Trends · Server hardware

The Rise of Arm Servers: What 2026’s Shift Away From x86 Means for Buyers

For a decade “Arm in the data center” was a slide in a keynote that never quite arrived. In 2026 it arrived. The shift is large enough that it changes how anyone procuring compute should think — not because x86 is finished, but because for a growing share of workloads Arm is now the default-sensible choice rather than the experimental one. This guide separates what genuinely changed from what is marketing arithmetic, and lays out when an Arm server is the right tool for the job.

How big is Arm in the data center now?

Bigger than most people assume, and the curve is steep. Arm-based server CPUs accounted for roughly 5% of the market in 2020 and about 15% by 2024; in the first quarter of 2026, IDC reported that Arm machines commanded more than 45% of data-center server revenue, since Arm made up over 95% of all non-x86 server revenue. Over the same window, x86 server revenue actually slipped — IDC measured a 2.9% decline to $63.9 billion in Q1 2026, which it attributed to supply constraints rather than collapsing demand. Arm Holdings itself had publicly targeted half of data-center CPU sales by the end of 2025.

The momentum is broad-based. AWS, the first hyperscaler to ship a custom Arm data-center CPU with Graviton, now drives more than half of its new CPU demand on Arm and serves something like 100,000 cloud customers on those instances. Google followed with Axion, Microsoft with Cobalt, and Ampere supplies merchant Arm silicon to everyone else. The competitive question shifted years ago from “can Arm servers work” to “whose Arm server chip is best,” which is the clearest possible sign of a mature market.

Why Arm is winning: performance per watt

The single reason Arm broke through is the same reason the whole industry is rethinking infrastructure — power. Arm server designs are credited with roughly 30 to 60 percent better energy efficiency than comparable x86 parts, and 30 to 50 percent better price-performance in many cloud workloads. When a facility’s binding constraint is megawatts and cooling rather than rack space, a processor that does the same work for less electricity stops being a nice-to-have and becomes a direct lever on both cost and capacity. We walk through that power-first reality in detail in AI infrastructure trends; Arm’s rise is one of its most concrete consequences.

There is a strategic dimension too. AWS deliberately priced Graviton below equivalent x86 instances to accelerate adoption, and it worked — Graviton became one of the fastest-growing compute products in AWS history. When the largest cloud operator runs its own infrastructure on Arm and passes the savings on, the architecture is validated at a scale no benchmark can match.

Is the “45% of the market” number what it looks like?

Not quite — and the gap is worth understanding before you repeat the figure. That 45% is a revenue share, and a large slice of it comes from Nvidia’s rack-scale AI systems. Each GB200 NVL72 pairs Grace Arm CPUs with Blackwell GPUs and can sell for up to about $6.5 million per unit, so a relatively small number of Arm host processors drags an enormous amount of dollars into the Arm column. By unit volume the picture is very different: analysts estimate Nvidia ships on the order of four million Grace and Vera CPUs in a year against roughly twenty million x86 server processors from AMD and Intel combined.

So the honest reading is layered. Arm’s rise in general-purpose cloud compute is real and substantial — Graviton and its peers genuinely run a growing share of everyday workloads. But the eye-catching “nearly half the market” headline is amplified by the AI GPU boom, where an Arm chip rides along inside a machine bought for its GPUs. Both things are true at once, and conflating them is how the topic gets oversold.

The vendors shipping Arm server silicon

Effectively all the major operators ship Arm now, plus a healthy merchant market — and the diversity is the point. AWS Graviton4 carries 96 Neoverse V2 cores with a 192-core Graviton5 in development; Google’s Axion is built on the same Neoverse V2 generation; Microsoft’s Cobalt 100 fields 128 Neoverse N2 cores for Azure’s own services. Ampere, the independent supplier, sells AmpereOne at up to 192 cores for anyone who does not run a custom silicon team. Nvidia’s Grace — and its successor Vera — anchor the GPU racks. Arm itself has even begun shipping its own server CPU, its first production silicon in decades, and Qualcomm is re-entering the server market. No one party controls the stack, which keeps pricing competitive and innovation fast.

When does an Arm server make sense for your workload?

When your software is portable and your priority is efficiency or cost — which today covers a lot of ground. Modern cloud-native stacks port to Arm with little friction: interpreted and JIT runtimes (Go, Java, Node, Python, .NET), containerised microservices, web and API tiers, caching layers, and many databases all have first-class arm64 builds. Multi-architecture container images mean the same pipeline can target both. For scale-out, stateless services the price-performance and power savings usually land straight on the bottom line.

# What am I actually running on?
$ uname -m
aarch64
$ lscpu | grep -E ‘Architecture|Model name|Core’
Architecture:        aarch64
Model name:          Neoverse-V2
Core(s) per socket:  96
 
# Will my image run here? Build for both, ship one manifest.
$ docker buildx build —platform linux/amd64,linux/arm64 \
-t registry.example.net/app:1.4 —push .
=> exporting manifest list … done
=> linux/amd64  OK
=> linux/arm64  OK
 
# If a dependency has no arm64 build, that’s your migration cost.

The migration cost is concentrated in a few places: native binaries compiled only for x86, niche ISV software without arm64 packages, anything leaning on x86-specific instruction sets like AVX-512, and Windows-centric estates where the Arm story is still thinner. The practical method is unglamorous — benchmark your actual workload on both architectures rather than trusting a vendor slide, start with a low-risk service, and measure cost-per-request, not just raw core counts. If you are weighing the broader build-versus-rent question, our bare metal vs cloud guide covers the surrounding decision.

One practical caveat for buyers: outside the big clouds, Arm bare metal is still a more specialised purchase than x86. Most single-tenant dedicated-server catalogues remain x86 by default, so if you specifically want Arm hardware you own rather than a Graviton instance you rent, you are shopping a narrower shelf. A handful of providers do offer it — Hetzner fields an Arm64 line in its dedicated range, and Scaleway’s Elastic Metal even includes an early RISC-V option alongside its Arm and Apple-silicon machines — but availability, core counts, and regions are thinner than the x86 mainstream. Plan for that: confirm the exact CPU, the arm64 support of every dependency in your stack, and the data-centre location before you commit, rather than assuming Arm bare metal is a drop-in for whatever x86 box you were going to buy.

Where Arm fits in a sovereignty and repatriation story

Arm’s rise is also part of a larger move away from a two-vendor x86 world, and that connects directly to the trends its siblings cover. A broader set of architectures — Arm, and the open RISC-V instruction set behind a growing class of servers — gives operators and nations more independence from any single supplier, which is one strand of the wider sovereign cloud movement. The same efficiency that makes Arm attractive in hyperscale also makes self-owned, in-region infrastructure more economical, which feeds the cloud repatriation trend of pulling steady workloads back onto controlled hardware. Architecture choice, in other words, is no longer just a performance question; it has become part of how organisations think about control.

The honest case for staying on x86

None of this means you should rush a migration. x86 still ships the majority of server CPUs by volume, and for good reasons: an enormous installed base, mature tooling, broad ISV certification, and workloads — legacy enterprise applications, certain HPC and database paths, Windows-heavy environments — where the x86 ecosystem remains simply easier. AMD and Intel are not standing still either; both have aggressive 2026 roadmaps with very high core counts, from Intel’s dense E-core designs to AMD’s latest EPYC generations. For a single-tenant dedicated server running established software, x86 is frequently the lower-risk, lower-effort answer, and choosing it is not falling behind.

The right frame for 2026 is not “Arm versus x86” as a loyalty test. It is a portfolio decision: put portable, scale-out, efficiency-sensitive workloads where Arm’s perf-per-watt pays off, keep x86 where compatibility and maturity matter more, and let the workload — not the hype cycle — decide. The teams getting value from Arm are the ones treating it as another tool in the rack, measured on their own numbers.