FinOps in the AI Era: A Critical Recalibration

Decision-makers with skin in the game

89% hold Director-level titles or above, including 36% at C-level. 92% have direct responsibility for cloud cost management. When they report on efficiency or forecasting accuracy, they’re assessing their own performance.

Industry composition

Software dominates at 58%, with Financial Services at 23%.

Within software: B2B SaaS (49%), Cyber Security (22%), Vertical SaaS (18%), and AI-Native (11%). That AI-Native segment proves particularly revealing throughout.

Company scale

The sample tilts toward mature companies: 27% exceed $250M in annual revenue, another 31% fall between $50-250M, and only 18% are early stage, earning under $10M. Nearly half (45%) have engineering teams exceeding 100 developers, up sharply from 26% last year.

Funding & geography

Private equity-backed companies represent 59% of respondents, with public companies at 20% and VC-backed at 16%. This PE concentration likely elevates the baseline for financial discipline. Geography skews North American at 95%.

AI is already embedded

The overall profile also speaks from experience. 89% have dedicated AI development teams.

And, 91% leverage AI in their products.

The numbers

Formal cloud cost programs nearly doubled (39% to 72%). Budget assignment jumped to 87%. Chargeback adoption rose to 64%.

FinOps functions now exist at 80% of organizations. The industry invested in discipline, and it took hold.

Every sign of maturity, from adoption to standardization, improved.

The Cloud Efficiency Rate

Meanwhile, the Cloud Efficiency Rate (CER) is based on a formula to help organizations gauge how efficient they are with their cloud spend.

Here’s how it’s calculated: (Revenue – Cloud Costs) / Revenue = CER. The lower your cloud costs are as a percentage of your total revenue, the higher your CER is. 

We go into depth on this later in the report (Section 5), but right away, we found that the Cloud Efficiency Rate dropped across every segment and every quartile, including top performers who fell from 92% to 85%. If growth alone explained the pattern, we’d expect mature, stable organizations to hold steady. They didn’t.

What changed

One variable entered the equation at scale: AI. 40% of organizations now spend $10M or more annually on AI, approaching the 47% spending that much on cloud, and closing fast. 

AI spend is catching up to cloud, and it’s the fastest-growing cost category we’ve seen.

There’s more nuance to it, though. AI changes the composition of spend, not just the total. Traditional cloud workloads will map cleanly to production or non-production, while AI portfolios include variables such as training, inference experimentation, model evaluation, safety testing, observability, and anything that might be classified as “non-production”, even when essential to delivering value.

At first glance, a declining CER looks like waste. But dig deeper, and it may reflect front-loaded investment that precedes revenue. It also may not include all AI spend as not every organization folds AI into cloud — so overall CER may be even lower than what we’re seeing here.

That’s the core problem: Without granular visibility, you can’t tell the difference.

The management gap

There’s also a gap in managing those costs. Four out of five (78%) fold AI costs into overall cloud costs.

That isn’t inherently problematic if you have the granularity to distinguish what’s driving spend. But most organizations are merging totals without the visibility to understand what’s inside. They’re watching a single number grow without knowing why.

However, the distinction isn’t as simple as “cloud is capacity-based, AI is consumption-based.” Modern cloud already includes consumption-driven services like Lambda, Fargate, and SQS. And many AI workloads, especially training, do run on reserved capacity and behave predictably.

The big shift, then, is in dominant cost drivers. Cloud costs, even those consumption-based, scale with infrastructure decisions controlled by engineering teams. AI costs introduce variables that weren’t there before, such as token volume, model selection, prompt efficiency, and user behavior patterns. A customer asking longer questions or a product team choosing GPT-5.2 over a lighter model can shift the needle significantly in ways that provisioning decisions didn’t before.

Emphasis: despite (or perhaps because of) all these nuances, our dataset shows that most organizations manage both through the same lens. That’s fine, until the new variables start dominating the bill. That’s when diagnosis gets expensive, and often wrong.

The fix is allocation; attributing costs to specific products, customers, features, and teams. Without it, you have a number. With it, you have insight.

The intentional bet

Some important context is needed to interpret the efficiency decline.

AI costs aren’t always mismanaged. In many organizations, looser controls are intentional. Every platform shift demands front‑loaded investment before returns materialize — from the dot‑com era’s “growth over profits” to the cloud migration’s “cloud first, optimize later,” which ultimately gave rise to FinOps itself once cloud stabilized.

AI is now following that same pattern, at light speed. Organizations are investing aggressively because customers expect AI features, competitors are shipping them, and opting out isn’t viable. In many cases, this spend is deliberately unprofitable. It’s a strategic bet made ahead of revenue.As CloudZero CTO Erik Peterson shared on a 2025 podcast, one customer described AI as the fastest‑adopted feature in their company’s history, and simultaneously the least profitable. Leadership’s stance was clear: “There’s no way we’re turning this off.” That’s how platform shifts work.

But here’s the line that matters: Strategic unprofitability only works if it’s measured. 

You need to know how much you’re losing, where you’re losing it, and for how long. Many organizations in this dataset can’t answer those questions. That’s the difference between intentional inefficiency and unquantified leakage. The former is a strategy. The latter is a risk.

This distinction is critical when interpreting the Cloud Efficiency Rate. Some portion of the CER decline likely reflects disciplined, front‑loaded AI investment. Some of it reflects waste. But, without granular visibility and allocation, organizations can’t tell which is which — and the same metric can mask two very different realities.

That’s the paradox, and the case for shifting focus from metrics to mechanisms.

Cloud: established, consolidated, expensive

First: 93% now use a public cloud provider, up from 89% in 2024. AWS leads at 76%, with Azure at 62% and GCP at 60%. Many organizations work with multiple providers. Nearly half (47%) spend $10M or more annually on cloud.

That bill extends well beyond compute and storage: 81% pay for AI and Advanced Services through their cloud providers, 63% for Data as a Service, 56% for Security as a Service.

Cloud has standardized. Teams know their providers, understand pricing mechanics, and have built reliable processes to manage spend.

AI: fast-scaling, fragmented, unpredictable

AI is different. Where traditional cloud consolidated around the Big Three, AI is fragmented across deployment models, provider relationships, and consumption patterns.

Where do AI workloads run? 52% use hybrid (on-prem and cloud), 42% public cloud, 38% private cloud, 24% third-party GPU providers, and 21% hosted LLM APIs. Most organizations use more than one.

Even when looking at AI workload locations through a software category lens, it’s still complex.

AI workloads resist consolidation and multiply complexity as they scale. Cloud AI services (Bedrock, Vertex, Azure OpenAI), direct API relationships (OpenAI, Anthropic), self-hosted models, specialized GPU providers — most organizations use several simultaneously, mixing and matching based on use case, latency, and cost.

The fragmentation tax

The flip side? This fragmentation creates management overhead that cloud teams didn’t have to deal with.

For example, a typical cloud cost stack might include two or three providers with relatively similar billing structures (as we stated above, there’s standardization in this area). But an AI cost stack may involve cloud providers for some workloads, direct API relationships for others, GPU providers for training, and multiple LLM vendors depending on the use case.

It gets more complicated. Each of these AI costs come with unique billing and separate monitoring. 

Just how fragmented is the landscape already? Consider the top five AI types utilized:

Plus, each category may involve different providers, different billing structures, and different cost models — all generating data in different formats, at different intervals, with different levels of detail.

The cost stack is splintered, with no single source of truth. Organizations are pulling data from cloud provider consoles, LLM billing portals, GPU invoices, and internal tracking systems, then attempting to reconcile manually or through cobbled-together integrations.

The AI conundrum is this: Choose simplicity and sacrifice capability, or choose capability and inherit chaos. Without a unified view, allocation becomes guesswork and optimization becomes impossible.

The maturity surge

We highlighted in Section 2 how organizations matured significantly in cloud cost management practices year over year — not only a significant jump in formal cloud cost programs, but also high numbers in formalized FinOps practices. 

Four out of five organizations now run formal FinOps functions. Adoption scales with company size, from 60% under $10M to 90% above $250M. It’s also strongest in Financial Services (90%) and PE-backed companies (85%).

These are strong signals of adoption — but adoption alone isn’t translating into performance.

Cloud costs already flow to pricing

Cloud costs already inform pricing for 80% of organizations, while an additional 6% plan to within the next 12 months.

This varies by segment:

That last bullet tells us that companies with products most defined by AI are least likely to have integrated cloud costs into pricing. Perhaps their cost structures are still too volatile.

Company size also factors in; 70% of companies with <$10M use cloud costs in their pricing, scaling to 89% of $100M-$250M ARR companies. 

Pricing is one form of accountability, but who actually owns the costs? Let’s look at how accountability has shifted inside organizations.

Ownership has consolidated

So, if adoption and accountability are mainstream, who’s held accountable? IT now owns cloud costs at 43% of organizations, up 10 points from last year. FinOps as a function grew from 13% to 18% in the same period. Infrastructure teams dropped from 14% to 6%. Product dropped from 5% to 1%. Cloud cost ownership is centralizing in IT and FinOps.

FinOps owning FinOps makes sense. So why is IT increasingly taking the reins? It could reflect the growing overlap between infrastructure and cost. That’s a natural shift as cloud complexity increases. 

Or it could reflect something more structural: FinOps teams are hitting bandwidth limits, and IT steps in as AI spend blurs the line between architecture and economics. Or FinOps teams are folded into IT teams.

Either way, there’s a deeper issue here. Ownership is consolidating, but usage decisions still live in product and engineering (and other departments). The people accountable aren’t necessarily the ones driving usage, and that’s a visibility and accountability gap.

Regardless of who holds the budget, the pressure to reduce it is real. Here’s how organizations are trying to cut costs.

How organizations are cutting costs

This doesn’t mean cost-cutting efforts aren’t happening in FinOps. Organizations are doing it (or trying). Half are turning to commitment-based discounts, 50% are optimizing their architecture, and 47% are negotiating enterprise discounts. Those are clear levers for most.

But in the middle tier, we see movement. Right-sizing instances jumped 13 points from 30% to 43% year over year. Third-party resellers climbed from 33% to 44%. And spot instance usage grew from 25% to 31%.

Code optimization remained flat (28% to 29%), despite proven outsized impact. At Kubecon 2025, OpenAI’s Fabian Ponce shared how a single line of code change cut CPU usage by half in their Fluent Bit platform. We repeat: That’s a 50% reduction from one fix. Engineering efficiency wins tend to be multiplicative where commercial levers are incremental. 

Why isn’t code optimization more widely adopted? It requires engineering investment when roadmaps are packed. It’s harder to measure without granular visibility into how services consume resources. And the wins are invisible until someone goes looking. Commercial levers are easy to pull. But in AI-heavy environments, the real leverage lives in the code.

It’s worth underscoring this: Engineering efficiency remains the least-utilized tactic despite offering the highest returns. The implication is sharp. In AI-heavy environments, FinOps process maturity without engineering leverage is insufficient. Organizations can perfect their budgets, chargebacks, and governance, and still lose the efficiency battle if the cost drivers live in code, prompts, and model selection.

At the same time, some organizations are going further: rethinking the very foundations of their infrastructure choices. A full 82% are either planning or evaluating repatriation for non-AI workloads (54% planning, 28% evaluating). Yes, that’s intent, not execution. But the signal goes beyond cost pressure, and reflects dissatisfaction with cloud-era abstractions for high-throughput, unit-economics-sensitive workloads.

That’s significant. Cost optimization is shifting down the stack from commercial levers and contract negotiations to infrastructure decisions and systems design. Repatriation carries its own costs. The math won’t work for everyone. But clearly, the optimization frontier has moved. Contracts and tooling aren’t enough. Architecture is back on the table.

What gets measured

Organizations are more closely tracking cloud costs against financial benchmarks than they were in 2024. Two thirds (63%) are measuring against revenue versus 55% a year earlier, while 43% measure against COGS, up 8 points from 35%. A little more than a third (36%) are measuring against R&D, up from 29%.

The jump in COGS measurement is significant. Cloud is increasingly understood as a direct input to product delivery, a core component of gross margin. For organizations where cloud costs are COGS, efficiency then means margin management.

Those trends suggest progress. But tracking against revenue or COGS is still relatively high-level. The real question is: How deep does that visibility go?

Granularity remains shallow

Granularity is a must for effective optimization. Without it, you’re measuring aggregate spend but don’t know whether individual customers or product lines are profitable.

That begs the question: How specifically are organizations measuring costs, and how?

We found that 55% track by product, 53% by business unit, and 48% by development team. These are expected, and are reasonable for infrastructure management.

Yet, just 43% track by customer, and 22% track by transaction. 

For organizations selling cloud-delivered products, customer-level visibility tells you which accounts are profitable. Without transaction-level data, you can’t tell whether your pricing model works or where it’s breaking (more on that in Section 6 under “The Pricing Paradox”). 

Those gaps are concerning, but they’re also understandable. Achieving this level of granularity for AI is much harder than it was for cloud. Attribution is structurally harder with shared models, pooled inference, and reused features. Frankly, you don’t just tag a server; you have to trace tokens through systems designed for efficiency, not observability.

The root problem isn’t tooling — it’s architecture. That’s whether systems were designed for traceability or not. Those closing the granularity gap are making deliberate architectural choices that enable accurate allocation. That’s engineering leverage again. While achievable, it’s not something that happens by accident.

Granularity levels differ across segments. For Financial Services, 63% track by business unit and 30% by transaction compared with 49% and 19% for Software.

AI-Native companies lead in customer tracking (58%) and B2B SaaS falls behind at 36% for the same.

Granularity and efficiency are linked. Organizations that can’t see costs at the customer and transaction level can’t diagnose what’s dragging down performance. With that context, let’s look at where efficiency actually landed.

Efficiency by segment

Let’s revisit the Cloud Efficiency Rate (CER) in depth. The CER measures the percentage of revenue spent on the cloud. A CER of 80% means 20% of your revenue goes to cloud bills — in other words, you want that CER to be as high as possible. In 2024, the median was 80%, top performers hit 92%, and the 25th percentile was 70%. This year, all three dropped: median to 65%, 75th percentile to 85%, 25th percentile to 45%.

CER needs more context before we get into the segmentation. First, the production value definition made sense for cloud. Provisioned capacity either serves production or it doesn’t. But AI introduces a third category: the training layer. 

Consider an ML-driven lending platform spending millions monthly on model simulation and training. That spend doesn’t serve production directly, but it is what makes production work. The cycle runs: from development to training to production, with a feedback loop from production back to training that continuously improves the model.

For organizations like this, a 55% CER might reflect disciplined investment in the training layer, not waste. CER still matters, but it wasn’t built for AI workloads where non-production spend drives production value. It’s just an indicator that we may need more granular ways to measure efficiency as well as spend attribution. Distinguishing training layer spend from true waste would also give organizations a clearer picture of where optimization opportunities exist. 

With that context, let’s look at how CER breaks down across segments — and where the gaps hit hardest.

By industry: 

Financial Services, despite higher FinOps maturity (remember, 90% have a dedicated FinOps function vs. 77% for Software vs. 80% overall), their median CER falls behind at 55% versus 66% for Software and 65% overall.

It’s worth understanding why the most FinOps-developed sector might still report the lowest efficiency: First, regulatory and compliance overhead may require non-production environments that inflate the denominator. It’s finance, after all.

Also, finance may have more stringent measures. If you’re better at seeing waste, you’ll report more of it. Or, perhaps, the complexity of financial workloads may simply demand more infrastructure that doesn’t map cleanly to “production”.

By software category:

Vertical SaaS leads with 70% median and 98% at the top quartile. AI-Native clusters in a narrow, lower band.

The pattern: Segments with the most complex cost structures show the lowest efficiency. Granularity gaps hit hardest where costs are hardest to untangle.

By company size:

The smallest companies (<$10M) achieve 70% median with 90% at top quartile. Largest companies (>$250M) show a troubling floor at 34%.

Scale creates complexity. Without granularity to allocate costs across hundreds of products, teams, and customers, waste hides.

Budgets exist, visibility doesn’t

A majority set aside budget for AI costs (85%).

PE-backed companies are higher at 92%, while VC-backed and public companies are behind at 78% and 75% respectively.

But when asked to rank their top three AI cost challenges, the responses are striking:

Visibility tops the list. The same problem FinOps solved for cloud a decade ago is now the top challenge for AI. Organizations have budgets but can’t see clearly what’s happening within them.

This is more than a tooling issue. It’s a signal of where organizations actually are in the AI cost maturity curve. They’re still diagnosing. And when the diagnosis is unclear, treatment becomes guesswork. You can’t optimize what you can’t see, and right now, most can’t see clearly. 

This means organizations are in diagnose-then-cure mode. When the diagnosis is clear, the cure emerges. But AI costs continue to scale while visibility lags behind.

So what does visibility look like today? In most cases: fragmented, improvised, and incomplete.

Monitoring is fragmented

Regarding visibility: Two-thirds (63%) use internal dashboards for AI cost monitoring, while 57% rely on billing tools provided by AI/LLM services. Nearly half (47%) use third-party FinOps platforms and 30% are still working off manual spreadsheets.

The overlap between approaches is telling. Most organizations are using multiple tools because no single tool gives them the complete picture that they want or need. They’re stitching together dashboards, consoles, spreadsheets, and third-party tools in a multisegmented setup without unified, single-source visibility or allocation back to business value.

Worth noting: AI-Native companies lead on internal dashboards (76%) but lag on provider tools (44%). They’re building their own solutions perhaps because off-the-shelf isn’t sufficient.

But even with all that instrumentation, many organizations still find out about overages the hard way.

Overages surface late

How do organizations discover AI cost overages?

That’s a full third of organizations who don’t know they’ve overspent until they receive the bill. That’s a serious control gap. 

The differences stand out at the funding level: PE-backed companies detect overages in real-time at 71%, compared with VC-backed companies at 45%. That 26-point gap represents the difference between managing AI costs and discovering them after the fact.

Beyond detection, there’s also the challenge of prediction. And here, the numbers tell an even tougher story.

Forecasting accuracy is poor across the board

Tracking and monitoring are crucial, but so is predictability. Only 20% forecasted their AI spend within ±10% of actual spend. Another 54% missed the mark by 11-25%. 

And one in five overshot their AI spend predictions by 50% or more.

When segmented, B2B SaaS has the worst precision, with only 8% forecasting accurately. AI-Native companies have the worst tail risk: 36% miss by 50%+ despite having the best real-time monitoring (72%) and the highest customer-level tracking (58%).

This tells us AI-Native companies have invested the most in visibility infrastructure and still can’t forecast reliably. They aren’t struggling with tooling or best practices; rather, it’s a volatility problem.

Why? Both explanations are likely true: They’re measuring something inherently more volatile, and they’re further out on the AI cost curve than anyone else.

AI-Native companies went deeper, faster. Their AI spend is larger, more central to their products, and more exposed to consumption variability. They’re not failing at forecasting; rather, they’re just trying to forecast the unpredictable. The monitoring infrastructure they’ve built will matter once patterns stabilize. Right now, they’re simply charting unmapped territory. AI consumption economics may remain difficult to forecast even with strong visibility, due to feedback loops between usage, UX, and cost.

Even with real-time monitoring, most are just watching the meter run without the granularity to see what’s driving it or where it’s going.

And yet, despite this limited visibility, organizations are already making pricing decisions based on those same costs. That’s where a deeper disconnect begins.

The pricing paradox

Organizations can’t see their AI costs clearly, but they’re pricing them anyway.

Visibility without allocation is watching the meter run. Without that connection, there’s no unit economics. Just aggregate trends. There’s a big difference between strategic pricing under uncertainty and pricing blind. 

We asked how organizations are pricing AI, and found that most are moving fast even if their visibility hasn’t caught up.

How organizations are pricing AI

Despite the visibility gaps, pricing models are emerging:

Not yet factored: 4%

Cost-plus pricing — where AI costs directly influence what customers pay — is the dominant model. And combining cost-plus with base pricing, we see 84% of organizations factoring AI costs into pricing while just 43% track costs by customers and 22% track by transaction. 

Organizations know AI spend isn’t optional, so it has to be priced in or absorbed somehow. But without customer-level visibility, they’re pricing on vibes, not unit economics.

Segment patterns in AI pricing

How AI costs flow to pricing varies widely by segment. AI-Native companies price most aggressively: 76% use cost-plus. AI is their product, so the cost-to-price relationship is direct and unavoidable.

Vertical SaaS shows the most experimentation. Nearly one in five (18%) offer AI as an uncharged add-on, while 11% charge separately. They’re testing what the market will bear.

B2B SaaS sits in the middle, with 7% still not factoring AI costs at all — comfortable for now, but potentially exposed as AI becomes a larger share of their cost base.

The maturity slope by company size is steep. Just 38% of companies under $10M use cost-plus pricing versus 76% of those above $250M.

And 25% of companies with less than $20M annual revenue haven’t factored AI costs into pricing at all, compared with just 1% of those above a quarter million in revenue. They may lack the data infrastructure to understand unit economics, or they’re absorbing costs to compete. Either way, they’re building margin exposure into their models.

These pricing models reflect intent. But without insight, intent can break down and strategic unprofitability can turn into unmanaged loss.

Where strategy meets risk

We discussed this above, and it’s an important reminder: Not every organization is trying to optimize AI pricing today, and that’s not necessarily wrong. Many are prioritizing adoption and market position, accepting near‑term losses in exchange for long‑term upside. We’ve seen this before with other disruptive trends.

But strategic unprofitability still requires clarity. Organizations still need to understand how much they’re spending, what’s driving those costs, and how or when that spend is expected to convert into value. Without customer‑level attribution, those questions go unanswered and you’re flying blind.

This is where intent breaks down. Some of the 18% of smaller companies that haven’t factored AI costs into pricing may be making deliberate bets on growth. Others may simply lack the visibility required to build a pricing model at all. The data doesn’t distinguish between the two right now, but the outcomes will diverge quickly as the AI cost landscape matures.

When costs are visible and allocation is clear, organizations can afford to invest ahead of returns. When they aren’t, losses accumulate without direction. That’s why insight is key.