Bedrock Brief 04 Nov 2026

Pleri AI

Pleri AI

— 4 min read
Share
Bedrock Brief 04 Nov 2026

This week brought us a tale of two very different AWS stories: one about the government finally moving at "the speed of business" (their words, not mine), and another about Amazon burning through cash faster than a GPU cluster training foundation models. Let's start with the Army's Project Griffin, where AWS and World Wide Technology just landed a $2.6 million pilot to build AI-powered cyber defense systems. The goal? Deploy "agentic AI" that can detect and counter threats faster than human operators, because apparently even the Pentagon is tired of waiting for security analysts to click through alerts. Bonus points for the Army now calling AI "super intelligence" to match the Trump administration's rebranding efforts, though I'm not sure anyone asked for that acronym change.

Meanwhile, Amazon just posted a jaw-dropping $42 billion free cash flow burn over the trailing 12 months, nearly double its previous record from the great warehouse overbuild of 2022. The culprit? A relentless AI infrastructure buildout that saw AWS capital spending more than double year-over-year to $77 billion in Q3 alone. Before you panic, the strategy seems to be paying off: AWS revenue grew 39% (its fastest clip in nearly five years), and CEO Andy Jassy is projecting over 40% growth through 2028. Anthropic's compute commitments alone represent about a fifth of AWS's backlog, which tells you just how hungry these AI labs are for inference capacity.

The broader picture here is fascinating. On one hand, you've got government buyers finally waking up to the potential of AI agents in real-world defense scenarios, even if the initial contract value barely covers a rounding error in AWS's infrastructure budget. On the other, you've got Amazon betting the farm (and borrowing heavily to do it) that the AI wave will justify record-breaking capital outlays. Bond traders are getting nervous, credit default swap prices are climbing, and Wall Street punished the stock after earnings missed expectations. But Jassy's doubling down, promising to "spend meaningfully more in 2027." Either this is prescient positioning for an AI-driven future, or it's the most expensive game of infrastructure chicken in tech history.

Fresh Cut

  • AWS Config can now track configuration changes for ElastiCache clusters and replication groups, plus 8 other resource types like QuickSight agents and billing views, so you can audit what changed in your infrastructure without manually checking each service. Read announcement →
  • Claude Haiku 5.5, the fastest model in Anthropic's Claude family with a 1M context window and adaptive thinking across five effort levels, is now available in AWS GovCloud regions through the Kiro IDE and CLI for developers who need to process high volumes of text quickly. Read announcement →
  • Amazon QuickSight's desktop app can now use the same login system as the web version instead of requiring admins to set up a separate OAuth application for desktop access. Read announcement →
  • Amazon Connect can pull data from third-party apps like CRMs or banking platforms directly in customer service flows using Model Context Protocol tools, eliminating the need to write Lambda functions for each integration. Read announcement →
  • Amazon Connect's Wait block can now pause customer service flows until a case field updates or an external tool finishes, letting you hold interactions until you have the data you need instead of transferring customers around unnecessarily. Read announcement →
  • You can use Amazon Textract's machine learning OCR to extract text and structured data from scanned documents in Tokyo and Zurich regions, though expense analysis and custom queries aren't available there yet. Read announcement →
  • Amazon Textract's document text detection model can now read documents in over 40 languages (including Arabic, Chinese, and Hebrew) and better handles handwriting, long words, and messed-up tables without you changing any code. Read announcement →
  • AWS can automatically route your OpenAI GPT model requests between two government cloud regions to handle traffic spikes while keeping your data in the US, and you'll see the usage logs and costs in the same place as your other AWS services. Read announcement →
  • OpenAI's GPT-5.6 Cyber model is designed specifically for security researchers to test exploits and find vulnerabilities with fewer AI safety refusals, available through AWS Bedrock with chip-level privacy guarantees that prevent your security research data from being used for training. Read announcement →
  • Amazon Connect lets you create immutable versions of customer service flows and use aliases to switch all phone numbers between versions with a single update instead of manually relinking each one. Read announcement →

The Quarry

Build efficient agents with knowledge distillation on Amazon SageMaker AI

Amazon's latest post tackles the wallet-draining reality of running agentic AI workflows where every tool-calling step hits your expensive frontier model, and they're pitching sequence-level knowledge distillation (SeqKD) as the escape hatch. Instead of distilling at the token level, SeqKD captures the teacher model's entire multi-step reasoning trajectory (the full sequence of tool calls and responses) and teaches a smaller student model to mimic that end-to-end behavior on SageMaker. The practical upshot: you can shrink your per-request costs dramatically while keeping your agent's decision-making quality intact, which matters when you're burning tokens on planning, tool selection, and result synthesis across dozens of interactions per user session. Read blog →

More posts:

Core Sample

GenerAIly Speaking: Let's build a research agent from scratch

Jay and Bharathi speedrun building a research AI agent in under an hour using Amazon Bedrock's AgentCore CLI, showing how to wire up Gateway tools for external data access, plug in Memory for conversation context, and orchestrate everything through the Strands Agents framework. The walkthrough covers the practical nitty-gritty of agent personalization and deployment, giving you a working blueprint rather than just architectural hand-waving. If you've been curious about how Bedrock's agent tooling actually fits together beyond the marketing slides, this is your crash course in making the pieces talk to each other. Watch video →

More videos:

Read more