Google Axion
Last reviewed
May 31, 2026
Sources
11 citations
Review status
Source-backed
Revision
v2 ยท 2,390 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
11 citations
Review status
Source-backed
Revision
v2 ยท 2,390 words
Add missing citations, update stale details, or suggest a clearer explanation.
Google Axion is Google's first custom Arm-based central processing unit designed for the data center. Google announced it on April 9, 2024 at the Google Cloud Next conference, and it is built on the Arm Neoverse V2 platform combined with Google's own silicon design and the company's Titanium offload system. Axion powers a growing set of Google Cloud virtual machines, starting with the C4A family, and it serves as a general-purpose host processor inside Google's AI Hypercomputer architecture, where it sits alongside the company's TPUs and GPUs. Google positions Axion as a higher efficiency, lower cost alternative to comparable x86 instances for everyday cloud work and for CPU-based AI training and inference.[1][2][3]
Axion is part of a broader shift among the large cloud providers, often called hyperscalers, who have started designing their own server chips instead of relying only on Intel and AMD. Amazon led the way with Graviton in 2018, Microsoft followed with its Cobalt CPU, and NVIDIA built the Grace CPU used in systems like the GH200. Axion is Google's entry into that group, and it reflects the same logic that drove the others. Owning the design gives a cloud operator more control over performance, power draw, and the total cost of running millions of cores.[3][4]
A hyperscale data center runs at a scale where small per-core differences in performance and power add up to very large numbers. At that size the economics change. Buying merchant x86 chips means paying a vendor margin and accepting whatever roadmap that vendor sets. Designing a chip in house, usually on the Arm architecture, lets a company tune the processor to its own fleet, its own software, and its own power and cooling limits.[3][4]
The Arm route is attractive for a few practical reasons. Arm licenses ready made CPU cores through its Neoverse line, so a cloud provider does not have to design a core from scratch. It can take a proven core, surround it with custom logic for memory, networking, and security, and ship a competitive part in far less time than a ground up design would take. Arm cores also tend to deliver strong performance for each watt of power, which matters when electricity and heat are among the biggest running costs in a data center. The result is better control over total cost of ownership, the all in figure that covers hardware, power, cooling, and operations over the life of a server.[3][5]
There is a software story too. Much of the cloud workload that hyperscalers run, things like web servers, databases, caches, and microservices, is written in portable languages or runs in containers, so it moves to Arm with little or no change. Google noted that many of its own large services already ran on Arm servers before Axion shipped, including BigTable, Spanner, BigQuery, Blobstore, Google Earth Engine, and the YouTube Ads platform. That existing base gave Google confidence that customers could move general-purpose work onto Axion without a painful rewrite.[1][2]
Axion uses the Arm Neoverse V2 core, which is based on the Armv9 architecture. Neoverse V2 is one of Arm's higher performance server cores, and Arm describes the Armv9 generation as adding gains in performance, power efficiency, and security, including support for confidential computing. The same V2 core underpins other recent server chips, including Amazon's Graviton4 and NVIDIA's Grace, which gives a sense of where Axion sits in the market. Google did not publish low level specifications such as the exact core count or clock speed at launch, so independent reviewers could not verify the design details on their own.[1][5][6]
Google's contribution goes beyond licensing the core. The company pairs the Neoverse cores with its own system design and with Titanium, a layer of purpose built silicon and offload logic. Titanium is a set of custom microcontrollers and scale out offloads that take networking, security, and input and output processing off the main CPU. Storage traffic is handled by Hyperdisk, Google's network attached block storage. By moving this housekeeping work onto dedicated hardware, Titanium frees the Axion cores to spend their cycles on the customer's actual workload, which improves both performance and consistency. This is the same general idea behind the data processing units and offload cards that AWS and other providers use.[1][2]
Axion reached customers through Google Cloud's Compute Engine machine families. The first was the C4A series, a general-purpose line aimed at workloads such as web and application servers, containerized microservices, open source databases, in memory caches, data analytics, and batch jobs. C4A went into preview in mid 2024 and became generally available in late October 2024, with full availability including Titanium SSD local storage following in November 2024. C4A predefined shapes scale up to 72 vCPUs and as much as 576 GB of memory in the high memory configuration, and they come in standard, high CPU, and high memory ratios. Certain shapes add Titanium SSD local storage for workloads that need fast local disk.[1][7][9]
Google expanded the Axion lineup in November 2025, when it announced N4A and C4A metal. N4A is described as Google's most cost effective N series virtual machine. It scales to 64 vCPUs and 512 GB of DDR5 memory, supports custom machine shapes, and targets flexible cost optimized work such as microservices, open source databases, batch jobs, and the data preparation that feeds AI applications. Unlike C4A, which uses the Neoverse V2 core, N4A is reported to use the newer and more efficient Arm Neoverse N3 core, which fits its cost optimized role. N4A reached general availability on January 27, 2026. C4A metal is a bare metal instance that gives customers direct access to the Axion hardware without a hypervisor, with up to 96 vCPUs and 768 GB of memory. It suits licensing sensitive software, Android development, automotive in car systems, and other specialized workloads, and it remains in preview.[7][8][2]
| Instance family | Type | Status | Configuration |
|---|---|---|---|
| C4A | Virtual machine | Generally available, October 2024 | Up to 72 vCPUs, up to 576 GB memory, Neoverse V2, Titanium SSD on select shapes |
| N4A | Virtual machine | Generally available, January 2026 | Up to 64 vCPUs, up to 512 GB DDR5, Neoverse N3, custom shapes |
| C4A metal | Bare metal | Preview | Up to 96 vCPUs, up to 768 GB memory, no hypervisor |
Google's headline numbers are vendor claims, not independently audited benchmarks, and they have shifted as the product matured. At the April 2024 announcement Google said Axion delivered up to 30 percent better performance than the fastest general-purpose Arm based instances then available in the cloud, up to 50 percent better performance than comparable current generation x86 based instances, and up to 60 percent better energy efficiency than those same x86 instances. Arm repeated these figures in its own materials about the chip.[1][6]
After the product matured, Google's marketing moved to price performance comparisons. For C4A, Google states that the virtual machines deliver up to 65 percent better price performance and up to 60 percent better energy efficiency than comparable current generation x86 instances, plus up to 10 percent better price performance than the latest generation of Arm based instances in the cloud, a group that includes Amazon's Graviton4. For the later N4A, Google claims up to 2 times better price performance and up to 80 percent better performance per watt than comparable current generation x86 virtual machines. The shift from a raw performance claim to a price performance claim is worth noting, because price performance folds in cost as well as speed and is not directly comparable to the earlier performance only figure.[3][8]
Google also points to customer numbers. It has cited Spotify reporting roughly 250 percent better performance on Axion based C4A machines for some workloads, Vimeo seeing about a 30 percent gain on a core transcoding workload versus comparable x86 machines, and ZoomInfo measuring about 60 percent better price performance on N4A for certain jobs. These are customer reported figures for specific workloads rather than standardized benchmarks.[6][8]
| Claim | Comparison baseline | Instance and timing |
|---|---|---|
| Up to 30% better performance | Fastest general-purpose Arm instances in the cloud | Axion, launch April 2024 |
| Up to 50% better performance | Comparable current-generation x86 instances | Axion, launch April 2024 |
| Up to 60% better energy efficiency | Comparable current-generation x86 instances | C4A |
| Up to 65% better price-performance | Comparable current-generation x86 instances | C4A |
| Up to 10% better price-performance | Latest generation Arm instances in the cloud | C4A |
| Up to 2x better price-performance | Comparable current-generation x86 VMs | N4A |
| Up to 80% better performance-per-watt | Comparable current-generation x86 VMs | N4A |
As always with vendor numbers, the phrase up to does a lot of work. The figures describe best case results on workloads Google selected, and real gains depend on the application, the comparison instance, and how well the software uses the hardware. Independent press coverage at launch pointed out that Google withheld detailed specifications and third party benchmarks, so outside parties could not check the claims when the chip debuted.[4][10]
Google markets its AI stack as the AI Hypercomputer, an integrated system that bundles accelerators, general-purpose compute, networking, storage, and software into one platform for AI infrastructure. The heavy lifting in training and serving large models is done by TPUs, Google's custom AI accelerators, and by NVIDIA GPUs. Axion fills the role of the host CPU in this picture.[3][2]
That role matters more than the name host suggests. Even in an accelerator heavy system, a general-purpose CPU does essential work. It runs the orchestration software, feeds data into the pipeline, handles preprocessing and tokenization, manages storage and network traffic, and runs the many supporting services that surround a model. A more efficient host CPU means more of the system's power budget and cost can go toward the accelerators rather than the surrounding infrastructure. Google frames the division of labor plainly. It describes its specialized accelerators, including the Ironwood TPU, as the parts that handle model serving, while Axion handles what it calls the operational backbone, meaning high volume data preparation, ingestion, analytics, and the application servers that host intelligent applications. Google also lists CPU based AI training and inference among Axion's intended uses, which covers smaller models and classical machine learning where a CPU is the practical choice. So Axion does not compete with TPUs and GPUs. It complements them, much as Grace pairs with the Hopper and Blackwell GPUs in NVIDIA systems.[3][8][2]
Axion arrived well after its rivals, and the comparison is the easiest way to understand it. Amazon's Graviton program started in 2018 and is now several generations deep, with Graviton4 built on the same Neoverse V2 core as Axion and offering up to 96 cores.[11] Microsoft announced its Cobalt CPU, an Arm based data center processor, in late 2023, paired with its Maia AI accelerator. NVIDIA built the Grace CPU on Arm Neoverse and ties it tightly to its GPUs in products such as the Grace Hopper GH200.[4][6]
Seen against that field, Axion is Google catching up rather than breaking new ground. Amazon has a multi year head start and a deep catalog of Graviton instances, and analysts generally regard Graviton as the most mature hyperscaler Arm effort. Google's advantage is less about being first and more about closing a gap in its own portfolio and gaining the same control over cost and efficiency that its competitors already enjoy. The fact that three of the four largest accelerator and cloud vendors now ship custom Arm CPUs is a strong signal that the merchant x86 model no longer fits every part of the data center.[4]
Axion is significant because it completes Google's vertical integration. With TPUs for acceleration, Titanium for offload, and now Axion for general-purpose compute, Google designs most of the major silicon in its cloud rather than buying it. That gives the company more room to optimize performance, power, and cost across its fleet, and it gives customers an Arm option that is tuned to Google's infrastructure.[1][3]
The limitations are real and worth stating plainly. Axion is an Arm CPU, so workloads that depend on x86 specific code or on software that is not yet ported will not run on it without effort, although the large and growing Arm software ecosystem keeps shrinking that gap. The public performance and efficiency numbers are Google's own, framed as best case results, and the company has not released the detailed specifications that would let outsiders verify them or compare core for core with Graviton4 or Grace. Axion is also confined to Google Cloud. Unlike a merchant chip, customers cannot buy it for their own data centers, so its reach is tied to Google's cloud business. And as a late entrant, it competes against an Amazon Arm program that is several generations more mature. None of this undercuts the chip, but it does set realistic expectations for what a first generation, cloud only, vendor benchmarked processor delivers.[3][4]