The Economics of AI: What CFOs Must Know in 2027

96% of SaaS CFOs had at least one AI cost category exceed forecast in the past 12 months. Most of the misses came from compute, data storage and networking: the cloud bill sitting underneath the model bill. As Finance teams build their 2027 budgets, that is the part of AI spend most likely to surprise them again.

2027 budgets are being built on AI cost assumptions most Finance teams can’t fully see yet. In conversation with Ben Murray of The SaaS CFO, Ed Barrow covers three things: the total cost of AI, a three-layer model for measuring it, and how Finance can secure better rates without locking in today’s architecture.

Ben Murray, founder of The SaaS CFO, joined me to work through what this means for CFOs heading into budget season. If you’re preparing for 2027 budget conversations, this is one to watch with your CTO. Below the replay you’ll find the substance: the data, the framework, and the moves Finance can make now. The complete transcript is at the end.

The scope of the problem

Together with Sapio Research, we surveyed 200 US and UK SaaS CFOs at companies with 50 to 1,000 employees. The results are published in The Cost of Compute: AI Edition and The Cost of Compute 2026.

• Compute is now a major share of revenue. Median compute spend runs 16 to 20% of revenue, and 31% of companies spend more than 30%. A few years ago, 1 to 3% was common for a software business.
• Margins are taking the hit. 89% of CFOs say rising compute costs hurt gross margin over the past year, and 38% say significantly.
• Spend is growing fast. 64% grew compute spend by 25% or more in 12 months. Among AI-native companies, that figure rises to 76%.

A software business with a 30% compute line runs on different economics, and the 2027 budget has to reflect that.

Five takeaways

1. Watch the quadrant where cost is heading.

Most CFOs are focused on workforce tokens: the coding agents, seats and overages engineering teams use to build the product. That spend belongs in OpEx and scales with headcount. As AI moves into the product, inference scales with customers, lands in COGS, and pulls production compute and data costs with it. For most companies, that bill will far exceed the 2026 engineering token bill.

Getting the split right is what makes forecasting possible:

• Workforce costs model forward from the headcount plan.
• Product costs model forward from customer and revenue assumptions.

Ben’s advice is to get more sophisticated with expense coding now, so every AI dollar lands in the right place on the P&L.

2. Watch the bill beneath the bill.

Tokens are the line everyone expects to break the AI budget. They did for 47% of the CFOs we surveyed. But compute and data storage exceeded forecast more often:

That’s the bill beneath the bill. Every model call your product makes also drives retrieval, data processing, storage and egress, and all of it lands on your AWS, Google Cloud or Azure invoice. It grows with every AI feature you ship.

This is good news for Finance. Managing the cost of AI largely means managing cloud infrastructure, and CFOs already have the tools for that: allocation, forecasting, rightsizing, and commitments that cut rates by 20 to 60% (more on that in #5).

Ben wants a GL account for each of these categories in the budget. He also wants engineering to explain how each one scales, whether linearly or in steps. “The last thing we want in January is a big budget variance that’s going to live with us for the rest of the year.”

3. This budget season, the CTO gets Finance’s attention.

CFOs have spent years pressing sales and marketing for pipeline and CAC data. Ben expects 2027 to shift that energy to engineering. Finance now needs:

• token usage by customer
• model choice by workload
• possibly a separate LTV to CAC for each AI product line, since usage and margin profiles differ

The data backs this up. Joint ownership of compute spend outperforms single ownership on every measure we tracked:

4. Measure AI by the unit, then by the dollar.

Connecting a token bill to business value is the same attribution problem CMOs have always faced. The bridge is the unit. We use a three-layer model:

• Layer 1, cost efficiency: commitment coverage, utilization, caching and model right-sizing.
• Layer 2, fully loaded unit cost: infrastructure plus model cost per interaction and per user, and the resulting effect on gross margin.
• Layer 3, business value: cost per closed ticket, time saved per employee, revenue per AI-assisted sale.

With units in place, every workload can face one test: does one more dollar of total AI spend return more than a dollar of outcome? If yes, scale it deliberately. If no, capture the discounts you’re missing or retire the workload. Several CFOs have told me they now see themselves as chief investment officer for this category of spend.

For internal AI, Ben recommends his ROSE metric: recurring revenue per dollar of employee, contractor and inference spend. He also tracks an inference efficiency ratio, which measures revenue generated per dollar of token spend.

5. Pay less without locking in.

There are two ways to save on anything: use less, or pay less for what you use. Engineering owns the first, and Finance can lead the second.

Cloud and model providers both offer committed pricing at 20 to 60% below on-demand rates. In exchange, you commit for one to three years. With AI innovation moving this fast, a wrong commitment becomes a real financial liability. Ben called overcommitting a CFO nightmare, and said the only way to commit with confidence is complete trust in your spend data.

Cloud Capital was built for this. We underwrite your commitments, so you get the committed rate and we carry the utilization risk. For many companies, that adds the extra couple of points of margin 2027 will demand.

What to do Monday morning

Ben laid out three steps, which he described as an AI maturity model for CFOs:

1. Audit your chart of accounts. Give internal AI subscriptions, API keys, internal inference and product inference each a GL home before the budget is set.
2. Build a governance framework for internal AI spend. Decide who gets which models, at what budget, and how new hires flow into the forecast.
3. Track customer usage on every AI product line. This lets you report AI margins and the unit economics behind them.

Get your total cost of AI mapped

Cloud Capital’s platform maps, allocates and forecasts your total cost of AI for free. If you want that in place before budget season, book a call with the Cloud Capital team.

Complete Event Transcript

Full transcript from the webinar “The Economics of AI: What CFOs Must Know in 2027,” hosted by Ed Barrow (Cloud Capital) and Ben Murray (The SaaS CFO). Edited for brevity and clarity.

Introduction [02:24]

Ed Barrow: Thank you, everyone, for joining us. Please pose questions throughout the session. We’ll tackle them as we go, and we’ll leave time at the end for anything we haven’t covered. We’re recording the session and will share it along with the slides.

I’m Ed Barrow, founder and CEO of Cloud Capital. We work with CFOs and CTOs of growth-stage technology businesses, providing a complete financial control layer for compute and AI infrastructure. I’m joined by the SaaS CFO himself, Ben Murray.

Ben Murray: Great to be here. I came up through the ranks of FP&A in the airline industry, then moved into software in 2004. I was an in-house SaaS CFO, and now I do fractional work and run my academy. I live and breathe SaaS finance metrics and AI metrics.

A New Metric for 2027: EBITTDA [04:10]

Ed: Because you’re the SaaS metrics guy, we thought we’d start with our favorite new metric. It solves every problem for CFOs struggling with ballooning AI costs. We had community-adjusted EBITDA. Now we have EBITTDA: earnings before interest, taxes, tokens, depreciation and amortization. You just take out the tokens and you’re done. Ben, can we adopt this as one of your core SaaS metrics?

Ben: It’s a mouthful. If you said that to your board, they’d ask what on earth you’re doing. But I loved this slide. What’s your EBITTDA to EBITDA gap? All those token costs you’d like to add back. That’s exactly why we’re here today: getting a handle on inference spend.

Ed: Until Ben formally adopts it, we’ll have to deal with these costs directly. We’ll cover four areas today. First, the total cost of AI: where these costs come from and how they behave. Second, how to measure, forecast and budget for this spend. Third, how to pay less for it. And fourth, a practical plan. It’s Tuesday, so you have the rest of the week. We’ll focus on what you can do next Monday.

What Kind of Business Are We Building? [07:24]

Ben: A year ago, everybody said SaaS metrics were dead and AI companies would never reach 80% gross profit. You don’t see that on social media anymore. That thesis is gone. But for CFOs, this is a great conceptual exercise. What business model are we building? Are we legacy SaaS moving into AI-infused revenue, or trying to be AI-native? Are we heading toward 50% margins or 60%? We’re in budget season for 2027, and boards will ask where AI margin and AI revenue are going. We have to get a handle on this now.

Ed: This has happened before. When software moved from perpetual licenses to cloud SaaS, infrastructure suddenly had a big impact on gross margin. We’re seeing that again. If you’re going fully AI-native, you have to build the entire business model around a different set of unit economic assumptions.

What interests me most is where CFOs’ attention is today versus where these costs are heading. Every CFO I speak to spends a lot of time on the tokens consumed by their engineering teams: the Claude Max plans and the overages on top. Those costs have been significant throughout 2026. They’re headcount-driven, in the same way development, testing and training environments always were. They’re OpEx. That’s where most of the attention is going today.

Where AI Cost Is Heading [10:48]

Ed: At a conference last week, many conversations were about a shift we’ll see in 2027. Companies used AI to build their product. Now they’re shipping it, customers are using it, and that usage drives token consumption.

A question came in: what is AI-native? You could build a software product with Claude Code that has no AI capabilities in it. Your workforce is consuming tokens, and you’re building more effectively. AI-native means embedding AI as a core capability within the product. That’s where we’ll see huge growth in token consumption in 2027, along with the underlying compute consumption. The tokens your engineering team uses will be eclipsed by the tokens your product consumes, and by the infrastructure that supports delivering that product.

Ben, how much of this transition are you already seeing?

Ben: It depends. For a lot of the legacy SaaS companies I help, it’s a slow transition. Everyone is using AI in development. Now it’s a slow infusion into the product, maybe an AI chatbot, and then the question becomes whether COGS is set up to track it. There’s no standard definition of AI-native. But these quadrants are great. We have dev AI spend, and now we have AI in our product lines. We need to know which departments use AI, and we need to get more sophisticated with expense coding.

The Stakes in Three Numbers [13:09]

Ed: Getting the allocation split right is incredibly important because these costs scale very differently. Workforce costs are driven by headcount: how many engineers you have and how much they use these tools. Once AI is embedded in the product, costs scale with customers. We can all see the potential for that to far eclipse the 2026 token bill from engineering.

We researched what CFOs are actually seeing, surveying 200 growth-stage technology CFOs in the US and UK. We published The Cost of Compute at the start of the year, then went deeper on AI costs in The Cost of Compute: AI Edition. If you’re wondering how your Anthropic or compute spend compares to your peers, the reports will tell you.

Three big takeaways stood out. First, people are spending a substantial share of revenue on compute and AI. When we started Cloud Capital, it was common for software businesses to spend 1 to 3% of revenue on the cost to deliver their product. That has ballooned in the last 12 to 18 months. The median is now 16 to 20% of revenue. That’s your bill from Google or AWS as a percentage of revenue.

More than 30% of companies now spend more than 30% of revenue on compute. A software business with 30% of revenue going to compute is a fundamentally different business. 89% of the CFOs we surveyed have seen rising compute costs hurt gross margin, and nearly two-thirds saw compute spend grow by more than 25% in the last 12 months. AI-native businesses are seeing even faster growth. Ben, does this match what you’re seeing?

Ben: Definitely. That’s why it’s so important to understand where your infrastructure is today and where it’s going. Our business model is changing as we adopt AI in our products and internally. This is a different budget season. We need to decide what we want to be when we grow up. CFOs need to work with their dev teams even more closely today.

The Bill Beneath the Bill [17:30]

Ed: The most surprising statistic came when we asked CFOs which AI and compute cost areas exceeded budget in the last 12 months. Going in, everyone would say tokens. And 47% did say foundation model APIs exceeded budget. But the underlying compute costs broke budgets more often: data processing, data storage, CPU, the bill from Amazon, Google Cloud or Azure.

These are the costs people didn’t expect to scale so fast. Coming into the year, attention was on tokens. But every time you call a large language model from your product, you incur more cost to process the prompt, send it to the model, get the response back, store the data, present it to the customer, and make sure the platform scales. For every dollar of token spend, there is substantial compute spend supporting it. Compute spend isn’t new. People have overspent on cloud for years. But our attention shifted to the token bill, and these other costs, driven by the same AI investment, are catching people out.

Ben: I’d want to see GL accounts for each of these categories under DevOps in your budget. Egress, the movement of data, is a new cost for a lot of CFOs. You may be moving a lot of data between the LLM and your network. We need to spend more time with the dev team to understand how these costs scale. Is it linear, or a step function? The last thing we want in January is a big budget variance that lives with us for the rest of the year.

The Token Treadmill [20:34]

Ed: There’s a lot of hope that token costs are coming down. The price per token, and the price for a given level of intelligence, has been dropping. Maybe gross profit fell in 2026, and cheaper tokens will save us in 2027. I think that’s a dangerous budget assumption. Usage is far outstripping the price drop. Token consumption is growing much faster thanks to agentic workflows, and it drags up CPU, data storage and egress along with it. Ben, you wrote about the token treadmill.

Ben: Per-token prices are going down, depending on whether you use frontier models or older ones. But context windows are bigger and we use more tokens per chat or per work unit, so the price per outcome goes up. If you have outcome-based pricing, how many tokens do your agentic workflows need? We have to understand the rate and volume economics of the models we use. In budget season, I’d want a governance framework for internal AI spend. And for AI product lines: do we need frontier models, or can we use older ones? Your board would love that budget clarity.

You Can’t Govern What You Can’t See [23:06]

Ed: You can’t control something if you can’t measure it. There are three basic elements to a governance framework, and none of them are new financial concepts.

First, visibility. How much are you actually consuming? When you used one cloud provider, that was one bill from one vendor. Now people spend on Anthropic, Snowflake, Databricks and their core compute provider, with costs coming from multiple directions. Seeing the total cost of AI is harder than it used to be.

Second, allocation. Some of this spend is driven by your workforce, and it belongs in OpEx. Getting that right for 2027 matters. Many of those tokens are an investment, with no tax on revenue. Isolating them gives you the right model for your business and shows you how each cost is driven. Workforce token spend can be modeled forward from engineering headcount. Production token and compute costs can be driven from revenue and growth assumptions. Allocation puts each cost in the right place in the GL, and it’s the basis for budgeting and forecasting.

Third, governance. Set the right budgets and guardrails so the business can invest in new capabilities that drive customer value, without runaway spend. We’ve seen people rebuild their financial model two or three times in 2026. Nobody wants to repeat that in 2027.

Finance and Engineering [26:25]

Ed: A vital foundation for all of this is a more collaborative relationship between finance and engineering. Ben, you had great examples of how working with sales compares to working with engineering.

Ben: CFOs usually spend a lot of time with go-to-market, pushing back on why CAC is too high and what the pipeline looks like. I think we’ll give sales and marketing a little break and spend more time with our dev leader or CTO. They may not like it, but we need more information from them now. If we offer AI product lines on a subscription basis, we need token usage by every customer, because the margin profile will be different. We need usage by customer and which models we’re using. You may end up with three LTV to CACs based on your AI product lines, because usage is so different. This budget season, we need time with the CTO.

Ed: The research proves the point. We asked CFOs whether engineering, finance, or both jointly own compute spend. Joint ownership produced far better results across forecast predictability, confidence in cost allocation and gross margin, and understanding of cost drivers. That shouldn’t be surprising. Anyone who has looked at the 70-page PDF from Amazon Web Services at the end of the month knows it might as well be in a foreign language.

Finance struggles to understand why these costs were incurred and what will happen to them next year. But finance has the customer and revenue projections and the headcount model. If you can work out with engineering which costs are driven by headcount and which by customer usage, forecasting into 2027 becomes much easier.

Allocating and Forecasting the Total Cost of AI [30:15]

Ed: Cost allocation and forecasting sound like exactly what everyone wants to do, but dissecting the Amazon or Anthropic bill into COGS and OpEx isn’t easy. Cloud Capital provides a free platform for finance leaders to allocate the total cost of AI. It takes spend in real time from Amazon, Google and Anthropic and allocates it automatically to the right GL line. That can be COGS versus OpEx at a high level, or by product, customer or component. It automates the process every month, so you close the books faster and more accurately.

We also provide a free forecasting platform that drives those allocated costs forward based on your financial projections, and crucially, factors in your engineering roadmap. If 2026 was about building and testing new AI technologies, 2027 will be about optimizing that spend. Model selection, re-architecting, moving between platforms: you want to know what those plans will do to your budget, so finance and engineering stay in lockstep.

There’s a lot that can be done to optimize this spend. If you’ve worked in FinOps, much of it is familiar. Have we chosen the right server sizes? Have we left things running? What data are we storing in which tiers? Are we making the right commitments to AWS or Google, and using them fully? Very similar concepts exist on the model side: model selection, caching so we don’t ask the same question twice, and routing between models. Ben, how much are CFOs leaning into these concepts?

Ben: I run a private CFO community, and we’re in education mode, absorbing webinars like this and learning terms like prompt caching and geo versus global routing. We don’t need to master them, but we need to know these concepts to talk with engineering. These questions will come up at the board. There will be a big focus on DevOps in COGS, which used to mean your AWS bill and now includes all your AI and infrastructure bills. If it’s depressing margins, expect a big budget discussion. CFOs need to move from listening and learning to influencing this discussion.

Measure What Matters [35:56]

Ed: That conversation happens at two levels. One is cost efficiency: are we using the right models and paying for them the right way? The other is the outcome we’re driving. I get asked a lot how to connect a big bill to revenue, customer outcomes and impact. With CMOs, the conversation was always attribution. We’re seeing something very similar here.

The bridge is the unit. We’ve talked about unit economics for years, and now it becomes critical for connecting costs to outcomes. Separate production costs in COGS from R&D costs in OpEx. Then go further and allocate costs to specific features, actions, orders, transactions and customers. You may not see a cost directly produce a revenue outcome, but connecting cost to unit gets you much closer, much faster.

Ben: We have compute metrics, and with outcome-based pricing, the cost to deliver that outcome. Internally, it could be pull requests. This isn’t easy as boards question AI ROI. If you offer outcome-based pricing, what’s your work unit, and how many tokens does it take? There’s also the inference efficiency ratio: how much revenue you generate for your token spend. Are we doing more for the customer with fewer tokens? This is all evolving, and your board may want a few more metrics.

The One More Dollar Test [39:38]

Ed: Ultimately, are we making the right investment decisions and driving a positive ROI? 2027 will remain a year of innovation and exploratory investment. We need a simple way to engage the board and our teams on whether to keep investing, or whether we’re burning budget without delivering an outcome. If we connect cost to unit, and unit to revenue or business impact, we can ask of each dollar we invest: is it delivering more than a dollar of outcome?

I love hearing CFOs say they’re the chief investment officer now, helping the business, and especially engineering, make the right investment decisions in a new area of spend. We’ll be spending more. The question is whether we have the right investment strategy.

Ben: CFOs have always asked where to invest and allocate capital, and this season it’s tokens versus people. On the revenue side: for a dollar of AI revenue, what’s the related COGS, and how will it scale? Internally, AI ROI is harder. Look up my ROSE metric. It’s recurring revenue generated for every dollar of employee, contractor and inference spend. At a macro level, it tells you whether your organization is becoming more efficient with AI. Is it moving up and to the right, or stalled?

Price Optimization: Commitments Without the Lock-in [42:30]

Ed: There’s a long list of engineering opportunities for 2027 to optimize what we use, and plenty of low-hanging fruit. Those are engineering decisions: model selection, architecture, workflows. Where finance can really lean in is price optimization. There are two ways to save money on anything: use less, or pay less for what you use. Engineering is leading on using less. Finance can add a lot of value on paying less.

This mirrors how cloud cost management evolved over the last ten years. Every model and compute provider starts you on pay-as-you-go. That’s smart business for Anthropic, Google or Amazon because it lowers the barrier to adoption. That’s been the world of 2026, whether you have bundled tokens in a seat license or on-demand compute.

As budgets come under pressure in 2027, a key finance lever is switching from on-demand pricing to committed pricing. AWS, Google Cloud and Azure have offered this for years. The same is available from Anthropic and OpenAI: commit to capacity in return for significantly lower costs. The same problems apply too. A commitment is a real financial liability, locking you into a service and spend level for one to three years. Innovation will be very rapid over that period, and if you get the commitment wrong, you’re paying for something you may not use. We have to be as efficient as possible with every dollar, and we can’t lock ourselves in when innovation needs to keep pace. Ben, how should CFOs make that trade-off?

Ben: Overcommitting is a CFO nightmare. That’s why we need complete trust in our cloud and AI spend data. Then we can commit to a level that might save a couple of points of margin, which is big. How can we commit to a number unless we know exactly how we’re spending today?

Ed: A question came in about where companies sit on the spectrum between SaaS with some AI features and aggressive AI investment. It’s a very broad spectrum. Across the organizations we work with, traditional SaaS companies dipped a toe into AI capabilities in the first half of this year and saw real customer value. So everyone has been investing more, partly for customers and partly for competitive reasons. I haven’t seen a single company try an AI feature and go back to a non-AI product.

That means these costs will grow. You’ll want to optimize price, and you’ll need to keep innovating. Your next feature may need different technologies. This is the commitment challenge Cloud Capital solves. We underwrite commitments to cloud and compute providers on your behalf, using the forecast we build with you. You get the optimized rate, and you avoid the financial and technology lock-in. That extra two points of margin will be critical in 2027, and your teams keep the flexibility to keep pace with the market.

Ben: On that margin question: I’ve helped an AI company raise a lot of money with great margins, so it’s up for debate. But running at 50% margins is so much harder. A services company runs 50% margins, but it doesn’t carry heavy R&D in OpEx. A software company with 50% margins and heavy R&D burns a lot of cash. You need a lot of volume or higher margins. I think investors will still expect software margins of 70 to 80% gross profit. If you’re stuck at 55%, you’ll need to explain why. And for measuring internal AI ROI, I’d use the ROSE metric to see whether you’re improving with all that internal AI spend.

What to Do Monday Morning [51:11]

Ed: It’s Tuesday, so everyone has until Monday. Ben, where would you recommend people spend their time to put the foundations in place?

Ben: First, something very tactical: assess your chart of accounts. Do you have the right GL accounts to track internal AI subscriptions, API keys, inference spend, and AI running through your product lines into COGS? Get the measurement in place so expenses land in the right spots on the P&L. Otherwise you’re presenting a skewed P&L, with AI hitting all over the place.

Ed: We can automate that cost allocation for you, for free. Plug in your compute spend and your Anthropic spend, probably coming from three different directions, and apply it to the GL. You don’t want to be reinventing your chart of accounts in January or February after the budget is set. Getting the cost allocation structure right as you start budgeting is critical.

Ben: Second, a governance framework for internal AI spend. Your board will want to know how you issue models to people, what their budget is, who gets what, and what happens with a new hire. Is all that baked into your forecast? Third, on AI product lines, can you track customer usage so you can track AI margins? You may have three different LTV to CACs across your AI product lines. You could call those three steps an AI maturity model for CFOs.

Ed: A question came in about how to measure the return, beyond the cost. The bridge is the unit: support tickets, customer engagement, actions your team takes, actions your customers take. Connecting costs to those units gives you a strong bridge to value and outcomes. Allocating your entire compute or token spend to specific features, components and customers isn’t always easy, but you can establish a framework for it. From there, you can see which metrics moved as a result.

Ben: The standard budget slides we use year in, year out won’t be enough. We’ll need a few more slides this year to prove what we’re doing internally and externally.

Closing [55:24]

Ed: Thanks, everyone, for staying with us. If you want to benchmark against your peers, take a look at The Cost of Compute 2026 and The Cost of Compute: AI Edition. Ben has written fantastic articles on TheSaaSCFO.com on COGS and these key metrics. And if you want your costs correctly allocated, with a budget driven by the right metrics and plans, go to cloudcapital.co. The platform is free to use as part of your budgeting process. We’re officially in budget season, the best time of year and the best time to get started.

Ben: I don’t know if CFOs will agree it’s the best time of year until the budget’s done. Ed demoed the platform to my community, and it was an eye-opener to see how the data flows through automatically once you integrate your cloud provider. Definitely take advantage of it.

Ed: Thanks for your questions. Ben runs the SaaS CFO community, and you can reach me through cloudcapital.co. If you have any questions at all during budget season, reach out.

Ben: Great timing for budget season and for preparing for next year. Thanks, Ed.