Graviton

Every computer contains a processor, the component that actually carries out the computing steps. Even the huge server halls where companies run their websites and programs consist of thousands of such processors. Amazon rents out such computing power through its AWS division: customers pay to have their software run on Amazon’s machines. For a long time, Amazon bought the processors for this from Intel and AMD. Since 2018, the company has also been building them itself, and this in-house design is called Graviton. Several generations now exist, from Graviton1 to Graviton4.

Why Amazon builds its own chips

Computing power is Amazon’s product. Every cent a server costs less per hour either stays as profit or makes the offering cheaper than the competition’s. A proprietary processor saves twice over: Amazon pays no profit margin to a chip manufacturer, and the chip contains only what Amazon actually needs. According to its own figures, Graviton servers deliver up to 40 percent more performance per euro than comparable machines with purchased processors.

The second point is power consumption. Data centers use enormous amounts of energy, and part of it is lost as heat that then has to be cooled away again. Graviton chips are designed for low consumption rather than peak performance on individual tasks. For typical cloud work such as serving websites, that is exactly the right design approach.

For the industry, Graviton is also a signal. When the world’s largest cloud provider builds its own processors, it buys less from Intel. Google and Microsoft have since followed suit and are also developing their own server chips. This is one of the reasons Intel has lost ground in recent years.

ARM instead of x86: what’s behind it technically

Processors understand different instruction sets, i.e. different basic vocabularies of computing commands. Intel and AMD use the x86 instruction set, which originates from the PC world. Graviton instead uses ARM, the instruction set found in almost every smartphone. ARM has been designed for efficiency from the outset, because phones only have a small battery.

Amazon does not design the chips from scratch. The company ARM sells ready-made blueprints for processor cores, and Amazon assembles its own processors from these. They are then manufactured at TSMC in Taiwan, the world’s largest contract chip manufacturer. So Amazon does not own a factory of its own, only the design.

One catch remains: a program compiled for x86 does not simply run on ARM. It must be recompiled for the other instruction set. For modern software written in languages such as Java, Python, or Go, this is usually a minor matter. For old programs whose source code no longer has anyone maintaining it, however, it can be impossible. This is exactly where some migrations fail.

Graviton in bills and quarterly figures

Anyone renting a server on AWS chooses from a list of instance types. Names containing a “g”, such as m7g or c8g, denote Graviton machines. They cost less per hour than the x86 variants. For operators of websites and apps, the choice is therefore purely a matter of cost, provided their own software runs on ARM.

As a private individual, you don’t notice Graviton directly. Nevertheless, some of the services you use every day run on these chips: streaming services, online shops, corporate software. In 2024, Amazon stated that more than half of the newly installed server capacity at AWS runs on Graviton.

One distinction is important: Graviton is a normal main processor, not an AI chip. For training large language models, Amazon uses other components, namely graphics cards from Nvidia or its own chips called Trainium and Inferentia. Graviton handles the everyday work surrounding such models, such as web servers, databases, and control logic. In stock market news, the name usually comes up when it’s about Amazon’s independence from external chip suppliers.