Restating an AI Startup's Gross Margin Before You Price the Round

Before you price an AI company's round, you need to know what its gross margin actually is once model inference sits in cost of goods sold and vendor credits are stripped out. This guide is for early-stage investors, platform and talent partners, and angels who have to defend a valuation. It carries one worked case all the way through the inference-to-COGS reclass and the credit-stripping math, with the real intermediate numbers, so you can follow along on your own deal.

The gross-margin line is where AI deals mislead. ARR pressure-tests, cap-table recasts, and traction reads are well-covered, but they stop at the top line. The margin itself gets quoted from a management deck and waved through. That is a mistake, because two accounting choices - where inference is booked and whether credits are netted - can swing the number by 18 to 20 points. A true 60 percent dressed up as a SaaS-like 78 percent is exactly the deception this guide is built to catch.

Why AI gross margins are structurally lower than SaaS

AI products run materially below classic SaaS, so a margin that looks SaaS-grade is the first thing to distrust. Bessemer's pricing playbook puts AI gross margins at 50 to 60 percent against 80 to 90 percent for traditional software, and a16z's framing lands in the same range.

The survey data agrees. ICONIQ's State of AI snapshot, covering roughly 300 software executives, restates the series as 41 percent in 2024, 45 percent in 2025, and a projected 52 to 53 percent in 2026. The spread inside that average is wide: Bessemer's "Supernovas" reached about $125 million of ARR in year two at gross margins near 25 percent, while "Shooting Stars" ran about 60 percent. So before you even open the data room, your prior should be that a credible AI company sits in the 45 to 60 percent band, and anything claiming 78 percent needs to earn it.

The reason the number is lower is structural, not temporary. Inference is the one COGS line that does not amortise with scale. It climbs from about 20 percent of AI product cost pre-launch to roughly 23 percent at scale, while talent's share falls from 32 percent to 26 percent, because adoption grows consumption faster than efficiency shrinks it. Mature SaaS COGS runs 10 to 25 percent of revenue; an AI company carries a usage-metered compute bill on top of that.

The worked case: Northwind, 72 percent on paper

Take one company all the way through. Call it Northwind, an application-layer AI product with a management deck showing 72 percent gross margin on $100 million of revenue. The CloudZero worked math gives us the shape: $20 million of traditional COGS and $8 million of AI COGS already disclosed, netting to a 72 percent gross margin. That is the headline I will try to break.

The first read is already a signal. A 72 percent margin on an application-layer company sits above the 45 percent ICONIQ prior for resellers and above the 53 percent balanced-differentiation mark. Either Northwind has genuine differentiation and in-house cost control, or the number is propped up. The rest of this guide is the method that tells the two apart.

I will walk the two load-bearing adjustments - the inference reclass and the credit strip - with real point swings, then the decile test that exposes what the average hides.

Reclassifying production inference into COGS

Production inference belongs in COGS, and moving it there is usually the first high-single-digit hit to a flattered margin. Under US GAAP the test is nature and purpose. Inference, LLM API fees, serving GPUs, embeddings, and vector databases scale with usage like hosting and serve current customer traffic, so they sit above the line. Model training and platform-wide fine-tuning are R&D under ASC 730, or potentially capitalizable internal-use software under ASC 350-40.

The practitioner default settles the argument cleanly. An AWS bill for production EC2 serving customer traffic is COGS. An OpenAI bill for production inference serving the same workloads is also COGS. There is no accounting principle that distinguishes them. If a company has parked inference in R&D, it is not reading the standard, it is dressing the margin.

On Northwind, the deck already shows $8 million of "AI COGS", which is encouraging, but the reclass question is whether all customer-facing compute is actually in that line or whether a chunk is sitting in engineering. The documented size of this error is well-attested: one case moved hosting to COGS and dropped gross margin from 76 percent to 68 percent, an 8-point swing, because the cost had been hidden in the wrong line item. In a full GAAP recast during Series A diligence, one modeled case saw gross margin drop by 21.7 percent once the chart of accounts was recast.

For Northwind, assume the diligence finds $8 million of serving GPU and third-party product API spend tagged as R&D. Reclassifying it lifts COGS from $28 million to $36 million and drops the margin from 72 percent to 64 percent. That is the first layer off the headline.

What stays in R&D

Leave training and fine-tuning where they belong. In most real-world AI development scenarios, training data costs do not meet the criteria for capitalization and must be expensed as R&D. The one thing to check is whether fine-tuning was capitalized under ASC 350-40 to keep it out of operating costs entirely - that flatters both COGS and operating expense at once.

Stripping credits and restating at rate

Credits do not reduce COGS. Free cloud or model credits hide a cost, so the restated margin has to show economics before credits and after any contractual discount. This is the second big swing, and it runs in the same 8-to-10-point range as the inference reclass.

The mechanism is straightforward and the size is documented. One worked narrative moved gross margin from 67 percent to roughly 76 percent purely from discounted credits - a 9-point gift that evaporates the moment the credits run out. Promotional credits can improve reported cash and hide mature unit cost, which is why underwriting should show the margin both ways.

The method: rebuild from the company's own usage ledger. Apply contracted meter rates to telemetry, reconcile the aggregate to the invoice, and retain any difference as allocation variance until resolved. Then produce two numbers, because sources genuinely disagree on which rate to use.

Line                         At list rate    At contracted rate
Metered usage (telemetry)    _______         _______
Less: committed-capacity     _______         _______
Less: negotiated discount    (excluded)      _______
Less: promotional credits    (excluded)      (excluded)
= Restated AI COGS           _______         _______
Reconcile to invoice:        variance _____  variance _____

Run both. One view restates at list price because many find it clearer to model margin before any discount. Another restates at the contracted rate after negotiated discounts but before free credits. Neither lets promotional credits reduce COGS. For Northwind, assume the $8 million of booked AI COGS was net of $1.5 million in expiring provider credits; grossing that back pushes AI COGS to $9.5 million. Combined with the reclassed $8 million, total COGS reaches $45.5 million and the margin falls to roughly 55 percent. The headline 72 has become a defensible 55.

Once you have found the ledger-to-invoice reconciliation is the choke point, finding the person who can speak to it is its own task. Compute-talent depth is a proxy for in-house cost control, and it is searchable.

Refolk turns that into a named list in plain English, which is faster than reconstructing it from scattered public profiles. The depth matters: in Refolk's index, 2,952 US profiles carry LLM plus Inference skills against 448 UK profiles, a 6.6x ratio, concentrated at employers like OpenAI and Atlassian and heavily in the San Francisco Bay Area. A UK target is far less likely to have hired the inference-optimisation talent that defends margin, which makes its restated number more fragile.

The decile test: what the blended average hides

A clean 55 percent blended margin can still sit on top of a negative top decile, so the decile plot is the load-bearing number, not the headline. Sort users by usage, group them into ten buckets, and compute gross margin per bucket. If the top decile is negative and the bottom decile is at 80 percent, pricing is not capturing value from power users.

This is not an edge case. Around 10 percent of users can generate 40 to 70 percent of the inference bill, and the top decile consumes 5 to 10 times the median. Under flat pricing, a single price covering all usage lets top users subsidize themselves straight into your loss column. The blended average is the deception vector precisely because it averages that away.

The inputs are specific: per-action cost at current provider rates with a 30 percent buffer, projected P50 and P90 monthly usage per active user from the last 30 days of logs, and a base price benchmarked to the P50 user. Watch for agent retries, which blow up cost per task - one ICONIQ-quoted workflow projected at $0.10 per run reached $1.50 or more once agents retried. Check the actual logged tokens, not the plan assumption.

For Northwind, assume the decile plot shows the top bucket running at negative 15 percent while the bottom sits near 80 percent. The company is one large power-user onboarding away from the blended 55 percent sliding further. That is the difference between "compresses" and "propped up".

The procedure, end to end

This is the full method, in order, from pulling the P&L to ruling the headline. Each step names who does it and what "done" looks like so you can run it on a live deal.

The point-swing arithmetic is what makes the ruling defensible. Each adjustment has a documented range, and your bridge should attribute the move explicitly rather than presenting one number out of thin air.

Cross-checking when the invoices are withheld

When a company will not hand over its usage ledger, model-provider economics and public filings let you triangulate dependency anyway. Position in the stack is the strongest free prior: the closer a company is to just passing through a frontier API, the thinner the margin it keeps. The 45 percent application-layer versus 53 percent balanced-differentiation split is your anchor when invoices are hidden.

Three public reads back that up:

• MD&A ratios. Public SaaS companies have started disclosing inference-related cost ratios, generally between 4 and 9 percent of revenue. A private target claiming materially lower is making a claim it has to defend.
• Committed-spend announcements. These read directly on dependency and concentration. Under one neocloud deal, Anthropic is paying roughly $1.25 billion per month through May 2029 and Google approximately $920 million per month. For your target, any multi-year committed-compute contract tells you the cost floor is locked, not flexible.
• Provider margin trends. Anthropic's inference margins rose from 38 percent to 70 percent in a single year per SemiAnalysis data. That gain does not automatically reach the app; falling token prices can be recaptured upstream by provider pricing power rather than passed through to the startup you are pricing.

One mechanic people reach for does not hold up here: there is no publicly established way to read inference dependency off a Form D filing. Use committed-compute announcements and MD&A ratios instead, and say plainly in your memo that the dependency estimate is triangulated rather than reconciled.

How this goes wrong

The failure modes below are where a restated margin gets faked or missed. Each has a tell and a check, and this is the part of diligence that actually protects the valuation.

Two more deserve their own note. Agent retries exploding cost per task is the quiet killer: a $0.10 plan line can become $1.50 or more in production, so read the actual logged tokens, not the pricing assumption. And customer-success staff can be hidden in either COGS or sales and marketing to flatter whichever line the company wants to look good; review job descriptions and compensation plans, as venture associates do, to see where the people actually sit.

One counterintuitive tell closes the set. A margin that improves with scale is a flag, not a comfort. Inference does not amortise, so genuine operating leverage on the compute line is rare. If the trend is up, probe whether it reflects credits, a reclass, or capitalized training before you credit it to the model.

Before you call the number defensible

Run this check before the restated margin goes in the memo. It is the difference between a number you can defend to the partnership and one that falls apart on the first question.

Keeping the method current

The thresholds in this guide move, so re-check the mechanism rather than the number. The margin benchmarks are survey-driven and reset each ICONIQ and Bessemer edition; pull the latest band before you anchor a prior. Provider pricing is the volatile input - token prices fall, but provider margin gains like Anthropic's 38-to-70-percent jump show that the cost floor is set upstream, so re-read committed-spend announcements and MD&A inference ratios each time you price a deal. And billing models shift under you: when a dominant tool moves to usage-based billing, every downstream reseller's unit economics change at once. The procedure holds; the inputs do not, so run the reclass, strip the credits, and plot the deciles fresh on every term sheet.