loop apart from the model call are sequential control logic that runs on CPUs.
According to company announcements, the server CPU market is expected to grow from $30-35bn in 2025 to more than $200bn in 2030, implying a CAGR of more than 45%.
The growth is supported by different elements for each company. Intel’s Q2-26 data center revenue growth of 59% YoY decomposed into a 9% YoY increase in server volumes and a 48% YoY increase in average selling price (ASP), driven mostly by a mix shift toward premium high-core-count CPUs. AMD, on the other hand, mentioned that two thirds of 70% YoY growth expected from Q2-26 CPU revenue will be due to unit growth. Traditional server CPUs are priced between $500 and $3k per unit (low-end to mid-range), while companies Arm expects next-generation, high-core-count AI CPUs to be priced at $3k-$5k per unit, and the flagship CPUs from AMD and Intel are listed at $13k-14k per unit, supporting the increase in ASP and growth rate of the AI CPU market.
Both AMD and Arm now describe three distinct CPU roles emerging within the data center infrastructure, each with different performance requirements and each driving CPU demand.
First general-purpose servers: the traditional enterprise and cloud workloads (databases, web services, application servers) that have run on CPUs and historically grew at high single-digit rates annually.
Second, AI host nodes (head nodes): CPUs that sit at the front of GPU/accelerator clusters, feeding data, scheduling jobs, and collecting results. GPU racks require host-node CPUs to operate, and the CPU-GPU ratio in these configurations is tightening from 1:8 toward to 1:4 and could eventually reach 1:1-1:2.
Third, agentic AI servers: a new tier of CPU-dense racks that sit between GPU clusters and general-purpose servers, orchestrating the workflow between them. These racks run coordination and decision-making logic such as calling the GPU clusters, involving external APIs, evaluating outputs, and deciding whether to iterate.
AMD CEO Lisa Su described agentic AI servers as early-stage components but likely to represent more than 50% of projected market opportunity by 2030.
The market is experiencing significant supply shortages. Intel stated at its Q2-26 results (July 2026) that data center customers are demanding more than it can produce, with industry-wide constraints across wafers, memory, and substrates expected to persist. Intel management guided to strong double-digit server growth for the industry in 2026 and 2027, with some momentum extended to 2028.
The server CPU market operates in three channels: 1) merchant CPUs sold openly to enterprises and cloud providers and historically dominated by x86 architecture (originated by Intel and produced by AMD under cross-licence); 2) bundled CPUs sold only inside GPU systems (Nvidia’s Arm-based Grace CPU); 3) captive CPUs designed by hyperscalers for their own data center workloads (Amazon’s Graviton, Microsoft’s Cobalt, and Google’s Axion).
Intel: Merchant
Intel historically has been the dominant merchant vendor and is reporting accelerating demand. Intel's data center and AI segment (majority from Xeon server CPUs) generated $6.3bn in revenue in Q2-26 with a 59% growth YoY. The growth decomposed into a 9% increase in server volume and a 48% increase in ASPs, driven primarily by a mix shift toward high-core count products such as Granite Rapids. Granite Rapids is selected as the host node for Nvidia’s DGX Rubin NVL8 systems, but Nvidia is replacing Intel at the top end with its own Vera CPU as the host node for the larger NVL72 racks.
Unlike every other player in the server CPU market, Intel is vertically integrated where it controls both chip design and fabrication. Though the fabrication unit is yet to prove its competitiveness to external customers, Intel’s own Granite Rapids with Xeon 6 cores ships on Intel 3 nodes. Clearwater Forest (general-purpose servers) with Xeon 6 is the first data center CPU on Intel 18A, and the next-generation Diamond Rapids is planned to use Intel’s 18A-P node.
Intel’s capex increase from $18bn to more than $20bn in 2026 partly funds this vertically integrated CPU capacity expansion. Intel has a foundry opportunity by signing external customers, but so far external customers are limited (Microsoft and Amazon’s networking silicon) and the unit continues to report an operating loss (-$2.1bn in Q2-26). With a foundry segment, Intel is also more exposed to a risk of oversupply.
AMD: Merchant
AMD has been gradually winning market share from Intel and projects 70% YoY CPU revenue growth in 2027 ($28.5bn revenue). In Q1-26, two thirds of the growth is expected from unit volumes rather than ASP, which is a different composition from Intel’s ASP-led story. AMD is targeting more than 50% server CPU market share by 2030, implying $110bn server CPU revenue per annum (more than 6x its estimated CPU revenue in 2026).
AMD launched EPYC Venice (6th Gen) family in May 2026, with purpose-built variants for each of the three data center sub-segments (general-purpose, head node, and agentic). Venice entered production ramp at TSMC on 2nm and is expected to be available through original equipment manufacturers (OEMs) in Q4-26, ahead of Intel’s competitive answer Diamond Rapids, which is expected by mid-2027 earliest.
Nvidia: Shift from bundled to merchant
NVIDIA’s CPU has historically been platform-integrated with 2.5mn Grace CPUs shipped until May 2026 that are inside GPU systems, pairing the CPU with Blackwell or Rubin accelerators via NVLink C2C. In May 2026, NVIDIA crossed into the merchant channel by offering Vera as a standalone CPU, with CoreWeave, Oracle, and ByteDance as first customers and OEMs Dell, HPE, Lenovo, and Supermicro providing configurations.
Built on custom Arm cores and co-designed with Rubin GPUs, Vera is claimed by the company to deliver 1.5× higher performance per core, 2× greater performance per watt, and 4× higher rack density compared with x86-based alternatives. Management claims Vera opens a ~$200bn addressable market in the long term and expects $20bn of CPU revenue by the end of its fiscal year (January 2027). Nvidia’s Q2 FY27 results on August 26 should provide first indication of standalone CPU shipment volumes and revenue.
Qualcomm: New merchant entrant
Qualcomm announced the Dragonfly C1000 data center CPU at its June 2026 investor day, with Meta as the anchor customer under a multi-generation collaboration to start production in H2-28. As the latest entrant to the CPU market, Qualcomm targets $5bn of total data center revenue by FY27 and more than $15bn by FY29, though CPU revenue is only a part of this revenue and begins in H2-28 while the near-term ramp is driven by connectivity and custom chips.
Arm: New merchant entrant
Arm occupies a unique position in the CPU market, providing IP architecture underpinning much of the captive and bundled channels and entering the CPU merchant market with its own silicon.
Arm's Neoverse IP is the foundation of various captive and bundled CPUs (AWS Graviton, Microsoft Cobalt, Google Axion, and NVIDIA Grace). Management puts CPU chips based on Arm architecture at approximately 50% of hyperscalers’ CPU deployments in 2025 (up from ~15% in 2024), largely reflecting hyperscalers’ captive designs plus NVIDIA's bundled Grace. This share reflects licensed IP/architecture deployed by others (primarily captive and bundled), rather than Arm's own products. Some research estimates put Arm’s CPU market share less than 25% of the broader global data center CPU market including enterprise.
In March 2026, Arm entered the merchant channel directly with the AGI CPU, its first own-brand data center chip, co-designed with Meta. Customer demand now exceeds $2bn across FY27–28 (up from $1bn at launch), with partners including OpenAI, Cerebras, Oracle, and ByteDance.
Arm's dual model (licensing IP to the industry while selling its own competing chip) is the central tension. Management describes the two businesses as complementary, projecting ~$15bn in AGI CPU revenue by FY2031 alongside ~$10bn in licensing. Whether customers who also license Arm IP will purchase Arm-branded chips at scale, rather than designing their own, is the open question.
Hyperscaler Silicon: Captive
Amazon (Graviton), Microsoft (Cobalt), and Google (Axion) design Arm-based CPUs for their own data center workloads. These chips are not sold as components today but are consumed by cloud customers as rented compute instances.
The strategic rationale for bringing chip design in-house include: 1) eliminating the merchant vendor margin on every CPU purchased, 2) tailoring the silicon chip to the specific workloads with higher performance per watt, and 3) reducing supply-chain dependence on vendors at a time when there are capacity constraints and rising prices.
Hyperscaler captive CPUs act both as a demand driver for Arm's licensing business and as a potential ceiling on the merchant addressable market, if they prove to be competitive. Hyperscalers do not separately report captive chip deployments, but Amazon disclosed in its April 2026 shareholder letter that its custom chip business (Graviton CPUs, Trainium AI accelerators, and Nitro networking) has exceeded a $20bn annual run rate, growing at a triple-digit percentage YoY. It also disclosed that Graviton is used across 98% of the top 1,000 EC2 accounts, though the share of capacity running on captive CPU is not disclosed.
Existing CPU vendors are reporting accelerating growth. Intel’s Data Center and AI (DCAI) segment grew 59% YoY in Q2-26 and AMD projects server CPU growth above 70% YoY in 2027. 2026 has seen multiple new entrants announce merchant CPU products that are expected to be shipped by 2026/early 2027: Nvidia’s Vera, Arm’s first own-brand CPU and Qualcomm’s C1000.
CPU Revenue Guidance and Estimates
EV/EBITDA Multiples
Valuation multiples across the group reflect opportunities from the potential CPU supercycle, especially in calendar Q1-26 and Q2-26, except Nvidia and AMD who are also active players in AI accelerators and have seen a decline in calendar Q1-26 due to AI bubble fears. Arm trades at the highest multiple by far, pricing its dual role as the IP/architecture provider to the broader market (data center and devices such as smartphones) and a new merchant CPU entrant. However, only 10-13% of shares are owned by entities other than SoftBank Group, amplifying price moves in either direction.
Intel’s multiple has re-rated significantly from its 2024 lows, supported by a strong data center segment performance in calendar Q2-26 mainly due to CPU demand, but also from Intel’s role as a foundry partner for the Tesla/SpaceX-backed Terafab chip manufacturing facility (though the project’s scale and timeline face significant skepticism from analysts).
Nvidia has the second lowest multiple after Intel, partly a function of the scale of its earnings base.
Qualcomm has the lowest multiple, consistent with its latest entry into the CPU market with C1000 and a steady-growth mobile and other-device chipset business.
Share Price Performance (01/02/2026 = 100)
Intel, Arm, and AMD share prices have surpassed +50% since calendar Q1 2026 after CPU demand-supply imbalance and new entry announcements. Intel’s share price performance is driven by government support, foundry expectations and data center CPU performance. Nvidia’s share price appreciation has been more moderate, partly reflecting broader AI valuation concerns and Qualcomm’s share price has been declining since its investor day on 24 June where Dragonfly C1000 was announced.
Server CPU Comparison (Current and Upcoming Architectures)
Source: Company announcements, Tom’s Hardware. Note: Google’s Axion and Qualcomm’s C100 are excluded due to lack of data. 1. Xeon 6 Granite Rapids are optimised for AI inference, whereas Clearwater Forest is optimised for general-purpose servers. 2. Intel and AMD have various server CPUs for AI applications, most advanced agentic AI server CPU that is available is shown. 3. According to Tom’s Hardware in Oct-2024. 4. Another Zen 6 variant called Verano with a smaller core count, optimised for head node applications will be launched later in 2027 (more comparable to Nvidia’s Vera CPU).
Intel and AMD use x86, while the remaining players (Nvidia, Arm, captive hyperscaler CPUs) build on Arm architecture, with Neoverse V3 core design serving as the latest common foundation.
A CPU core is an independent processor etched into the chip where each core can execute its own task. A thread is a stream of instructions a core handles. Intel and AMD have the highest number of threads (256) through simultaneous multithreading (SMT), which enables a core to share resources between two threads at once and effectively doubles the task capacity.
High-core-count CPUs are usually used for general-purpose cloud and agentic workloads, where running many concurrent tasks matter more than single-thread speed. AMD’s Venice is expected to have the highest core count (256) once shipped in Q4 2026. By contrast, Nvidia’s Vera (88 cores) and AMD’s upcoming Verano (72 cores) are optimised for head node workloads, designed to sit in front of AI accelerators and GPU clusters and feed them data at maximum bandwidth with minimum latency through deeper caches and wider memory channels.
Arm’s early-stage AGI CPU (136 cores) sits between the two categories according to the firm's management, serving as a head node for clients such as Cerebras and OpenAI and as an agentic/general-purpose CPU for Meta and Cloudflare.
Looking ahead, the competitive landscape is set to intensify through late 2026 and into 2027, easing the demand-supply imbalance, with several new merchant chips reaching volume shipments for the first time. Nvidia's standalone Vera is expected to begin shipping in H2 2026, with its August 26 earnings result providing the first indication of merchant CPU volumes. Arm's AGI CPU production revenue starts in Q1 2027, and Qualcomm's Dragonfly C1000 follows in H2 2028. By the end of 2028, the merchant server CPU market will have grown from two vendors (Intel and AMD) to at least five.
This article was co-written by Guniz Kama and Sweety Gupta.
This blog post is for informational purposes only. The information contained in this blog post is not legal, tax, or investment advice. FactSet does not endorse or recommend any investments and assumes no liability for any consequence relating directly or indirectly to any action or inaction taken based on the information contained in this article.