New AWS rates for EC2 Capacity Blocks, the product customers use to reserve NVIDIA GPU clusters for a fixed period, take effect today, October 7. It is the latest in a run of 2026 increases on a pricing model that sits outside the commitments AWS makes for On-Demand and Savings Plans. The same week brought Microsoft's account of an Azure networking outage across 18 regions, and an open-source tool from Google built to move Kubernetes workloads off Amazon EKS.
These are separate events. Each one bears on a question cloud buyers keep coming back to: how much of their cost and availability is decided by one provider acting on its own.
AWS reprices reserved GPU capacity
The Capacity Blocks pricing page lists new hourly rates per accelerator, effective October 7, 2026, across all available AWS Regions: $16.146 for P6-B300, $14.208 for P6-B200, $7.895 for P5en, $6.866 for P5e, $5.970 for P5, $2.546 for P4de and $1.696 for P4d. In GovCloud, the two Blackwell rates are higher, at $16.819 and $14.801. AWS states that all other Capacity Block prices, and fixed pricing models such as On-Demand and Savings Plans, are unchanged (AWS, 2026).
AWS has not published a percentage for the change. Wells Fargo analysts put it at 15%, following a 20% increase in July (roic.ai, 2026). The pattern started in January, when the hourly price of a p5e.48xlarge Capacity Block rose from $34.61 to $39.80 in most regions (InfoQ, 2026).
The headline number matters less than where the exposure sits. AWS documentation says Capacity Block prices are dynamic and updated periodically based on supply and demand, and that a block is charged at the rate in force when it is purchased, even if it starts after a price change (AWS documentation). A team that books training runs a quarter at a time pays whatever the rate is on the day it books. For forecasting, that makes Capacity Blocks a variable line, not a reserved one. Whether a particular enterprise agreement's price protections reach Capacity Blocks is a contract question; AWS's public commitment names On-Demand and Savings Plans.
Azure traces an 18-region outage to a gateway change
Between 20:30 UTC on September 30 and 02:15 UTC on October 1, a subset of Azure customers using gateway services saw degraded or interrupted network connectivity, according to Microsoft's status history. Affected services included ExpressRoute Gateway, VPN Gateway, Azure Firewall, Application Gateway and Web Application Firewall, and Azure VMware Solution. The 18 affected regions ran from West US and North Europe to Japan West, South Africa North and Jio India Central.
Microsoft's stated cause is specific. A recent change to a regional gateway management service generated more load than expected when an unrelated operating system maintenance process moved gradually through multiple regions. Dependent services could not scale to meet the demand. Microsoft reverted the gateway manager change, which reduced the load and let services recover.
That incident began the day after a separate one. From 10:03 to 15:58 UTC on September 29, Azure OpenAI Service, Foundry Agent Service, Foundry Models and Cognitive Services in Sweden Central returned intermittent HTTP 5XX errors and higher latency (Microsoft Azure status, 2026; Windows Forum, 2026). Microsoft's status history records the two as separate incidents.
The gateway failure is the kind of problem a multi-region design is supposed to absorb. Here it crossed region boundaries, because both the trigger and the maintenance process were rolling across regions at the same time. ExpressRoute and VPN gateways are how many enterprises connect on-premises networks to Azure, so the impact reached hybrid connectivity, not only cloud-native workloads. The useful question for any provider is how changes to shared regional management services are staged, and whether a second region in the same cloud depends on the same ones.
Google builds a tool for leaving EKS
Google Cloud this week released GKE agentic migration, an open-source agent plugin for moving Kubernetes workloads from Amazon EKS to Google Kubernetes Engine, in public preview. It runs locally inside a team's existing development tooling. It indexes the source infrastructure-as-code, translates Terraform and Kubernetes manifests, and maps AWS-specific components to GKE equivalents, for example Karpenter autoscaling configuration to GKE Custom Compute Classes. Google says the LLM translation is checked with deterministic validation such as terraform validate, and that output arrives as pull requests and data-migration runbooks for human review, with no changes made to live clusters.
The plugin is part of a wider Google Cloud Modernize portfolio. It also includes a Modernization Hub that maps dependencies in Java, .NET and mainframe code, a generally available Agentic Quick Estimator that projects Compute Engine costs from VMware inventory exports, and X5 VMs with up to 43 TiB of memory in a single node (CIO, 2026).
Kubernetes is often sold as portable. In practice, migration effort concentrates in the parts that differ by provider: autoscaling, storage classes, load balancing and identity bindings. A tool aimed at those mappings goes after the actual switching cost rather than the container layer, which was already portable. Two caveats apply. It is a vendor tool built to move customers in one direction, and it is in preview, so its accuracy on large production estates has not been shown independently.
What to watch
Microsoft's final post-incident review for the September 30 gateway outage, if it publishes one, should explain how the gateway manager change was rolled out and why region-by-region staging did not contain it. That detail will tell Azure customers more about the independence of their regions than the status entry does.
For AWS, the next signal is whether Capacity Block rates move again at the start of the next quarter, and whether AWS begins publishing the size of each change rather than leaving analysts to calculate it.
For Google, the test is adoption outside its own announcement: public accounts from EKS customers who have run GKE agentic migration against a real estate, and whether the plugin reaches general availability.
Sources
- AWS. "Amazon EC2 Capacity Blocks for ML Pricing." 2026. https://aws.amazon.com/ec2/capacityblocks/pricing/
- AWS. "Capacity Blocks pricing and billing." 2026. https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/capacity-blocks-pricing-billing.html
- roic.ai. "AWS Hikes GPU Prices Again as AI Demand Stays Strong." 2026. https://www.roic.ai/news/aws-hikes-gpu-prices-again-as-ai-demand-stays-strong-10-01-2026
- InfoQ. "EC2 ML Capacity Price Hike." 2026. https://www.infoq.com/news/2026/01/ec2-ml-capacity-price-hike
- Microsoft. "Azure status history." 2026. https://azure.status.microsoft/en-us/status/history
- Windows Forum. "Azure Outages Hit Sweden Central AI and 18-Region Gateways: Root Causes and Monitoring Lessons." 2026. https://windowsforum.com/news/azure-outages-hit-sweden-central-ai-and-18-region-gateways-root-causes-and-monitoring-lessons.447008/
- Google Cloud. "Introducing GKE agentic migration for AI-assisted EKS-to-GKE migrations with built-in governance." 2026. https://cloud.google.com/blog/products/containers-kubernetes/gke-agentic-migration
- CIO. "Google overhauls cloud modernization portfolio with new AI agents." 2026. https://www.cio.com/article/4231151/google-overhauls-cloud-modernization-portfolio-with-new-ai-agents-2.html