- AI was the catch of the day at AWS re:Invent, with the company announcing new foundation models and Bedrock upgrades
- Silicon was second string but general availability of Trainium2 still generated some noise
- S3 updates were among the sleeper announcements, but are part of AWS' broader AI strategy
AWS RE:INVENT, LAS VEGAS – AI, chips and storage, oh my! We’re totally making a cloud pun when we say the big news from AWS’ flagship conference fell into three big buckets.
Among other things, the company debuted so-called “UltraServers” comprised of its Trainium2 AI chip, S3 advancements that will make it much easier for developers to manage the oodles of data sitting in data lakes and unveiled six brand spankin’ new AI models. But that’s just scratching the surface.
Here’s what you need to know about the major announcements in each category – and what analysts had to say about whether and how they will impact the market.
1) Artificial intelligence overload
AWS went hard on the AI front, really hard. Its six new models include four foundation offerings – Nova Micro, Lite, Pro and Premier – that provide benchmark capabilities that are as good or better than today’s leading options like Meta’s Llama, Google’s Gemini and ChatGPT Mini. Amazon CEO Andy Jassy, who trotted out during a keynote to announce the Nova lineup, claimed these are 75% less expensive than other models available on Amazon’s Bedrock platform.
AWS also came out with Nova Canvas, an image generation model, and Nova Reel for video generation. The latter can currently create clips up to six seconds, but Jassy said it is aiming to provide clips up to two minutes long soon.
But it’s far from done. Looking ahead, Jassy said Amazon is working toward the launch of a speech-to-speech model in Q1 and an “any-to-any” multimodal model in the middle of 2025.
“We’ve going to give you the broadest and best functionality you can find anywhere. And what that’s going to mean is it’s going to mean choice,” Jassy said. “The reality is all of you are going to use different models for different reasons at different times.”
Of course, AWS is aiming for its Bedrock platform to be home to many of these. And it’s also looking to make it easier for enterprises to build the custom models they need. The company unveiled Model Distillation capabilities for Bedrock that streamline the process of creating customized mini models from large scale LLMs.
Fun fact: this announcement earned an enthusiastic “woo” from one of the many many men sitting in the row behind me at the keynote. (So much for women in tech!?) It also got a thumbs up from Gartner VP Analyst Sid Nag, who said he was “struck” by what the tool could mean for enterprises.
“Building industry specific models out of a general purpose frontier model is so hard,” he explained. Making it possible to input a few parameters and have Bedrock spit out a customized mini model could be a game changer, he added.
AWS also rolled out Automated Reasoning checks which will allow Bedrock users to ensure their models aren’t hallucinating and Multi-Agent Collaboration tools which will allow users to create a supervisor agent to oversee a cluster of mini expert agents working on complex problems.
AvidThink founder Roy Chua noted, "The distillation features with teacher-student models, and choice of teacher models will be helpful to those looking to create smaller models with improved performance, lower latency, and that use much less resources. Automated reasoning checks can be useful to improve quality of output, and the multi-agent collaboration looks really interesting." Chua added that as far as telcos are concerned "the Bedrock enhancements can be used by telcos experimenting with GenAI and trying to figure out how to get ROI and lower the cost of inferencing."
All in all, AWS essentially positioned AI capabilities as the next building block – alongside compute, storage and databases – that customers will need, and set out to secure a leadership position on that front.
As SiliconANGLE Principal Analyst Shelly Kramer put it: “In a nutshell, Garman's keynote indicated a return to AWS cloud's roots, focusing on its core, fine-tuning and innovating with building-block services and leaning deeply into its dedication to developers and startups.”
But there were some unanswered questions left by the keynote, namely what the future holds for AWS' Titan line of models, Nag said. Chua similarly said the Nova announcements were a surprise, as he was expecting updates to the Titan lineup.
2) Silicon strides
There was a shorter list of headliners in this category, among them the general availability of the Trainium2 AI chip AWS announced last year, news that it’s working on a Trainium3 chip due out in late 2025 and the debut of Trainium2-based UltraServers. The latter are comprised of four Trainium2 instances. AWS CEO Matt Garman announced that leading AI company Anthropic is building an ultracluster with “hundreds of thousands” of Trainum2 chips.
Oh and Apple also came out of its hermit cave to make a rare appearance on the keynote stage. Senior Director for AI and Machine Learning Benoit Dupin talked up its work with AWS and its use of the cloud giant’s Graviton and Inferentia chips. Apple is also in the “early stages” of evaluating Trainium2, Dupin said.
So, what did analysts think?
Kramer highlighted the Apple cameo, calling it “noteworthy” because “Apple doesn’t really ever do that.”
Similarly, Chua pointed to Apple's presence as a major win for AWS. "Benoit revealed Apple's use of Amazon Graviton and Inferentia to achieve more than 40% efficiency gains in its machine learning inference for search, and also said they are evaluating Trainium 2. Apple usually doesn't reveal their vendors nor endorse them, so that was surprising and a coup for AWS."
Meanwhile, Moor Insights and Strategy VP and Principal Analyst Matt Kimball told Fierce UltraServers are a “good landing zone for enterprise customers training larger models to frontier (models). This feels very much like the approach that others are taking in building finely tuned stacks for training. The partnership with Anthropic is a good billboard that demonstrates Trainium’s ability to support the enterprise.”
He added the biggest challenge AWS faces in the silicon market is that it’s “not NVIDIA, and many enterprise IT organizations have been force fed NVIDIA for the last 18 months or so.” The upside? “This is where a Graviton-like approach will pay off.”
What does that mean?
Well, Kimball explained that like Graviton, AWS can put Trainium out into the market, see where the demand is and “lean in.”
“I don’t see Trainium as competing with NVIDIA Blackwell. Frankly, I don’t think AWS wants Trainium to compete,” he continued. “The market is growing so large so quickly that there’s opportunity for both.”
3) Storage wars
Among the less sexy but immensely useful announcements (I’m not touching the database news – you can’t make me!) made during the keynote were upgrades for Amazon’s S3 storage service. The biggies here were S3 tables, which will enable faster data query performance, and S3 Metadata, which will basically help customers make their metadata searchable at scale. Garman called S3 Tables a “game changer for data lake performance.”
"From a telco perspective, I see the improvements in data handling will help in any telco data modernization efforts — those can be employed immediately," Chua said.
But why did these make the list and not, say, SageMaker Unified Studio or Aurora DSQL? It all comes back to AI.
As Nag explained: “They’re trying to all this find connective tissue between other AWS capabilities and the GenAI strategy. They don’t want these data stores and the S3 type of technology being sort of standalone anymore.”
So, for example, these new S3 offerings could make it much easier for a customer with tons of company-specific data sitting in S3 buckets to leverage that information using Amazon’s new AI tools.
And voila, there’s your full circle moment.