TOGETHER WITH KION
Who's best positioned to govern AI token spend? (Hint: It's you)
AI is introducing a new category of technology spend most organizations aren't prepared to manage.
Which teams are driving AI token costs? Why did spending spike last month?
FinOps teams are uniquely positioned to solve this, because we've been here before.
Kion's Tatum Tummins joined theCUBE to recap FinOps X and share how to apply proven FinOps principles to AI token spend.
AWS
How CrescoNet cut AWS costs by 40% without sacrificing performance
CrescoNet processes 4.5 billion meter readings every day for Australia's energy sector. They cut their AWS bill by over 40 percent while keeping their systems fast and reliable. Here's how they did it.
CrescoNet runs smart meter data for electricity, water, and gas companies. As Australia adds more solar panels and electric cars, the amount of data keeps growing. Small waste at this scale adds up fast.
CrescoNet used dashboards to find where money was going. They looked at each opportunity and asked: What do we gain? What do we give up? Then they made changes and checked the results.
Switched from AWS Glue to Amazon EMR on EKS. This saved over $50,000 per month on data processing.
Fixed the small file problem with Apache Iceberg. CrescoNet had 450TB of meter data split into millions of tiny files.
Made Lambda functions faster and cheaper. They updated Python from version 3.8 to 3.11.
Set expiration dates on DynamoDB data. This cut storage by 66 percent and saved $3,000 per month.
Batched SQS messages. Instead of sending 140 million separate messages per day, they packed multiple messages together. This cut SQS costs by 20 times.
Bought reserved instances for databases. They committed to using certain database sizes and got a big discount.
Turned off test systems at night. Development and testing systems now shut down outside business hours automatically.
These changes didn't just save money. Performance actually got better in several areas. The system can now handle more growth without breaking the budget.
CLOUD PROVIDERS
Amazon EC2 Capacity Reservations can now be cancelled + New CUD coverage updates at GCP
AWS
Amazon EC2 Future-dated Capacity Reservations can now be cancelled, giving teams more flexibility when plans change, with an upfront cost quote before cancelling.
S3 Vectors cuts query charges by up to 80% for vector indexes over 10 million entries; no code changes needed to save.
Amazon RDS adds Graviton5 M9g instances for PostgreSQL, MySQL, and MariaDB with 30% better performance and 23% better price-performance than Graviton4.
Amazon Redshift RG Graviton instances expand to more regions, offering up to 4.2x better price-performance versus other warehouses.
Read All AWS Updates
Google Cloud
Google Cloud CUD recommendations now cover GPUs, Local SSDs, and premium OS licenses, making it easier for FinOps teams to find savings on more resource types.
Read All GCP Updates
Azure
No Updates
EVENTS
Meetups
Our next meetup will explore a topic many organizations are already facing: how to integrate AI, automation, and new operating models without losing control, efficiency, or visibility.
This session is designed for professionals working in Cloud, FinOps, AI, Digital Transformation, and Operations. Expect technical talks, engaging discussions, and networking with the community.
October 7 – Madrid (Utopicus Nuevos Ministerios)
https://luma.com/036wqk98
Webinar
On September 22, we'll be joined by Luis Guirigay and Alex Scharifker to discuss performance, AI governance, and architectures built to scale in the era of agentic AI.
📅 September 22
🕚 5:00 PM CEST (Spain) / 11:00 AM EDT
VIDEOS & PODCASTS
FinOps X 2026 Interviews: Part 2
FinOpsX Interviews continue to drop every day, here’s what you haven’t watched already:
AI FINOPS
AI agents can quietly become one of your biggest cost drivers.
AI agents that fail don't just fail once and stop. They keep running, burning tokens, long after the first sign something went wrong.
A recent study of 165 agent traces found that failed runs spent 58.1% of their tokens after the first warning signal appeared. That means more than half the cost came after the agent already had enough information to know it was off track.
A developer built a simple 40-line Python script to measure this waste in your own agent logs. The tool is completely offline - no API keys, no network calls, no vendor SDKs.
You feed it a JSON trace of an agent run, and it tells you what percentage of tokens were spent after the first detectable problem. It looks for two signals: explicit tool errors, and when an agent calls the same tool with identical arguments it already tried.
Anthropic's Dynamic Workflows can run up to 16 concurrent subagents and 1,000 total per run.
The script won't stop a runaway agent in real time. It's a post-mortem tool that measures what already happened.
But that measurement tells you where to focus: the retry logic that keeps going when it should stop. Most agent frameworks assume errors are transient and worth retrying.
CLOUD GOVERNANCE
What to Expect When You're Expecting Direct CSP with Microsoft
Changing your Cloud Solution Provider doesn't move your Microsoft tenant, users, licenses, or Azure resources. What changes is who bills you, provides support, and has delegated access to your environment.
For organizations already using CSP, Azure subscriptions and Microsoft 365 licenses can usually transfer seamlessly without disrupting users. However, pay-as-you-go Azure subscriptions require a full resource migration, including rebuilding RBAC permissions and updating resource references.
The biggest risk comes after the switch. Billing may transfer automatically, but old partner access doesn't. Teams should manually remove legacy GDAP roles, Azure Lighthouse delegations, AOBO permissions, service principals, and any Conditional Access exceptions to prevent former partners from retaining privileged access.
A CSP switch is primarily an operational and security exercise, not a technical migration, provided you're not moving from pay-as-you-go Azure.
BEST PRACTICES
Falling AI token prices don't necessarily mean lower AI costs
While the cost per million tokens dropped 67 percent in one year, total AI spending is exploding because companies are using far more tokens than they ever planned.
The FinOps Foundation now calls AI the fastest-growing category of enterprise spend.
A traditional server has predictable costs because you control what it does. An AI agent writes its own bill every time it runs. It decides how many reasoning steps to take, which models to call, and how much context to carry forward.
Every step resends the entire conversation, so costs climb with each decision the agent makes. A single agent stuck in a loop can cost fifty thousand dollars before anyone notices.
Three cost classes make the ceiling real
High-volume, low-value work like support deflection gets a ceiling of cents.
Standard knowledge work worth fifty dollars of labor gets a five dollar budget.
High-value work worth five thousand dollars gets five hundred, because starving the agent is more expensive than feeding it.
In most companies, engineering owns the deployment and finance owns the invoice. The agent's cost falls between them, which is why it grows. Every agent needs a named owner with authority to set limits and trip the breaker, just like you name an owner for its output.
The price of intelligence will keep falling, but the cost of deploying it will keep climbing, and that gap is where the next wave of canceled projects already lives.
🎖️MENTION OF HONOUR
How one SaaS company cut its AWS bill by 45%
A security SaaS company cut their AWS bill from $30,000 to $16,500 per month without removing features or slowing down deployments. The team at Matters.AI spent two quarters finding and fixing specific cost problems.
They committed to paying $3.21 per hour for three years with no money upfront. Result: $22,000 saved per year.
They had one NAT gateway in one zone. They added a second NAT gateway so traffic stays local. This cut cross-zone transfer costs almost completely.
Database traffic was going through the NAT gateway to the public internet, paying fees twice. They set up VPC peering instead. Now traffic goes directly from pods to the database with no NAT and no internet charges.
They had 2,400 secrets costing $0.40 each per month. Many were orphaned from old customer integrations that no one deleted. They also moved AWS service credentials to IRSA, which uses temporary credentials instead of stored secrets.
Auto Mode for Short Jobs. EKS Auto mode picks the cheapest available instance type automatically from 30+ options.
Removing Unused Public IP Addresses. A simple audit and cleanup saved $174 per month.
In February, they turned on a security scanner that did a full scan of everything at once. The bill jumped from $270 to $16,806 in one month. The real work was putting these fixes into their infrastructure code.
The Hidden Costs
Data transfer costs were as big as compute costs. But they don't show up clearly in basic reports. You have to drill down three levels in Cost Explorer to see them. The team recommends opening Cost Explorer, picking your biggest account, grouping by service, then drilling into usage type for EC2 and VPC.
Save 20% on FinOps Certifications
The job market is hungry for certified professionals who can prove results. Don't let your company's budget leak due to a lack of specialization.
Use code: FINOPSWEEKLY_20 to get an instant 20% discount on the most prestigious certification bundles:
FinOps Certified Practitioner
FinOps Certified Engineer
FinOps Certified FOCUS Analyst
FinOps for AI
COMMUNITY CONTRIBUTORS
Top 10 contributors of the community:
1. @Afor Linda Odoma - 697.22 (Level 4)
2. @Emma - 307.5 (Level 3)
3. @James Johnson - 294.8 (Level 3)
4. @Ben - 251.54 (Level 3)
5. @Ashley Bar-Shay - 199.56 (Level 3)
6. @Stuti Sharma - 188.06 (Level 3)
7. @Karna - 175.48 (Level 3)
8. @Sam Nord - 172.02 (Level 3)
9. @Madhavi Yamani - 139.38 (Level 3)
10. @Jackie MacRobert - 133.12 (Level 3)
How to get points
Replying to other people
Reacting to other people
Sharing knowledge and asking questions
Full ranking at: https://finopsweekly.com/community