The Performance Story: Graviton 5 Delivers
Amazon’s Graviton 5 is here, and it’s a beast. Announced at AWS re:Invent in December 2025, this fifth-generation Arm-based CPU delivers a 35% performance boost over its predecessor, Graviton 4, across applications, machine learning inference, and databases. That’s not just incremental progress—it’s a significant leap in a market where every percentage point counts.
What’s striking is how AWS is positioning this chip. Instead of celebrating it as a general-purpose workhorse, the company and its partners keep slapping the "AI chip" label on it. But let’s be clear: Graviton 5 is a CPU, not an AI accelerator. It’s designed to handle a wide range of workloads, from running your backend services to crunching data in databases. Calling it an "AI chip" is like saying Excel can be used as a database—technically true, but not its primary purpose.
The performance gains are real, though. AWS’s own benchmarks show Graviton 5 outperforming Graviton 4 by 35% in applications and ML inference, and 30% in databases. That’s impressive, especially when you consider that Graviton 4 was already a strong performer. But here’s the catch: Graviton 5 instances are 9% more expensive than their Graviton 4 counterparts. For customers running high-CPU-utilization workloads, the cost per performance might still make sense. But for those with fixed-node setups—like database replicas or multi-AZ deployments—the price hike is just that: a hike.
What Graviton Actually Is
Graviton 5 is a general-purpose Arm CPU designed by Annapurna Labs, an AWS subsidiary acquired in 2015. It’s part of a lineage that started with the first Graviton chip in 2018, and it’s built to handle everything from web servers to data analytics. But here’s where things get messy: AWS and its partners keep calling it an "AI chip."
This mislabeling isn’t just a minor semantic issue. It obscures what Graviton 5 actually is—a versatile, high-performance CPU. AWS already has a dedicated AI accelerator: Trainium, a systolic array designed specifically for machine learning workloads. Graviton 5 isn’t that. It’s a CPU that can run AI workloads, sure, but so can any modern CPU. That doesn’t make it an "AI chip."
The confusion isn’t helped by AWS’s own marketing. The press release for Graviton 5 calls it a chip "for the Agentic AI era." Meanwhile, The Wall Street Journal reported that Snowflake’s $6 billion AWS commitment was for "agentic computing chips"—a reference to Graviton. This kind of language might sound flashy, but it risks undermining the credibility of a business that’s now pulling in over $20 billion in annual revenue.
The AI-Labeling Problem
The insistence on calling Graviton 5 an "AI chip" isn’t just misleading—it’s a problem. For one, it dilutes the meaning of what an AI chip actually is. Trainium, AWS’s actual AI accelerator, is designed from the ground up for machine learning workloads. It’s a systolic array, optimized for the kind of matrix math that powers modern AI. Graviton 5, on the other hand, is a general-purpose CPU. It’s great at running a wide range of workloads, but it’s not specialized for AI.
This kind of mislabeling isn’t just a pet peeve of industry watchers. It has real implications. For a business that’s now generating over $20 billion in annual revenue, credibility matters. If AWS and its partners keep calling everything an "AI chip," it risks confusing customers and undermining the unique value of its actual AI hardware.
The issue isn’t just about semantics. It’s about clarity. Customers need to know what they’re buying and what it’s best suited for. If Graviton 5 is marketed as an "AI chip," they might expect it to perform like one—only to be disappointed when it doesn’t deliver the same kind of performance as a dedicated accelerator like Trainium.
The Pricing Shift
Graviton 5 isn’t just a performance upgrade—it’s also a pricing shift. Unlike previous generations, where newer chips often came with a price drop, Graviton 5 instances are 9% more expensive than their Graviton 4 counterparts. That’s a notable change, especially for customers who’ve come to expect that newer generations would be cheaper.
AWS argues that the price increase is justified by the performance gains. For customers running high-CPU-utilization workloads, the cost per performance might still work out in their favor. But for those with fixed-node setups—like database replicas or multi-AZ deployments—the price hike is just an added cost.
This shift reflects a broader trend in the industry. Component costs are rising, and hyperscalers like AWS are feeling the pinch. The days of "newer = cheaper" might be coming to an end, and customers will need to adjust their expectations accordingly.
The $20 Billion Chip Business
Graviton isn’t just a side project for AWS—it’s a major business. The chip division, led by Annapurna Labs, is now generating over $20 billion in annual revenue, with triple-digit year-over-year growth. That’s a far cry from the early days of Graviton, when it was just another experiment in AWS’s portfolio.
The scale of this business is impressive. AWS now has over 100,000 customers using Graviton-based instances, and some of the biggest names in tech are betting big on the platform. Meta, for example, is deploying tens of millions of Graviton cores for CPU-intensive agentic AI workloads. That’s a vote of confidence in Graviton’s performance and reliability.
But with scale comes responsibility. As Graviton becomes a more central part of AWS’s business, the company needs to be careful about how it markets the chip. Mislabeling it as an "AI chip" might grab headlines, but it risks confusing customers and undermining the credibility of a business that’s now a major player in the semiconductor industry.
Availability and Demand
Despite the hype and the demand, Graviton 5 remains widely available. The Register reports that two companies even tried to buy out AWS’s entire Graviton capacity for the year—but AWS still had plenty to go around. That’s a testament to the scale of AWS’s operations and its ability to meet demand, even in a competitive market.
For customers, this means that upgrading to Graviton 5 isn’t just a theoretical possibility—it’s a practical one. The Register’s author even upgraded their own dev node to Graviton 5 and reported that "It Just Worked." That’s the kind of seamless experience that AWS is known for, and it’s a big part of why Graviton has become such a success.
But availability isn’t just about having enough chips to go around. It’s also about ensuring that customers can get the performance they need when they need it. As Graviton becomes a more central part of AWS’s business, the company will need to continue investing in its supply chain and infrastructure to meet the growing demand.