Data Center CPU Dominance Is Shifting To AMD And Arm

Fig. 1: Created by ChatGPT from a text prompt.

The data center processor market has seen two major tectonic shifts in the last decade.

It used to be that all data center compute was x86, and well more than 90% of that was Intel. GPUs first appeared in the data center in 2016 (Pascal GPU).

Now, the majority of computation is done on GPUs. AMD is looking to pass Intel in x86 share, and Arm-based CPUs are growing fast with boosts from NVIDIA and the hyperscalers.

AMD is displacing Intel in data center CPUs

AMD is grabbing share from Intel in x86 data center processors at a rapid rate.

Fig. 2: AMD server share. Source: AMD Advancing AI Event, June 2025

As the chart above shows, AMD’s share is now 40%, up from almost nothing in 2018. This is because of:

1. Lisa Su, AMD’s president and CEO;
2. Su’s decision to shift to TSMC;
3. Her team’s great execution, and
4. Intel’s loss of semiconductor process technology dominance.

These underlying factors aren’t likely to change anytime soon, so it’s likely AMD will become the largest supplier of x86 CPUs for the data center in 2026. (Intel’s 18A seems viable on paper, but the company appears to be struggling with scaling the node and associated financial constraints).

AMD’s data center segment revenue in Q1 2025 was $3.7 billion – this is CPUs and GPUs. Best estimate is most of this is their server CPUs, ~$2.5 billion to $3 billion for Q1 2025.

In terms of units, for Q3 24 total server CPU shipments for AMD and Intel combined totaled 5.5 million units. That is an ~22 million/year run rate. The top 4 hyperscalers buy about half of this, or about 2.5M units/year each. (Note: ASPs for different CPUs can be very different. For example, AMD server CPUs sell for almost 2X Intel’s. Dollar revenue is the better indicator, but to compare to Arm we only have unit information.)

Hyperscalers buy about half of x86 data center processors, and all of them are using Arm-based CPUs

Arm’s Mohamed Awad, senior vice president and general manager for data center infrastructure, said in March that at the end of 2024, Arm-based CPUs had ~15% share in data centers. But he expected he expected that for the top Hyperscalers, half of the compute shipped to them would be Arm-based – presumably this is share in units.

This is because of two more tectonic shifts. First, hyperscalers are designing their own CPUs for general applications. And second, NVIDIA is designing its own CPU for GPU co-processing.

About half of data center CPU/GPU purchases are made by the top hyperscalers.

Arm’s Mohamed says the motivation for hyperscalers to build their own CPUs is primarily to optimize them for their specific workloads in order to get better power/performance. Cutting out the margins of Intel/AMD are a consideration, but probably not enough alone to justify the shift.

Of course, much of the workloads the hyperscalers run are their customers’ software, which is primarily optimized for x86 today. The Arm processors will be used first for workloads that the hyperscalers control — search, photos, shopping, Facebook, LLMs, etc. But look on the websites of the hyperscalers. They are aggressively promoting their Arm processors, so this will change.

Amazon was the first Arm-CPU in the data center. Graviton is in its 4th generation, starting from 2018. As of mid 2024, Amazon built >2 million Gravitons. Assuming say 500K units of Graviton were built in 2024, that’s roughly 20% of total x86 CPU purchases. Graviton is used as a standalone compute instance in their cloud, and in conjunction with their GPUs. Andy Jassy, Amazon’s CEO, said on CNBC on June 30th that Graviton is 30% to 40% more price performant than x86, and better on power.

Fig. 3: The Graviton 4. Source: Amazon AWS

Google announced its Axion processors in April 2024. Google Cloud’s website claims “Axion delivers industry-leading performance and energy efficiency,” specifically “up to 50% better performance and up to 60% better energy-efficiency than comparable current-generation x86-based instances.”

Microsoft announced its Azure Cobalt CPU in 2023 (along with its Azure Maia AI accelerator). Microsoft claims ~30% savings vs. x86.

The dollar savings are important, but power ultimately will become the most important factor. Power is the biggest challenge in building the increasingly large AI data centers for Frontier LLM models.

All of these Arm-CPUs are available to customers on the Cloud, which if the price/performance is superior to x86, will result over time in customers shifting their workloads to Arm from x86.

Rumors are that Meta and Arm may be working on an Arm data center CPU.

NVIDIA has shifted to controlling their GPUs using Arm-CPUs

In 2022 NVIDIA developed Grace, its first Arm-based CPU for data centers. It was first deployed with Hopper GPUs, and this year it is being deployed with Blackwell GPUs. NVIDIA does not support x86 with Blackwell in an NVLink72 configuration. And because NVL72 delivers the most performance, it’s what most will deploy.

Comparing the Grace CPU to Intel Xeon Platinum 8480+ and AMD EPYC 9654, NVIDIA claims Grace’s relative performance is 1.2X to 2.4X on most non-AI benchmarks, and is 1.5X to 3.0X more energy efficient.

JPMorgan expects 5 million Blackwell GPU shipments in 2025 (and 7.5 Million Blackwell/Rubin units in 2026). Two Blackwell GPUs are controlled by 1 Grace CPU, meaning Grace CPU shipments in 2025 would be ~2.5 Million in 2025. One industry expert in data center infrastructure said that for AI workloads, the x86 servers are only for running control plane operations, so there will be 5 to 10 x86 CPU racks for 100,000 GPUs. A CPU rack holds 42 1U servers, and each 1U server can have 1 or 2 x86. Doing the math, that means 210 to 840 x86 CPUs for 100,000 GPUs, which have 50,000 Grace CPUs. So Grace CPUs outnumber x86 by 50X to 100X for AI workloads.

Earlier we saw x86 data center shipments are ~22 million. With 2.5 million Grace and maybe 1 million Arm-based CPUs made by the hyperscalers, Arm’s 2025 share is ~15%, not the near 50% Arm was predicting.

Arm will grow share by 2030, but even then it seems x86 will hang on for unit share majority

NVIDIA’s roadmap, presented at GTC 2025 in March, shows a new Arm processor named Vera to pair with the Rubin GPU and then the Feynman GPU. So NVIDIA is sticking with Arm+GPU.

CPUs attached to GPU will grow much faster than other data center CPUs. These will be primarily Arm-based, unless AMD grows GPU share rapidly (because AMD will connect their GPUs to their x86 CPUs).

The rest of CPUs in data centers, not attached directly to GPU, ship half to hyperscalers — which all have Arm CPUs. Those are seeing better price/performance, and will likely shift their customers to Arm with lower prices.

McKinsey in April of this year published their projections for data center growth through 2030.

Fig. 4: AI vs. non-AI workload in global data centers, measured by GW. Source: McKinsey

AI workloads, measured in watts, will grow 3.5X by 2030. Non-AI workload growth is 1.7X.

If the non-AI workload in 2025 is done by about 22 million x86 CPUs and about 1 million Arm CPUs, and the unit count grows as fast as the GW (rough guess), 2030 units will be about 37 million x86 CPUs and about 2 million Arm CPUs. For now, we are assuming the hyperscalers don’t shift their mix to Arm. If the top four hyperscalers who buy about half of CPUs shifted to 50% Arm, then in 2030 units will be about 29 million x86 CPUs and about 10 million Arm CPUs. The latter scenario seems more, likely given the hyperscaler investment and the promotion they are doing on their websites.

If the non-AI workload in 2025 is done by about 5 Million GPUs that will scale up to roughly 18 million GPUs by 2030 requiring about 9 million Arm CPUs. (Depending on how GPUs are counted it could be fewer Arm CPUs).

So in total by 2030 there will be roughly 48 million data center CPUs, with Arm accounting for up to 19 million of the total. These are rough numbers, but they show the direction is definitely toward Arm challenging for data center CPU dominance later this decade. The two big swing factors driving this are: 1) Hyperscalers shifting workloads from x86 to Arm in their clouds, and 2) AI Nvidia-based data centers using 50 to 100X more Arm CPUs than x86.

The dollar share is more important, but we don’t have Arm ASP data. It’s likely the x86 dollar share will be higher than their unit share, but the trends are still moving things in the direction of Arm share steadily growing over time. AMD taking significant share in AI from Nvidia would slow down Arm’s share growth since AMD will presumably use x86 for GPU co-processing, not Arm.