# Amazon’s Newest Gambit: Selling AI Chips > Source: > Published: 2026-06-19 12:19:36+00:00 Nvidia’s $5 trillion chip empire apparently has a new challenge: Two of its largest customers are planning to sell their in-house AI chips directly to data center companies. In April this year, Google unveiled plans to sell its custom tensor processing units (TPUs) to select customers, and now Amazon is in talks to sell its Trainium AI chips directly to other companies for use in their data centers. In fact, Amazon CEO Andy Jassy signaled this move back in April in a letter to shareholders. “It is quite possible Amazon would sell racks of its chips to third parties,” he wrote. He also noted that the company’s in-house chip business had surpassed a $20 billion annual revenue run rate. Amazon’s semiconductor portfolio comprises Graviton CPUs, Trainium and Inferentia AI accelerators, and Nitro networking cards. Graviton is an Arm-based CPU that provides a low-power alternative to traditional server workhorses from Intel and AMD. Trainium is a purpose-built AI accelerator designed to train and run massive machine learning models. On the other hand, Inferentia chips are custom-designed, purpose-built processors specifically optimized for machine learning inference. Amazon claims Inferentia AI chips are 40% cheaper to run for generating responses from AI models. [View All](https://www.eetimes.com/category/sponsored-content/) There are reports that Uber is among the first external partners to adopt the Trainium3 AI chip. Then, there is Anthropic, which has committed to deploying over one million Trainium chips in its data centers hosted by Amazon Web Services (AWS). It’s worth noting that Amazon has invested $4 billion in this startup, best known for its Claude AI assistant. In return, Anthropic declared AWS as its primary cloud and training partner. But Amazon’s chip business is more of an enigma: A $20 billion semiconductor business no one talks about. It also mirrors a semiconductor business model pioneered by Apple and Google, which developed and deployed custom chips in their hardware at various scales. Below is a sneak peek at Amazon’s semiconductor journey that began in 2015 with the acquisition of Annapurna Labs. **Amazon’s chip trajectory** Amazon’s $350 million acquisition of Annapurna Labs set the foundation for its bespoke silicon solutions. With this in-house semiconductor expertise, the company began developing custom AI silicon around 2019. A year later, Amazon unveiled the Trainium AI accelerator. Trainium1 was a modest chip primarily designed to handle inference workloads within AWS. Trainium2, launched in 2024, aimed to deliver 4× the performance and 3× the memory capacity of its precursor, Trainium1. It incorporated advanced features such as enhanced heat management and fewer internal components, which boosted its compute performance. These improvements were especially instrumental in training AI models. Trainium3, launched in late 2025, claims to deliver up to 4× the performance of its predecessor, Trainium2. It’s a 3-nm AI accelerator designed for training and running large-scale generative AI models in the cloud. The third version of the Trainium chip is largely sold out, according to Amazon AI chief Peter DeSantis. Annapurna Labs in Austin has also helped develop several generations of Graviton; Amazon has recently started shipping this general-purpose processor to Meta. **An Nvidia challenger?** Are Amazon’s chip ambitions a side hustle or a serious challenge to Nvidia? Trade media is abuzz with commentaries that see the sale of custom chips from Amazon and Google as a competitive threat to Nvidia. Especially when Amazon and Google have been developing alternatives to Nvidia’s GPUs. However, it’s important to note that Nvidia is selling a full stack, not just GPUs. So, while custom AI accelerators from Amazon and Google are likely to win in specific workloads and cost optimization, Nvidia will continue to dominate general-purpose AI infrastructure. Nvidia’s edge isn’t merely in silicon; it also lies in CUDA, developer tooling, partnerships, and support. That’s why Amazon is developing the AWS Neuron—its AI model development platform—around Trainium chips. But building hardware, software, and support ecosystems around silicon requires an enormous effort and significant time. And besides Nvidia’s established ecosystem, the bespoke nature of Amazon and Google’s AI chips may also hinder rapid adoption. So, for now, it seems less of a competitive threat to Nvidia and more of a “seize the moment” move to take advantage of an unprecedented demand for AI chips. However, the fact that nobody likes one company—Nvidia—to dominate the entire market could also help the chip ambitions of Amazon and Google along the way. ##### Read also: [AWS Rolls Out AI Inference Chip](https://www.eetimes.com/aws-rolls-out-ai-inference-chip/) [The Trillion-Dollar Race to Fragment the Nvidia Monopoly](https://www.eetimes.com/the-trillion-dollar-race-to-fragment-the-nvidia-monopoly/)