Annapurna Labs

The Haifa venture that became Amazon's silicon arm

Annapurna Labs was founded in May 2011 in Yokneam Illit by three former Marvell Technology engineers: Hrvoje Bilic, Nafea Bshara and Ronen Boneh, with seed investment from Avigdor Willenz. In 2016 the company moved to the Matam park in Haifa. In January 2015, less than four years after it was founded, Amazon acquired it for about $370 million. According to an official AWS blog post from 2018, this was the starting point of Amazon's entire custom-silicon project: "We acquired Annapurna Labs in 2015 after working with them on the first version of the AWS Nitro System."

From Nitro to Graviton: how Amazon's first chip was born

According to the same official post, the team built and released, together with Amazon, two generations of Nitro chips that move all the system functions of EC2 servers off the main processor and onto dedicated hardware. Only after several years of working together did the team begin to think about a custom CPU of its own. The result, AWS Graviton, was announced in November 2018 and is based on the ARM architecture. Its fourth generation was announced in November 2023.

Trainium and Inferentia: the silicon that trains and runs models

Alongside Graviton, Annapurna Labs developed two lines of AI-focused chips. Inferentia, announced in December 2018, is designed for inference: running a trained model on new input. Trainium is designed for training large language models, the stage that consumes most of the computing power. The second generation of Trainium was announced in November 2023, and according to English Wikipedia it was used, among other things, to train Anthropic's Claude models.

Why a cloud provider builds its own chips

A cloud provider that wants to avoid total dependence on Nvidia, in both price and availability, needs silicon of its own. Amazon's bet proved successful. Annapurna Labs is described today as one of the largest "fabless" chip companies in the world (a company that designs chips but outsources the manufacturing itself to an outside maker), and it is one of the five largest customers of the Taiwanese chipmaker TSMC.

A company almost no one knows, despite its impact

As of 2026, the company is almost unknown to the general public, even in Israel, although it is one of the most influential Israeli companies in AI hardware. This is because its products are not sold to end customers. They are used only in the data centers of its parent company. When a large language model is trained or run in Amazon's cloud, the computation often takes place on a chip that was designed in Haifa.