Bedrock Brief 28 Oct 2026

Pleri AI

Pleri AI

— 4 min read
Share
Bedrock Brief 28 Oct 2026

Verizon just dropped $2 billion on an AWS infrastructure deal, and if you squint, you can see the future of AI infrastructure taking shape. It's not sexy, but fiber optic cables and data center connectivity are now the hot commodity everyone's chasing. Verizon's CEO is practically giddy about signing deals with hyperscalers, mentioning an "active pipeline" of more partners lining up for dark and lit fiber. Meanwhile, AWS is taking a victory lap in Germany with a €4.8 billion investment to beef up its Frankfurt region, because apparently Europe still wants more cloud capacity despite all the sovereignty hand-wringing. The lesson? AI doesn't just need fancy chips and clever algorithms. It needs boring, reliable pipes to move all that data around, and telecoms are finally having their moment.

AWS growth is supposedly going to hit 40% year over year, according to Bank of America, fueled by AI lab contracts and inference workloads that are making Bedrock and Trainium chips look like actual revenue drivers instead of science projects. That's the kind of number that makes CFOs sit up and pay attention, especially when enterprise customers are moving beyond the "let's kick the tires on generative AI" phase into "we need this running at scale yesterday." The catch? Analysts are whispering that expectations might balloon to 42% or higher, which means even a strong quarter could feel like a letdown. Nothing like setting yourself up for disappointment by being too successful.

On the ground, the AI pitch competition scene is alive and well in Little Rock, where a restaurant phone AI agent just won $50,000 in AWS services. It's a reminder that while hyperscalers and telecoms are playing big money games, the real adoption story is still about solving unglamorous problems like answering phones during dinner rush. High operating costs and razor-thin margins make restaurants a perfect AI testing ground, and if the tech works there, it'll work almost anywhere. The gap between billion-dollar fiber deals and a scrappy startup trying to help a pizzeria is where the actual innovation happens, even if it doesn't make the front page of CNBC.

Fresh Cut

  • MLflow 3.16 gives you a customizable trace explorer with reorderable columns and adjustable rows to debug your AI applications, plus faster loading of metric charts when you're tracking thousands of experiment results. Read announcement →
  • Aurora serverless databases can now scale up to 768 GB of memory (384 ACUs at 2 GB each) before automatically scaling back down to zero, which matters if you're building apps with unpredictable traffic spikes and want to avoid paying for idle database capacity. Read announcement →
  • You can flip ElastiCache clusters between two persistence modes—synchronous writes that guarantee zero data loss across availability zones with single-digit millisecond latency, or asynchronous writes that keep microsecond latency but risk losing up to 10 seconds of data during failures—turning your cache into a durable database when you need it. Read announcement →
  • OpenAI's GPT-6.1 Terra is available through AWS Bedrock with a 1 million token context window, costs 50% less than competing models, and runs three times faster than GPT-6 Astra while scoring nearly as well on complex agent tasks. Read announcement →
  • Amazon Connect's testing tool can now check the actual data returned by individual flow actions like Lambda functions or chatbot intents, letting you spot which specific step failed instead of debugging backwards through your entire workflow. Read announcement →
  • Amazon Bedrock's Web Search tool lets you whitelist or blacklist up to 100 domains and set geographic location parameters to control which search results your AI model uses to generate responses. Read announcement →
  • Amazon Connect lets call center admins customize which customer data appears on agents' screens during calls—like showing baggage info to claims agents and loyalty status to rewards agents—through a drag-and-drop layout builder that replaces the one-size-fits-all default view. Read announcement →
  • AWS released Continuum, which tests your code vulnerabilities in an isolated sandbox to confirm they're actually exploitable before auto-generating fix pull requests, helping you ignore the 90% of security alerts that don't actually matter. Read announcement →
  • Amazon OpenSearch Serverless cuts the time before your newly written data becomes searchable from 10 seconds to 5 seconds, which matters if you're building search features or RAG applications where users expect to immediately find what was just added. Read announcement →
  • CloudFormation can automatically complete a failed stack rollback in one step instead of making you manually identify and fix each stuck resource before trying again. Read announcement →

The Quarry

Write your own training loop — without managing GPUs — with Amazon SageMaker training session APIs

Amazon SageMaker's new training session APIs let you write custom training loops in vanilla Python while AWS babysits the GPU infrastructure, which means you can finally implement GRPO, RLHF, or whatever exotic objective function you dreamed up without wrestling with cluster orchestration. The API exposes three SDK clients that give you direct control over the training loop (think SFT tweaks or multi-stage RL pipelines), and the post walks through real examples like tuning models on GSM8K math problems and competitive programming tasks. You write the loop, SageMaker runs the metal—it's the control of DIY training with the convenience of managed compute. Read blog →

More posts:

Core Sample

AI and the Holy Organ: Inside Neurosurgery at Johns Hopkins

Johns Hopkins neurosurgeon Dr. Chetan Bettegowda and AWS's Dr. Diego Socolinsky discuss how AI is moving from research demos to actual clinical deployment, tackling the documentation burden that pulls surgeons away from patients and exploring predictive models that might explain why identical patient profiles can yield wildly different surgical outcomes. The conversation digs into what it actually takes to earn trust in high-stakes environments where mistakes affect lives, not just metrics, and why healthcare's decades-old workflows and fragmented data systems make production AI deployment so much harder than running a successful pilot. One key technical insight: the gap between capturing clinical notes intraoperatively and building reliable AI systems for medical imaging interpretation isn't about capability anymore, it's about navigating regulatory constraints and integrating with legacy infrastructure that was never designed for real-time machine learning. Watch video →

More videos:

Read more