Skip to content

Deal economics

Accepted and collected: an AI data startup capital efficiency test

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Judge AI data startup capital efficiency by what buyers accepted and paid for, divided by every dollar the operation consumed, capitalized automation included. Track four ratios by delivery cohort: acceptance yield, accepted value per outlay dollar, cash conversion and cash per outlay dollar. Then compare the latest spending increment with the average: a marginal ratio far below it signals that hires and tools are scaling cost faster than accepted delivery.

Value drivers

Accepted and collected

This page covers accepted and collected: an AI data startup capital efficiency test.Illustrative

Key takeaways

  • Activity is not efficiency: value counts once a buyer accepts a delivery, and cash counts once it arrives.
  • A SaaS burn multiple divides by new recurring revenue, which most data deliveries are not, so replace it with accepted value over full operating outlay.
  • Add capitalized automation, rejected candidates and partner fees back into the denominator, or the ratio flatters spending.
  • The marginal ratio on the latest spending increment shows cost outrunning acceptance quarters before the average does.
  • Release capital through five questions and a milestone table judged on frozen definitions and the full-outlay denominator.

What counts as capital efficiency in AI data infrastructure?

Capital efficiency in AI data infrastructure is the accepted value and the cash a company produces for each dollar its operation consumes, measured on delivery cohorts and on the latest increment of spending. Records processed, datasets shipped and contracts signed are activity. They count toward efficiency only once a buyer has accepted the delivery and paid for it.

The definition matters because a data business spends long before it is paid. Sourcing, rights work and preparation consume money first; value is created when a buyer signs off on a delivery against its acceptance test; cash arrives after that, on the contract's payment terms. Each of those gaps is a place where an efficiency figure can be flattered without anyone writing a false number.

The evidence that supports a useful answer is narrow: acceptance records for each delivery, invoices and receipts tied to them, and a ledger of everything the operation spent, including software development that was capitalized rather than expensed. This article turns that evidence into four ratios, an Illustrative worked example, a decision tree for the next tranche of capital and a milestone table. It estimates no company's ratios, SourceX's included, and quotes no benchmark for what a good ratio is, because the research for it found none published for this kind of business.

Why a SaaS burn multiple misreads a data delivery business

Because its denominator is recurring revenue, and most data deliveries are scoped projects that a buyer accepts once. A metric built for subscriptions either has nothing to divide by or divides by a number that overstates what the business has earned.

David Sacks defines the burn multiple as net burn divided by net new annual recurring revenue, presents it as an inverted version of Bessemer's efficiency score, and calls it a catch-all, because weak gross margins, poor sales efficiency or churn all push it up. For a software company selling subscriptions, that is a sound way to ask how much cash each increment of growth consumes.

A data licensing or preparation business breaks both ends of the ratio. Annualizing a single large delivery manufactures recurring revenue that may never recur, and counting contracts at signature puts value in the denominator before a buyer has accepted anything. On the other end, net burn leaves out development that the company capitalized, which is exactly where automation spending tends to go.

Keep the catch-all virtue and replace the inputs: accepted value instead of new recurring revenue, and full operating outlay instead of net burn. The ledger below does that.

Accepted value over every dollar spent: numerator and denominator

Count value only when a buyer accepts it, and count every dollar the operation consumed, including the ones that left the income statement. Both rules come from how contracts and accounting standards already treat acceptance and software spending.

Acceptance is a contractual event, not a shipment. In federal procurement, the Federal Acquisition Regulation says acceptance "constitutes acknowledgment that the supplies or services conform with applicable contract quality and quantity requirements," and elsewhere in the same subpart it allows acceptance before, at or after delivery, depending on the contract. A private data contract with an acceptance test works the same way: until the buyer signs off, the delivery is work in progress with a risk of rework.

Revenue figures do not settle the question on their own. In the revenue standard, FASB lists the customer's acceptance of an asset as one indicator that control has passed and directs companies to separate guidance on how an acceptance clause affects timing. Depending on the terms, revenue and acceptance can fall in different periods, so an investor should ask for acceptance records directly rather than infer them from the revenue line.

On the cost side, FASB's Accounting Standards Update 2025-06 says companies start capitalizing software costs once management has authorized and committed funding and it is probable the project will be completed. Capitalized automation then sits on the balance sheet instead of in operating expense, so an efficiency ratio built on operating expense quietly leaves it out. Add it back in the period it was spent, along with spending on candidate datasets that never shipped and fees paid to outside preparation partners.

The capital efficiency ledger: four ratios and their inflated twins

Compute four ratios on each delivery cohort, plus a marginal version, and reject the inflated twin of each. A cohort here is every delivery shipped in one quarter, followed until acceptance and payment.

Cash trails acceptance by design, so measure cash conversion only after the payment terms have run. Even the federal payment rule sets an invoice's due date at the later of 30 days after the billing office receives a proper invoice or 30 days after the government accepts the work. Measuring collections before open invoices age out makes the newest cohort look worse than it is, which can invent a decline or hide a real one.

Public companies already have to answer the cash question in their annual reports: Regulation S-K requires them to analyze their ability to generate and obtain adequate amounts of cash for the next 12 months and beyond. An early data company should be able to answer the same question from this ledger. The table sets out each ratio:

The capital efficiency ledger: four ratios and their inflated twins
RatioNumeratorDenominatorInflated twin to reject
Acceptance yieldValue of the cohort's deliveries that buyers acceptedValue of the cohort's deliveries shippedTreating shipped or invoiced deliveries as accepted
Accepted value per outlay dollarAccepted valueFull operating outlay, capitalized development includedOperating expense only, after automation was capitalized
Cash conversionCash collected on accepted deliveriesAccepted value, read after payment terms have runCollections read before open invoices age out
Cash per outlay dollarCash collectedFull operating outlayBookings or an annualized run rate in place of cash
Marginal accepted value per outlay dollarChange in accepted value between periodsChange in full operating outlayA lifetime average that hides the latest increment

Average or marginal: catching cost that outruns acceptance

Compare the ratio on the latest increment of spending with the average, because a marginal ratio falling well below the average is how hiring and automation that scale cost faster than accepted delivery show up in the numbers. The average alone moves slowly and can stay respectable for several quarters after new spending stops paying.

At a glance

Average or marginal: catching cost that outruns acceptance

A low marginal ratio for one period is not proof. New hires take time to ramp and automation takes time to settle, so compare after the ramp, and treat two consecutive post-ramp periods below the company's preset band as the signal (an Illustrative rule). The reverse pattern, a marginal ratio above the average, is what genuine operating leverage looks like, and a company claiming leverage should be able to show it this way.

The worked example is Illustrative and describes no real company. Outlay is in index points, with period A's full operating outlay set to 100 and every other figure on the same scale:

  • Period A: full outlay 100, nothing capitalized; deliveries shipped 72, accepted 60, cash collected 54.
  • Period B: full outlay 140, of which 15 went to capitalized automation; shipped 90, accepted 70, cash collected 56.
  • Acceptance yield: 83.3 percent in A, 77.8 percent in B.
  • Accepted value per outlay dollar: 0.60 in A; in B, 0.56 on operating expense alone, 0.50 on full outlay, and 0.25 on the increment.
  • Cash conversion: 90.0 percent in A, 80.0 percent in B. Cash per outlay dollar: 0.54 in A, 0.40 in B, and 0.05 on the increment.
  • Reading: the operating-expense view shows a small dip, the full-outlay view a larger one, and the increment produces accepted value at half of period B's full-outlay average and under half of period A's. Cash is sharper still: 40 extra points of outlay brought in 2 extra points of cash.
Horizontal bars comparing accepted value per outlay dollar: 0.60 in period A, then 0.56, 0.50 and 0.25 for period B on operating expense, full outlay and the latest increment.
One period read three ways. Illustrative: invented index-point inputs, no real company.

Five ledger questions that decide the next tranche

Ask five questions in order before releasing more capital; the first answer that comes back no names what to fund instead. The questions turn the ledger into a decision tree an investor can check against records rather than plans, and the bands they refer to are set by the company in advance and labeled Illustrative until it does.

The figure in the previous section shows the readings behind questions 2 and 4. The questions are:

  • 1. Can the company show acceptance by cohort, separate from shipment and invoicing? If not, fund acceptance records and buyer sign-off before any spending meant to scale.
  • 2. On the full-outlay denominator, is accepted value per outlay dollar steady or rising across recent cohorts? If not, find the stage that consumes outlay without producing acceptance, such as rework or rejected candidates, before adding capacity.
  • 3. Does the proposed spending go to the stage where work ready for acceptance is waiting longest? If not, rewrite the plan, because capacity added elsewhere raises cost without raising accepted output.
  • 4. After the ramp, is the marginal ratio inside the preset band around the average? If not, release capital in smaller tranches, each tied to the next milestone.
  • 5. Does cash follow acceptance within the contract's payment terms? If not, fix terms and collections first, because growth will consume cash faster than it returns it.
  • Five yes answers: release the next tranche and measure it with the same ledger and the same definitions.

Milestone table: what each tranche has to show

Release capital in tranches, each tied to a ratio the previous tranche had to hold and a stop signal written down before the money moves. The entries are Illustrative; the structure is what transfers.

Two rules keep the table honest. Definitions are frozen when the first tranche is released, so a ratio cannot improve because accepted started to mean delivered. And every tranche is judged on the full-outlay denominator, so capitalizing a cost cannot make a milestone easier to reach.

Milestone table: what each tranche has to show
TrancheMilestone evidence (Illustrative)Ratio that must holdStop signal
MeasurementAcceptance dates and a full-outlay ledger cover every delivery in the last two cohortsAll five ratios computable from recordsA cohort with deliveries but no acceptance dates
First added capacitySpending lands on the stage where accepted-ready work waits longestAccepted value per outlay dollar stays inside its bandMarginal ratio below half the average after ramp
Automation of one stepThe automated step's output passes buyer acceptance at the manual rateAcceptance yield unchanged or higherYield falls while shipped volume rises
Broader buyer baseAccepted deliveries from more than one buyer in a cohortCash conversion holds as volume growsCollections slip past contract terms
ScaleTwo post-ramp periods with the marginal ratio at or above the averageCash per outlay dollar risingAny ratio recomputed under a changed definition

What could break the capital efficiency case

A company can pass every milestone and still disappoint, and an investor should hear how before hearing the upside. Each risk below shows up in the ledger before it shows up in cash, which is the reason to keep the ledger:

  • Buyer concentration. A single buyer's acceptance pattern can keep every ratio steady until that buyer pauses, so read the ratios by buyer as well as by cohort.
  • Acceptance terms that tighten. Federal rules let acceptance fall before, at or after delivery depending on the contract; when buyers move it later or raise the bar, acceptance yield falls without any change inside the company.
  • Accounting policy changes. ASU 2025-06 changed when software capitalization begins, so comparisons across periods can shift with policy; ask whether the capitalization policy changed between the cohorts compared.
  • Price pressure and substitutes. Synthetic data or cheaper sources can lower the value buyers accept per delivery even when volume holds; no public measure of that effect exists for this market, so watch accepted value per delivery directly.
  • Payment terms that lengthen. Cash conversion depends on terms the buyer negotiates; longer terms tie up cash in every cohort and raise the funding the operation needs.
  • Definitions that drift. A ratio reported under a new definition is a new ratio; insist on a frozen definition sheet and restated history whenever a definition changes.

Capital efficiency questions for any AI data company, SourceX included

Put the same questions to every company in the space and ask for the records behind each answer rather than the ratios alone. SourceX should face them too.

For companies that license operational data to AI developers, SourceX takes on the whole transaction: it finds and qualifies suppliers, reviews rights, writes the contract, prepares the data, delivers it and settles payment. Because it sources datasets for buyers on request rather than holding them in stock, the questions worth asking it concern acceptance and collections rather than unsold inventory. Its acceptance, cost and cash figures are not public; the company discusses company and commercial metrics with investors one to one. Its Rule 506(c) raise is limited to verified accredited investors, and the investor page at sourcex.si/investors is where to request materials and send a specific diligence question.

  • For the last four delivery cohorts, what share of shipped value did buyers accept, and when?
  • What is total operating outlay per period with capitalized development and partner fees added back?
  • What was accepted value per outlay dollar on the most recent spending increment, after ramp?
  • How long after acceptance does cash arrive, by buyer, and how has that changed?
  • Which definitions changed in the last year, and has history been restated to match?
  • Which stage currently holds the oldest work ready for acceptance, and does the plan fund it?

Methods note: definitions, the worked example and what was not measured

The ledger, the decision tree and the milestone table are editorial frameworks drawn up for this article on 2026-10-27. Definitions rest on David Sacks's description of the burn multiple, FASB's revenue standard (ASU 2014-09) and internal-use software update (ASU 2025-06), the Federal Acquisition Regulation's acceptance and payment due-date rules (FAR 46.501 and 32.904) and the liquidity requirement in Regulation S-K Item 303, each checked on its publisher's site on that date.

The worked example is Illustrative: its inputs are invented to show how the readings diverge, are stated in index points rather than money, and are computed as exact fractions rounded once (ratios to two decimals, percentages to one). The thresholds in the decision tree and milestone table, including half the average after ramp and two consecutive periods, are Illustrative defaults for a company to replace with its own.

Not measured: any company's acceptance yield, outlay, cash conversion or marginal ratio, including SourceX's, and any market-wide benchmark for these ratios. The research for this article found no published benchmark for capital efficiency in AI data businesses, so none is quoted.

Frequently asked questions

What is a good capital efficiency ratio for an AI data startup?

No published benchmark exists for this kind of business, and thresholds built for subscription software do not transfer to project deliveries. Compare a company with itself: accepted value and cash per full outlay dollar across recent cohorts, and the ratio on the latest increment of spending against the average.

Should capitalized software count when measuring capital efficiency?

Yes, in the period the money was spent. Under ASU 2025-06, qualifying software costs are capitalized rather than expensed, so a ratio built on operating expense leaves automation spending out. Adding it back keeps a company from improving its ratio simply by capitalizing more.

Why measure acceptance instead of revenue?

Acceptance is the buyer's confirmation that a delivery meets the contract, and revenue can be recognized in a different period depending on the terms. Acceptance records by cohort show how much shipped work was taken and how much came back for rework, which revenue alone cannot show.

When should cash conversion be measured for a delivery cohort?

After the contract's payment terms have run, so open invoices that are not yet due do not count as failures. The federal payment rule, for example, sets the due date at the later of 30 days after a proper invoice arrives or 30 days after the government accepts the work.

Can automation improve an AI data company's capital efficiency?

Yes, when the automated step limited accepted output and its rules are stable. The evidence is a marginal ratio that rises after the ramp, on a denominator that includes the capitalized build. If the marginal ratio falls while volume rises, the automation is adding cost faster than acceptance.

Sources

  • David Sacks defines the burn multiple as net burn divided by net new ARR, describes it as an inverted Bessemer efficiency score and calls it a catch-all metric. Source
  • FAR 46.501 says acceptance constitutes acknowledgment that supplies or services conform with applicable contract quality and quantity requirements. Source
  • FAR subpart 46.5 allows acceptance before, at or after delivery depending on the terms and conditions of the contract. Source
  • FASB's revenue standard (ASU 2014-09) lists customer acceptance as an indicator that control has transferred and points to paragraphs 606-10-55-85 to 55-88 for acceptance clauses. Source
  • FASB ASU 2025-06 requires companies to start capitalizing software costs once management has committed funding and completion is probable. Source
  • FAR 32.904 sets the payment due date at the later of 30 days after receipt of a proper invoice or 30 days after government acceptance. Source
  • Regulation S-K Item 303 requires registrants to analyze their ability to generate and obtain adequate amounts of cash for the next 12 months and beyond. Source

Request SourceX investor materials

SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.

See if your company qualifies