AI products · Software and SaaS
Gross margin for AI products, measured per customer and per model call
An AI product can grow revenue and lose margin in the same month, because every customer request carries a model or compute cost that a seat-based product never had. We attribute that cost to the customer, feature and model that incurred it. Pricing then reflects what serving each customer actually costs.
In one paragraph
Companies building on large language models or their own trained models — copilots, agents, vertical AI tools and APIs — are usually early-stage to growth-stage, with revenue up to ₹100 crore. Model APIs and GPU time dominate their cost of revenue. Founders and finance heads see the difference from conventional SaaS quickly: cost of revenue moves with usage rather than headcount. A single heavy customer can turn a profitable plan into a loss-making one.
Founder or CEO, with the CTO for usage data
The economics
How the money moves in this trade
Revenue is often a mix of platform fees, seat licences and usage charged per token, per call or per credit. Cost arrives as invoices from model providers and cloud platforms, billed in dollars, usually monthly and in arrears. Prepaid credits sold to customers create a liability; committed spend with a cloud provider creates an obligation in the other direction. Both have to be tracked if gross margin is to mean anything.
Training and fine-tuning add a second layer. Compute spent building a model is different in nature from compute spent serving customers. Whether any of it is capitalised is a judgement that needs a written policy. Most of the large vendors are overseas, so each invoice also raises reverse-charge GST on imported services and tax withholding on payments abroad, questions a domestic vendor never would.
The problem
Where the numbers usually break down
Inference cost sits in one cloud line
Model API and GPU spend is posted as a single expense, so gross margin cannot be split by customer, plan or feature.
Credits sold but not held as a liability
Prepaid usage credits are booked as revenue on sale, though most of the usage — and the cost of serving it — is still ahead.
Training and serving compute mixed together
Experiments, fine-tuning runs and production inference share the same cloud accounts, so neither research spend nor serving cost is visible on its own.
Overseas vendor invoices without a tax position
Model-provider and cloud bills paid by card or wire are booked without a documented view on reverse-charge GST and withholding.
Plans priced before cost was known
Tiers were priced at launch on an estimate of usage, and the unlimited or high-usage plans have never been checked against actual cost per customer.
What we do
The work, and the result each part produces
Usage-based cost attribution
Model calls, tokens and GPU hours tagged to customer, feature and model from product logs, and reconciled to vendor invoices.
Gross margin is reported per customer and per plan.
Credit and usage liability ledger
Prepaid credits, rollovers and expiries tracked as a liability and released as they are consumed.
Revenue follows usage, and unused credits are a known balance.
Training and serving split
Compute separated into research, training and production by project tag, with a written capitalisation policy applied consistently.
Research spend and cost of revenue are reported apart.
Overseas vendor tax mapping
Each overseas vendor mapped to its reverse-charge GST and withholding treatment, with open positions confirmed by a tax professional.
Every payment abroad carries a documented tax position.
Plan profitability review
Actual cost per customer compared with plan price, with heavy users and loss-making tiers identified.
Repricing and usage limits rest on observed cost.
Committed-spend tracking
Cloud and model-provider commitments, credits and discounts tracked against actual draw-down.
Under-used commitments are seen before they lapse.
Where we start
The first steps of an engagement
Vendor invoices and usage logs
We ask for model-provider and cloud invoices, billing and usage exports from your product, and the logs or tags that link calls and GPU hours to customers.
Usage logs reconciled to invoices
Usage in the logs is reconciled to each vendor invoice. Test traffic, retries and internal tools are separated out.
Credit liability and compute tags
Prepaid credits move to a liability ledger released on consumption, and compute is tagged as research, training or serving.
Gross margin by customer
The first report shows gross margin by customer, plan and feature, with the credit liability and committed spend alongside.
How it is measured
The reports the work is judged on
Each is recorded as a baseline before the work starts, so later reports have an honest comparison.
- Gross margin by customer, plan and feature
- Inference cost per thousand requests, by model
- Credit liability: opening, sold, consumed and expired
- Research, training and serving compute, reported apart
- Committed spend against draw-down, by vendor
- Overseas vendor payments with tax position recorded
Who it suits
Written for, and not the right fit for
- You need only the accountant's certificate for a foreign remittance; an independently engaged CA signs that.
- You want engineering help to cut model cost; we measure the cost, and your engineers change the architecture.
- Companies whose cost of revenue is dominated by model APIs, GPU time or both.
- Products sold on usage, credits or hybrid seat-plus-usage plans.
- Teams that need to show investors a gross margin that holds up customer by customer.
Services
Services this work draws on
Costing and profitability
What each product, line and customer really contributes.
Explore service ↗MIS and financial planning
Reporting built around the questions you actually need answered.
Explore service ↗Fractional CFO services
Senior financial leadership on the days you need it, not a full-time hire.
Explore service ↗GST, TDS and tax compliance
GST, TDS and income-tax prepared from reconciled data, not a working file.
Explore service ↗Every other service, also available in this trade
Common questions
Questions owners in this trade ask
Why do our usage logs never agree with model-provider invoices?
The differences tend to come from a few places — test traffic, retries, internal tools and vendor rounding. Once those are separated, the remaining gap is usually small enough to allocate on a stated basis.
Should we capitalise model training costs?
It depends on what the training produces and whether the recognition criteria are met, and it's a judgement your auditor will examine. We help you set a written policy, tag compute so it can be applied, and keep the evidence the auditor will ask for.
Do payments to overseas model providers need tax withheld?
It depends on the nature of the payment and any treaty position, and the answer isn't the same for every vendor. GST on those imported services is usually yours to pay under reverse charge. We map each vendor, document the position, and route open questions and any required certificate to an independently engaged CA.
5th Quadrant prepares cost attribution, vendor tax mapping and supporting records. Certificates for foreign remittances, the statutory audit and any other prescribed certification are signed by an independently engaged licensed professional, coordinated through 5th Quadrant. Full detail is on our Professional Scope page.
Explore further