ARM Server Market Share 2026: Why There's No Single Number

ARM's Market Share, Properly Scoped for Mid-2026

The biggest flaw in most 'Arm is winning' coverage is treating cloud instance share, CPU shipments, deployed servers and OEM channel availability as one interchangeable percentage. They are four separate measurements, published by different organisations on different timelines, and merging them produces a market-share figure that describes nothing real. For anyone deciding where to spend a 2026 infrastructure budget, the distinction matters more than any single headline number.

On the cloud-fleet side, Network World reported in January 2025 that half of all instances spun up on AWS over the prior two years used Graviton silicon. A separate 2026 article *estimated* Graviton4's share of AWS's general-purpose compute instances at approximately 40%; this is an article-level claim, not an AWS-published statistic, and it covers a narrower instance category over a different window than the 50% figure above.

On the shipped-hardware side, Nvidia vice president Ian Buck said in July 2026 that Nvidia had shipped over 2.5 million Grace CPUs to date, alongside hundreds of thousands of Grace standalone servers. That is a cumulative shipment count across Nvidia's own customer base — not an instance-share metric, and not directly comparable to AWS's internal cloud numbers.

On the on-prem channel side, Ampere Computing's own 2026 materials show availability through OEM and branded partner platforms rather than a broad UK retail catalogue. That's a supply-channel fact, not a market-share statistic — but it's arguably the most important one for a UK buyer trying to work out what they can actually purchase.

AWS Graviton: From Capacity Crunch to Graviton5

AWS's own disclosures in 2026 tell a story of demand outrunning supply expectations. In April, the company said two large customers had already asked to buy all of its Graviton instance capacity for 2026 — a rare admission of internal fleet pressure from a hyperscaler that rarely discusses capacity constraints publicly.

Two months later, in June 2026, AWS made Graviton5 generally available, first through the EC2 M9g and M9gd instance families. At launch, these were available in four regions: US East (N. Virginia), US East (Ohio), US West (Oregon), and Europe (Frankfurt). AWS also confirmed M9g and M9gd could be purchased through Savings Plans, On-Demand, Spot Instances, Dedicated Instances and Dedicated Hosts — the same full purchasing toolkit AWS offers its flagship x86 families, which signals Graviton5 is being treated as mainstream fleet compute rather than a niche SKU.

Older Graviton generations remain very much alive alongside the new flagship: as of August 2026, AWS was still advertising a Graviton2 t4g.small free trial (750 hours per month), currently scheduled to run through 31 December 2026. A five-generation-old architecture still onboarding new developers at zero cost, next to a brand-new fifth-generation chip rolling out with full enterprise purchasing options, illustrates just how deep Arm now runs inside AWS's fleet.

Performance Proof Points: Graviton5 and Redshift RG

AWS reports its Graviton5-based R9g and R9gd memory-optimized instances deliver **up to 25% better compute performance** than the prior Graviton4-based R8g generation in published benchmarks.

A more concrete, service-specific proof point arrived in May 2026 with Amazon Redshift RG instances, launched in two sizes (rg.xlarge and rg.4xlarge). AWS said RG instances powered by Graviton are up to 2.4x as fast as the previous RA3 instances and cost 30% less per vCPU. Both figures are specific to Redshift as a managed data-warehouse service — they are not a proxy for how Arm performs across general-purpose EC2 compute, and shouldn't be read as such.

Nvidia Grace: Hyperscale Silicon at a Different Scale

In July 2026, Nvidia vice president Ian Buck stated that Nvidia had shipped over 2.5 million Grace CPUs to date, alongside hundreds of thousands of Grace standalone servers. While reported across industry press, this figure represents Nvidia's internal disclosures rather than an audited third-party census.

The key context is workload profile: this shipment volume reflects Nvidia's AI-adjacent infrastructure builds and hyperscale deployments rather than off-the-shelf enterprise server racks. It demonstrates that Arm-based CPU pairing functions reliably at massive scale in AI environments, but does not indicate widespread adoption for general-purpose enterprise applications.

Ampere and the On-Prem Reality Check

Ampere positions its Arm server CPUs as general-purpose, non-hyperscaler silicon. Its 2026 product lines remain targeted at OEM and branded platform partners rather than broad retail component distribution, making server sourcing fundamentally different from standard x86 procurement.

For UK buyers, this means sourcing hardware through certified OEM platforms—such as Gigabyte, Supermicro, or Wiwynn—rather than building white-box systems from retail channels. Procurement teams must verify application-level aarch64 support, OS kernel compatibility, and local warranty backing with the integrator, as field replacement units and engineer call-outs depend directly on the specific OEM partnership rather than a universal distributor network.

The UK Angle: Renting vs Owning Arm Compute

For most UK organisations in mid-2026, renting Arm compute through the cloud is the practical default over owning on-prem hardware, given the narrow server channel compared to hyperscale cloud capacity. AWS's Graviton5 launch included Europe (Frankfurt) among its four initial regions — the closest of AWS's launch regions to UK enterprises that need low-latency European deployment without running their own Arm hardware.

On-prem procurement in the UK runs through specialist system integrators rather than standard distribution catalogues. Platforms like Supermicro's MegaDC Arm series and Gigabyte's Ampere lines are supplied via UK integrators such as Boston Limited and Broadberry Data Systems. Entry-level 1U single-socket Ampere Altra servers typically start between £3,500 and £5,500 depending on memory and storage, rising above £10,000 for dense multi-socket configurations. Nvidia Grace systems, by contrast, are integrated into high-end HPC nodes from OEMs like HPE and Supermicro, where complete nodes sit comfortably above £25,000 and standard list prices do not exist outside custom project quotes. Buyers modelling both paths side by side can use our server configuration tools to price a comparable x86 baseline before deciding how much of a workload moves to Arm.

Migration Strategy, Software Fit and the Bigger Strategic Picture

Moving an existing x86 workload to Arm is primarily a due-diligence exercise rather than a like-for-like swap. The sensible sequence is: confirm the application and its dependencies have an aarch64-compatible build; verify the specific vendor support statement for any database, middleware or ISV software involved, especially for regulated workloads; and run a representative performance test before committing to capacity purchasing arrangements. AWS's own purchasing flexibility for Graviton5 instances — spanning Savings Plans through to Dedicated Hosts — gives buyers room to trial before locking in long-term commitments.

Beyond the migration mechanics, Arm's rise raises genuine strategic questions for IT decision-makers: energy efficiency per workload, supply-chain diversification away from a single CPU architecture, and the risk of trading one form of vendor dependency (a single x86 supplier) for another (a single cloud provider's custom silicon). Readers building a TCO model should benchmark Arm options against a current **Xeon vs EPYC** baseline, weigh workload-level power draw using recent server energy efficiency data, and review broader enterprise use cases via dedicated Arm-in-enterprise coverage before treating Arm as a default rather than a considered choice.

For UK buyers, the takeaway is straightforward: Arm has achieved critical mass in hyperscale cloud fleets and specialised AI platforms, while remaining a niche, build-to-order option for on-prem data halls. Evaluating it requires assessing each workload against those operational realities rather than relying on generalised market narratives.

Methodology

This study compiles publicly disclosed figures from vendor announcements (AWS's News Blog, AWS 'What's New' posts, About Amazon, and Ampere Computing), Nvidia executive disclosures, and trade press coverage (Data Center Dynamics, Network World), gathered through early September 2026.

Each figure was verified against its original primary source and operational scope—distinguishing cloud instance share inside a single provider from cumulative manufacturer shipments or OEM channel distribution. Nvidia's 2.5 million Grace shipment figure is tracked directly to executive statements by Ian Buck.

Where a figure originated as an article-level claim rather than a vendor-published statistic — the VaasBlock estimate of Graviton4's roughly 40% share of AWS general-purpose compute — it is labelled as such throughout rather than presented as confirmed data. No published shipment or instance-share figures for Azure Cobalt or Google Axion appeared in the sources examined at the time of research, and none have been estimated or inferred for this piece.

Sources

Every figure in this article traces to the sources below.

• •Data Center Dynamics — AWS Graviton capacity demand, April 2026
• •About Amazon — Graviton5 general availability, June 2026
• •AWS Blog — M9g/M9gd availability, regions and purchasing options
• •AWS News Blog — Graviton5 R9g/R9gd performance uplift
• •AWS — Redshift RG instance launch, May 2026
• •AWS — Graviton2 free trial offer terms
• •Network World — Graviton share of new AWS instances
• •VaasBlock — Graviton4 general-purpose instance share estimate
• •Gadget Otaku — Nvidia Grace CPU shipment disclosure
• •MLQ AI — Nvidia Grace standalone server shipment figures