For most of the last decade, a Series A in software was a growth argument. You showed a few million in recurring revenue, a steep slope, and a credible plan to spend the money on sales. AI companies in 2026 walk into the same meeting with a different problem: the revenue line looks excellent and the cost of delivering it is visible in a way it never was for seat-based software. The round now hinges on whether the gross profit behind that revenue can fund the go-to-market motion the next stage requires.
This is not a story about a higher or lower bar in the abstract. It is a story about which numbers get interrogated, in what order, and what evidence makes an investor comfortable. Founders who prepare the margin, the retention cohorts, and the concentration table before the first meeting spend their time on the vision instead of on defense.
The bar moved from growth to gross profit
Classic software has long been underwritten against a blended gross margin near 72%, and that figure still shows up as the comparison point in current venture writing on AI economics. The reason is structural: at 72% gross margin, a dollar of new revenue can pay for sales and still leave room for the research and development that keeps the product ahead. At 45%, the same dollar pays for inference, support, and infrastructure before it pays for anything else, and the payback math on every new customer changes.
The size of the market is not in question. Inference spending is estimated at roughly $25 billion in 2025 and about $130 billion in 2026 in one investor's compilation, with forecasts well above that for 2027. A large market with compressed margins produces a specific kind of company, and the Series A bar is calibrated to find the ones that can reach scale without needing a second margin miracle.
ARR quality: contracted beats annualized
The first diligence question is not how much revenue you have but what kind. Annualized run-rate is a snapshot of the most recent month multiplied by twelve, and in a usage-priced AI product that number can move sharply in either direction from one quarter to the next. ARR should mean contracted, invoiced, recurring revenue with a defined term, and the run-rate should be disclosed separately with the assumption behind it.
Three tests separate the two. Is the commitment contractual beyond a monthly rollover? Is the revenue invoiced rather than merely consumed? Does it recur without a new purchase decision? A company that passes all three has revenue an investor can model. A company that fails the first test has revenue that will be discounted, and the discount compounds when the customer list is short.
Usage-based pricing is not disqualifying, and in AI it is often the right model. What matters is whether usage is tied to a durable workflow or to a burst of experimentation. The diligence signal is cohort behavior: customers acquired nine months ago should still be consuming, and consumption should be stable or growing after the initial pilot ends.
Net retention: separate expansion from inference inflation
Net revenue retention measures revenue from an existing cohort divided by the same cohort a year earlier. It is the cleanest single indicator that a product is becoming more valuable to the people already paying for it, and it is the number most often flattered by AI pricing models.
Here is the trap. If your price is per token, per seat, or per outcome, a customer that merely uses the product more will expand the account. That is real revenue and real expansion, but it is not evidence of value capture, because it can be driven by heavier usage of a workflow that has not deepened. Gross retention, which strips expansion out, tells you whether the base stays at all. Show both. A company with 118% net retention and 82% gross retention is telling a different story than one with 118% and 94%.
David Skok's long-standing treatment of SaaS metrics makes the economic point that still governs here: retention is what turns customer acquisition cost into a payback calculation instead of a bet. In an AI product, the same logic applies, with the additional requirement that the retention cohort be measured on revenue after inference cost. A cohort that retains 100% of revenue while its margin falls is a cohort that is getting more expensive to keep.
Gross margin after inference: the number that decides the round
Work out the margin line by line, and label the assumptions. Consider a company billing $100 per month per account, with $32 of model inference, $6 of vector storage and orchestration, and $4 of support and hosting attributable to the account. Cost of revenue is $42 and gross margin is 58%. Now assume inference is cut in half through caching, smaller models on easy requests, and batch processing: cost of revenue falls to $26 and margin rises to 74%, which is the difference between a company that needs a margin story and one that has already told it.
Those numbers are an illustrative model, not a benchmark. What matters is that the model exists, that the inference figure is measured per account rather than blended, and that the trend is visible by quarter. One useful reference point: market commentary in late 2026 put frontier model margins in the 40% to 70% range and noted that the cost per completed task varies substantially between comparable models. An application company that resells inference inherits a version of that spread, and investors know it.
Three levers are credible in a diligence conversation. Routing: send easy requests to cheaper models and keep the expensive model for the hard cases. Caching and context discipline: pay once for the same input instead of on every call. Architecture: own the narrow, high-volume parts of the workflow with smaller models or classical software. A fourth lever, passing cost to the customer through usage pricing, is legitimate but moves the risk to retention, which is why both numbers belong on the same page.
Concentration and defensibility
Concentration comes in two directions and both are checked. On the customer side, top-three concentration near a third of ARR is a reasonable place to start asking questions, and anything above half is a story problem rather than a spreadsheet problem, because a single renewal decision can reshape the next twelve months. On the supplier side, the question is how much of your inference runs through one model provider. A dependency above roughly 60% of volume invites the follow-up: what happens to your margin and your product if that provider raises prices or changes terms?
Defensibility is where AI companies most often answer with the wrong evidence. Saying that you have proprietary data invites the question of what makes it proprietary. Saying that you have a fine-tuned model invites the question of what happens when the next base model is better at the task without fine-tuning. The durable answers look like switching costs and workflow depth: the product holds the system of record for a process, the customer has rebuilt an internal routine around it, and the retention cohort shows the effect. A useful test is whether a competent competitor with the same model access could displace a customer in a quarter. If yes, the moat is the team and the go-to-market, which is a legitimate answer at seed and a weak one at Series A.
A benchmark checklist with labeled assumptions
The table below is a starting checklist, not survey data. Every bar in the third column is an assumption chosen to be defensible in a 2026 diligence conversation, and each should be replaced with the number your own market supports.
| Metric | What to show | Illustrative bar (assumption) | Why it matters |
|---|---|---|---|
| ARR definition | Contracted, invoiced, recurring revenue; run-rate disclosed separately | 70% or more of ARR under 12-month terms | Usage spikes inflate annualized numbers |
| Gross retention | Cohort dollar retention with expansion removed | 90% or higher | Expansion can hide churn |
| Net revenue retention | Trailing twelve-month cohort, revenue basis | 110% to 125% | Evidence that the product deepens |
| Gross margin after inference | Fully loaded: model calls, accelerators, storage, support | 55% to 70% blended | Sets the ceiling on sales spend |
| Inference cost trend | Cost of revenue as a share of revenue, by quarter | Flat or falling for four quarters | Shows operating leverage is real |
| Customer concentration | Top three and top ten as a share of ARR | Top three at or below 35% | One renewal should not decide the year |
| Supplier concentration | Share of inference volume from the largest provider | 60% or less | Pricing and availability risk |
| CAC payback | Months to recover acquisition cost on gross profit | 18 months or less | The efficiency bar for the next round |
| Defensibility evidence | Cohort retention by vintage plus a switching-cost test | Improvement across four cohorts | A moat that model upgrades do not erase |
Two rows are worth arguing about in advance. Gross margin is the row founders most often present as a single blended number, and the row investors most often re-cut by customer segment, because a handful of heavy accounts can be carrying an unprofitable long tail. Customer concentration is the row most often understated through a footnote about a reseller or a design partner. Disclose both plainly; the discovery of a surprise is worse than the surprise.
What to do next
- Rebuild the revenue schedule so that contracted ARR, annualized run-rate, and usage overage are three separate lines instead of one.
- Produce twelve-month cohorts with gross and net retention computed on revenue after inference cost.
- Measure inference cost per account, not blended, and identify the three accounts that would move the margin most if optimized.
- Build the concentration table for customers and model providers, and write the mitigation for the largest line in each.
- Write one page on defensibility that cites cohort behavior and switching costs rather than data or model claims.
The Series A bar is not higher in 2026 so much as it is more specific. Investors want to see a business whose gross profit can pay for its own growth, whose revenue survives a definition test, and whose retention does not depend on a single customer or a single model vendor. Founders who bring those numbers to the first meeting get to spend the rest of it talking about what they are building.
