Arm Neoverse grabs hyperscalers' AI-ready share: Spotify saw 250% better Axion performance

Spotify’s cloud processor test is a pretty blunt signal about where cloud computing is going next. In its evaluation of next-generation cloud processors, the company found that workloads running on Google Cloud Axion processors built on Arm architecture delivered roughly 250% better performance. That is not “a little faster.” It’s the kind of jump that makes you rethink migration plans, capacity planning, and cost models when your software is already running 24/7 at scale.

And Axion is not just a single product. The article frames it as part of a bigger Arm-based shift built on Arm’s Neoverse datacenter architecture, which “has been adopted across all major hyperscale cloud platforms.” AWS says its Arm-based Graviton processors have accounted for over half of new CPU capacity deployed over the past three years, and it reports that 98% of its top 1,000 Amazon EC2 customers running production workloads on Graviton benefit from Graviton’s price-performance advantages compared to x86. In other words, this move is not living in a lab. It’s monetized in production.

So what’s driving the urgency? The article makes it clear that AI workloads change the optimization game. Traditional enterprise deployments often focused on predictable CPU utilization and storage throughput. AI, especially when you combine training and inference with networking demands and storage access, forces simultaneous tuning across multiple parts of the stack while minimizing energy consumption and latency. At hyperscale, small inefficiencies do not stay small. They scale into real money and real operational constraints, especially because power consumption is becoming a significant portion of datacenter operating costs.

The physical constraint side is getting louder, too. The article cites an IDC report saying AI-ready datacenters are seeing rapid increases in power density, with rack requirements rising from typical levels of 5-10 kW to 30 kW or more, and in some cases exceeding 100 kW per rack. When racks get that power-hungry, the bottleneck is no longer just compute availability. It becomes how you design compute, networking, storage, and cooling systems as integrated rack-level subsystems. That pressure is also collapsing the old boundaries between compute, networking, storage, and acceleration. The “efficiency” metric is shifting from isolated component performance to end-to-end system behavior.

This is where Arm’s Neoverse story connects. Neoverse is described as Arm’s evolution from a mobile-first architecture to a platform purpose-built for cloud and AI infrastructure. It provides a common foundation that hyperscalers can use to design custom silicon optimized for their own workloads, tailoring performance, power, and system behavior. The article also points to hyperscalers moving from simply deploying Arm processors to designing silicon and infrastructure together based on real usage patterns.

Azure Cobalt and Google’s Axion are examples. The new Cobalt 200 processor is built on Arm Neoverse technology and was engineered using telemetry from real Azure workloads plus an internal suite of benchmark variants meant to reflect production behavior. On Google’s side, Axion processors are pursuing price-performance and energy efficiency gains, with C4A instances delivering up to 65% better price-performance and up to 60% greater energy efficiency than comparable x86 systems.

The measurable results from production migrations are where this becomes more than a tech trend. Databricks is using Azure Cobalt 100 virtual machines, built on Microsoft’s Arm-based CPU architecture, to optimize data-intensive and AI workloads, with claims of up to 50% better price-performance compared to previous generations, plus improvements in query speed and latency for analytics applications. Pinterest provides one of the most dramatic outcomes: by migrating workloads to AWS Graviton-based instances, it achieved 38% savings on compute resources and 47% cost savings for key workloads, while reducing carbon emissions by 62%. Uber’s transition shows the operational side of Arm adoption at scale, with the company migrating more than 2,800 services and shifting nearly 20% of its infrastructure capacity from x86 to Arm, requiring updates to codebases, dependencies, and deployment pipelines. Atlassian’s migration of Jira and Confluence involved moving more than 3,000 instances to Graviton, with production instance counts dropping by around 30%, throughput improving by up to 30%, and latency decreasing across key metrics.

Zooming out, the article connects these migrations to the rise of agentic AI. It describes a converged AI datacenter where CPUs act as the control plane, coordinating scheduling, data movement, and system services, while accelerators handle compute-intensive training and inference. In this model, efficiency is measured across the entire rack and datacenter footprint, not just a CPU benchmark. AI workloads also demand higher compute density under fixed power and cooling limits, which makes “compute output per unit of space” a key design objective. The article argues that Arm’s architecture spans layers, enabling providers to optimize the full stack while keeping software compatibility and ecosystem consistency.

It even gestures at how this is shaping rack-level solutions, citing NVIDIA’s Grace Blackwell and Vera Rubin platforms that combine Arm CPUs with high-performance GPU accelerators, plus AWS examples like Trainium3 UltraServers. The second-order implication is the quiet one executives can’t ignore: cloud strategy is becoming system design strategy. If your organization depends on predictable performance-per-watt and constrained power envelopes, you are no longer just choosing a processor. You are choosing an integrated architecture that can coordinate CPUs, accelerators, memory, and networking without wasting energy on bottlenecks and unnecessary data movement.