Good Morning!
I spend most of my time telling companies they’re spending too much money on AWS. But here’s the dirty secret: most of them don’t even know what they’re spending money on. That’s the actual problem.
Duckbill is building Skyway — software that makes enterprise cloud bills legible to the humans who are theoretically responsible for them. When you’re spending $200 million a year on AWS, “what are we paying for” shouldn’t be a philosophical question, but for most companies it absolutely is.
We’re hiring in San Francisco. If you want to build software that turns the world’s most incomprehensible invoices into something a CFO can look at without weeping, come join us.
Things I Found on the Internet
AWS quietly emailed customers this last to say they’re on their own when it comes to video patent infringement claims. The alternative, Amazon says, would be raising prices. Florian Mueller’s detailed breakdown on ip fray compares this to what Azure and Google Cloud offer their customers. Spoiler: it’s not a great look for AWS.
Joe Ruscio makes a compelling case that we’re headed toward a world where most production code is never read by a human. Not skimmed, not reviewed. This piece on “Write-Only Code” is a great frame shift around what engineering work actually looks like when the review bottleneck disappears entirely.
What AWS Has For Us This Time
Amazon Aurora DSQL is now available in additional AWS Regions
Aurora DSQL expanding to 14 regions is great, but calling it “the fastest serverless, distributed SQL database” is doing some real heavy lifting when you’re competing primarily against yourself. Seriously, what other contenders are there here? Two Canadian regions though – finally, Aurora apologizes to Canada for something. Free tier included, so your regret can start at $0; only God knows what Aurora DSQL will cost past that, because I sure don’t.
Amazon Bedrock adds support for six fully-managed open weights models
Six open-weights models land on Bedrock via “Project Mantle,” which is apparently an entire distributed inference engine they built just so they could onboard models faster. Translation: AWS got tired of watching OpenRouter eat their breakfast, lunch, and dinner. OpenAI-compatible API endpoints is a particularly delicious touch, but it’s not like AWS had a choice here; it’s the OpenAI way or the highway.
AWS Config now supports 30 new resource types
Thirty new resource types you can now track in AWS Config, which means thirty new resource types generating configuration change events you’re paying for. If you left “record all resources” enabled, congratulations: your Config bill just went up without you doing a single thing. And it highlights my entire problem with Config: I see these “x services now supported” announcements all the freaking time. So what is vs. isn’t covered? You can’t build a compliance system out of a patchwork quilt like this.
Announcing new Amazon EC2 general purpose M8azn instances
EC2 instance naming has reached the point where “M8azn” looks like a vanity license plate for someone with a very particular fetish. Five gigahertz of clock speed targeted at high-frequency traders, because if there’s one group that truly needs a break on infrastructure costs, it’s hedge funds. At least the rest of us get nine sizes to overprovision.
AWS Network Firewall announces new price reductions – Charging extra to inspect encrypted traffic always felt like a protection racket with better branding. Now they’ve dropped the TLS inspection surcharge and extended NAT Gateway discounts to secondary endpoints. No customer action required, which is AWS-speak for “we were quietly overcharging you and hoped you wouldn’t notice.”
Amazon S3 Tables add partition and sort order definition in the CreateTable API – S3 Tables now lets you define partition and sort order at creation time instead of after. You know, the thing every other table format has let you do from day one. Congratulations to whatever team had to file a six-pager to justify adding a parameter to an API call.
Amazon Athena adds 1-minute reservations and new capacity control features – Dropping from 24 DPU minimums to 4 and from 60-minute reservations to 1 minute is genuinely useful. “Up to 95% savings” is doing some heavy lifting in that sentence, though. Also, the autoscaling solution requires Step Functions, CloudWatch, and presumably a prayer circle. Serverless simplicity at its finest.
Building fault-tolerant applications with AWS Lambda durable functions – Lambda finally got a memory. Durable functions let your code checkpoint, pause for up to a year, and resume without you duct-taping DynamoDB and Step Functions together like some kind of serverless Frankenstein. Of course, Step Functions still exists, so now you get to argue with your team about which orchestration to use forever.
Simplify cross-account stream processing with AWS Lambda and Amazon DynamoDB – “Simplify” doing a lot of heavy lifting here, but honestly? Resource-based policies for cross-account DynamoDB Streams is in fact useful. Previously you needed a Kinesis relay or custom IAM spaghetti just to read a stream from another account. Now it’s just… a policy. Almost suspiciously reasonable for AWS.
Automated Reasoning checks rewriting chatbot reference implementation – Mathematically proving your chatbot isn’t lying feels like installing a breathalyzer on a self-driving car. Points for honesty about the hallucination problem, though. Nothing inspires confidence in your AI strategy like needing a second system to fact-check the first one you’re paying Bedrock for.
Mastering Amazon Bedrock throttling and service availability: A comprehensive guide – Nothing screams “production-ready AI platform” like needing a comprehensive guide just to handle all the times it tells you to go away. A 3,000-word blog post about retry strategies is AWS admitting that “429 Too Many Requests” is a core feature of Bedrock, not a bug.
Reservoir computing on an analog Rydberg-atom quantum computer
Quantum reservoir computing on Rydberg atoms for pharmaceutical ML tasks sounds like someone fed a physics textbook into a buzzword blender. Points for honesty though: it works best on “small datasets,” which is quantum computing’s polite way of saying classical hardware still eats its lunch at any meaningful task or scale.
… and that’s what happened Last Week in AWS.
