Amazon is building one of the world’s most powerful artificial intelligence supercomputers in collaboration with Anthropic, an OpenAI rival that is working to push the frontier of what is possible with artificial intelligence. When completed, it will be five times larger than the cluster used to build Anthropic’s current most powerful model. Amazon says it expects the supercomputer, which will feature hundreds of thousands of Amazon’s latest AI training chip, Trainium 2, to be the largest reported AI machine in the world when finished.

Matt Garman, the CEO of Amazon Web Services, revealed the supercomputer plans, dubbed project Rainer, at the company’s Re:Invent conference in Las Vegas today, along with a host of other announcements cementing Amazon’s rising dark-horse status in the world of generative AI.

Garman also announced that Tranium 2 will be made generally available in so-called Trn2 UltraServer clusters specialized for training frontier AI. Many companies already use Amazon’s cloud to build and train custom AI models, often in tandem with GPUs from Nvidia. But Garman said that the new AWS clusters are 30 to 40 percent cheaper than those that feature Nvidia’s GPUs.

Amazon is the world’s biggest cloud computing provider, but until recently, it might have been considered a laggard in generative AI compared to rivals like Microsoft and Google. This year, however, the company has poured $8 billion into Anthropic, and it has quietly pushed out a range of tools through an AWS platform called Bedrock to help companies harness and wrangle generative AI.

At Re:Invent, Amazon also showcased its next-generation training chip, Trainium 3, which it says will offer four times the performance of its current chip. It will be available to customers in late 2025.

“The numbers are pretty astounding” for the next-generation chip, says Patrick Moorhead, CEO and chief analyst at Moore Insight & Strategy. Moorhead says that Trainium 3 appears to have received a significant performance boost from an improvement in the so-called interconnect between chips. Interconnects are critical in developing very large AI models, as they enable the rapid transfer of data between chips, a factor AWS seems to have optimized for in its latest designs.

Nvidia may remain the dominant player in AI training for a while, Moorehead says, but it will face increasing competition in the next few years. Amazon’s innovation “shows that Nvidia is not the only game in town for training,” he says.

Garman told WIRED ahead of the event that Amazon will also introduce a range of tools to help customers wrangle generative AI models that he says are often too expensive, unreliable, and unpredictable.

These include a way to boost the capabilities of smaller models using larger ones, a system for managing hundreds of different AI agents, and a tool that provides proof that a chatbot’s output is correct. Amazon builds its own AI models, for recommending products on its ecommerce platform and other tasks, but it primarily serves as a platform to help other firms build their own AI programs.

While Amazon does not have a ChatGPT-type product to advertise its AI capabilities, the scope of its cloud services will give it an advantage selling generative AI to others, says Steven Dickens, CEO and principal analyst at HyperFRAME Research. “The breadth of AWS—that’s going to be an interesting thing,” he says.

Amazon’s own line of chips will help it make the AI software it sells more affordable. “Silicon is going to have to be a key part of the strategy of any hyperscaler going forward,” says Dickens, referring to cloud providers that offer hardware for building the very largest, most capable AI. He also notes that Amazon has been developing its custom silicon for longer than competitors.

Garman says a growing number of AWS customers are now moving on from demos to building commercially viable products and services incorporating generative AI. “One of the things that we’re quite excited about is having customers move from having their AI experiments and proof of concepts,” he told WIRED.

Garman says that many customers are far less interested in pushing the frontier of generative AI than in finding ways to make the technology cheaper and more reliable.

A newly announced AWS service called Model Distillation, for instance, can produce a smaller model that is faster and less expensive to run while still having similar capabilities to a larger one. “Let’s say you’re an insurance company,” Garman says. “You can take a whole set of questions, feed those into a really advanced model, and then use that to train the smaller model to be an expert on those things.”

Another new cloud tool announced today, Bedrock Agents, can be used to create and manage so-called AI agents that automate useful tasks such as customer support, order processing, and analytics. It includes a master agent that will manage a team of AI underlings, providing reports on how they function and coordinating changes. “You can basically go create an agent that says you’re the boss of all the other agents,” Garman says.

Garman says he expects that companies will be particularly excited about Amazon’s new tool for ensuring that a chatbot’s outputs are accurate. Large language models are prone to hallucinating, and existing methods for keeping them on track are imperfect. Customers such as insurers, which cannot afford to make errors with their AI model, are clamoring for this kind of safeguard, Garman told WIRED. “When you ask, ‘Is this covered by my insurance?’ you don’t want the model to say no when it is or yes when it’s not,” Garman says.

Amazon’s new verification tool, called Automated Reasoning, is different from a similar product that OpenAI announced earlier this year. It relies on logical reasoning to parse a model’s output. In order for it to work, a company needs to turn its data and policies into a format that allows for logical analysis. “We take the natural language, we translate it into logic, we prove or disprove the statement, and then we can provide an argument as to why the statement is true or not,” Bryon Cook, a distinguished scientist at AWS and vice president of the company’s Autonomous Reasoning Group, told WIRED.

Cook says the same kind of formal reasoning has been used for decades in areas like chip design and cryptography. He adds that the approach could be used to build chatbots that handle airline ticket refunds or that provide human resources information without getting facts wrong.

Cook adds that companies can combine multiple systems that feature Automated Reasoning to build more sophisticated applications and services, including ones that incorporate autonomous agents. “Now you have communicating agents that are doing formal reasoning and communicating their rationale,” he says. “Reasoning will become a very important thing.”