Amazon.com Inc.’s cloud division, Amazon Web Services, has accelerated the release of its newest artificial intelligence chip, Trainium3, renewing the company’s push to compete more directly with hardware from Nvidia Corp. and Alphabet Inc.’s Google. The accelerator was recently deployed in a limited number of data centers and will become available for customers beginning Tuesday, according to Dave Brown, vice president at Amazon Web Services.
Brown said AWS plans to move quickly in expanding availability, with rapid scale-out expected early next year. The launch marks an unusually fast development cycle by semiconductor industry standards, arriving roughly one year after the prior Trainium generation. Brown noted that the company is aiming to offer strong price-performance advantages over Nvidia’s graphics processing units, which are widely considered the market standard for AI model training and deployment. He added that Trainium hardware continues to show efficiency benefits in internal benchmarks.
The aggressive rollout aligns with AWS’s broader strategy to strengthen its foothold in AI computing. While AWS remains the dominant global provider of rented computing power and cloud infrastructure, it has struggled to achieve the same leadership position with developers building frontier AI models. Many startups and enterprises have opted for Microsoft Corp., which maintains deep ties to OpenAI, or Google, whose Tensor Processing Units are used widely among advanced model developers.
Despite Amazon’s claims about Trainium performance, the company still faces an entrenched challenge: Nvidia’s extensive and mature software ecosystem. Bedrock Robotics, a company that builds autonomous capabilities for construction equipment, runs its infrastructure on AWS but relies on Nvidia chips for model development. Kevin Peterson, the company’s chief technology officer, said Nvidia continues to provide the fastest and easiest path to high-quality results, highlighting the gap Amazon must overcome.
A significant share of Trainium chips in operation today is dedicated to Anthropic, with AWS deploying more than 500,000 units across data centers in Indiana, Mississippi and Pennsylvania. AWS said earlier this year that it plans to allocate up to 1 million Trainium chips to Anthropic by the end of the year. The startup also maintains partnerships with both Google and AWS, securing access to large volumes of computing power from each provider. Analysts say the lack of other major Trainium customers makes it hard to evaluate the technology’s competitiveness.
Amazon unveiled the Trainium3 launch at re:Invent, its flagship cloud customer conference, which has increasingly centered around AI products and services. As part of the event, the company also introduced updates to its in-house AI model family, known as Nova. The latest Nova 2 generation includes a multimodal variant called Omni, which can process text, images, speech and video, and produce both text and image outputs. Although earlier Nova models have not ranked among industry leaders in standardized benchmarks, Rohit Prasad, head of Amazon’s Artificial General Intelligence efforts, said real-world performance is what ultimately matters.
In addition, Amazon announced Nova Forge, a new platform allowing advanced customers to customize partially trained versions of Amazon’s models using proprietary data. Reddit Inc. is currently using Nova Forge to develop a model that evaluates whether user posts violate the platform’s safety policies. Chris Slowe, Reddit’s chief technology officer, said the ability to tailor a model to a specific domain delivers more value than relying solely on the most advanced general-purpose system.
