Amazon Web Services custom AI chip Trainium2 is fully subscribed and Trainium3 will feature about a 40% performance boost and the ability to reach more customers, said Amazon CEO Andy Jassy.

Jassy used a strong third quarter earnings report and AWS launch of Project Rainier, a cluster of up to 1 million Trainium2 processors used to train Anthropic's next Claude model, to lay out a few big themes you'll hear from AWS re:Invent 2025 in less than a month.

Leading up to Amazon's quarter, the narrative was that AWS is trailing in AI behind Microsoft Azure and Google Cloud. The latter features its own TPU custom AI chips that have also landed Anthropic as a customer. Note that the AWS is trailing in AI narrative played out last year leading up to re:Invent.

With that backdrop, Jassy made it clear that its custom AI chips are going to broaden the market for enterprise AI and give them more price-to-performance options. AWS also sees Amazon Bedrock with Trainium3 as the lead inference engine and a business that'll be as big as EC2.

Jassy also said the tipping point for more AWS customers will be AI agents that are secure. Like Google Cloud and its integrated AI stack, AWS has its own version that'll feature Amazon SageMaker and Amazon Bedrock along with AgentCore running on Trainium. However, Jassy was quick to note that AWS buys a lot of Nvidia, AMD and Intel too.

This AI integrated stack, Trainium3 and building of secure AI agents will be the core themes of re:Invent, which appears to be a little more speeds and feeds than recent years. Here's a look.

Trainium scales

Now that Project Rainier is operational, Jassy has a lot more to talk about when it comes to AI infrastructure, Trainium prospects and capacity.

He said AWS has added more than 3.8 gigawatts of power in the past 12 months, double capacity of AWS in 2022. AWS will double again by 2027 and plans to add at least another 1 gigawatt of power in the fourth quarter.

Jassy added that additional capacity includes, power, data centers and chips, mostly Trainium and Nvidia.

The Anthropic effort also gives AWS a flagship customer where it can expand.

"Trainium2 continues to see strong adoption, is fully subscribed is now a multibillion-dollar business that grew 150% quarter-over-quarter," said Jassy. "Today, Trainium is being used by a small number of very large customers but we expect to accommodate more customers starting with Trainium3."

Jassy said AWS is monetizing as soon as capacity comes online. Amazon has spent nearly $90 billion on capital expenditures in 2025 and most of it relates to AWS infrastructure and Trainium. Some of that sum, expected to hit $125 billion for the full year, goes to fulfilment and transportation for commerce.

Trainium2 has a few very large customers, noted Jassy, and Trainium3 will broaden the base. "As customers start to contemplate broader scale of their production workloads, moving to being AI-focused and using inference, they badly care about price performance," said Jassy. "We have a lot of demand for Trainium. Trainium3 should preview at the end of this year with much fuller volumes coming in the beginning of '26, we have a lot of customers, both very large, and medium-sized who are quite interested in Trainium3."

Amazon's custom chips are made by Annapurna Labs, a chip designer acquired in 2015 for $350 million. Along with 2012's $750 million purchase of Kiva, a robotics company that enabled Amazon to automate its warehouses, Annapurna has to be the company's smartest acquisition.

AI agents: Secure, horizontal

"We're bringing the same building block approach to AI. SageMaker makes it much simpler for companies to build and deploy their own foundation models. Bedrock gives customers leading selection of foundation models and superior price performance to deploy inference into their next-generation applications," said Jassy. "A lot of the future value companies will get from AI will be in the form of agents. AWS is heavily investing in this area and well positioned to be a leader."

Jassy said enterprises will both create agents and use agents from other companies. AWS will be for the builders that can leverage platforms like Strands, Kiro, Transform and AgentCore. Behind the scenes, AWS will be a key player for the ecosystem offering packaged agents.

While the focus is on builders, AWS is also targeting business users with efforts like Amazon QuickSuite.

"I think that the number of companies who are working on building agents is very significant. I do believe that a lot of the value that companies will realize over time and AI will come from agents," said Jassy. "When you talk to enterprises or companies that care a lot about security and scale. They're starting to build agents, and they don't really feel like they've had building blocks that allow them to have the type of secure, scalable agents that they need to bet their businesses and their customer experience and their data on."

Efforts like AgentCore are designed to be a primitive building block for AI agents, akin to compute, storage and database.

Inference everywhere

The strategy for AWS revolves around playing for inference, which will be a much larger category than training. Jassy said:

"We're building Bedrock to be the biggest inference engine in the world and in the long run, believe Bedrock could be as big a business for AWS as EC2, and the majority of token usage in Amazon Bedrock is already running on Trainium. We're also continuing to work closely with chip partners like NVIDIA, with whom we continue to order very significant amounts as well as with AMD and Intel. These are very important partners with whom we expect to keep growing our relationships over time."

Simply put, a proliferation of AI agents is going to require a lot of inference. AWS wants the bulk of those workloads.

And enterprises are increasingly going to look for the best performance at the right price. That reality is why Jassy is so bullish on its custom silicon, which includes Graviton as well as Trainium.

"For our customers to be able to use AI as expansively as they want. Remember, it's still relatively early days at this point. Customers are going to need better price performance and they care about it deeply," said Jassy.

Constellation Research's take

Holger Mueller, analyst at Constellation Research, said:

"It's good to see AWS getting into the supercomputer business. Remarkably, AWS is doing this with its in-house chips, which are both proof points for the R&D chops of Amazon and frugality when it comes to infrastructure spending. The frugality makes for competitive AI spending that CxOs welcome. AWS is on a hardware level with Google Cloud when it comes to running custom silicon.

AWS and Google Cloud are the only cloud players running custom silicon at scale. Why? They have unique scalability and cost performance needs in their core business. Amazon has razor thin retail margins. Google has its freemium offerings. Where Google is ahead is putting custom algorithms on custom hardware (TensorFlow on TPUs). Amazon is still recovering from missing out on the initial AI wave in that sense. It's questionable AWS would have come out with something like TensorFlow - as it usually is an adapter of what is being used. That's Amazon's retail DNA."

Research:

 

Business Research Themes

Data to Decisions, Future of Work, New C-Suite, Next-Generation Customer Experience

Roles

Chief Information Officer

Tags

Amazon, Artificial Intelligence (AI), agentic AI, hyperscale data centers