This article originally appeared in the Forbes Technology Council section of Forbes.com
In its 2025 State of FinOps report, the FinOps Foundation showed survey evidence that 63% of companies are actively managing AI spending, an increase of 31% from 2024. This shift is likely more than enough to make corporate finance, IT and procurement professionals a bit anxious. Many of today’s enterprises are either still struggling with managing cloud spend or are very sensitive to avoiding reliving episodes of sticker shock that accompanied many an invoice from their cloud service providers. If they have implemented and are following a FinOps program at their organization, they have hopefully wrangled much of the runaway costs of their multicloud estates.
With the arrival of AI, the specter of uncontrolled costs is once again threatening. Is FinOps going to be the solution that helps organizations avoid a repeat of the wild cost overruns of the cloud era? If unmanaged, AI will propel companies into a cost challenge even greater than when the cloud first hit. This time, however, we can be ready.
FinOps for AI has a number of factors that are similar to cloud, SaaS and UCaaS, such as the need to normalize how costs are named, captured, categorized, distributed, and, of course, how to go about proving ROI. But cloud and AI usage diverge based on a few different factors and behaviors, making AI costs land on organizations in some unpredictable ways.
Let’s take a closer look at how AI is creating a new age of rising, hard-to-control service costs and where organizations can apply FinOps principles and best practices to meet the moment.
AI Isn’t “Cloud,” And The Bill Proves It
While the promise of AI is speed, insight and automation, the reality for many enterprises is a growing cloud bill driven by the very innovation they hoped would save them money. Let’s examine some of the high-priced habits of AI workloads.
GPU-Driven Computing
GPUs have become the default for AI workloads, but they’re hard to get (demand far outweighs supply right now) and cost much more per hour to rent than regular central processing units (CPUs). On a cloud bill, this shows up as specialized, high-rate compute line items that can dwarf the cost of equivalent CPU time.
High-Volume Data Transfer & Storage
AI is like turning on a firehose inside your infrastructure: Constant, high-volume data flows unlike anything traditional cloud apps demand. Every extra drop comes with a price tag in GPU time, storage capacity and network bandwidth. On the bill, this can look like charges for premium storage tiers, high-throughput networking and data transfer between regions or availability zones.
AI Experimentation
Teams are spinning up pilots for everything from automated customer support to code generation, creating a sprawling ecosystem of AI workloads. This generates spiky, unpredictable consumption patterns that can trigger unexpected on-demand rates or overage fees. Unlike predictable SaaS subscriptions, AI experimentation often appears as sudden jumps in compute, storage and API usage from one billing cycle to the next.
Inference Versus Training
For AI to work the way it needs to, it must be fed massive datasets it can learn from. This training phase is resource-heavy and done in large bursts.
Then comes inference, when the trained AI puts its skills to work (answering questions, making intelligent recommendations, etc.). It’s a non-stop process where the meter is always running, so to speak. These costs can spiral quickly depending on what “flavor” of AI you’re using. Take agentic AI, where a single request can snowball into long, multistep workflows that can run indefinitely based on external inputs or outcomes.
AI APIs
An AI API (Application Programming Interface) is a fast way to add AI features without building your own, but every request adds to the tab. In billing terms, these are per-call or per-token charges that can pile up as adoption grows.
Provider-Powered AI Premiums
Public cloud providers sell their own AI services, from pre-trained models to custom AI pipelines. Using them can accelerate projects, but you’re essentially at the mercy of their pricing. Rates can shift quickly, and not always in your favor.
How FinOps Cracks The AI Cost Paradox
Nearly eight in 10 companies use AI, according to McKinsey, yet just as many report little to no bottom-line impact. Both can’t be true if AI is really the transformational force it’s claimed to be. McKinsey calls this the “AI paradox”: the result of a growing imbalance between AI “quick wins” that rarely pay off and deeper investments that produce measurable, long-term value. Funds often flow into broad applications of AI that don’t really move the needle, while the initiatives that could actually drive results get sidelined. This cloud cost disparity is exactly what FinOps was built to fix, now factoring in the complexities of AI.
Without question, a mature practice will prevent history from repeating.
Assign clear cost ownership. Continuously allocate spend across teams, projects and departments with granular visibility into AI versus cloud costs. This enables true cost ownership and reveals what is driving value (and what isn’t).
Link usage to outcomes. Tie GPU hours, API calls and associated cloud services directly to business objectives, ensuring AI isn’t just consuming but delivering measurable impact.
Control experimentation before it spirals. Forecast and benchmark usage so innovation happens responsibly, not at the expense of runaway budgets.
Prove ROI with precision. Normalize how costs are named, captured and reported so ROI can be tracked consistently over time instead of being guessed after the fact. Deliver the hard numbers that prove innovation is paying off or whether an initiative is worth the investment.
Your organization is likely using AI, but can you honestly say your investments are paying off? Consider what it would take to gain that visibility, and what you could unlock if you did. Will you work to gain control, or will it be déjà vu all over again?
Ready to ensure your innovation is financially sustainable? Talk to Tangoe.