Bedrock Brief 09 Sep 2026

Amazon's AWS just turned a few industry heads by inking a multi-generation silicon deal with Qualcomm, complete with $4 billion in stock warrants and a roadmap that stretches into 2028 and beyond. This isn't just another procurement win for Qualcomm. It's a public signal that AWS is diversifying its custom-chip strategy beyond its own Trainium and Inferentia accelerators, and pulling in a new player to tackle inference workloads at scale. The collaboration spans both compute and optical connectivity up to 1.6T, which means Qualcomm gets a shot at selling silicon and networking fabric into AWS data centers while Amazon keeps feeding its $496 billion backlog. Qualcomm's stock jumped 5% on the news, and the deal handed the chipmaker a marquee customer credential in a market where Nvidia has been eating most of the pie.

Why does this matter? Because AWS is now the loudest buyer in AI infrastructure. The cloud giant posted $42.2 billion in AWS revenue last quarter (up 36.7% year-over-year) and burned through $53.1 billion in capex in three months. CEO Andy Jassy said AWS could eventually become a trillion-dollar annual revenue business, and most AI capacity is already locked up in five-year contracts. That kind of demand pressure doesn't just validate Qualcomm's $15 billion data center revenue target for 2029. It also puts pressure on Broadcom, the incumbent in custom AI accelerators and hyperscaler networking, whose shares ticked up 3% on Tuesday as investors tried to gauge read-through. The fine print will be whether Qualcomm can actually convert this credential into meaningful revenue, or whether the deal mostly serves as a hedge for Amazon while Broadcom keeps the bulk of the business.

Meanwhile, in less headline-grabbing but still useful news, Ramp announced it's now available in AWS Marketplace with AI agents that code expenses and process invoices around the clock. The fintech company is pitching AI-powered finance operations for the "agentic era," including token spend management that gives teams visibility into LLM and API costs before they spiral out of control. It's a reminder that not every AWS AI story this week involves billions in chipmaker warrants. Sometimes it's just about letting a bot handle your expense reports so your finance team can go home on time.

Fresh Cut

• OpenAI's GPT-6 Astra can process up to 1 million tokens of input at once through AWS Bedrock, letting you build AI agents that analyze massive documents or codebases that would overwhelm previous models. Read announcement →
• You can now update individual columns in your machine learning feature database instead of rewriting entire rows, which cuts latency and lets multiple data pipelines modify the same record without overwriting each other's work. Read announcement →
• AWS Bedrock AgentCore Memory lets you skip storing chat logs as short-term events and feed content directly into long-term memory using a new IngestData API, useful if you only need persistent storage without the temporary event history. Read announcement →
• CloudFront's image transformation tool can now detect and preserve custom objects like logos during smart cropping by combining face, text, and label detection in a single request, plus it automatically resizes images for devices that don't support Client Hints (about 30% of traffic). Read announcement →
• You can protect your free AWS Builder ID account with multi-factor authentication even when signing in through Google, Apple, or GitHub, and recover access yourself using a backup email instead of contacting support. Read announcement →
• An open-source Nx plugin generates full-stack application code with built-in AWS infrastructure (CDK or Terraform), security rules, logging, and type-safe connections between your APIs and frontends so breaking changes fail at build time instead of in production. Read announcement →
• You can connect SharePoint, OneDrive, and Confluence to Amazon Bedrock's knowledge bases by logging in with your own credentials instead of needing IT admins to create service accounts, cutting setup from hours to minutes. Read announcement →
• You can build AI chatbots that answer questions from your ServiceNow knowledge base without writing custom code to sync the data, just connect with credentials and it automatically crawls articles and attachments. Read announcement →
• AI coding assistants like Claude can automatically diagnose your broken Lambda functions by checking error patterns across the last 7 days and analyzing all connected AWS services in one go instead of making you check each service separately. Read announcement →
• AWS's Graviton4 processors deliver 30% better performance than the previous generation and are now available in Taipei, New Zealand, and US GovCloud regions for compute-heavy tasks like batch processing, video encoding, and ML inference. Read announcement →

The Quarry

Build a Physical AI model factory with NVIDIA Cosmos 3 on SageMaker HyperPod

NVIDIA's Cosmos 3 on SageMaker HyperPod turns Physical AI development into a persistent model factory where synthetic data generation, post-training, and closed-loop evaluation run continuously rather than as one-off jobs. The setup uses a resilient EKS-based HyperPod cluster that tracks GPU goodput (actual productive compute time, not just uptime) as the north star metric, which matters because your expensive H100s sitting idle during job transitions or failures is money literally evaporating into the cloud. This architecture treats Physical AI like the iterative manufacturing process it actually is, where your robot or autonomous system gets smarter through repeated cycles of generating synthetic training scenarios, fine-tuning models, and testing them against real-world edge cases. Read blog →

More posts:

• Take on your most ambitious work with GPT-6 Astra on Amazon Bedrock
• Pathway’s brain-inspired architecture development on Amazon SageMaker HyperPod
• Amazon SageMaker Feature Store introduces UpdateRecord for feature-level writes
• Govern models with MLflow and Amazon SageMaker AI Model Registry sync: Part 2
• Govern models with MLflow and Amazon SageMaker AI Model Registry sync: Part 1

Core Sample

In The Field – We go inside TBC.co's lab to see the future of AI efficiency

TBC.co is building custom silicon called Monolith that promises to deliver 160 TOPS (tera-operations per second) of AI inference performance while sipping just 5 watts of power, a power efficiency ratio that would make AWS Inferentia chips jealous. Their secret sauce involves processing everything at the edge with a hardware architecture optimized for transformer models, eliminating the latency tax of round-tripping to the cloud for every AI decision. The pitch is simple: if you can run sophisticated AI locally on battery-powered devices without melting them, you unlock entirely new categories of always-on intelligence that aren't feasible when you're tethered to a data center. Watch video →