—— AI Startup Benchmarks · 2026
The numbers your inference bill put on the table.
Generic startup benchmarks were written for a business with no marginal cost. Yours has one, it posts to COGS, and it moves every quarter. These are the AI-native figures investors are actually underwriting — margin after compute, revenue per head, burn against a falling cost curve — with the sources named and the disagreements left in.
52%
Average AI product gross margin, 2026
ICONIQ
~23%
Of revenue consumed by inference at scaling stage
ICONIQ
$1.13M
ARR per employee, AI-native leaders
Bessemer
<2.0x
Burn multiple expected by Series A
CRV
~42%
AI valuation premium at seed
Carta
01 · Margin after compute
Gross margin is now an engineering output, not a business-model property.
The 80% software benchmark does not survive contact with a live model call. What replaces it is a band — and where you sit inside that band says more about your architecture than your market.
52%
average AI product gross margin, 2026
The new center of gravity
Up from 41% in 2024 and 45% in 2025 — a real glide path, but one that plateaus well below the SaaS standard. Quote the trajectory, not the snapshot.
ICONIQ State of AI, Jan 2026 · ~300 software execs
~60%
gross margin, capital-efficient AI cohort
What "good" looks like
Bessemer's fast-growing, strong-PMF AI companies clear the average by roughly eight points — earned through routing and caching, not pricing.
Bessemer Venture Partners, 2026
~25%
gross margin, thin-wrapper products
The bottom of the band
Explosively scaling wrappers and early-stage AI-first companies can sit here. Growth at 25% gross margin is a financing problem wearing a growth costume.
Bessemer · ICONIQ, 2026
$230k
inference cost per $1M of AI revenue
The line item that didn't exist
Roughly 23 cents of every revenue dollar walks out as inference at scaling-stage AI B2B. Know this number by cohort before an investor asks for it blended.
ICONIQ, 2026
4–9%
inference disclosed as % of revenue
Public comps are starting to tell
Public software companies began breaking inference out separately in Q1 2026 filings. Disclosing early earns analyst credit; discovering it late does not.
Q1 2026 public-company MD&A
50–70%
achievable inference cost reduction
The four levers
Model routing (send the routine 80% to small models), prompt caching (~90% discounts on major APIs), batching where latency allows, and context compression — combined, without measurable quality loss.
Industry practice, 2026
02 · The cost curve you're standing on
Your largest variable cost deflates on a schedule.
This is the fact that makes AI financial planning different from every prior software cycle — and the one most static burn models get wrong in the same direction.
≈1,000× in three years for constant capability — a16z's estimate of roughly 10× annually. Alongside it, H100 rental fell 64–75% in fourteen months, while flagship APIs still price at $5 per million input tokens and $30 per million output. The planning implication: a model that holds unit costs flat across five years systematically overstates the raise. Deferring compute-heavy features has real option value.
$0.30–$14.90
per GPU-hour, by provider and tier
A 50× spread on the same input
Reserved versus on-demand is a futures book: you are trading price risk against utilization risk. Treat it as treasury management, not procurement.
Provider pricing, 2026
$100k → $2.3M
monthly inference spend, pre-launch to scaling
The step function nobody budgets
Roughly $100k/month before launch, $1–1.6M at general availability, $1.1–2.3M once scaling. The jump lands between two board meetings.
ICONIQ Builder's Playbook
Rented margin
that only exists at today's API prices
Stress-test against your supplier
Model repricing, rate limits and deprecation are all live risks. Show the resilient number and the contract mechanics that hold it — commitments, price locks, multi-provider fallbacks.
Editorial position
03 · Growth & retention
AI-native companies grow roughly twice as fast — and churn far less predictably.
The growth premium is well documented. The retention picture is the one founders get ambushed on, because the headline number is an average across two completely different businesses.
~55%
median AI-native growth rate
Roughly 2× legacy SaaS
Within the same ARR band; wider still at aggregate level. Below top-quartile growth, raising in 2026 gets materially harder.
BVP State of the Cloud · ICONIQ
5.7 yrs
to $100M ARR, top AI performers
Versus 7.5 years overall
Nearly two years compressed off the classic path — which is precisely why investors underwrite the trajectory rather than the current snapshot.
Bessemer Cloud 100
85% vs 48%
annual NRR: enterprise-priced vs blended AI-native
The average is a lie
Products above ~$250/month retain around 85% annually. The blended AI-native median of 48% is dragged down by prosumer tools. Never present this metric blended — segment it or an investor will.
ChartMogul · BVP · ICONIQ
120%
NRR, best-in-class AI-native
The top of the range
Achievable, and increasingly the Series A expectation for enterprise AI. Usage-based pricing helps here — expansion happens without a renegotiation.
BVP · ICONIQ, 2026
~20mo
typical CAC payback, private software
With an AI-specific caveat
LTV/CAC of 3:1 remains the floor. But when gross margin is 52% rather than 80%, the same CAC takes materially longer to earn back — payback math must run on contribution, not revenue.
CRV, 2026
$3.5M ARR
now expected at Series A for AI
The bar moved up
Paired with ~120% NRR and 60%+ gross margin in the composite investors describe. Two companies at identical ARR land in different rooms on the strength of the other three numbers.
CRV · Carta composite, 2026
04 · Capital & team efficiency
Revenue per head is the metric that broke its own scale.
AI-native outliers are running four to five times the classic software benchmark. Which means the old target now reads as underperformance in the rooms where AI companies get priced.
| Cohort | ARR per employee | Read |
|---|---|---|
| SaaS under $25M ARR | $130k | The median private software company. Where most benchmark decks stop. |
| SaaS $25–100M ARR | $172k | Scale-ups now run 20–25% leaner than the 2017–19 cohort at the same band. |
| SaaS above $100M ARR | $249k | Approaching public-company productivity. |
| Median public SaaS | $283k | The reference point AI companies are now measured against — and past. |
| AI-native leaders | $1.13M | Four to five times the SaaS benchmark. Individual outliers run several times higher again. |
Sources: ICONIQ Compass Benchmarks; Bessemer Cloud 100 (2025–26); High Alpha SaaS Benchmarks (800+ companies). Note the composition trap — the AI-native figure is a leaders' cohort, not a median, and comparing your company to it is comparing yourself to a survivor sample.
<2.0x
burn multiple expected at Series A
Strongest run near 1.0x
Below $1M ARR, multiples well above 2.0x are common and forgiven. By Series A they are not. Net burn per dollar of net new ARR is the tightest capital-efficiency proxy in use.
CRV, 2026
50+
AI companies expected to reach $10M ARR with under 10 staff in 2026
A structural change, not a stunt
Coding tools, AI support and automated GTM compress every function that historically scaled with revenue. Headcount plans built on 2021 ratios overstate hiring by a wide margin.
Industry tracking, 2026
Rule of 40
adjusted for AI companies
Same rule, different starting line
With gross margins structurally 20–30 points lower, the profit half of the equation starts from a deficit — so AI companies need more growth to clear the same bar. Some investors now apply a margin-adjusted version.
Practitioner consensus, 2026
05 · Capital & valuation
The premium is real, concentrated, and someone eventually has to grow into it.
Two markets are running at once. Knowing which one you're pricing into determines whether your last round was a win or a hurdle.
~42%
AI seed valuation premium · $17.9M pre-money
Two markets running at once
Against roughly $13M for comparable non-AI teams — a gap that widened toward 46% into 2025. Widest for teams shipping genuinely AI-specific products rather than AI-labelled SaaS.
Carta
~70%
AI premium at Series A
The premium is a debt
Every point of premium raises the growth you must deliver before the next round. Over-marked seeds are where the company-specific risk sits, not in the index.
Carta, 2026
>60%
of venture dollars going to AI
Concentration is the mechanism
Over 60% of every venture dollar on Carta in Q1 2026; US VCs allocated roughly 64% of deal value to AI/ML. Too much capital chasing too few fundable teams — price is the release valve.
Carta Q1 2026 · PitchBook-NVCA
10–30×
revenue multiples, seed through Series A
With a caveat about the denominator
Roughly 10–25× at seed, 15–30× at Series A. But at seed most companies sit at $0–1M ARR, which makes the multiple close to meaningless — seed is priced on team and trajectory.
Qubit Capital · Carta, 2026
11.4%
down-round rate, Q1 2026
No broad reset yet
Down from a 2023 peak above 20%. The correction, when it comes, will likely be company-specific rather than sector-wide.
Carta Q1 2026
40%
of 2026 seed and Series A dollars in rounds of $100M+
The barbell to watch
Fewer companies closing rounds, at higher prices. The headline median describes a market that the median founder is not raising in.
PitchBook, 2026
06 · Apply it
Where do your numbers actually sit?
Move the five sliders. Each track shows the benchmark zones from the sections above — below par, in band, top cohort — with your position marked.
Method
Questions worth asking of any benchmark page, including this one.
Where do these numbers come from?
Published research rather than proprietary client data: ICONIQ's State of AI and Compass Benchmarks, Bessemer's Cloud 100 and State of the Cloud, Carta's private-market data, CRV's seed and Series A KPI work, PitchBook-NVCA, High Alpha's SaaS benchmarks, ChartMogul retention data, and a16z's work on inference cost decline. Where sources disagree, both figures appear rather than an averaged number that belongs to nobody.
Why do published AI gross margin figures conflict so much?
Because they measure different populations at different moments. The 52% average, the ~60% capital-efficient cohort, the ~25% wrapper floor, and the 60–70% figure public companies now describe are all defensible — they simply describe different businesses. Any single number quoted without its population is a marketing number. Ask which cohort, which quarter, and what sits inside COGS.
Should I treat these as targets or as guardrails?
Guardrails, with one exception. Missing a benchmark is not fatal if you can name the mechanism and the timeline that closes it — investors fund trajectories. Missing one you can't explain is the actual problem. The exception is burn multiple at Series A: below 2.0× functions closer to a gate than a guideline.
How fast do these go stale?
Unevenly, which is the trap. Inference pricing and the valuation premium move quarterly and should be re-based every two quarters. Gross margin bands and retention move annually. ARR-per-employee for AI-native cohorts is the least stable of all, because the cohort definition is still being argued over.
Do these apply outside enterprise AI?
Partly. The margin and inference figures generalize well, since they follow architecture rather than market. Retention and NRR do not — the gap between prosumer and enterprise-priced AI is the single widest split in this dataset. Consumer AI, agents sold on outcomes, and vertical AI with heavy human-in-the-loop review each need their own bands.
Benchmarks tell you which conversation you're in.
They don't tell you what to do next. Modelling the margin trajectory, sequencing the migrations that bend it, and building a number you can defend in diligence — that's the work.
Book a call