Three hyperscalers now build their own Arm server chips, and by September 2026 all three have shipped a new generation within the same twelve months. AWS pushed Graviton5 into general availability in June 2026 with 192 cores per socket. Microsoft followed with Cobalt 200, a 132-core Neoverse V3 design built for the same class of general-purpose workloads. Google’s Axion, running on Neoverse N3 cores, anchors the newer C4A instance family on Compute Engine. Each company claims its chip delivers the best price-performance in the cloud, but the three use different core architectures, different process nodes, and different pricing models, so “best” depends heavily on what you’re running.
This comparison breaks down the actual silicon, the instance families each chip powers, published pricing for comparable configurations, and the workloads where each option wins. If you’re deciding between EC2 M9g, Azure’s Cobalt-based Dpsv7 family, or Google’s C4A instances, here’s what the numbers say as of mid-September 2026.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Why cloud providers build their own Arm chips
AWS, Microsoft, and Google all license Arm’s Neoverse cores rather than designing CPU cores from scratch. That distinction matters: none of these companies is competing with Arm Holdings directly, they’re competing with Intel and AMD for margin on the instances they rent out. A custom Arm chip lets a cloud provider control the full stack, from the silicon to the hypervisor, and capture the profit that would otherwise go to a third-party CPU vendor.
AWS was first to this game, launching the original Graviton in 2018 and iterating annually since. According to Amazon’s own Graviton5 announcement, more than 120,000 AWS customers were already running workloads on some generation of Graviton by June 2026, the same month M9g and M9gd instances reached general availability, according to Amazon’s official announcement. Microsoft’s Cobalt line is younger, with Cobalt 100 arriving in 2024 and Cobalt 200 following in late 2025. Google’s Axion, tied to the Neoverse N3 core and available through Compute Engine’s C4A family, is the newest of the three from a productization standpoint, even though Google’s Arm efforts trace back further through Tau T2A instances.
The pitch is consistent across all three: lower cost per unit of compute, lower power draw per rack, and better throughput on the kind of scale-out, stateless workloads that make up most of a modern cloud bill. Web tiers, microservices, container fleets, open-source databases, and CI runners are the default targets. What differs is how aggressively each company has pushed the architecture, and how far along each is in convincing enterprise customers to actually move.
Why this fight matters to the broader cloud price war
Custom Arm silicon isn’t a side project for any of the three companies, it’s a direct lever on the margin structure of their core business. AWS held roughly 30% of global cloud infrastructure spend in Q1 2026, with Azure at approximately 25% and Google Cloud at around 13%, according to figures cited from Synergy Research Group and reported by cloud analytics firm usage.ai. Every percentage point of compute margin a provider can claw back by moving customers off licensed x86 chips and onto proprietary Arm silicon flows straight to that provider’s cloud segment profitability, which is exactly why AWS, Microsoft, and Google have all been willing to spend years and, in some cases, multiple chip generations refining their in-house Arm programs rather than treating them as a niche SKU.
That’s also why the marketing language from each company leans so heavily on price-performance rather than raw benchmark supremacy. A cloud provider doesn’t need its Arm chip to be the fastest processor on the planet, it needs the chip to be cheap enough to produce and license that passing a meaningful discount to customers still improves the provider’s own margin versus selling an equivalent x86 instance. That dynamic is a big part of why all three companies report Arm instances priced 10-20% below comparable x86 options: the discount is real for customers, but it’s calibrated to still be profitable for the provider, not simply a loss-leader to drive adoption.
AWS Graviton5: 192 cores and a four-chiplet design
Graviton5 is the newest and, on paper, the most aggressive of the three chips. AWS built it around four chiplets, each carrying 48 Arm Neoverse V3 cores, for a total of 192 cores per socket â specifications, including its PCIe Gen 6 connectivity, that Amazon disclosed in June 2026 â according to a detailed architecture breakdown from The Next Platform. The chip runs on a TSMC 3-nanometer process, up from the 4nm node used for Graviton4, and clocks at 3.3GHz base frequency. Each socket carries a reported 180MB of L3 cache, which Tom’s Hardware described as large enough to directly challenge high-end AMD EPYC and Intel Xeon parts rather than just AWS’s own previous generation.
Memory bandwidth got a significant bump too. Graviton5 pairs with 12 channels of DDR5-8800 memory, which AWS describes as the fastest DDR5 currently deployed in a public cloud, plus eight PCIe Gen 6 controllers delivering 96 lanes with CXL 3.0 support, according to AWS’s own launch post for EC2 M9g and M9gd instances. AWS says the combination cuts inter-core latency by up to 33% versus Graviton4, a change the company frames as specifically useful for agentic AI workloads that spend a lot of time coordinating between concurrent processes rather than running a single long computation.
On raw generational improvement, AWS claims M9g instances deliver up to 25% better compute performance than the equivalent Graviton4-based M8g family, with narrower workload claims of 35% faster web application performance, 35% faster ML inference, and 30% faster database throughput at similar pricing. Those are AWS’s own numbers, measured against its own prior generation, not independent third-party benchmarks against Cobalt or Axion. No neutral cross-vendor SPECint or Geekbench comparison of all three current-generation chips has been published as of this writing, which is worth flagging before treating any single vendor’s price-performance claim as the final word.
Azure Cobalt 200: Microsoft’s second-generation Arm play
Cobalt 200 is Microsoft’s answer, and it shares more architectural DNA with Graviton5 than either company advertises. Both chips use Arm Neoverse V3 cores built on a TSMC 3-nanometer-class process. Cobalt 200 uses a dual-chiplet design totaling 132 physical cores, a smaller number than Graviton5’s 192, according to hardware coverage tracking the chip’s rollout, with Microsoft reporting that Cobalt 200 preview virtual machines became available in June 2026. Azure VMs built on Cobalt 200 expose up to 128 vCPUs per virtual machine, with 3MB of L2 cache per core and 192MB of system-level L3 cache shared across the socket.
One differentiator Microsoft leans on is default memory encryption, implemented through a custom memory controller baked into the Cobalt 200 design rather than bolted on as an optional feature. That matters for regulated industries running on Azure who need encryption-at-rest guarantees for memory contents, not just disk. Cobalt-based VMs also support up to 85 Gbps of networking bandwidth on the top-tier SKUs, positioning them for network-heavy workloads like content delivery edges and high-throughput API gateways.
Microsoft has been more conservative than AWS about publishing headline performance percentages for Cobalt 200 against its own predecessor, focusing marketing instead on total cost of ownership and power efficiency per rack. That’s consistent with Azure’s broader positioning: rather than chasing the highest core count, Microsoft is betting that Cobalt’s efficiency profile and default security features will win enterprise workloads that are more sensitive to compliance requirements than to raw throughput benchmarks.
Google Axion and the C4A family: efficiency over core count
Google’s approach with Axion diverges from both AWS and Microsoft at the core-architecture level. Where Graviton5 and Cobalt 200 both use the higher-performance Neoverse V3 core, Axion is built on Neoverse N3, a core designed for better performance-per-watt at the expense of some peak per-thread throughput. Google’s Axion-branded N4A instance family reached general availability in August 2026, according to Tech Insider, and Axion-derived instances top out at 64 vCPUs per VM with support for up to 512GB of memory on the largest configurations, specifications Tech Insider confirmed in September 2026. The related, newer C4A instance family (Compute Engine’s current general-purpose Arm line) reaches up to 100 Gbps of network bandwidth on select SKUs, ahead of the older Axion-branded N4A line’s 50 Gbps ceiling.
Google has published one of the few concrete, currently-listed hourly prices among the three providers: a c4a-highmem-16-lssd instance, carrying 16 vCPUs and 128GiB of memory, lists at $1.1907 per hour on-demand in us-east5, us-west1, and us-east4, a rate confirmed current as of September 11, 2026 by cloud pricing tracker CloudOptimo. That gives a concrete anchor point that AWS and Microsoft haven’t matched with equally granular public pricing pages for comparable Graviton5 or Cobalt 200 configurations.
Google’s public case for Axion and C4A rests less on chart-topping core counts and more on Neoverse N3’s efficiency curve translating into better performance-per-dollar at scale, particularly for horizontally-scaled stateless services, which is exactly the workload profile most likely to run in a large fleet of smaller instances rather than a handful of giant ones.
Specs comparison: Graviton5 vs Cobalt 200 vs Axion
| Spec | AWS Graviton5 | Azure Cobalt 200 | Google Axion / C4A |
|---|---|---|---|
| Core architecture | Arm Neoverse V3 | Arm Neoverse V3 | Arm Neoverse N3 |
| Max cores per socket | 192 (4 chiplets x 48) | 132 (2 chiplets) | Not publicly disclosed at chip level |
| Max vCPUs per VM | Up to 192 (bare-metal shape) | Up to 128 | Up to 64 (Axion); C4A scales higher |
| Process node | TSMC 3nm | TSMC 3nm (N3P) | Not publicly disclosed |
| Base clock speed | 3.3GHz | Not publicly disclosed | Not publicly disclosed |
| L3 cache | ~180MB | 192MB (system-level) | Not publicly disclosed |
| Memory support | 12-channel DDR5-8800 | DDR5 (channel count undisclosed) | Up to 512GB per VM (Axion) |
| PCIe / I-O | PCIe Gen 6, CXL 3.0, 96 lanes | Not publicly disclosed | Not publicly disclosed |
| Peak network bandwidth | Varies by instance size | Up to 85 Gbps | Up to 100 Gbps (C4A) |
| Default memory encryption | Via Nitro Isolation Engine | Yes, custom memory controller | Not highlighted as a default feature |
| Instance family | EC2 M9g / M9gd | Azure Dpsv7 / Epsv7-class VMs | Compute Engine C4A |
| GA date | June 2026 | Late 2025 | 2025 (C4A expansion through 2026) |
A caveat worth stating plainly: Microsoft and Google have not published the same level of granular chip-level detail that AWS has for Graviton5. AWS’s marketing and technical documentation are unusually specific, down to per-chiplet core counts and named memory bandwidth figures. Azure and Google’s public materials focus more on VM-level specs and less on the underlying silicon, which makes some rows above harder to verify independently. Where a figure isn’t publicly confirmed, it’s marked as such above rather than estimated.
Pricing: what a comparable instance actually costs
Getting an apples-to-apples price comparison across all three clouds is harder than it should be, because AWS and Microsoft don’t publish the same granular per-instance pricing tables that Google does for its C4A line. Here’s what’s confirmed as of mid-September 2026 for a roughly comparable general-purpose configuration.
| Provider | Instance | vCPUs | Memory | On-demand price |
|---|---|---|---|---|
| Google Cloud | c4a-highmem-16-lssd | 16 | 128GB | $1.1907/hour (us-east5, us-west1, us-east4) |
| Google Cloud | c4a-standard-4 | 4 | 16GB | Scales down proportionally from the highmem tier |
| AWS | M9g (approx. 16 vCPU shape) | 16 | 64GB | Not published in a static table; requires AWS pricing calculator |
| AWS | M9g.8xlarge | 32 | 128GB | Not published in a static table; requires AWS pricing calculator |
| Azure | Cobalt-based Dpsv7-class VM (approx. 16 vCPU) | 16 | 64GB | Not published in a static table; requires Azure pricing calculator |
The takeaway isn’t that AWS and Azure are hiding their prices, it’s that Graviton5 and Cobalt 200 pricing is distributed across region-specific, commitment-tier pricing calculators rather than a single static list price, while Google publishes fixed regional rates for C4A directly. If you’re doing a real procurement comparison, budget time to run each provider’s own calculator with your exact region, commitment term, and storage class, because list prices can shift 15-20% just by switching between on-demand, 1-year committed use, and 3-year committed use discounts. Public commentary from cloud cost analysts consistently notes that all three Arm families price at a discount to equivalent x86 instances on the same cloud, generally in the 10-20% range, though the exact gap depends heavily on instance size and region.
Performance claims: what’s vendor-published versus independently verified
This is the section where a comparison article has to be honest about the limits of what’s currently public. As of September 2026, no independent lab has published a controlled, apples-to-apples SPECint, SPECfp, or Geekbench benchmark run across all three current-generation chips side by side. What exists instead is a set of vendor-published, intra-generational comparisons: AWS comparing Graviton5 to Graviton4, and various third-party writeups estimating Cobalt 200 and Axion’s positioning based on core architecture alone.
What can be said with reasonable confidence based on core architecture: Graviton5 and Cobalt 200 both use the newer, higher-performance Neoverse V3 core, which Arm designed for stronger per-thread throughput than the Neoverse N3 core underpinning Axion. That architectural choice suggests Graviton5 and Cobalt 200 should outperform Axion on single-threaded or lightly-parallel workloads, while Axion’s N3 core is tuned to deliver more total throughput per watt when a workload can be spread across many smaller, cheaper cores. Graviton5’s edge is raw core count: at 192 cores per socket versus Cobalt 200’s 132, AWS’s chip has a meaningful density advantage for workloads that scale near-linearly with core count, like containerized microservices or parallel build pipelines.
Independent industry analysis comparing all three chips for enterprise migration planning has cautioned that the marketing percentages each vendor publishes are not directly comparable across vendors, since each uses a different baseline and a different workload mix for its own claims. That’s a fair characterization of where the public benchmark landscape stood as of late August 2026. Any team making a real procurement decision should run its own representative workload on trial instances from all three providers rather than relying on any single vendor’s percentage claim.
Total cost of ownership: beyond the hourly rate
The hourly compute rate is the number everyone fixates on, but it’s rarely the number that decides a real migration. Storage IOPS, network egress, and commitment discounts routinely move the actual bill by 20-30% in either direction, and each provider structures those charges differently around its Arm instance families.
AWS bundles Graviton5 pricing into the same EBS, S3, and data-transfer pricing structure used across the rest of EC2, so switching from an x86 M8i instance to an Arm-based M9g instance changes only the compute line item, not storage or networking costs. That’s a meaningful advantage for teams doing incremental migration, since it isolates the compute savings cleanly without forcing a parallel storage re-architecture. Azure follows a similar model for Cobalt-based VMs sitting inside existing virtual networks and managed disk tiers. Google’s C4A pricing is quoted per-instance inclusive of local SSD in some configurations (the “lssd” suffix in c4a-highmem-16-lssd denotes local solid-state storage bundled into that specific SKU), which can simplify budgeting for storage-heavy workloads but makes direct line-item comparison against AWS or Azure trickier, since the storage cost isn’t broken out separately.
Commitment discounts are where the real savings show up for any team planning to run Arm instances for more than a few months. All three providers offer multi-year committed-use pricing, and the discount curve is steep: a 3-year commitment on any of the three clouds typically cuts the effective hourly rate by 40-60% versus on-demand pricing, according to general cloud commitment-pricing patterns documented across all three providers’ public rate cards. That means the headline on-demand numbers in the tables above, including Google’s confirmed $1.1907/hour C4A rate, will look very different once a procurement team layers in a 1-year or 3-year reservation. Anyone building a serious cost model should run the comparison at the commitment tier they actually intend to buy, not at list price, since list-price comparisons between providers can be misleading by double-digit percentages once discounts are applied unevenly.
Power efficiency and sustainability claims
Power draw per unit of compute is the other half of the pitch each provider makes for its Arm silicon, and it’s arguably a bigger deal for the hyperscalers themselves than for most customers, since power and cooling are a direct line item on their own data center budgets. Arm’s Neoverse core family was designed from the outset around a lower power envelope than comparable x86 cores, and all three custom chips inherit that baseline advantage before any provider-specific optimization is applied.
AWS frames Graviton5’s efficiency case around density: packing 192 cores into a single socket means fewer physical servers are needed to hit a given aggregate compute target, which reduces the number of chassis, power supplies, and cooling loops required per unit of delivered compute. Microsoft’s public framing for Cobalt 200 leans more explicitly on power efficiency per rack as a headline selling point rather than raw core count, consistent with Azure’s broader data center sustainability commitments. Google’s Neoverse N3-based approach with Axion is the most efficiency-first of the three architecturally, since N3 cores were designed by Arm specifically to prioritize performance-per-watt over peak per-thread throughput, the same tradeoff that shows up in Axion’s lower per-VM vCPU ceiling relative to Graviton5 and Cobalt 200.
None of the three providers has published a directly comparable watts-per-core or performance-per-watt figure that allows a clean three-way efficiency ranking, so efficiency claims should be read the same way as performance claims: directionally useful, but not independently verified against competitors. For a workload where power efficiency genuinely matters, such as a large batch job billed at a fixed compute budget, the more actionable move is testing against your own workload’s actual power-adjusted cost per request, which each cloud’s billing data can approximate once you have live production traffic on the instance.
How to benchmark these chips against your own workload
Given the absence of an independent, neutral three-way benchmark, the responsible path for any engineering team is to build a small, repeatable test harness rather than trust a vendor’s percentage claim. A few practices make that comparison meaningful rather than misleading.
Test with production-representative traffic, not a synthetic loop. A generic compute benchmark like SPECint measures the CPU in isolation, but most real workloads spend meaningful time waiting on memory, network I/O, or downstream service calls, and Graviton5’s larger cache and faster memory bandwidth may matter more or less depending on how memory-bound your specific application actually is. Run the same container image, same dependency versions, and same request mix across trial instances on all three providers, and measure end-to-end latency and throughput under your actual load pattern, not a CPU-only microbenchmark.
Hold instance size roughly constant when comparing across clouds, since a 16-vCPU instance on one provider isn’t guaranteed to deliver the same per-vCPU performance as a 16-vCPU instance on another, given the underlying core-count and clock-speed differences documented in the specs table above. Also test at the concurrency level you’ll actually run in production. Graviton5’s lower inter-core latency, for example, is specifically optimized for workloads running many concurrent lightweight processes, so a single-threaded synthetic test would understate its advantage for an agentic AI or microservices workload, while overstating it for a workload that’s genuinely single-threaded.
Finally, price the trial at the commitment tier you’ll actually purchase, not on-demand. A chip that looks marginally more expensive on-demand can flip to the cheaper option once a 1-year or 3-year committed-use discount is applied, since discount depth varies by provider and instance family.
What’s next: the Arm roadmap through 2027
All three hyperscalers are on annual or near-annual Arm chip release cadences at this point, which means the comparison in this article will need revisiting well before Graviton5, Cobalt 200, and the current Axion/C4A generation reach end of life. AWS has shipped a new Graviton generation roughly every 12-18 months since 2018, and Graviton5’s June 2026 general availability puts the company on track for a successor sometime in 2027 if that cadence holds. The jump from Graviton4’s 4nm process and 96 cores to Graviton5’s 3nm process and 192 cores suggests AWS is willing to make large generational leaps rather than incremental refreshes, which raises the bar for what a Graviton6 would need to deliver to look like a meaningful upgrade again.
Microsoft’s Cobalt line has moved from Cobalt 100 to Cobalt 200 in roughly 18 months, and Google has been steadily broadening its C4A family’s instance shapes and regional availability throughout 2026 even without a headline “Axion 2” branding event. The practical implication for anyone planning a multi-year infrastructure strategy: don’t treat whichever chip wins this comparison in September 2026 as a permanent choice. Build your migration tooling (multi-architecture container images, infrastructure-as-code templates that parameterize the instance family) so that switching to next year’s chip generation, on the same cloud or a different one, doesn’t require re-doing the entire Arm compatibility audit from scratch.
Software and OS compatibility
All three chip families implement the Armv9 instruction set with SVE (Scalable Vector Extension) support, and all three run mainstream Linux distributions without source-level changes for most application code, provided your dependencies already ship Arm64 binaries. Graviton5 specifically exposes Armv9.2-A with a 128-bit SVE vector width and one hardware thread per core, according to CPU information pulled directly from live M9g instances by Japanese cloud consultancy Classmethod in a hands-on breakdown of the instance launch. That configuration is consistent with Graviton4, which means most workloads already tuned for Graviton4 should run on Graviton5 with no code changes, only a recompile or container rebuild targeting the newer instance type.
The practical blocker for all three Arm families isn’t the OS, it’s the dependency graph. Any team with a Python, Node.js, or Java stack built entirely from packages that publish Arm64 wheels or binaries can usually migrate in days. Teams with legacy C/C++ extensions, proprietary licensed software without an Arm build, or GPU-adjacent tooling tightly coupled to x86-specific instruction sets face a longer validation cycle regardless of which cloud they’re on. This is true whether you’re moving to Graviton, Cobalt, or Axion; the migration friction is architectural, not vendor-specific.
Real-world adopters and use cases
AWS has the deepest public adoption story simply because Graviton has existed the longest. Amazon’s own materials cite more than 120,000 customers running some Graviton generation, spanning SaaS platforms, financial services firms, and machine learning inference pipelines. That scale gives AWS an ecosystem advantage: more managed services (RDS, ElastiCache, EMR, and others) offer Graviton-backed instance options as a default choice, not an afterthought.
On Azure, data and analytics platform Databricks has been cited as an early adopter of Cobalt-class infrastructure for its Azure-hosted workloads, drawn by the combination of Neoverse V3 throughput and the default memory encryption Microsoft bakes into Cobalt 200 VMs. That pairing is a natural fit for regulated-industry analytics workloads where encryption-at-rest for memory is a compliance requirement, not just a nice-to-have.
Google’s Axion and C4A adoption skews toward cloud-native SaaS vendors and internal Google services migrating stateless microservices off x86. Because C4A instances are newer to Compute Engine’s broader lineup than Graviton is to EC2, the third-party case study base is thinner in public reporting, though the underlying use case pattern (horizontally-scaled web and API tiers) mirrors what AWS and Azure customers report doing with their own Arm fleets.
Five use cases and which chip fits best
- Containerized microservices at scale: Graviton5’s 192-core density gives it an edge for teams packing many small containers onto fewer, larger hosts, particularly on EKS where Graviton support is mature and well-documented.
- Regulated-industry analytics and finance workloads: Cobalt 200’s default memory encryption and Neoverse V3 throughput make Azure the stronger default when compliance requirements demand encryption guarantees beyond disk-level encryption.
- Cost-sensitive horizontal web tiers: Google’s published, fixed C4A pricing (like the confirmed $1.1907/hour for a 16-vCPU/128GB instance) makes budgeting easier for teams that want price certainty without running a custom calculator scenario for every SKU.
- Agentic AI orchestration and multi-process coordination: AWS specifically engineered Graviton5’s lower inter-core latency for exactly this pattern, positioning M9g as the default choice for teams running agent frameworks that spawn many concurrent lightweight processes.
- Mixed x86/Arm fleets during a gradual migration: All three clouds support running Arm and x86 instances side by side behind the same load balancer, so teams unsure which chip fits their specific workload can pilot a subset of traffic on each provider’s Arm family before committing.
Migration guide: moving a workload to a cloud Arm instance
The migration path looks broadly similar whether you’re targeting Graviton5, Cobalt 200, or Axion/C4A. Here’s the general sequence teams use in practice.
- Audit your dependency tree for Arm64 (aarch64) compatibility. Check whether every package, container base image, and compiled binary in your stack publishes an Arm64 build.
- Rebuild your container images with multi-architecture manifests (Docker buildx or an equivalent) so the same image tag works on both x86 and Arm hosts during the transition period.
- Spin up a small trial instance on your target Arm family (M9g on AWS, a Cobalt-based Dpsv7 VM on Azure, or a C4A instance on Google Cloud) and run your actual application workload, not a generic benchmark.
- Profile CPU, memory, and I/O under realistic load, then compare against your current x86 baseline for the same traffic pattern.
- Check managed service compatibility: confirm your database, cache, and message queue services support Arm-backed compute if they’re self-managed on the same instance family.
- Update your infrastructure-as-code templates (Terraform, CloudFormation, or Bicep) to reference the new instance type, keeping the x86 definition available as a rollback path.
- Roll out gradually behind a load balancer, shifting a small percentage of production traffic first and monitoring latency and error rates before scaling the migration further.
- Recalculate your cost model using the provider’s actual commitment-tier pricing (on-demand, 1-year, or 3-year) rather than list price, since committed-use discounts materially change the cost comparison.
Pros and cons of each platform
AWS Graviton5
Pros: Highest core density at 192 cores per socket, deepest managed-service integration after eight years of Graviton iteration, largest publicly cited customer base, and the most granular technical documentation of the three providers.
Cons: Per-instance pricing isn’t published in a simple static table the way Google’s C4A pricing is, requiring use of AWS’s pricing calculator for exact figures, and the highest-performance claims are self-reported against AWS’s own prior generation rather than independently verified against competitors.
Azure Cobalt 200
Pros: Default memory encryption baked into the chip design, strong fit for regulated industries already standardized on Azure, and Neoverse V3 cores matching AWS’s architecture choice for per-thread performance.
Cons: Lower maximum core count than Graviton5, less public case-study depth than AWS, and Microsoft has published fewer granular architecture specs than AWS has for Graviton5, making independent verification harder.
Google Axion / C4A
Pros: Clear, fixed, publicly listed hourly pricing for specific instance shapes, strong network bandwidth ceiling on C4A (up to 100 Gbps), and an efficiency-first Neoverse N3 core that can be cost-effective for horizontally-scaled workloads.
Cons: Lowest maximum vCPU count per VM among the three (64 vs. 128-192), thinnest public case-study base, and the least architectural detail disclosed publicly about the underlying chip.
The verdict: which chip should you actually pick
There’s no single winner here, and any comparison claiming otherwise is oversimplifying. If your team is already deep in the AWS ecosystem and running containerized, horizontally-scaled workloads, Graviton5’s core density and mature managed-service support make it the lowest-friction path, and the newly reduced inter-core latency specifically helps agentic AI workloads that are becoming a bigger share of cloud compute demand heading into 2027. If your organization is Microsoft-centric and operates under regulatory requirements around data encryption, Cobalt 200’s default memory encryption and comparable Neoverse V3 core performance make Azure the more defensible choice, even with a lower maximum core count than Graviton5. If cost predictability matters more than raw peak performance, and your workload is genuinely horizontal (many small instances rather than a few large ones), Google’s fixed and publicly listed C4A pricing removes a lot of the pricing-calculator guesswork the other two providers still require.
The honest caveat, worth restating: nobody has published an independent, side-by-side SPECint or Geekbench run of Graviton5, Cobalt 200, and Axion on identical workloads as of this writing. Every performance claim currently in public circulation is either vendor-published against that same vendor’s prior generation, or a third-party architectural inference based on which Neoverse core each chip uses. Until that independent benchmark exists, the responsible move for any team weighing a real migration is to trial your actual workload on all three, not to pick a winner off a spec sheet.
Frequently asked questions
Is AWS Graviton5 faster than Azure Cobalt 200?
Graviton5 has more cores per socket (192 versus 132) and both use the same Neoverse V3 core architecture, so Graviton5 likely has a throughput edge on workloads that scale with core count. No independent cross-vendor benchmark has confirmed this directly, and AWS’s own published performance percentages are measured against Graviton4, not Cobalt 200.
What instance families use Google Axion?
Axion originally powered Google’s N4A instance family, while the newer, broader C4A family is Google Compute Engine’s current general-purpose Arm line, with higher network bandwidth ceilings than the original N4A instances.
Does Azure Cobalt 200 support memory encryption by default?
Yes. Cobalt 200 includes a custom memory controller that enables memory encryption as a default feature rather than an opt-in configuration, which distinguishes it from both Graviton5 and Axion in how each vendor discloses that capability.
How much cheaper are Arm cloud instances than x86 equivalents?
Public commentary from cloud cost analysts generally puts the discount for Arm-based instances at 10-20% versus comparable x86 instances on the same cloud, though the exact gap varies by instance size, region, and commitment tier. Google’s C4A pricing is the most transparently published of the three, while AWS and Azure require using each provider’s pricing calculator for exact figures.
Can I run the same container image on Graviton5, Cobalt 200, and Axion?
Yes, provided the image is built as a multi-architecture manifest targeting Arm64/aarch64. All three chips implement the Armv9 instruction set with SVE support and run standard Linux distributions, so a properly built Arm64 container should run on any of the three without source changes, though performance will vary based on the underlying core architecture.
Which cloud has the most mature Arm ecosystem?
AWS, by a meaningful margin, based on tenure. Graviton has existed since 2018, giving AWS the longest runway to integrate Arm support across managed services like RDS, ElastiCache, and EMR. Azure’s Cobalt and Google’s Axion are both newer efforts with less time to build out equivalent managed-service coverage.
Is there an independent benchmark comparing all three chips directly?
Not as of September 2026. All publicly available performance claims are either vendor-published comparisons against that vendor’s own prior generation, or third-party architectural analysis based on which Neoverse core each chip uses. Teams making procurement decisions should run their own representative workloads on trial instances rather than relying on a single vendor’s benchmark claim.
