Skip to content

Deal economics

Dataset refresh contract economics: renewal odds that beat a lump sum

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Dataset refresh contract economics turn on one number: the renewal odds at which a refresh contract's expected present value equals the larger one-time license. Normalize both offers to the same rights first, compound the odds because one declined refresh ends the stream, net out per-batch costs, and discount for time only. In this article's illustrative model, a 168-unit refresh headline needs 83 percent odds at every refresh to beat 100 upfront.

License terms

Dataset refresh contract

This page covers dataset refresh contract economics: renewal odds that beat a lump sum.Illustrative

Key takeaways

  • A refresh contract's headline total assumes every refresh is taken; at 80 percent odds, six scheduled refreshes yield 3.0 on average and a 26 percent chance of reaching the sixth.
  • The number to test is the break-even renewal odds, where expected present value equals the one-time license; in the illustrative base case it is 83 percent.
  • Normalize rights before cash: future records, exclusivity on refresh batches, model retention, per-batch rights review and acceptance terms each move the break-even.
  • Put renewal risk in the odds or in the discount rate, never both; doing both moved the illustrative break-even from 83 to 91 percent.
  • Renewal odds need evidence of a task that uses data up on a clock, like RealTime QA's 226 weekly question sets, not a stated intent to renew.

When does a refresh contract beat a larger one-time license?

A refresh contract beats a larger one-time license only when the buyer's odds of taking each refresh are higher than the break-even rate, the rate at which the contract's expected present value equals the one-time payment. Below that rate the larger upfront license is worth more, however large the refresh contract's headline total looks.

The decision is narrower than it sounds. Two offers sit on the table for the same historical records: a larger payment now, or a smaller payment now plus scheduled payments for later batches of new records, each of which the buyer can decline. An investor valuing a data supplier, a strategic buyer pricing a data asset and a company choosing between two term sheets all face the same comparison.

Three pieces of evidence settle it: both offers normalized to the same rights and scope, a renewal assumption stated per refresh rather than per contract, and a discount rate that carries only the risks the renewal odds leave out. This article builds each piece and joins them in one worksheet, with every model input labeled illustrative.

Why the headline total overstates a refresh contract

Because renewal odds compound. When one declined refresh ends the contract, the chance of reaching the sixth scheduled refresh is the per-refresh odds multiplied by themselves six times: at 80 percent each time that is 26 percent, and the expected number of refreshes actually taken is 3.0, not 6.

The illustrative base case shows the gap. A one-time license pays 100 units at signing. The refresh offer pays 60 units at signing for the same snapshot, then 18 units for each of six half-yearly batches of new records, a headline total of 168 units. Each batch costs the licensor 4 units to extract, prepare, de-identify and re-review for rights, and money is discounted at 5 percent per half-year. At 80 percent odds per refresh, the refresh offer's expected present value is 96.0 units, below the one-time license despite a headline 68 percent larger.

A second shortcut does almost as much damage. Treating each refresh as an independent yes-or-no that never ends the contract, the way an option is sometimes valued period by period, puts the same offer at 116.8 units at 80 percent odds. That method fits a buyer that can skip a batch and come back later; it overstates a contract that lapses at the first decline. Write down which kind of contract it is before modeling it.

The break-even test, run on seven illustrative cases

Solve for the renewal odds p at which the refresh offer's expected present value equals the one-time license. Expected value is the upfront payment plus, for each refresh t from 1 to T, p to the power t times the discount factor to the power t times the net value of one refresh. Net value is the accepted refresh payment minus the licensor's cost of producing the batch and minus any alternative sale the batch rules out.

Then read the answer against evidence, not hope. If the renewal odds you can document sit well above the break-even, the refresh contract earns its place; if they sit near or below it, take the larger license or ask for more money upfront. In the base case the refresh offer needs 83 percent odds at every refresh.

Three results run against intuition. Stretching the schedule from six refreshes to ten adds 72 units of headline value but lowers the break-even only from 83 to 79 percent, because distant refreshes are unlikely to be reached. Moving money from the upfront payment into the refreshes raises the headline from 168 to 176 units and makes the deal slightly worse, lifting the break-even to 84 percent. Making refresh batches exclusive to the buyer, at an assumed 3 units of alternative sales given up per refresh, lifts it to 90 percent.

Every row was computed by the script described in the methods note; inputs are illustrative units, and the one-time license is 100 in every case:

The break-even test, run on seven illustrative cases
CaseWhat changes from the baseHeadline total (units)Break-even renewal oddsExpected value at 80 percent odds
A. Base caseNothing: 60 upfront, 18 per refresh, cost 4 per refresh, 6 refreshes, discount 5 percent per half-year16883 percent96.0
B. Refresh ships only new and changed recordsCost per refresh falls from 4 to 216879 percent101.2
C. Refresh batches exclusive to the buyer3 units of alternative sales given up per refresh16890 percent88.3
D. Longer schedule10 refreshes instead of 624079 percent101.8
E. Renewal risk also loaded into the discount rateDiscount rises from 5 to 15 percent per half-year16891 percent88.4
F. Some batches fail acceptanceOne batch in ten is rejected and unpaid; its cost is still incurred16887 percent91.4
G. Smaller upfront payment, larger refreshes50 upfront and 21 per refresh17684 percent93.8
Line chart of expected present value against renewal odds per refresh, in illustrative units: a flat one-time license line at 100, a refresh contract curve that crosses it at 83 percent, and an exclusive-refreshes curve that crosses it at 90 percent, with a side panel showing the 168 headline value.
Break-even renewal odds for the illustrative base case and for exclusive refresh batches. Illustrative model values, not measured data.

Eight terms to equalize before the two offers can be compared

Normalize rights before comparing cash, because two offers that look like a price choice are often a choice between different bundles of rights. The worksheet names the eight terms that most often differ between a refresh contract and a one-time license, the model input each one moves and the document that settles it.

Future records deserve the first look. Under SourceX's published deal terms, exclusivity for AI training runs from Closing and covers only the selected data, with future data and unselected records left out, while the term and any alternatives are set in each deal's Dataset Deal Confirmation. A one-time license written that way leaves next year's records free for a separate, later deal. A refresh contract that commits those records to one buyer gives that freedom away, and the foreclosure input f is its price.

Model retention comes second. SourceX's supplier terms make its license to a de-identified dataset permanent once paid, surviving termination, and models already trained on it are not retrained or deleted. When a buyer keeps both the snapshot and its models, renewing buys only new information, so the renewal odds must rest on what the new records add. Under a term license that requires deletion at expiry, part of the renewal motive is keeping access, and the same odds mean something else.

The worksheet, with the input each term moves:

Eight terms to equalize before the two offers can be compared
Term to equalizeQuestion to ask of both offersModel input it movesDocument that settles it
Snapshot scopeDo both offers cover the same systems, history window, fields and exclusions?The one-time payment and the upfront payment: compare like-for-like snapshots onlyThe dataset definition in each offer
Future recordsDoes either offer reach records created after the cutoff, and on what terms?f, when later records are committed to the buyerThe clause defining what the license covers
Exclusivity on refresh batchesAre refresh batches exclusive for AI training, and for how long?f for each refresh while the contract runsThe exclusivity clause and its term
Model retention and survivalCan the buyer keep the snapshot and trained models if it stops renewing?p: renewal must be justified by new records aloneSurvival and termination clauses
Work per batchWhat extraction, preparation, de-identification and rights re-review does each batch need?cA costed run of the first delivery and the rights review of any new source
Payment conditionsIs each refresh payment conditional on acceptance of that batch?a, the share of batches accepted and paidAcceptance and payment terms
Buyer purposeTraining, evaluation or both, and on what clock does the task use data up?p and the length of a refresh periodThe buyer's stated use and its retraining or evaluation cadence
Discount basisWhich risks does the discount rate carry, and are any already in p or a?rThe valuation memo's stated rate build-up

Where should the renewal odds come from?

From evidence that the buyer's task uses data up on a clock, and from contract terms that make each renewal a live decision. Stated intent to renew is not evidence, and U.S. federal acquisition rules make the point in contract form: FAR clause 52.217-9 lets the government extend a contract's term by written notice after a preliminary notice of intent, and says that preliminary notice does not commit the government to an extension.

Some tasks consume data on a fixed clock, and those can support high odds. RealTime QA, an academic benchmark that asks questions about the present, published 226 weekly question sets between 17 June 2022 and 23 October 2026, by a count of its public repository made for this article. Of the 225 gaps between releases, 223 were exactly seven days and two were two-week breaks around the new year, and 77 of its 5,277 questions (1.5 percent) repeated an earlier one. For a task like that, last month's set has little continuing use, and the next batch is the product.

Data for slower-moving training tasks loses value more gradually, and there the refresh unit matters. The TemporalWiki authors built training sets from the differences between consecutive English Wikipedia snapshots; their EMNLP 2022 paper reports, in a paraphrase of its abstract, that continuing to train on those differences matched or beat training on whole snapshots in their tests at a fraction of the compute. A refresh that ships only new and changed records costs less to produce and to use, which is case B in the table.

Synthetic data cuts both ways. Alemohammad and colleagues reported in 2023 that generative models trained in self-consuming loops progressively lose quality or diversity unless enough fresh real data enters each generation (paraphrased from the abstract). That supports demand for refreshed real records; it says nothing about how much, or from whom.

Then count the buyers. A refresh contract with one buyer has one renewal decision behind every future payment, so the per-refresh odds are only as good as that buyer's budget and strategy. Price a refresh book concentrated in one or two buyers as one correlated bet, not many independent ones.

Put renewal risk in the odds, not in the discount rate

Count each risk once. IFRS 13, the fair value standard, says in paragraph B14(c) that discount rates should reflect assumptions consistent with those already in the cash flows, so that risk factors are neither double-counted nor omitted, and it illustrates the rule with loans: a rate that prices expected defaults suits contractual cash flows, not probability-weighted ones. Renewal odds are a probability weighting, so the rate applied after them should carry time value and the risks the odds leave out, such as the buyer's ability to pay.

The error is not small. Loading renewal risk into a 15 percent half-yearly rate while also applying renewal odds moved the illustrative break-even from 83 to 91 percent in case E, enough to swing a close decision toward the lump sum. Write the rate build-up into the valuation memo, item by item, beside the odds and the evidence for them.

Can posted prices or payout ranges settle the comparison instead?

No. Asking prices and published payout ranges are offers or estimates, not market-clearing prices, and neither says how often buyers actually renew. The comparison here is between two concrete offers, so the inputs that matter are the terms in those offers and the evidence behind the renewal odds.

The strongest version of the alternative holds that posted subscription and one-time prices already encode renewal expectations. They would if posted prices cleared, but a posted refresh price shows what a seller hopes to collect each period, not how many periods buyers pay for. A company's own early estimate is no better anchor: SourceX's published terms call any earlier estimate indicative and fix the exact payout only in the signed Dataset Deal Confirmation.

Use posted prices for one job only, as a ceiling check on the refresh payment. Executed contracts with comparable rights, normalized with the worksheet above, are the only price evidence that belongs in the model.

What could make the refresh comparison wrong

Six things can make the test mislead, and most of them flatter the refresh contract on paper while hurting it in practice:

  • Correlated renewals. Renewal odds borrowed from a portfolio of contracts assume independent decisions; a refresh book that rests on one or two buyers renews or lapses together.
  • Refresh cost creep. New records can arrive under new customer contracts, notices or platform terms, so the rights re-review inside c can grow over time instead of shrinking.
  • Substitutes. If synthetic or public data pass the buyer's acceptance test for the newer period, renewal odds fall whatever the quality of the records; the self-consuming-loop research supports demand for real data but sets no floor under it.
  • Reported revenue trails the economics. FASB's ASU 2016-10 bars recognizing license revenue before the customer can use and benefit from the license, and the Board chose to make renewals wait until the renewal period begins, so a refresh-heavy licensor's revenue line lags its signed contracts and expected renewals are not backlog.
  • Mismatched retention terms. A survival clause that ends model use at termination raises renewal odds through lock-in rather than value; compare offers under the same retention terms or the comparison is not like for like.
  • Sensitivity near the line. Between 80 and 90 percent odds, small errors in c, a or r move the break-even by several points, so report low, central and high cases rather than one figure.

How SourceX's published terms fit the test, and what to ask

SourceX's published supplier terms sit on the one-time side of this comparison. Its deal terms say it usually pays a one-time fee for a dataset, that some agreements instead set recurring deliveries paid as each agreement states, and that there are no royalties or usage fees. A returning supplier keeps its Master agreement and signs a new Dataset Deal Confirmation for each new deal, so outside an agreement that sets recurring deliveries, a later batch of records is a new, separately approved deal rather than an automatic renewal.

For an investor, that moves the refresh question to the supply side: how often suppliers come back with new records and sign again, and what each repeat costs to review. SourceX has not published those figures; its investor page says company and commercial metrics go to investors directly. Its current raise is open only to verified accredited investors under Rule 506(c), with materials available on request at sourcex.si/investors. Questions that apply the test to any data licensing company, SourceX included:

Checklist · 6 items

  1. Which contracts end at the first declined refresh, and which let a buyer skip a batch and return?
  2. What share of scheduled refreshes was taken in each cohort of contracts, with the count beside each rate?
  3. What did the latest refresh batch cost to extract, prepare, de-identify and re-review for rights, compared with the first delivery?
  4. Which refresh batches are exclusive to one buyer, and which alternative sales did that exclusivity rule out?
  5. Do buyers keep the snapshot and their trained models if they stop renewing, and how does that change the renewal evidence?
  6. Which discount rate does the valuation use, and which risks does it carry that the renewal odds do not?

0 of 6 ticked. Tick each item as you go.

Methods note: the break-even model, the RealTime QA count and their limits

The break-even model is refresh_breakeven.py, written for this article on 28 October 2026. It values the refresh offer as the upfront payment I plus, for refreshes t from 1 to T, (p times d) to the power t times (a times R minus c minus f), where d is 1 divided by (1 plus r), and values the one-time license as L paid at signing. The contract ends at the first declined refresh, so reaching refresh t has probability p to the power t, and the expected number of refreshes taken is the sum of those probabilities. The headline total is I plus T times R. Break-even odds are found by 60 rounds of bisection between 0 and 1. The independent-option figure replaces p to the power t with p. Values are rounded once, at output, to one decimal place and odds to the whole percent, half up. Every input is an illustrative unit chosen to show sensitivity, with L set to 100; nothing is drawn at random.

The RealTime QA count is rtqa_cadence.py, run on the data files of github.com/realtimeqa/realtimeqa_public at commit e288dcf (a sparse checkout; no repository code was run). It counts files named past/<year>/<YYYYMMDD>_qa.jsonl and latest/<YYYYMMDD>_qa_public.jsonl, matched case-sensitively, and lists the 577 other jsonl files in those folders that it skips, such as date-range compilations, retrieval results and public copies. Release dates come from file names under strict date parsing, and two files with the same date would stop the run; files are read with any byte-order mark removed, and every non-blank line must parse as a question with an ID and text; gaps are exact day differences between consecutive release dates. A question counts as a repeat when its ID, or its text after lower-casing and collapsing white space, appeared earlier, including earlier in the same file; the repeat share is computed from exact counts and rounded once.

Sources: SourceX's live how-it-works and investors pages, IFRS 13, FASB ASU 2016-10, FAR 52.217-9 and the RealTime QA repository; the TemporalWiki and self-consuming-loop papers were checked through search excerpts from arxiv.org, which could not be opened here, and are paraphrased. Not measured: any company's renewal rates, refresh costs or contract terms, SourceX's included; how buyer purpose maps to renewal odds beyond the one benchmark counted; and what buyers pay for refreshed records. RealTime QA is a free academic project, so its cadence shows how fast one task uses data up, not what anyone would pay for it.

Frequently asked questions

What is a dataset refresh contract?

A dataset refresh contract licenses a snapshot of historical records and adds scheduled later batches of new records, each paid when delivered and usually each one the buyer can decline. Its value depends on how many batches are actually taken, what each costs to produce and whether the batches carry exclusivity. A one-time license pays once for the snapshot.

How do you calculate break-even renewal odds?

Set the refresh contract's expected present value equal to the one-time license and solve for the per-refresh odds. Expected value is the upfront payment plus each refresh's net value, weighted by the chance of reaching that refresh and discounted for time. A spreadsheet goal-seek or a few lines of code finds the rate.

Should renewal risk go into the discount rate?

Not when renewal odds already carry it. IFRS 13 says discount rates should be consistent with the assumptions inside the cash flows, so risks are neither double-counted nor omitted. Applying both moved this article's illustrative break-even from 83 to 91 percent.

Does a longer refresh schedule make a contract worth more?

Less than the headline suggests. In the illustrative model, going from six refreshes to ten added 72 units of headline value but cut the break-even odds only from 83 to 79 percent, because late refreshes are rarely reached unless renewal odds are very high.

Does SourceX sign refresh contracts with suppliers?

SourceX's published terms say it usually pays a one-time fee for a dataset and that some agreements set recurring deliveries, each paid as the agreement states. A returning supplier signs a new Dataset Deal Confirmation for each new deal, and each deal's term is set in that document.

Sources

  • SourceX's deal terms: it usually pays a one-time fee for a dataset, some agreements set recurring deliveries paid as each agreement states, there are no royalties or usage fees, exclusivity for AI training covers only the selected data and excludes future data, returning suppliers sign a new Dataset Deal Confirmation, earlier estimates are indicative, and the license to the de-identified dataset is permanent once paid. Source
  • SourceX's investor page says it shares company and commercial metrics with investors directly and offers investor materials on request. Source
  • IFRS 13 paragraph B14(c): discount rates should reflect assumptions consistent with those inherent in the cash flows, to avoid double-counting or omitting the effects of risk factors. Source
  • FASB ASU 2016-10: revenue from a license of intellectual property cannot be recognized before the customer can use and benefit from it, and revenue for a renewal waits until the renewal period begins. Source
  • FAR 52.217-9 lets the government extend a contract's term by written notice and states that the preliminary notice of intent to extend does not commit the government to an extension. Source
  • TemporalWiki (EMNLP 2022) reports that continual training on the differences between consecutive Wikipedia snapshots matched or beat training on whole snapshots at much lower computational cost. Source
  • Self-Consuming Generative Models Go MAD (2023) finds that without enough fresh real data in each generation, self-consuming loops progressively lose quality or diversity. Source
  • RealTime QA's public repository: weekly question sets about current events, counted for this article from its data files at commit e288dcf. 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