AWS’s Priorities in AI and Beyond

In contrast to Garman’s opening keynote to launch the event, the Q&A session was much more intimate and informal. A small group of industry analysts were invited to ask Garman questions about all things AWS. The session lasted about an hour, and provided a great opportunity to hear his thoughts directly, outside of the spectacle of the keynote.

As you might expect, the analysts’ questions covered a gamut, from infrastructure lifecycles to community concerns about the impacts of datacenters. Consistent across Garman’s answers were big commitments to R&D and infrastructure. He talked specifically about continuing AWS’s huge CapEx for datacenter infrastructure. He talked about AWS partnerships and how they are working to create better AI models or expand satellite networks. He also talked about AWS’s involvement in forward-looking technologies like physical AI and quantum computing. Although much of what he said on that last topic could be summarized as “We’re working on it,” it was still clear that AWS wants to push the envelope in multiple areas.

I particularly enjoyed hearing about AWS’s pivot on multi-cloud. Not too long ago, “multi-cloud” was a dirty word for leaders at AWS — they wouldn’t say the word or allow it to be said at sessions like this one.Thankfully, they’ve changed their course. As I recently said on X in response to AWS’s announcement about its multi-cloud service with Google, I’ve been a long-time supporter of multi-cloud for the simple reason that enterprises want it. When I started talking about this a decade ago, a lot of people thought I was coming out of left field, but it always seemed so clear to me based on my own background as an operating executive and on everything I heard from the cloud providers’ enterprise customers. These days, multi-cloud is much closer to being conventional wisdom. So I’m glad to see that AWS is finally bowing to reality, and that multi-cloud is no longer “that which shall not be named.”

Also worth mentioning is that Garman didn’t seem to want to rule out any potential area for AWS. When asked about hardware or SaaS coming from AWS, he kept those options on the table. “We don’t sell hardware,” he said. “We don’t have plans to, although I wouldn’t rule it out in the future.” His answer for SaaS was similarly open-ended. Perhaps something to look out for some years down the line.

AI Factories and Trainium

Some of the big-ticket items at Reinvent were developments around AWS’s AI Factories and its Trainium chips. Unsurprisingly, these came up in the Q&A, where Garman gave more insight into how AWS is thinking about these new offerings.

AI Factories are a sovereign AI solution — on-prem, at the customer datacenter. This lets customers use the datacenter and hardware they already have, enhanced by AWS cloud services. This way, a customer doesn’t have to invest in a massive AI cluster, and can rely on AWS for that instead. This is potentially a big deal for public-sector customers especially, because they have additional concerns and regulations around privacy and data sovereignty. This is something I’m researching more, because a while back AWS rolled out “sovereignty,” but only of a sort, because it was logical rather than physical. Assuming AWS is billing its AI Factories accurately, this new offering should establish proper physical sovereignty, which is important for a number of reasons — not least to address many of the stiffer data regulations out there.

Some at the Q&A noted the similarity between AWS’s new AI Factories and AWS’s Outposts, something the company has offered for years. When asked about Outpost, Garman made it clear that AI Factories are not a replacement for Outpost. “It’s a doubling down of the Outpost strategy,” he said. “Think about [Outposts] on steroids. That’s kind of the AI Factory.” AWS has been in and out of this sort of thing for years, so I’m interested to learn more about exactly how this new approach differs from Outposts, and to see where this “Outpost on steroids” concept takes us.

Trainium was another item on several people’s minds, myself included. The Trainium product line has been hugely successful for AWS. On the last earnings call, Amazon announced that Trainium2 had been fully subscribed. During the Reinvent Q&A, Garman said that Trainium3 was likewise fully subscribed, albeit with a much lower quantity of chips available at this point. He also mentioned that Trainium3 production is expected to ramp up greatly over the coming year, and he anticipates that all of the chips will continue to be consumed as soon as they are available. (This also matches the broader market picture — high demand guaranteeing full consumption — for NVIDIA’s GPUs, Google’s TPUs, and other leading-edge AI chips for the datacenter.)

Trainium also came up in the context of Anthropic, one of AWS’s major AI partners. More specifically, people wanted to know what AWS’s relationship with Anthropic is like and, by extension, what AWS thinks of its competitors — especially Microsoft — growing closer to Anthropic. Garman highlighted the “fantastic” relationship between AWS and Anthropic, mentioning that all of Anthropic’s models run on Trainium chips. As for Microsoft getting cozier with Anthropic, he seemed unconcerned, pointing out “They don’t have Trainium.” In a nod to Anthropic’s long-time use of Google chips alongside Trainium, he also said, “They don’t have TPUs.” So I got the sense that Garman has plenty of confidence in his own silicon, and in AWS’s position with Anthropic. (He also expressed confidence about AWS’s relationship with OpenAI, which, despite its closeness to Microsoft, uses AWS for a lot of compute and storage.)

Worth mentioning is that no analyst at the session asked about Graviton. Perhaps we were waiting for the update on it that would come later that day, but it does seem a little strange that one of AWS’s key product lines would get only two offhand mentions from Garman over the course of an hour. In the bigger picture, I don’t understand why AWS saved the Graviton5 announcement for last at the conference, because frankly it’s the most differentiated chip AWS has on the market. And that’s not to mention that 55% of all new CPUs in the AWS ecosystem are Gravitons. Seems significant enough to underline more heavily, whether the audience is industry analysts or potential enterprise customers.

AWS’s Attitude and Outlook

The Q&A touched on a lot of topics about AWS and the industry that are inherently uncertain because they extend out into the future, but one thing that is certain is that AWS is extremely confident. Garman was bullish about almost everything he addressed, from Trainium chips to satellite infrastructure. And I don’t think that’s undeserved. AWS is the biggest player in the cloud space and one of the biggest in AI, and it keeps driving ahead with new developments and new chips that the industry is demanding. What we see with Trainium exemplifies this. AWS is fully subscribed on Trainium2 and 3, and projects to land every new Trainium3 chip it can make, while still selling a ton of Trainium2s as well. So the confidence is certainly earned.

That confidence also extends to, or rather against, the competition. It seems clear Garman thinks that the way AWS does things is outright better than its competitors. And he isn’t afraid to say as much: “I’m not gonna do the same thing as our competitors, because I actually think they have a poor strategy. Frankly, I think ours is much better.”

Bullish About AWS’s Future

I always appreciate opportunities like the Reinvent analyst Q&A to engage more directly with CEOs in a small-group setting, and this one definitely offered further insight into how AWS is approaching innovation and execution, and where it sees offerings like AI Factories fitting in. Of course, we will also need to see how its innovations play out in the market, for instance as AI Factories contend with more established offerings from Oracle and others.

In AWS, I see a company that is extremely bullish and looking to push the envelope wherever it can. With developments like the ones announced at Reinvent, it’s clear the company is well suited to do this. I look forward to seeing how the new offerings play out between now and Reinvent 2026.