Amazon EC2 M9g and M9gd Instances Powered by AWS Graviton5 Are Now Generally Available
AWS announced general availability of Amazon EC2 M9g and M9gd instances, the first to run on the fifth generation of AWS's custom Graviton processors. Compared to the previous Graviton4-based generation, AWS is citing up to 25% better overall compute performance, up to 30% faster database workloads, and up to 35% faster performance for both web applications and machine learning tasks.
M9g instances target general-purpose workloads — application servers, microservices, gaming backends, caching layers, containers, and increasingly, agentic AI applications. M9gd instances add local NVMe SSD storage, aimed at use cases like media processing, batch processing, and other workloads that benefit from fast, low-latency temporary storage.
Notably, these are the first instances built on AWS's sixth-generation Nitro System, which introduces the "Nitro Isolation Engine" with formal verification technology for stronger security guarantees at the hypervisor level. The instances are available now in US East (N. Virginia, Ohio), US West (Oregon), and EU (Frankfurt), and can be purchased via Savings Plans, On-Demand, Spot, Dedicated Instances, or Dedicated Hosts. Teams running compute- or memory-intensive workloads on Graviton4 today have a clear, drop-in upgrade path to evaluate for both cost and performance improvements.
Read more — AWS What's New
A Dynamic Document Extraction Pipeline with Amazon Bedrock
AWS published a reference architecture for extracting structured data from large volumes of scanned documents — the kind of problem teams hit when digitizing archives of PDFs like land lease agreements, contracts, or forms. The design combines two parallel processing paths so teams can balance latency and cost depending on the request.
The on-demand path uses an SQS FIFO queue feeding a Lambda function that processes a single document within seconds, converting scanned PDFs to images so they can be handled by Claude's multimodal capabilities — useful for time-sensitive lookups. The batch path uses a standard SQS queue with EventBridge Scheduler to group 100+ documents into a single Bedrock batch job, with separate pre- and post-processing Lambda functions. AWS reports the batch pipeline can process roughly 1,000 documents in 15 minutes using Python multiprocessing, at around 50% lower cost than the on-demand path.
A key piece of the design is Amazon Bedrock's Prompt Management service: both pipelines pull stored prompts and prompt versions dynamically, so a single pipeline can apply different extraction prompts and even different Claude models per document type. For documents exceeding Claude 4 Sonnet's 20-page limit, the pipeline automatically splits them into chunks before processing, and results land in DynamoDB with full metadata tracking. For teams building document-heavy ingestion pipelines, this is a concrete pattern for getting both fast single-document lookups and cost-efficient bulk processing out of the same Bedrock setup.
Read more — AWS Machine Learning Blog
