CloudFinOpsKit
@cloudfinopskit.com
One-click FinOps cost assessments for Azure, AWS and Google Cloud. Find the waste, prove the savings. cloudfinopskit.com
Send a thousand rows to a model as a JSON array and you just paid for eight field names a thousand times.
As CSV the headers are declared once. Models read delimited data fine.
Same answer, a fraction of the input tokens.
Token count is a property of your code, not the model.
As CSV the headers are declared once. Models read delimited data fine.
Same answer, a fraction of the input tokens.
Token count is a property of your code, not the model.
September 29, 2026 at 3:10 PM
Send a thousand rows to a model as a JSON array and you just paid for eight field names a thousand times.
As CSV the headers are declared once. Models read delimited data fine.
Same answer, a fraction of the input tokens.
Token count is a property of your code, not the model.
As CSV the headers are declared once. Models read delimited data fine.
Same answer, a fraction of the input tokens.
Token count is a property of your code, not the model.
A 20-step agent does not cost twice what a 10-step agent costs. About four times.
Models are stateless, so the orchestrator re-sends the whole history every step. Step 10 pays for steps 1-9 again.
Cost grows with the SQUARE of the step count.
max_tokens does not bound a loop.
Models are stateless, so the orchestrator re-sends the whole history every step. Step 10 pays for steps 1-9 again.
Cost grows with the SQUARE of the step count.
max_tokens does not bound a loop.
September 24, 2026 at 8:01 PM
A 20-step agent does not cost twice what a 10-step agent costs. About four times.
Models are stateless, so the orchestrator re-sends the whole history every step. Step 10 pays for steps 1-9 again.
Cost grows with the SQUARE of the step count.
max_tokens does not bound a loop.
Models are stateless, so the orchestrator re-sends the whole history every step. Step 10 pays for steps 1-9 again.
Cost grows with the SQUARE of the step count.
max_tokens does not bound a loop.
Most common AI finding we see: a model deployment that served zero requests in 30 days and billed anyway.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
September 24, 2026 at 7:57 PM
Most common AI finding we see: a model deployment that served zero requests in 30 days and billed anyway.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
Most common AI finding we see: a model deployment that served zero requests in 30 days and billed anyway.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
September 24, 2026 at 1:51 PM
Most common AI finding we see: a model deployment that served zero requests in 30 days and billed anyway.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
Someone ran a POC. The project moved on. The deployment did not.
No quality trade-off, no code change, no A/B test. Just delete it.
Cleanest money on any AI bill.
Terminate an EC2 instance and any volume with DeleteOnTermination=false survives it.
It drops to "available", attached to nothing, billed in full - EBS bills provisioned capacity, not usage.
A forgotten 500GB io2 with 5k IOPS is about $4,700 a year.
It drops to "available", attached to nothing, billed in full - EBS bills provisioned capacity, not usage.
A forgotten 500GB io2 with 5k IOPS is about $4,700 a year.
September 22, 2026 at 1:43 PM
Terminate an EC2 instance and any volume with DeleteOnTermination=false survives it.
It drops to "available", attached to nothing, billed in full - EBS bills provisioned capacity, not usage.
A forgotten 500GB io2 with 5k IOPS is about $4,700 a year.
It drops to "available", attached to nothing, billed in full - EBS bills provisioned capacity, not usage.
A forgotten 500GB io2 with 5k IOPS is about $4,700 a year.
Ask three clouds what a resource cost last month and you used to get three schemas, three definitions of "cost", three discount models.
FOCUS ends that. One billing schema, and every major cloud now emits it natively - Microsoft, AWS, Google, Oracle.
One query, any provider.
FOCUS ends that. One billing schema, and every major cloud now emits it natively - Microsoft, AWS, Google, Oracle.
One query, any provider.
September 11, 2026 at 12:57 PM
Ask three clouds what a resource cost last month and you used to get three schemas, three definitions of "cost", three discount models.
FOCUS ends that. One billing schema, and every major cloud now emits it natively - Microsoft, AWS, Google, Oracle.
One query, any provider.
FOCUS ends that. One billing schema, and every major cloud now emits it natively - Microsoft, AWS, Google, Oracle.
One query, any provider.
Before you panic about a cloud bill spike, check the date range.
The most recent 1-2 days of cost data are ESTIMATES, and they often revise down.
A good share of "our bill exploded" turns out to be unfinalized days.
Rule out the false alarm before you start the hunt.
The most recent 1-2 days of cost data are ESTIMATES, and they often revise down.
A good share of "our bill exploded" turns out to be unfinalized days.
Rule out the false alarm before you start the hunt.
September 11, 2026 at 12:57 PM
Before you panic about a cloud bill spike, check the date range.
The most recent 1-2 days of cost data are ESTIMATES, and they often revise down.
A good share of "our bill exploded" turns out to be unfinalized days.
Rule out the false alarm before you start the hunt.
The most recent 1-2 days of cost data are ESTIMATES, and they often revise down.
A good share of "our bill exploded" turns out to be unfinalized days.
Rule out the false alarm before you start the hunt.
Kubernetes is where cloud cost goes to hide.
You do not pay for pods. You pay for nodes.
The scheduler packs by what pods REQUEST, not what they use - so most clusters run at 30-40% real CPU while paying for 100% of the nodes.
That gap is the bill.
You do not pay for pods. You pay for nodes.
The scheduler packs by what pods REQUEST, not what they use - so most clusters run at 30-40% real CPU while paying for 100% of the nodes.
That gap is the bill.
September 11, 2026 at 12:57 PM
Kubernetes is where cloud cost goes to hide.
You do not pay for pods. You pay for nodes.
The scheduler packs by what pods REQUEST, not what they use - so most clusters run at 30-40% real CPU while paying for 100% of the nodes.
That gap is the bill.
You do not pay for pods. You pay for nodes.
The scheduler packs by what pods REQUEST, not what they use - so most clusters run at 30-40% real CPU while paying for 100% of the nodes.
That gap is the bill.
Azure Cost Management tells you what you SPENT.
It does not tell you what you WASTED - and that is not a flaw, it is not its job.
It will never surface an orphaned disk, an idle gateway or missing Hybrid Benefit.
Budgets and alerts there on day one. Then go looking.
It does not tell you what you WASTED - and that is not a flaw, it is not its job.
It will never surface an orphaned disk, an idle gateway or missing Hybrid Benefit.
Budgets and alerts there on day one. Then go looking.
September 11, 2026 at 12:56 PM
Azure Cost Management tells you what you SPENT.
It does not tell you what you WASTED - and that is not a flaw, it is not its job.
It will never surface an orphaned disk, an idle gateway or missing Hybrid Benefit.
Budgets and alerts there on day one. Then go looking.
It does not tell you what you WASTED - and that is not a flaw, it is not its job.
It will never surface an orphaned disk, an idle gateway or missing Hybrid Benefit.
Budgets and alerts there on day one. Then go looking.
"We stopped those VMs months ago."
Stopping a VM stops the compute bill. It does not stop the disk bill.
Every disk on every deallocated VM is still billing you at full rate, for capacity nobody has read since.
Deallocated is not deleted.
Stopping a VM stops the compute bill. It does not stop the disk bill.
Every disk on every deallocated VM is still billing you at full rate, for capacity nobody has read since.
Deallocated is not deleted.
September 11, 2026 at 12:56 PM
"We stopped those VMs months ago."
Stopping a VM stops the compute bill. It does not stop the disk bill.
Every disk on every deallocated VM is still billing you at full rate, for capacity nobody has read since.
Deallocated is not deleted.
Stopping a VM stops the compute bill. It does not stop the disk bill.
Every disk on every deallocated VM is still billing you at full rate, for capacity nobody has read since.
Deallocated is not deleted.
The AI pillar is where a CCoE's arbitration job gets sharpest, because the trade-offs are unfamiliar:
A cheaper model that fails more often can cost more overall.
Self-hosting shifts cost from marginal to fixed.
A stricter data policy may rule out the best-performing provider.
A cheaper model that fails more often can cost more overall.
Self-hosting shifts cost from marginal to fixed.
A stricter data policy may rule out the best-performing provider.
September 11, 2026 at 12:54 PM
The AI pillar is where a CCoE's arbitration job gets sharpest, because the trade-offs are unfamiliar:
A cheaper model that fails more often can cost more overall.
Self-hosting shifts cost from marginal to fixed.
A stricter data policy may rule out the best-performing provider.
A cheaper model that fails more often can cost more overall.
Self-hosting shifts cost from marginal to fixed.
A stricter data policy may rule out the best-performing provider.
Start a new CCoE with cost. Not because it matters most - because it has the least ambiguous scoreboard.
Nobody argues about whether the bill went down.
Security and reliability wins are real and hard to prove. Cost wins buy you the credibility to go after them.
Nobody argues about whether the bill went down.
Security and reliability wins are real and hard to prove. Cost wins buy you the credibility to go after them.
September 11, 2026 at 12:54 PM
Start a new CCoE with cost. Not because it matters most - because it has the least ambiguous scoreboard.
Nobody argues about whether the bill went down.
Security and reliability wins are real and hard to prove. Cost wins buy you the credibility to go after them.
Nobody argues about whether the bill went down.
Security and reliability wins are real and hard to prove. Cost wins buy you the credibility to go after them.
FinOps vs CCoE, since they get conflated:
FinOps owns one pillar in depth - cost, and the business value of spend.
A CCoE owns the operating model, and is where cost trades off against reliability, security and speed.
Small org: same people. Large org: FinOps sits inside it.
FinOps owns one pillar in depth - cost, and the business value of spend.
A CCoE owns the operating model, and is where cost trades off against reliability, security and speed.
Small org: same people. Large org: FinOps sits inside it.
September 11, 2026 at 12:53 PM
FinOps vs CCoE, since they get conflated:
FinOps owns one pillar in depth - cost, and the business value of spend.
A CCoE owns the operating model, and is where cost trades off against reliability, security and speed.
Small org: same people. Large org: FinOps sits inside it.
FinOps owns one pillar in depth - cost, and the business value of spend.
A CCoE owns the operating model, and is where cost trades off against reliability, security and speed.
Small org: same people. Large org: FinOps sits inside it.
The real test of cloud maturity isn't whether you can scale up.
It's whether scaling something DOWN is routine rather than exceptional.
Most orgs have a well-worn path for growing a resource and no path at all for shrinking one.
It's whether scaling something DOWN is routine rather than exceptional.
Most orgs have a well-worn path for growing a resource and no path at all for shrinking one.
September 11, 2026 at 12:53 PM
The real test of cloud maturity isn't whether you can scale up.
It's whether scaling something DOWN is routine rather than exceptional.
Most orgs have a well-worn path for growing a resource and no path at all for shrinking one.
It's whether scaling something DOWN is routine rather than exceptional.
Most orgs have a well-worn path for growing a resource and no path at all for shrinking one.
Performance efficiency and cost look like the same pillar. They aren't.
Performance asks: are resources matched to demand?
Cost asks: are we paying the best rate for what we use?
A workload can be perfectly sized and still overpriced, because nobody bought a commitment.
Performance asks: are resources matched to demand?
Cost asks: are we paying the best rate for what we use?
A workload can be perfectly sized and still overpriced, because nobody bought a commitment.
September 11, 2026 at 12:53 PM
Performance efficiency and cost look like the same pillar. They aren't.
Performance asks: are resources matched to demand?
Cost asks: are we paying the best rate for what we use?
A workload can be perfectly sized and still overpriced, because nobody bought a commitment.
Performance asks: are resources matched to demand?
Cost asks: are we paying the best rate for what we use?
A workload can be perfectly sized and still overpriced, because nobody bought a commitment.
Ask five teams which tier their service is.
All five will say tier one.
That's why workload tiering can't be self-assessed. Without someone owning the classification, redundancy gets applied uniformly and expensively, or inconsistently and dangerously.
Usually both.
All five will say tier one.
That's why workload tiering can't be self-assessed. Without someone owning the classification, redundancy gets applied uniformly and expensively, or inconsistently and dangerously.
Usually both.
September 11, 2026 at 12:53 PM
Ask five teams which tier their service is.
All five will say tier one.
That's why workload tiering can't be self-assessed. Without someone owning the classification, redundancy gets applied uniformly and expensively, or inconsistently and dangerously.
Usually both.
All five will say tier one.
That's why workload tiering can't be self-assessed. Without someone owning the classification, redundancy gets applied uniformly and expensively, or inconsistently and dangerously.
Usually both.
Two ways a CCoE fails:
Publish standards without paving the road -> documents nobody reads.
Become an approval gate -> a queue teams route around.
The working version makes the compliant path the easiest path. A landing zone nobody uses is a project, not an operating model.
Publish standards without paving the road -> documents nobody reads.
Become an approval gate -> a queue teams route around.
The working version makes the compliant path the easiest path. A landing zone nobody uses is a project, not an operating model.
September 11, 2026 at 12:53 PM
Two ways a CCoE fails:
Publish standards without paving the road -> documents nobody reads.
Become an approval gate -> a queue teams route around.
The working version makes the compliant path the easiest path. A landing zone nobody uses is a project, not an operating model.
Publish standards without paving the road -> documents nobody reads.
Become an approval gate -> a queue teams route around.
The working version makes the compliant path the easiest path. A landing zone nobody uses is a project, not an operating model.
"We have a policy for that."
Where does it run?
A control that only exists in a spreadsheet is a preference, not a guardrail.
If it isn't deny-by-default at deployment time, you don't have a control. You have a strongly worded opinion and a quarterly audit.
Where does it run?
A control that only exists in a spreadsheet is a preference, not a guardrail.
If it isn't deny-by-default at deployment time, you don't have a control. You have a strongly worded opinion and a quarterly audit.
September 11, 2026 at 12:53 PM
"We have a policy for that."
Where does it run?
A control that only exists in a spreadsheet is a preference, not a guardrail.
If it isn't deny-by-default at deployment time, you don't have a control. You have a strongly worded opinion and a quarterly audit.
Where does it run?
A control that only exists in a spreadsheet is a preference, not a guardrail.
If it isn't deny-by-default at deployment time, you don't have a control. You have a strongly worded opinion and a quarterly audit.
A CCoE is not six specialists advocating for six pillars.
It's the one place those pillars are allowed to argue.
Reliability wants another replica. Cost wants fewer. Security wants an inspection layer; performance wants it gone.
Arbitration, not evangelism.
It's the one place those pillars are allowed to argue.
Reliability wants another replica. Cost wants fewer. Security wants an inspection layer; performance wants it gone.
Arbitration, not evangelism.
September 11, 2026 at 12:53 PM
A CCoE is not six specialists advocating for six pillars.
It's the one place those pillars are allowed to argue.
Reliability wants another replica. Cost wants fewer. Security wants an inspection layer; performance wants it gone.
Arbitration, not evangelism.
It's the one place those pillars are allowed to argue.
Reliability wants another replica. Cost wants fewer. Security wants an inspection layer; performance wants it gone.
Arbitration, not evangelism.
Your AI cost per request is lying to you.
Not by much, if things are healthy. By a lot, if they aren't - and the number looks identical either way.
Here's the metric that doesn't lie.
Not by much, if things are healthy. By a lot, if they aren't - and the number looks identical either way.
Here's the metric that doesn't lie.
September 11, 2026 at 12:48 PM
Your AI cost per request is lying to you.
Not by much, if things are healthy. By a lot, if they aren't - and the number looks identical either way.
Here's the metric that doesn't lie.
Not by much, if things are healthy. By a lot, if they aren't - and the number looks identical either way.
Here's the metric that doesn't lie.
The most expensive thing in an AI estate isn't what serves the most traffic.
It's the GPU endpoint someone spun up for a POC in March, still holding 2 replicas, serving nothing.
Replicas provisioned, zero predictions. Four figures a month.
It's the GPU endpoint someone spun up for a POC in March, still holding 2 replicas, serving nothing.
Replicas provisioned, zero predictions. Four figures a month.
September 11, 2026 at 12:47 PM
The most expensive thing in an AI estate isn't what serves the most traffic.
It's the GPU endpoint someone spun up for a POC in March, still holding 2 replicas, serving nothing.
Replicas provisioned, zero predictions. Four figures a month.
It's the GPU endpoint someone spun up for a POC in March, still holding 2 replicas, serving nothing.
Replicas provisioned, zero predictions. Four figures a month.
You measure cost per API call.
Your business gets value from successful outputs.
If 12% of your calls fail, those are two different numbers - and only one of them means anything.
Almost no dashboard computes the second one.
Your business gets value from successful outputs.
If 12% of your calls fail, those are two different numbers - and only one of them means anything.
Almost no dashboard computes the second one.
September 11, 2026 at 12:47 PM
You measure cost per API call.
Your business gets value from successful outputs.
If 12% of your calls fail, those are two different numbers - and only one of them means anything.
Almost no dashboard computes the second one.
Your business gets value from successful outputs.
If 12% of your calls fail, those are two different numbers - and only one of them means anything.
Almost no dashboard computes the second one.
If someone sells you "complete AI cost visibility", ask about self-hosted models.
An open-weight LLM on your own endpoint emits no token metrics. Not to Azure Monitor, CloudWatch, or Cloud Monitoring. The counts live inside your model server.
Cost is measurable. Tokens aren't.
An open-weight LLM on your own endpoint emits no token metrics. Not to Azure Monitor, CloudWatch, or Cloud Monitoring. The counts live inside your model server.
Cost is measurable. Tokens aren't.
September 11, 2026 at 12:47 PM
If someone sells you "complete AI cost visibility", ask about self-hosted models.
An open-weight LLM on your own endpoint emits no token metrics. Not to Azure Monitor, CloudWatch, or Cloud Monitoring. The counts live inside your model server.
Cost is measurable. Tokens aren't.
An open-weight LLM on your own endpoint emits no token metrics. Not to Azure Monitor, CloudWatch, or Cloud Monitoring. The counts live inside your model server.
Cost is measurable. Tokens aren't.
I ran my own FinOps tool against my own Azure OpenAI account.
It flagged my prompt cache hit rate as too low.
Wrong. 22 requests in 30 days, and Azure's cache clears after 5-10 min idle. Nothing was ever going to cache.
It was measuring cadence, not misconfiguration.
It flagged my prompt cache hit rate as too low.
Wrong. 22 requests in 30 days, and Azure's cache clears after 5-10 min idle. Nothing was ever going to cache.
It was measuring cadence, not misconfiguration.
September 11, 2026 at 12:47 PM
I ran my own FinOps tool against my own Azure OpenAI account.
It flagged my prompt cache hit rate as too low.
Wrong. 22 requests in 30 days, and Azure's cache clears after 5-10 min idle. Nothing was ever going to cache.
It was measuring cadence, not misconfiguration.
It flagged my prompt cache hit rate as too low.
Wrong. 22 requests in 30 days, and Azure's cache clears after 5-10 min idle. Nothing was ever going to cache.
It was measuring cadence, not misconfiguration.
Built AI cost tooling across 3 clouds. They can't agree what a "request" is:
Bedrock - Invocations counts successes only
Azure OpenAI - the total includes failures
Vertex AI - split by response_code
Same question, three denominators. Get it wrong and cost-per-call is fiction.
Bedrock - Invocations counts successes only
Azure OpenAI - the total includes failures
Vertex AI - split by response_code
Same question, three denominators. Get it wrong and cost-per-call is fiction.
September 11, 2026 at 12:47 PM
Built AI cost tooling across 3 clouds. They can't agree what a "request" is:
Bedrock - Invocations counts successes only
Azure OpenAI - the total includes failures
Vertex AI - split by response_code
Same question, three denominators. Get it wrong and cost-per-call is fiction.
Bedrock - Invocations counts successes only
Azure OpenAI - the total includes failures
Vertex AI - split by response_code
Same question, three denominators. Get it wrong and cost-per-call is fiction.