In this episode of What’s New in Cloud FinOps / September 2026 Cloud FinOps News, Frank Contrepois and Stephen Old review updates across AWS, Google Cloud, Microsoft, Oracle, Alibaba Cloud, and the wider cloud financial management ecosystem.
Compute and infrastructure
AWS Elastic Beanstalk cluster mode lets multiple applications share pooled EKS-backed infrastructure. This can reduce per-application costs as deployments grow, although the underlying EKS Auto Mode resources remain chargeable.
Bedrock AgentCore Runtime v2 now bills for actual memory usage instead of retaining a session’s peak allocation until it ends. Billing closer to consumption reduces the financial impact of temporary spikes.
AWS Lambda adds Graviton5-powered EC2 support through Lambda managed instances, including C9g, C9gd, M9g, and M9gd. The hosts distinguish better price performance from lower unit prices: savings may require resizing workloads to exploit faster hardware.
Amazon EC2 T8i burstable instances use sixth-generation Intel Xeon processors and offer improved price performance over T3. Nano, micro, small, and medium sizes target small websites, test environments, and lightweight workloads, with smaller sizes eligible for the free tier.
Amazon RDS adds AMD-based M8a instances using fifth-generation AMD EPYC processors. The hosts compare these with Graviton, noting that managed databases can make processor changes easier than migrating self-managed applications.
AWS Direct Connect introduces flat-rate pricing for 10/100 Gbps connections across five geographic scopes. It removes data transfer out charges within the selected tier and includes an optional redundant port pair at no extra cost. Large data movers gain greater predictability alongside potential savings.
CloudFront Flat Rate Plans become programmable through APIs and infrastructure as code, including subscriptions, upgrades, downgrades, and cancellations, with optional paid-plan approvals. This supports automated governance of CDN, DDoS protection, DNS, logging, and edge compute bundles without usage-based overages.
Amazon WorkSpaces adds NVIDIA Blackwell GPUs through G7 graphics instances. These offer up to 2.2 times G6 performance, 32 GB GPU memory per GPU, and 2.67 times faster bandwidth. Configurations span 1–8 GPUs, 8–192 vCPUs, and 32–768 GB memory. Initial availability in Ohio and Oregon targets CAD, 3D, visualisation, editing, and AI-assisted design.
Lambda also adds direct read configuration for S3 File High Performance Storage and standard buckets, helping optimise throughput and latency for data-intensive serverless workloads.
Google Cloud Run Jobs introduces deferred execution in preview. Non-urgent jobs can wait up to 12 hours for reduced pricing, offering savings where execution timing is flexible, including batch AI workloads.
Billing, chargeback, and visibility
AWS Billing Conductor adds custom rate and usage-tier pricing configurations, supporting more granular internal chargeback across services, teams, products, and business units.
AWS Billing and Cost Management dashboards gain anomaly widgets, bringing detected anomalies into views alongside budgets, cost efficiency, Savings Plans, and Reserved Instance utilisation.
The hosts examine proportional RI and Savings Plans chargeback. Centrally purchased commitments can be difficult to allocate fairly. AWS guidance combines Glue, Athena, Lambda, SSM Parameter Store, Billing Conductor, and EventBridge Scheduler to distribute realised savings proportionally across accounts.
The new ListBillingViewSegments API exposes historical billing context, including payer/member relationships, Billing Conductor groups, and whether billable or pro forma rates applied. This matters when accounts move: using today’s structure to interpret past spending can distort attribution, reconciliation, and audits.
AWS also publishes guidance on understanding Enterprise Support charges through billing APIs, including monthly totals, calculation rules, and account-level breakdowns. Stephen highlights the historical difficulty of allocating support costs across multiple payer accounts.
Google Cloud introduces an Incentives page for spend-based milestone credits, ramp incentives, Rapid Migration and Modernization Program credits, discounts, and other conditional incentives. It replaces the Spend-based Milestones tab for customers with custom pricing contracts.
Amazon Q gains payment capabilities to investigate overdue balances, identify failed payments and their causes, and resolve issues within a console conversation, potentially reducing manual reconciliation.
AI economics and tokenomics
An Oracle Cloud Infrastructure GenAI update from 23 September enables automatic routing of on-demand inference across approved regions. Initially focused on capacity resilience, routing profiles could eventually help balance cost, capacity, latency, and availability.
AWS and OpenAI Bedrock Managed Agents are in preview, with no additional Managed Agents charge beyond underlying AWS resources. Pricing may change at general availability, making this a FinOps watch item.
Google Cloud extends Gemini Enterprise overage controls to all invoiced cloud billing accounts and introduces session and memory bank compute metering for the Gemini Agent Platform.
From 2 November 2026, new Microsoft 365 Copilot business licences purchased through CSPs will have usage-based billing enabled by default. Seat licences can therefore expose organisations to variable charges for experiences including Copilot, Work IQ APIs, and GitHub Copilot-related capabilities.
Stephen sees this as a convergence of FinOps and IT Asset Management: licence optimisation alone is insufficient when subscriptions unlock pay-as-you-go usage. Microsoft’s FinOps for AI approach connects everyday subscriptions and advanced consumption-based capabilities with business outcomes.
AWS’s “Getting Started with Tokenomics” blog frames AI financial management around visibility and attribution, optimisation, governance, and value/ROI. The hosts welcome the focus on value rather than assuming fewer tokens always means a better result.
The discussion of Claude Sonnet 5.5, available through AWS Bedrock and Google Cloud, reinforces the shift from cost per token to cost per task. A higher-priced model may deliver a cheaper outcome through fewer tool calls, fewer interactions, faster completion, or less human intervention.
Stephen recommends “The Economics of Agent Optimization: Context Engineering for Enterprise AI Agents”, exploring how context windows affect costs and how organisations decide what agents can access. He also highlights Steve Sweetman’s Microsoft blog on agent governance, observability, spending limits, model and provider policies, budgets, and ROI.
Alibaba Cloud introduces dedicated cache-hit pricing below the corresponding input token price. Token counts alone cannot explain effective costs: cache status, context length, processing priority, and execution mode also matter.
The episode cites Reuters reports from 24 and 29 September about Anthropic reportedly signing a seven-year, $11.6 billion cloud agreement and anticipating at least $518 billion in AI infrastructure obligations over a decade, around 80% binding regardless of usage. The hosts explore the risks of long-term, largely non-cancellable capacity commitments in a fast-changing market.
Automation and governance
An AWS Cloud Financial Management blog from 10 September outlines four levels of autonomous FinOps. The discussion shifts from whether AI can find savings to which changes organisations will permit it to make. Recommendations, approvals, assisted action, and autonomous remediation carry different operational and financial risks.
An AWS Budgets pattern from 1 September combines Budgets, Identity Center, IAM, and Lambda as an automatic cost circuit breaker. The hosts discuss warnings at 80–90% and stronger intervention at 120%, potentially including hard stops or cheaper models, while protecting legitimate work.
An AWS case study reports $825,000 in annual recurring savings from a four-day FinOps Experience-Based Acceleration, with a path to $1.4 million. Actions included RDS rightsizing, Graviton migrations, EC2 optimisation, removing idle replicas and non-production instances, and improving RI/Savings Plans coverage. The emphasis is on scaling implementation across thousands of accounts.
Oracle’s Sovereign AI eight-layer blueprint by Ravi Tulapati covers infrastructure, GPU orchestration, trusted model supply chains, inference optimisation, managed GenAI services, security and governance, enterprise data grounding, and agent lifecycle operations. The hosts consider it useful beyond Oracle environments.
Storage, compliance, and sustainability
Amazon S3 Object Lock adds event-based variable retention. Objects can remain on hold until an event, such as a contract closing or an audit completing, triggers the required retention period. This can reduce unnecessary over-retention while supporting compliance. Stephen encourages direct discussions with risk and compliance teams instead of assuming every dataset needs maximum retention.
Google Cloud’s Storage Intelligence Advisor reached general availability on 10 September. Stephen argues that storage optimisation often begins with identifying data and its owners. Centralised intelligence can reveal where lifecycle policies and governance need attention.
Google Cloud Carbon Footprint methodology version 17, released on 15 September, updates Scope 1 and Scope 3 allocation factors using 2026 environmental report data. The hosts note that GreenOps baselines and
D'autres épisodes de "What's new in Cloud FinOps?"
Ne ratez aucun épisode de “What's new in Cloud FinOps?” et abonnez-vous gratuitement à ce podcast dans l'application GetPodcast.
