A shrinking compute bill can look like a win unless the money spent on it doesn’t deliver a return on investment.
Every FinOps report can show what a team spent against budget. Anderson Oliveira wants them to also tell how much value was created.
That knowledge gap creates more risk when it comes to the potentially massive compute costs created by AI initiatives. AI projects are more than just another line item on the balance sheet. These are high visibility endeavors that need to clearly produce return on investment (ROI).
Oliveira, author of Ultimate FinOps Earned Value Management, has spent more than 25 years navigating major shifts in enterprise IT. On The Compute Layer, Oliveira explained why cost visibility keeps getting harder to maintain, and why measuring cost alone is the wrong approach to managing AI projects.
This article is for the FinOps or platform leader whose AI spend has outpaced the tools built to track it, and for the finance team trying to differentiate a healthy compute spend versus one that has run amok.
Cost cycles used to have a lot more slack than today. AI requires us to measure costs in shorter cycles — much shorter. In the before times of traditional IT finance, a missed monthly invoice would be caught the next month. The cost would not be that much of a surprise.
"In the old project management times where we were running costs on a monthly basis, if we missed an invoice, that would be something we could recover."
FinOps tightened that cycle to daily. AI closes it entirely, because the failure mode is no longer a missed invoice. It can happen over the weekend.
"We're listening to very frightening stories, like people on a pet project or just trying some prompts spend thousands of dollars in a weekend. I heard the other day someone that spent fifty thousand dollars, almost the cost of an FTE or half an FTE, in one weekend."
Oliveira's take on what comes next is pointed.
Now AI is leading us to, ‘I need to see cost in real-time.’ Absolutely real-time, and that's the big change.
🚀 Bring your biggest pipeline. Run it on Falcon in your own environment with a free 2–4 week Test Flight, and see what the same workload costs on a fraction of the compute.
On IT projects, you need to manage both AI and cloud cost. Part of what makes AI cost harder to manage is that it does not respond to the same controls as traditional compute. Compute is parametric; use more, pay more. Technical measures can shape that curve. AI costs are not so easily correlated to volume as it shows up on the dashboard. The correlation, Oliveira says, is on the other side of the glass.
The lever is on the mind of the people, is on the behavior of the people, is how they think about their job.
That reframes the optimization problem. A dashboard alone cannot manage AI cost. People implementing AI need to understand when a cheaper model is enough and when a more expensive one saves three rounds of iteration.
Managing AI and cloud spend the same way you manage any other budget line feels natural. Oliveira argues it is also where most organizations go wrong.
When we look just at costs, we set ourselves up for failure, not only at the beginning but in the long run.
A rising bill is not automatically bad news. A system that is well built and elastic will cost more as demand grows. If the value produced outpaces the spend, that is value — though he emphasizes that is not always a profit that shows up in dollars.
📣 Want more value out of the same compute budget? Run your heaviest pipeline side by side in your own environment with a free Haevek Test Flight.
Oliveira tells a story from his book about a team that looked like a clear success. They tracked costs against budget in the normal way, and they were spending less than planned. Everyone was pleased.
Then the team added a third line: earned value. Alongside the plan and the actual spend, they measured how much value the work was producing and translated it into monetary terms.
The team was under budget but underdelivering on value. The issue was caught in the middle of a financial cycle, so the team reworked their architecture and process, reversing the trend before the year closed out.
Looking only at the budget line answers whether you overspent the budget. What you need to know is if the spend was worth it.
Oliveira's framework asks for two things most traditional systems are not built to give. One, cost visibility fast enough to catch a fifty thousand dollar weekend before it happens. Two, a value measurement precise enough to catch a team that is under budget but also under-delivering.
Haevek’s Falcon is built to create efficient infrastructure stacks that are a boon to FinOps looking to get more value for budget. It runs the same batch, streaming, and AI/ML workloads at up to 90% less compute and memory, changing the ratio between what a workload costs and what it can prove in value.
Capacity-based annual licensing, priced by maximum concurrent cores with no burst pricing, eliminates unpredictable compute costs and provides finance with a reliable number they can actually plan against.
None of that replaces the value discipline Oliveira describes. It gives a team more room inside the budget to build it.
A smaller compute and AI bill is not a complete business goal. Measure the value and track the spend in relationship to it.
The fix is a value line next to the cost line, tracked close to real time.
Two to four weeks, in your environment or ours, benchmarked side by side.