AI Startup Funding: Reading the Numbers Without Being Misled

No figures for round sizes, valuations or totals appear on this page. They change quarterly, they are reported selectively, and repeating one without its source and date would be worse than useless. What does not change is why the financing requirement is different when serving a customer costs real money.

Three cash drains classic software lacks

A conventional software company burns cash on people and on acquisition. Both are discretionary in the short run: stop spending and the burn stops. A company with metered cost of goods has three additional drains, and none of them is discretionary while customers are being served.

Serving costs before revenue

Trials, free tiers and onboarding all consume inference before anyone pays. In classic software these cost approximately nothing; here they are a per-user cash cost with a conversion rate applied to it. The arithmetic is worked through in the free-tier page, and the summary is that a free tier is an acquisition line item that scales with signups rather than with spend decisions.

Payback stretched by margin

months to recover acquisition cost = CAC / (ARPU * m)

Assume CAC = $900 and ARPU = $80 per month:

  m = 0.85  ->  900 / 68 = 13.2 months
  m = 0.60  ->  900 / 48 = 18.8 months

Every customer is funded for five and a half months
longer at the lower margin, and the capital needed
to add customers at a given rate rises in the same
proportion.

Working capital that scales with volume

Cost of goods is incurred when the work happens and collected when the invoice is paid, so a permanent slug of cash is tied up in the gap and it grows with the business. The calculation is in the business-models page. None of these three is a sign of anything being wrong; they are properties of the cost structure, and a plan that does not include them will run out of money while every operating metric looks healthy.

Recurring revenue is not comparable

Comparing companies on recurring revenue implicitly assumes their gross margins are similar. Once one of them has a metered input, that assumption fails, and the comparison stops meaning anything:

Company A: revenue R, margin 0.85 -> gross profit 0.85R
Company B: revenue R, margin 0.55 -> gross profit 0.55R

Identical headline revenue. B has 35% less money to
pay for everything that is not cost of goods —
engineering, sales, support, and the next customer.

The comparable unit is gross profit, not revenue.

Two habits follow. Quote gross-profit growth alongside revenue growth, and be suspicious of any comparison, benchmark or multiple that does not state the margin it assumes. A revenue multiple is a margin assumption wearing a disguise: applying the same multiple to two businesses with different cost structures says they are worth the same per unit of revenue, which is only true if their gross profits per unit of revenue are the same.

Burn multiple, and why it reads worse

burn multiple = net cash burned / net new recurring revenue

Assume two companies each add $1,000,000 of new
recurring revenue in a year.

  Company A, margin 0.85:
    COGS on the new revenue = $150,000
  Company B, margin 0.55:
    COGS on the new revenue = $450,000

If everything else about them is identical, B burned
$300,000 more for the same headline growth, and its
burn multiple is worse by exactly that amount divided
by the new revenue.

This is worth internalising because it is frequently misread as inefficiency. The company with the metered input is not worse run; it is buying revenue that costs more to deliver. Judging both on the same burn-multiple threshold penalises a cost structure rather than a management team. The right adjustment is to compute the multiple against new gross profit, which makes the two comparable again.

The same adjustment applies to efficiency rules of thumb generally. Targets that combine growth with profitability were calibrated on businesses whose cost of goods was small and mostly fixed, and applying them unmodified to a business with a metered input imports an assumption that no longer holds. This does not make the targets wrong; it makes them targets for a different cost structure. State the margin the benchmark assumed, and compare like with like or not at all.

Survivorship, and the denominator problem

Every statistic about funding is drawn from a biased sample, and the bias runs one way. Announcements are voluntary: rounds get publicised when they are flattering, down rounds and bridge financings often do not. Failures are silent — companies rarely announce that they stopped. Databases are assembled from announcements, so they systematically over-represent success.

What you want:
  P(outcome | started)

What a funding database gives you:
  P(outcome | started AND announced AND still
             visible enough to be recorded)

The correction requires the denominator — every
company that started, including the ones that
vanished — and that denominator is exactly what
nobody collects.

A related distortion works through time rather than through selection. Any statistic about a young cohort is measured before most of its outcomes have happened, so the observed failure rate is always lower than the eventual one, and the newer the cohort the more flattering it looks. Comparing a two-year-old cohort to a ten-year-old one and concluding that the recent one is doing better is a comparison of elapsed time, not of quality.

The practical defences are simple and rarely applied. Insist on a cohort: “of companies that raised a first round in year X, what share had raised again by year X+2” is answerable and comparable, while “the average round size this quarter” is a statement about which companies chose to announce. Check whether the source counts only disclosed rounds. Check whether currency, geography and stage definitions changed between periods. And treat any comparison across years as suspect unless the classification of what counts as an “AI company” was held fixed — that definition has been widening, which alone can manufacture a trend.

What to look up, and how to record it

The current figures have to come from primary sources, and this page will not pretend otherwise. When you collect them:

• Record the period, the source and the date you read it. A funding figure without a period is not a fact about anything.
• Prefer disclosures over aggregations. A regulatory filing has a defined scope; an industry roundup is a sample with undocumented inclusion rules.
• Separate committed from deployed. Announced capital, drawn capital and spent capital are three different numbers and are routinely conflated.
• Ask what a valuation is a claim about. A price paid for a small share, often with liquidation preferences attached, is not the same as a value for the whole company, and the difference is large enough to change conclusions.

That last point is the bridge to the broader argument about whether current investment is proportionate, which needs the same discipline and is taken up in the bubble question.