TOGETHER WITH CLOUDZERO
40% spend $10M+ on AI. What's the ROI?
Our 2026 research asked 475 leaders to put a number on AI ROI. The recalibration starts here.
What you'll take away:
Where AI spend outruns measurable return
The metrics that actually track ROI now
A framework to recalibrate for the AI era
TOKENOMICS
Tokenomics in Enterprise AI
Token costs in enterprise AI work a lot like cloud computing bills. Every time your team sends a prompt to an AI model, you pay for input tokens (the question and context you send) and output tokens (the answer you get back).
And just like cloud costs, these bills can spiral out of control if nobody is watching. Most companies waste money on AI in four predictable ways.
First, they send huge chunks of unnecessary context with every request—full conversation histories, entire policy documents, and oversized data retrievals when a small snippet would work fine.
Second, they use expensive premium models for simple tasks like formatting data or answering basic questions, when cheaper models would do the job.
Third, they let AI responses run wild with no length limits, generating paragraphs when a few bullet points would suffice.
Fourth, they build automated systems that retry failed requests over and over without proper validation or caching.
The fix starts with treating tokens like any other cloud resource. You need visibility into who is using what, budgets for different teams and projects, and clear rules about which tasks deserve premium models versus basic ones. Smart companies segment their AI workloads the same way they segment compute workloads.
The best organizations track token metrics the same way they track cloud spend. Reviews should show token consumption per workflow, cost per successful outcome, cache hit rates, and what percentage of requests use cheaper models.
Teams that overspend shouldn't just get blocked—they should get help redesigning their prompts and workflows.
CLOUD PROVIDERS
GCP default Shared CUDs across Projects + AWS Billing History Data + more
AWS
AWS Cost Explorer now keeps historical billing data intact even when accounts are grouped via Billing Conductor, so old reports still match original charges.
Amazon S3 Vectors query costs dropped by up to 80% for large indexes with 10M+ vectors — no code changes needed.
RDS for SQL Server gets new X2m instances with up to 4 TB memory and 50% lower SQL Server licensing costs vs. older instance types.
Read All AWS Updates
Google Cloud
Google Cloud now shares Committed Use Discounts across projects by default, so discounts are less likely to go unused when one project has more capacity than another.
Cloud Billing Reports added new filters and group-by options for Products and Originating Services, making it easier to see what's driving your spend.
Read All GCP Updates
Azure
GitHub Pre-Purchase Plans let you prepay for Copilot, Actions, and Codespaces over 12 months with up to 15% built-in savings — turning surprise bills into a fixed budget.
Azure Savings Plans can now be right-sized using real hourly usage data from the Cost Management API, so commitments are based on facts, not guesses.
Read All Azure Updates
VIDEOS & PODCASTS
FinOpsX Interviews Part 1
FinOpsX Interviews Started to roll out and we had wonderful content waiting for you:
Alon on how to raise $60M for FinOps
Roni on how to close the execution gap AI agents
Adedeji on how FinOps is an opportunity in the current tech market.
FINANCE
ABC: How to do FinOps Accounting
Activity-Based Costing sounds like an accounting textbook topic, but it's actually a practical way to understand why your technology costs are what they are.
The core idea is simple: work drives cost. When something needs more effort, time, or resources, it costs more.
Think of ABC as a pipeline with three stages:
Activities consume resources (like engineering hours, compute time, storage, and licenses).
Products and services consume those activities (your apps, platforms, and AI models need those activities to run).
Cost flows through this chain: Resources → Activities → Products or Services.
Start by identifying your major activities, not every single one, just the meaningful ones. Figure out what each activity costs using your existing accounting data for labor, cloud, software, and hardware. Find your cost drivers, the things that explain why an activity uses more or less resources.
Calculate the rate: if an activity uses 40% of a resource, it gets 40% of the cost.
Then assign those activity costs to the products or services that depend on them.
ITFM, TBM, and FinOps all use this same underlying logic.
ITFM provides financial structure around technology spend.
TBM organizes costs using a standard taxonomy.
FinOps brings cost awareness into daily operations where activity patterns change fast. ABC is the foundation they all build on.
Understanding that cost follows activity helps you explain why your cloud bill jumped, why that new AI project costs more than expected, or why supporting one customer costs three times more than another.
AI
How to Control your AI Agent Costs
AI agents with access to cloud budgets can rack up surprise costs fast, and most teams only find out when the monthly bill arrives.
The problem isn't usually malicious – it's a retry loop that never stops, an agent calling the same tool forty times in a row, or someone who accidentally pointed their test suite at the most expensive model available.
Spending caps don't solve this because they're too blunt. Your legitimate power users hit the cap just as often as the runaway processes do.
The better approach: build a baseline for each developer
Instead of setting one limit for everyone, track what "normal" looks like for each person on your team. When someone blows past their own normal pattern, you get an alert the same day – while you can still stop the runaway process.
Start with a simple table that tracks four things: date, developer name, model used, and cost in dollars. The easiest way to get this data is to give each developer their own API key, so spending automatically groups by person.
Most AI providers will also return cost data in the API response if you ask for it. Calculate a baseline using the last 14 days of spending for each developer. Use basic statistics – the mean and standard deviation – to figure out what's normal for that person.
Three things to add so you don't get false alarms:
Don't alert until someone has at least seven days of history, because two data points aren't enough to know what's normal.
Set a dollar floor so you're not getting pinged about someone who went from twelve cents to forty cents.
Give new developers a grace period, because their first week of heavy use is just onboarding.
Make the alerts show up where people actually look
Run this check once a day and send results to Slack or wherever your team already communicates.
A good alert tells you who went over, by how much, which model they were using, and how unusual this is compared to their baseline.
That one message gives you everything you need to start a conversation and fix the problem while it's still running.
🎖️ MENTION OF HONOUR
[CODE] The Context Tax: Why Step 12 Costs 42x Step 1 + Script FIX
A single AI agent session can cost 42 times more at step 12 than at step 1, even though you're using the same model and doing the same task.
The reason is simple but expensive. Every time an AI agent takes a step, it has to re-send the entire conversation history as input. The AI model doesn't remember what happened before. So your framework sends turns 1 through 11 again when you make turn 12.
Step 1 bills a short user message. Step 12 bills that message plus a file dump plus a grep output plus a stack trace plus every assistant reply in between.
This means total cost grows with the formula n(n+1)/2. When token prices drop, you're just lowering the unit price on a number that keeps climbing.
A researcher named Alexey Spinov built a 40-line Python script to measure this "context tax" on real sessions.
The script reads your session transcript and reports four things: how much input you billed at each step, the multiplier between your first and last step, how much dead weight you're carrying, and the total cost.
On a realistic debugging session, the script measured a 42.8x multiplier. It also found that 19.3% of the tokens were dead weight - old content that never got referenced again but kept getting re-billed anyway. The script can run in your CI pipeline and fail the build if a session balloons past your threshold.
Research backs up why you need to measure instead of guess. A recent paper found that AI agents burn roughly 1,000 times more tokens than plain code chat. The same task can vary up to 30 times in cost from run to run.
Models can't accurately predict their own token usage - the correlation was only 0.39, barely better than random.
Three fixes actually work: reset the context between separate tasks, trim or summarize large tool outputs instead of re-sending them verbatim, and collapse old turns into short recaps once they're settled.
The lean example session kept its multiplier at 9.4x by resetting scope between tasks. The bloated session hit 42.8x because it kept re-sending a verbose dependency log on every step after it appeared.
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 (The foundation for success).
FinOps Certified Engineer (For high-level technical profiles).
FinOps Certified FOCUS Analyst (Specializing in data standards).
FinOps for AI (The frontier of modern efficiency).
COMMUNITY TOP 10 CONTRIBUTORS
Top 10 contributors of the community:
1. @Afor Linda Odoma - 697.22 (Level 4)
2. @Emma - 297.5 (Level 3)
3. @James Johnson - 294.8 (Level 3)
4. @Ben - 244.04000000000002 (Level 3)
5. @Ashley Bar-Shay - 199.56 (Level 3)
6. @Stuti Sharma - 168.06 (Level 3)
7. @Sam Nord - 167.02 (Level 3)
8. @Karna - 155.48 (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