AS | Ankit Sarawagi|Founder, CFOmatrix·July 2026·11 min read | SaaS Finance |
- Traditional SaaS runs at ~80 percent plus, AI SaaS at ~65 percent plus. The gap is real compute, not a mistake, and it is expected to narrow.
- COGS means every direct cost of delivery. All compute (your own models, databases, pipelines), third-party providers tied to delivery, and enterprise customer success belong here.
- Segregating R&D from cost of services is the hard part. The same engineers build the product and run production, so split by activity, not by team name.
- Per-customer and per-geo margin is the new center of gravity. With variable compute, blended margin hides which customers are actually profitable.
- AI SaaS margins should rise. Falling model prices and proven ROI both push the ~65 percent upward over time.
| ~80%+ Traditional SaaS gross margin | ~65%+ AI SaaS gross margin today | Per-unit The margin you now track by customer |
| 1. What is a good SaaS gross margin 2. Traditional 80 vs AI SaaS 65 3. What belongs in cloud COGS | 4. R&D vs cost of services 5. Per-customer and per-geo margin 6. Will AI SaaS margins improve |
What Is a Good SaaS Gross Margin
Gross margin is revenue minus the direct cost of delivering the service, expressed as a percentage. For software it is the cleanest signal of how scalable the business really is: the higher it is, the more of every rupee of revenue is left to fund growth, and the more the company is worth per rupee of revenue. For traditional SaaS a good gross margin is around 80 percent or higher. The product is already built, so serving one more customer costs very little: some cloud hosting, a little support, a few third-party tools.
For AI SaaS the honest benchmark today is around 65 percent or higher. That is not a failure of discipline, it is the arithmetic of a business where every user request can trigger real, variable compute. Both figures assume you have classified cost of goods sold correctly and are not quietly parking delivery costs in operating expenses to flatter the number.
Traditional 80 Percent vs AI SaaS 65 Percent
The single biggest change AI brings to a SaaS P&L is that cost of delivery becomes variable. In classic software the cloud bill is real but relatively flat: you have paid to build the product, and running it for one more user barely moves the needle. In AI SaaS, every request can call your own hosted model or a paid large language model, hit a vector database and run a data pipeline. Cost now scales with usage, and it lands in the same place revenue does: at the point of delivery.
So why does AI SaaS sit lower right now? Two reasons. Compute is still expensive, and the frontier models that make a product impressive are the priciest to call. And the market has not fully repriced for AI value yet: buyers are still learning what an AI outcome is worth, so price has not caught up with cost. The margin is a snapshot of an early market, not a permanent ceiling. To see how this reshapes the whole cost base, read the hidden costs of SaaS.
In the AI SaaS companies I have worked with, the blended gross margin was almost useless on its own. Because compute is variable, one heavy-usage customer can quietly run at a loss while the average still looks like 65 percent. The center of gravity has moved to per-customer margin: until you meter compute by customer, you do not actually know which accounts are paying for themselves. Build that metering into the finance SOP from day one; it is not complicated, and doing it late is painful.
What Belongs in Cloud Cost of Goods Sold
The judgement in a SaaS P&L is almost entirely in what you put in cost of goods sold. The rule is simple to state: if a cost rises when you serve one more customer or one more request, it is usually COGS. AI makes the list longer and heavier, because so much of delivery is now variable compute. Here is the full stack.
Two classic mistakes flip the number the wrong way. Burying delivery-side customer success in operating expenses inflates gross margin and makes an AI SaaS look more like classic software than it is. Lumping a large implementation or services team into the “software” line deflates it and hides a genuinely healthy product margin. A clean split gives you a true software (or AI-service) gross margin and a separate, lower-margin services line. When you are ready to lay it all out, use the SaaS P&L template, and see how pricing has to track this variable cost in SaaS and AI pricing models.
The Hard Part: R&D vs Cost of Services
Here is the classification problem that trips up almost every product company. You are doing two things at once: building new capability (research and development, an operating expense) and running the live service (cost of services, part of COGS). The trouble is that the same engineers and often the same compute environment do both. Get the split wrong and your gross margin is either flattered (too much cost pushed into R&D) or crushed (build cost dumped into COGS).
The answer is to split by activity, not by team name. Do not ask “is this the engineering team?” Ask “is this cost building the future, or serving today’s customers?”
- Building new features and products
- Training and fine-tuning future models
- Experiments, prototypes, evaluation runs
- Compute in training and staging environments
- Operating, monitoring and supporting the live product
- Inference and compute the running service consumes
- Production databases and data pipelines
- The engineers on call to keep it up
You do not need a perfect study. Ask engineers to tag their time even at a rough percentage, meter compute by environment so production is separated from training and experimentation, and then apply the same split every period so gross margin stays comparable. The goal is not accounting theatre; it is a gross margin an investor can trust and a cost base you can actually manage. This is a core discipline of a SaaS finance function, and it is one of the first things a fractional CFO installs.
Pushing the entire engineering and compute bill into research and development makes gross margin look spectacular, but a diligence team will unwind it in an afternoon. If a real chunk of your engineers and compute keep the live service running, that portion is cost of services. An honest 70 percent beats a flattering 90 percent that collapses under scrutiny.
Per-Customer and Per-Geo Margin: The New Center of Gravity
In classic SaaS, one gross margin described the whole business, because cost of delivery barely varied between customers. AI breaks that assumption. When compute is variable, a light user and a power user can have completely different margins on the same plan, and a ₹12 lakh account that hammers the model can be less profitable than a ₹4 lakh account that sips it. Blended margin averages them and tells you nothing useful.
This is not a nice-to-have report; it is the lever that decides pricing and expansion. Dig into it in per-customer and per-unit margin for AI SaaS, and connect it to how you charge in SaaS and AI pricing models.
Will AI SaaS Margins Improve
Most likely, yes, and the direction of travel matters more than today’s exact figure. Three forces push the AI SaaS gross margin up from roughly 65 percent over time.
None of this is automatic. A team that keeps adding the most expensive frontier-model features will hold its margin down, and that can be a deliberate, correct choice while it is winning the market. But for a disciplined AI SaaS that meters its compute and prices with intent, the 65 percent of today is a floor to build up from, not a ceiling. Watch it the right way and you will see it move; the guide to the whole picture sits in our SaaS finance pillar.
“In classic SaaS, gross margin was a number you reported. In AI SaaS, it is a number you manage, customer by customer, because compute makes it move.”
Ankit Sarawagi, CFOmatrixFrequently Asked Questions
What is a good SaaS gross margin?
For traditional software, around 80 percent or higher, because once the product is built the cost to serve one more customer is very low: mostly cloud hosting, some support and a few third-party tools. For AI SaaS, around 65 percent or higher today, because heavy per-request compute (your own models, the large language models you call, databases and data pipelines) is a real variable cost of delivery. Both assume you have classified cost of goods sold correctly and are not hiding delivery costs in operating expenses.
Why is AI SaaS gross margin lower than traditional SaaS?
Two reasons. First, compute is expensive and variable: every request can trigger inference on your own models or a paid large language model call, plus vector databases and pipelines, so cost scales with usage rather than sitting flat. Second, the market is not yet paying a full premium for AI outcomes, so price has not caught up with cost. In the AI SaaS companies I have worked with, this is exactly why margins print near 65 percent today. It is expected to rise as return on investment is proven and model prices keep falling.
What goes in SaaS cost of goods sold?
Every direct cost of delivering the service: all compute (your own model hosting, inference, databases, data pipelines), third-party providers tied to delivery (the large language models you call, hosting, embedded software), payment processing, and the customer-facing delivery team such as enterprise customer success and technical support. Sales, marketing, general overhead and research and development are operating expenses. The test: if the cost rises when you serve one more customer, it usually belongs in cost of goods sold.
Is customer success part of SaaS COGS?
The delivery and support part of customer success belongs in cost of goods sold, because it is a direct cost of keeping the service running for paying customers, especially in enterprise where onboarding and technical support are hands-on. The expansion and upsell part, where the team is effectively selling more, is closer to sales and sits in operating expenses. Many teams split one customer success function across both lines by time or headcount, which honestly reflects that it does two different jobs.
How do I split R&D from cost of services?
Split by activity, not by team name, because the same engineers often build new features (research and development, operating expense) and keep production running (cost of services, part of COGS). Time spent building future capability is research and development; time spent operating and supporting the live service, plus the compute that service consumes, is cost of services. Ask engineers to tag their time even roughly, meter compute by environment (production versus training and experimentation), and apply the split consistently so gross margin is comparable period to period.
Will AI SaaS gross margins improve over time?
Most likely yes. Model and compute prices keep falling, so the cost of the same output drops. As buyers see proven return on investment, pricing shifts toward the value delivered and the market pays a premium it does not pay today. Better engineering (caching, smaller fine-tuned models, routing cheap requests to cheap models) also cuts unit cost. It is not guaranteed, and a company that keeps adding expensive frontier-model features can hold its margin down by choice, but the direction of travel is upward.
This is general educational information for founders, current to mid-2026, and not legal, tax or investment advice. Gross-margin benchmarks (traditional SaaS ~80 percent plus, AI SaaS ~65 percent plus) are indicative and vary by stage, usage intensity and business model. Compute and model prices change quickly; verify the current position before acting on a specific matter.
The Hidden Costs of SaaS
SaaS and AI Pricing Models
AS | Founder, CFOmatrix | Finance Strategy & Equity Compliance CFOmatrix is a knowledge platform focused on how finance actually works inside growing companies, from SaaS and AI SaaS metrics and unit economics to structure, fundraising and exits. |