# 98% of FinOps Teams Now Manage AI Spend. Two Years Ago It Was 31%

> AI spend went from a side concern to nearly universal in two years, and teams are being told to fund it out of cloud savings that have already been harvested.

Two years ago, 31% of FinOps practitioners managed AI spend. In the 2026 State of FinOps survey, it is **98%**.

That is not growth. That is a category arriving all at once. And the same report contains the fact that makes it awkward: the easy cloud savings are already gone.

## The number in context

The FinOps Foundation's 2026 report is its sixth annual survey, drawing on 1,192 respondents who between them represent more than $83 billion in annual cloud spend. So this is not a vendor poll of a hundred people.

Inside that population, managing AI spend went from a minority activity to a near-universal one in twenty four months. And AI cost management is now the **number one skillset teams say they need to develop**, which is a polite way of saying almost nobody currently has it. The discipline is being invented in real time by the people practising it, which is exactly the phase where bad habits calcify.

## Why AI spend breaks the FinOps playbook

The classic cloud cost problem is waste. An instance nobody turned off. An environment somebody spun up for a demo in 2023. A cluster sized for a traffic spike that never came. The work is to find the idle thing and switch it off, and the emotional character of that work is pleasant, because nobody defends a machine doing nothing.

AI spend is different in a way that matters. Almost none of it is idle. It is consumption that somebody actively wanted: a feature calling a model, a retrieval pipeline embedding documents, an agent looping until the task is done. There is no idle GPU to reclaim and no forgotten environment to kill.

Which means the question changes from **"what is wasted"** to **"is this worth what it costs."** That is not an infrastructure question. It is a product question, and it has to be answered by people who can say whether a feature earns its inference bill. Most organisations have no forum where that conversation happens, so it does not happen, and the bill grows.

## The trap in the same report

Here is where it gets uncomfortable, and this is the finding I would put in front of a CFO.

The Foundation reports that many organisations are being asked to **self-fund AI investments through optimization savings**. Find the waste, use the savings to pay for AI. It is a tidy story and every executive likes it.

The same report also says, in its own words, that teams have **hit the "big rocks" of waste** and now face a high volume of smaller opportunities that require more effort to capture.

Read those together. The plan is to fund AI with cloud savings, and the cloud savings have largely been harvested. What remains is a long tail of small, effortful optimizations that cost engineering time to capture. You can absolutely still find money there. You will just spend real money finding it, and the yield per hour is falling.

So the honest position for most companies is that **AI is a new cost, not a free one financed by yesterday's efficiency drive**. Budget it as a new cost. The alternative is a plan that quietly assumes a savings pipeline that has already been drained, and those plans fail in Q4 rather than in the planning meeting.

## FinOps also stopped being about the cloud bill

One more shift worth naming, because it reframes what the function is.

The 2026 data shows scope expanding well past public cloud. Around 90% of practices now manage SaaS spend or plan to within the coming year, up sharply. 64% manage licensing, 57% private cloud, 48% data centre.

FinOps is quietly becoming the discipline of technology spend in general, rather than the discipline of the AWS invoice. If you run a SaaS platform, that is the more useful framing anyway: your gross margin is made of infrastructure, third-party SaaS, licences, and now model calls, and optimizing one of those four while ignoring the rest is how a business with healthy-looking cloud costs still has terrible unit economics.

## What to actually do

The practices that work are unglamorous and they start with visibility, because you cannot govern what you cannot attribute.

- **Attribute AI spend to features, teams, and customers.** Most organisations can tell you their total model bill and nothing else. That is the same maturity level as knowing your total AWS bill in 2015. Until you can say which feature and which customer drove the cost, every conversation about it is vibes.
- **Stop defaulting every call to the most capable model.** Route by task. Classification, extraction, and routing rarely need your best model, and the price difference across a tier is not subtle. This single habit is usually the largest available saving and it changes nothing the user can perceive.
- **Cache like you mean it.** A surprising share of production model traffic is the same question asked again.
- **Cap the loops.** An agent that retries until it succeeds is, from a billing perspective, an agent that retries until you notice. Put a ceiling on it.
- **Give every AI feature a cost budget, the way you give it a latency budget.** If a feature cannot justify its inference cost, that is a product decision, and it should be made deliberately rather than discovered in the invoice.

## The summary

The industry went from barely tracking AI spend to nearly all tracking it, in two years, without building the skills to do it, while being told to pay for it with savings that no longer exist.

That is a normal way for a new cost centre to arrive. It is not a reason for panic, but it is a reason to stop treating AI spend as an infrastructure rounding error and start treating it as a line item with an owner, an attribution model, and a product manager who can defend it.

If you are running AI in production and cannot yet say which features and customers are driving the bill, that is where the work starts, and it is work our [Cloud Infrastructure](/services/cloud-infrastructure) practice does. Tell us what you are running and [book a discovery call](/contact). [Track'em](/work/trackem) is one build where that cost visibility had to be designed in rather than reported after the fact. The architectural half of this decision, self-hosting an open model against paying per token, is covered in [open-source LLMs for your product](/blog/open-source-llms-for-your-product).

_Part of our guide to [AI security](/blog/ai-security)._

_All figures from the FinOps Foundation's State of FinOps 2026 report (1,192 respondents, $83bn+ in annual cloud spend). Verified against the primary source on 12 July 2026._