TOGETHER WITH AI FINOPS TOOLKIT
Control AI Costs Without Slowing Down Innovation
A self-serve execution toolkit that gives finance and engineering teams a shared playbook to audit, forecast, and govern AI/GPU spend before it's too late.
Diagnostic audit matrix
ROI calculator
30-day action plan
Includes:
Model Routing & Cascading
Prompt & Semantic Caching
GPU Utilization & Batching
Cost Attribution & Tagging
Audit.
Plan.
Reinvest.
AI COST OPTIMIZATION
Who Captures the Most Margin in AI?
Your AI bill can rise even when your usage stays flat. Tomasz Tunguz’s model shows how higher compute prices could squeeze profits across AI services.
The article puts power, data centers, GPU rental firms, model services, and apps on one cost scale: dollars per watt per year.
At an assumed 40% app gross margin, apps take 40% of the modeled gross profit pool, GPU rental firms take 28%, and model services take 24%.
The GPU rental share excludes the cost of hardware losing value over time. Adding that cost cuts estimated gross profit from $7 to roughly $0–$2 per watt per year.
CoreWeave’s cited new long-term contract rates are $20–$25 per watt per year, versus about $9 across its current fleet.
For FinOps teams, the useful test is cost per successful task, not just cost per token.
Track whether the work an AI tool saves still covers its bill. Try smaller models and send simple tasks to cheaper options. These figures are estimates, not audited results.
Plan for higher compute rates, count hardware costs, and tie AI spend to clear business gains.
AI PROVIDERS
Claude Fable 5.2 pricing, new Claude budget tools, ChatGPT Enterprise on AWS Marketplace
Anthropic
Claude Fable 5.2 is now the default Fable model, and its prices are listed. It costs $8/$40 per million tokens, and cache reads cost $0.20 per million tokens. Use these numbers to update your cost forecasts.
Claude Code v2.1.319 adds new spend tools for gateway users. /usage shows spend totals, pooled budgets help teams share one budget, and spend-limit warnings flag budget problems.
Claude Code release notes show better spend visibility. Updates include better /usage views, JSON gateway logs, and fixes for OpenTelemetry export. These help you track costs across developers.
Gemini Enterprise model access now follows location policies. If a policy blocks global use, models that use the global endpoint may be turned off. Check your workloads and costs before this affects your team.
Gemini 4 Carbon is now generally available in Gemini Enterprise. Many tools can use it, so decide who owns the spend. Google did not share pricing yet.
OpenAI
ChatGPT Enterprise is now in AWS Marketplace. Teams can buy and pay through their AWS accounts. It is a new way to buy, not a lower price.
GPT-6 Luna inference can now stay in-country in India, Japan, and Australia. This helps with compliance and regional cost planning.
PODCAST
How to Optimize On-Prem Costs and Hybrid Cloud Spending
Discover how to track on-prem costs, improve resource allocation, reduce infrastructure waste, optimize GPU utilization for AI workloads, and compare cloud vs. data center spending.
Learn why unified visibility, accurate tagging, and daily financial feedback are essential for modern FinOps teams.
AI COST MANAGEMENT
GPU utilisation is a timesheet, and AWS is rightsizing from it
A GPU can show 100% use while much of its chip sits idle. That can leave costly waste out of your cloud cost review.
FinOps advisor Ben de Mora explains that AWS Compute Optimizer tracks whether a GPU ran a task, not how much of the chip that task used.
NVIDIA’s own example shows how a task using just 20% of the chip can still score 100% on that measure. This means “no finding” is not proof of “no waste.”
Before you approve the next GPU bill:
Track SM activity, which shows how broadly work uses the GPU’s processing units.
Check memory use and other chip activity measures too.
Compare average SM activity over a typical work period with the reported GPU use.
Ask your cost tool vendor which measure drives its advice.
NVIDIA says SM activity below 50% likely points to poor use. Even 80% does not prove the GPU is used well. The article rests mainly on AWS and NVIDIA docs, not a production fleet test by the author.
Treat automated findings as a starting point, not a clean bill of health for GPU spend.
EVENTS
100th Edition FinOps Weekly Event
We’re bringing the FinOps community together for a special end-of-year event filled with inspiring talks, meaningful conversations, and new connections.
This is your chance to connect with fellow FinOps professionals, share ideas, exchange experiences, and celebrate everything our community has achieved.
Be part of this special milestone and help us make our 100th edition an unforgettable experience!
AI INFRASTRUCTURE OPTIMIZATION
A 4% cache hit rate was costing us money. Here is the arithmetic I should have run first.
An AI cache meant to cut costs raised one team’s bill by about 5%. The team’s contract charged less for cache reads but added a fee for writes when the cache missed.
With those rates, the cache needed a hit rate above 21.7% to save money. It launched at just 4%.
A date and time field near the top of each prompt stopped the cache from reusing text that stayed the same. Moving that field below the fixed text raised the hit rate to 71%.
Three checks matter:
Check both read and write prices before you fund cache work.
Work out the savings limit: the share of input that can be cached × the read discount × input’s share of the total bill.
Track cached tokens and actual billed rates, not just how often the cache hits.
Here, even a perfect cache could cut the bill by only 21.3%. The fix saved 13.4%, or about $4,100 a month, compared with no cache.
Base the business case on savings against no cache—not just the drop from a costly setup.
AI COST ALLOCATION
When the Central Bank of AI Changes Chairs
Your AI budget may depend on more than chip prices. The supplied article reports Jensen Huang’s exit as Nvidia CEO and argues that tighter funding under new leadership could slow AI growth.
The leadership claim and financial figures are not independently verified here. The key issue is Nvidia’s role as both a chip seller and a source of cash for its customers.
The article cites $61 billion in equity investments and $164.5 billion in lease guarantees and cloud capacity support. It also says some customers have up to a year to pay.
If that support shrinks, cloud vendors could face cash pressure, putting service plans and future deals at risk.
Three steps can help:
Ask key AI vendors how they fund their growth and what happens if that funding falls.
Check contracts for price changes, minimum spend, and exit terms.
Test other chips on suitable tasks, comparing total cost, speed, and the work needed to switch.
Treat vendor funding risk as part of your AI cost plan, not just a supplier issue.
🎖️ MENTION OF HONOUR
Observations on Tokenomics and FinOps
A lower AI bill does not always mean a better business result. For FinOps teams, that gap matters.
J.R. Storment draws a line between two linked areas: AI FinOps tracks how well you use AI spend, while tokenomics looks at how AI changes products, prices, profit, and staffing.
These are his personal views, not formal guidance. His advice points to three checks:
Count the full cost of AI, including compute, storage, data work, security, software, and staff time. Storment says tokens often make up just 10–25% of that total.
Look past cost per task to cost per business outcome. More code or more completed tasks do not, on their own, prove business value.
Link cost data to results finance leaders can verify, such as higher sales, lower costs, or more work done with the same team.
FinOps supplies the cost data and tracking that these wider business talks need. For your next AI review, bring both the bill and proof of what the business gained.
Save 20% on AI Value & FinOps Certifications
The job market is hungry for certified professionals who can prove results. Don't let your company's budget leak because of a lack of specialization.
Use code: FINOPSWEEKLY_20 to get an instant 20% discount on the most prestigious certification bundles:
FinOps Certified Practitioner
FinOps AI Value
FinOps Technology Value
FinOps Certified Engineer
FinOps Certified FOCUS Analyst
AI Economics Community
Join the Slack community where the real AI Money conversations happen
