Amazon Developing Custom AI Chips to Cut NVIDIA Dependence

Amazon Plans to develop its own custom AI chips in order to reduce reliance on NVIDIA. The company has previously created some custom processors for data center workloads, and this effort expands on its 2015 purchase of chip design startup Cerebras. In particular, Amazon is expected to share more information on these custom AI processors as it relates to its Trainium chip family next month.

Amazon Developing Custom AI Chips to Reduce Dependence on NVIDIA

Amazon’s Annapurna Labs, which has provided chips for AI company Anthropic (a Microsoft-backed Oracle-partnered startup and one of Amazon’s top AI partners) is designing these new AI chips. Access to its foundational AI model, Claude, for Amazon, is provided by Anthropic. It is one part of a larger industry push by big tech to lessen their reliance on NVIDIA’s high-performance, and most expensive, GPUs for AI workloads as demand outstrips supply.

This is an effort by Amazon to reduce costs and become more self-reliant from NVIDIA. Through the acquisition of Annapurna Labs, it has experience designing custom chips which lowers costs by reducing dependency on AMD and Intel products for typical enterprise compute workloads. These include the processors from Amazon itself, such as Graviton chips and Trainium chips designed for large language models. Amazon announced its Trainium2 in November 2023 but demand was outstripped by supply chain constraints.

Amazon’s custom circuitry for its chips is based on licensing agreements with Alchip and powered through its silicon partner Taiwan Semiconductor Manufacturing Company (TSMC). And over 50,000 AWS customers are using Amazon’s Graviton chips. Determined to lower their dependence on NVIDIA chips, other tech giants such as Google, Meta, and Apple are working towards developing in-house AI chips. Google has announced the introduction of its new AI Chip Trillium, boosting higher performance in training as well as inference levels.

FAQs