Amazon has begun development of its third-generation AI chip, Trainium2, at a facility north of Austin, Texas, according to Bloomberg, which interviewed several senior company executives. Trainium2 is a key element of Amazon’s strategy to take greater control over its chip production and reduce reliance on external suppliers such as Nvidia, which currently supplies many of the AI accelerators used in Amazon’s data centers.
For years Amazon has relied on Nvidia chips for AI-powered services, but the company now aims to replace those with its own custom processors, specifically optimized for Amazon’s workloads. This effort is part of a broader push to develop in-house hardware that can deliver greater cost-efficiency and operational flexibility for its cloud and AI services.
“Nvidia is a very, very capable company doing excellent work, so they will have a great solution for many customers for a long time to come,” says James Hamilton, Senior Vice President at Amazon and a central figure in the push to design Amazon’s own chips. Hamilton persuaded Amazon founder and then-CEO Jeff Bezos to invest in developing proprietary chip technology to better meet future demands.
“We believe we can build a component that competes with the best from end to end,” Hamilton adds. That statement underscores Amazon’s confidence that it can challenge established suppliers like Nvidia by designing and producing its own AI accelerators.
Trainium2: A pivotal milestone in Amazon’s plan
Industry experts describe Trainium2 as a potential “make-or-break” moment for Amazon. According to the industry’s informal “three-generation rule,” a third-generation product must deliver substantially better performance and value to justify the investments made in its development. If Trainium2 fails to meet performance and industry expectations, Amazon may need to reassess its strategy and consider alternatives.
“I have literally never seen a product deviate from the three-generation rule,” says Naveen Rao, a chip industry veteran and AI expert at Databricks. Rao closely follows chip development and has noted that Databricks recently agreed to use Trainium2 to support its AI tooling as part of a broader collaboration with Amazon.
This moment is therefore critical for Amazon’s long-term goal of becoming more self-sufficient when it comes to the advanced chips required to train and deploy large-scale AI systems.
Partnership with Databricks and improved performance
In October 2023 Databricks announced an agreement with Amazon to train certain AI models on Trainium2 as part of their ongoing partnership with Amazon Web Services (AWS). While Databricks’ AI tools currently rely largely on Nvidia hardware, the new agreement envisions a phased transition in which parts of the platform will run on Trainium2.
Amazon says Trainium2 offers impressive specifications: the chip is claimed to deliver four times the performance and three times the memory of the previous generation. Those improvements position Trainium2 as a much more capable solution for the enormous computational loads and data volumes needed to train and operate advanced AI models. Amazon expects the chip to strengthen its competitive position against Nvidia and other major chipmakers such as Intel and AMD.
Amazon’s goal extends beyond producing a competitive product; it also aims to make its internal AI infrastructure more efficient and cost-effective over the long term. If Trainium2 succeeds, it could influence how companies design and deploy specialized accelerators for AI and cloud services in the future.
Amazon and the future of AI accelerators
Amazon’s investment in proprietary AI chips aligns with its broader ambition to remain a leader in both cloud computing and AI. Developing in-house accelerators can yield both technical advances and economic benefits by reducing dependence on external suppliers and lowering long-term operating costs.
By advancing Trainium2, Amazon is positioning itself to control more of the hardware stack that powers its services. Success with this generation would validate the company’s strategy of building custom silicon tailored to its specific workloads and could reshape competitive dynamics in the AI hardware market.