7 Things Matt Garman Announced AWS Is Focusing On
Analytics India Magazine (Mohit Pandey)
AWS used re:Invent 2025 to signal that the next phase of its AI strategy will be built on speed, scale and a new layer of agent-based computing.
The company launched its Trainium3 UltraServers with huge jumps in performance and energy gains, shared first details of Trainium4, expanded Bedrock into the largest neutral model hub, and pushed deeper into agent tooling across policy, evaluation and autonomous workflows.
AWS said it has already deployed more than 1 million Trainium chips, and the new stack is meant to cut costs, reduce latency, and support systems that run on massive parallel agents.
Matt Garman’s keynote outlined where AWS sees the world going and what it wants to build for it.
A Deeper Push into NVIDIA and Large Scale AI Training
Garman opened by framing the NVIDIA partnership as central to AWS. He said AWS and NVIDIA have worked together for more than 15 years and that nothing about the collaboration is accidental.
“Nothing’s too small for us to really work together to make sure that we have the most reliable performance.” He added that NVIDIA itself trains its largest systems on AWS, calling it “a testament to work together.”
The message was that AWS wants to be the most stable home for frontier model training, with NVIDIA hardware blending into AWS silicon and networking.
AI Factories for Customers that Want Hyperscale Training Inside Their Own Walls
Garman said many governments and large enterprises have the data centre footprint but lack the know-how to run giant AI clusters. He said the idea came from working with players like OpenAI, Humane and the Saudi AI initiative.
“Why can’t we help more customers, the ones who really need this large-scale infrastructure, see what our expertise, our services, are understanding?” he said.
AI factories let AWS place its infrastructure, software and governance controls inside a customer environment while meeting rules on sovereignty and policy. AWS wants to make hyperscale AI feel like owned infrastructure rather than a remote service.
The Next Generation of AWS Silicon with Trainium3 and Trainium4
Garman previewed Trainium4 while Trainium3 went live. He said Trainium4 delivers “over 6x the FP, 4x performance, 4x more memory family and 2x more high bandwidth memory capacity” and doubles power efficiency compared to the earlier generation.
He also showed how fast inference now resembles training loads. “There’s not going to be an experienced application of a system built that doesn’t rely on inference.” AWS is positioning its chips as the backbone of low-cost training, low-latency inference and huge agent workloads.
Bedrock Becoming the World’s Largest Mix and Match Model Platform
Bedrock now has more than 100,000 customers. Garman said it has doubled the number of models in a year and will add 18 more new open weight models, including Google, Minimax, Mistral, NVIDIA and OpenAI GPT OSS.
He said customers are increasingly running many models at once. “This mix and match is going to be normal.” AWS also refreshed its own Nova family. Nova Light is for cost-efficient reasoning. Nova Pro targets complex reasoning across documents and video. Nova Sonic adds multilingual speech-to-speech.
The unified Nova multimodal model handles text, images, video and speech as inputs. Garman said this solves a real need for creative teams who want one model that “can output different forms of text and imagery” without juggling multiple systems.
Nova Forge, a New Way for Companies to Build Their Own Frontier Model
This was one of Garman’s biggest claims. Customers want their models to reflect their own language and systems, but fine-tuning breaks when pushed too far. Garman said the team asked a simple question. “Why not make that possible? Why can’t that be true?”
Nova Forge gives access to Nova checkpoints and lets customers blend their own data with Amazon-curated sets, then deploy the resulting frontier model on Bedrock with full guardrails. This lets enterprises create models that act like internal experts rather than generic assistants.
The Full Stack for Building, Governing and Monitoring Billions of Agents
Garman said the world is entering a time “where there were literally billions of agents working together.” AWS wants to make those agents safe, fast and easy to build. Bedrock Agent Core brings building blocks like memory, gateway and identity. New upgrades include policy and evaluations.
With policy, he said, customers can set rules in simple language. “We do the hard work to translate it into policy code.” Evaluations monitor an agent’s behaviour in the real world and raise alerts when quality slips. “
You’re going to get an alert that says the agent review isn’t acting as it should.” AWS sees this as the missing layer for running agent systems at scale.
Frontier Agents, AWS’s Next Step in Autonomous Software
Garman said the company learned from internal use of Hero that teams were still treating agents like simple assistants. AWS then changed the design. Agents should be autonomous, handle long tasks, work across hundreds of parallel actions, and improve without human babysitting.
“I don’t have to overwork, I don’t have to babysit.” The result is Frontier agents.
The Hero autonomous agent keeps persistent context, pulls requests, improves code and learns how a team works. The security agent embeds a security expert in every step of development and can run pen tests on demand. The DevOps agent handles incident triage and recovery.
Garman said it offers “fewer alerts” and faster recovery across multi-cloud and hybrid setups. AWS wants these agents to give step-change productivity, not small gains.
The post 7 Things Matt Garman Announced AWS Is Focusing On appeared first on Analytics India Magazine.
Generated by RSStT. The copyright belongs to the original author.