Skip to content

Deal economics

Fresh dataset pricing: five proofs before a recency premium counts

By SourceX Editorial · Reviewed by Noah Loul ·

10 sections

Short answer

Fresh dataset pricing holds up only when a seller can show that recent records beat the best historical alternative on the buyer's own task, by more than cheaper routes recover, for long enough to matter, on records that are genuinely new and covered by the rights. Price the historical baseline first, credit recency as a decaying increment, and treat any surcharge beyond that as unproven.

Value drivers

Fresh dataset pricing

This page covers fresh dataset pricing: five proofs before a recency premium counts.Illustrative

Key takeaways

  • A fresh dataset's price bundles three lines: the value of the best historical alternative, a recency increment and the cost of producing the next batch.
  • Markets already price delay apart from content: under Nasdaq's policies, delayed usage of an eligible data product may carry no charge.
  • EU law ties dynamic data's economic value to immediate availability and regular updates, and excuses public bodies when instant release costs too much.
  • Research cuts both ways: target-period labeled data beats unlabeled temporal adaptation, yet training on Wikipedia diffs matched full snapshots at 12 times less compute.
  • In this article's illustrative worksheet, a 160-unit asking price on a 100-unit baseline supports only 114.6 once decay, substitutes and rights are applied.
  • Freshness is a flow, not a stock: each batch ages into history, so the case rests on producing the next batch for less than its increment.

What does a freshness premium actually buy?

It buys the gap between what a buyer's task scores with current records and what it scores with the best historical alternative, for as long as that gap lasts, minus whatever a cheaper route could recover. Anything a seller charges above that gap is an asking price, not evidence of value.

The distinction matters to an investor because fresh is easy to claim and hard to prove. A dataset sold as current can hold mostly old material, serve a task that barely changes, or answer questions the buyer could look up when the question is asked. Each of those deserves the historical price, not a recency premium.

This page sets out what must be proven before fresh dataset pricing enters a valuation: decay, substitutes, novelty, duration and rights. It ends in a worksheet in illustrative units, a worked decomposition of one asking price, the risks and the questions to put to any data company, SourceX included. No figure here is a measured result or a SourceX number.

Where markets already price recency apart from content

Recency already carries its own price wherever the same records are sold at different delays. Nasdaq's data policies let distributors delay certain information and recognize fixed delay intervals, among them 15 minutes, 24 hours and midnight local time, and they state that when a product is eligible for delayed pricing and the data is delayed as the policy requires, "there may not be a charge for the usage of the delayed data." The records are identical; only their age changes the price.

Lawmakers have written the same idea into public-sector data rules. The EU Open Data Directive, (EU) 2019/1024, defines dynamic data as documents subject to frequent or real-time updates because of their volatility or rapid obsolescence, and its recitals say the economic value of such data depends on the immediate availability of the information and of regular updates. Article 5(5) requires public bodies to release dynamic data immediately after collection, and Article 5(6) allows a delay where that would exceed their financial and technical capacities.

Both cases carry two lessons. Recency becomes a separate price line only where records age fast, as quotes, traffic counts and sensor readings do. And producing recency costs money, which is why even a law demanding immediate release excuses bodies that cannot afford it. Most business records licensed for AI training sit far from a stock quote on that spectrum, so their premium has to be shown case by case.

Three prices inside one fresh quote

Split every fresh quote into three lines before comparing it with anything: the historical baseline, the recency increment and the refresh obligation. The baseline is what the best historical alternative is worth for the same task, scope and rights. The increment is what recent records add on top. The refresh obligation is what it costs the seller to produce the next fresh batch, because today's batch becomes tomorrow's history.

Rights normalization comes first, as in any comparison of deals. The historical alternative has to cover the same use, the same fields and the same preparation, or the baseline is wrong before freshness enters the picture. Older records licensed for a narrower use, or delivered raw where the fresh set arrives de-identified and linked, are a different product, not the same product at an earlier date.

The split also shows where the money is. A seller that can offer the increment once has a one-off premium; a seller that can produce each new batch for less than its increment has a flow. An investor valuing a data company should know which of the two it is buying.

Five proofs before a recency premium counts

Credit a recency premium only when all five proofs hold; one failed proof sends the price back to the historical baseline. Each can be tested with evidence a serious seller should already hold, and the live protocol on measuring what a refresh adds covers the experiment itself. This page asks what the result is worth.

The five proofs, in the order a diligence team can check them:

  • Decay on the buyer's own task. A model built on the historical alternative must score worse on a holdout from the latest period than one built on recent records. Luu and colleagues found across eight tasks that temporal misalignment degrades end-task performance, and that gains from continued pretraining on recent unlabeled text were small next to finetuning on labeled data from the target period (paraphrase). Recent records with outcomes are therefore the ones that can carry a premium.
  • No cheaper route closes the gap. More historical volume, public data or retrieval of current facts at answer time can each recover part of the increment. TemporalWiki's authors report that continual training on the differences between Wikipedia snapshots reached similar or better perplexity than training on the full snapshot at 12 times less compute (paraphrase). When only the changes matter, the buyer should pay for the changes, not for the whole archive again.
  • The records are actually new. A fresh delivery can repeat old material. Cheng and colleagues found that language models' effective knowledge cutoffs often differ from the cutoffs their developers report, and traced the gap partly to non-trivial amounts of old data in new Common Crawl dumps and to deduplication that misses near-duplicates. Count the genuinely new share of a fresh batch before crediting all of it.
  • The increment lasts long enough. Recency wears off as a batch ages, so a premium measured at delivery has to be averaged over the window the buyer actually uses the data, up to its next retrain.
  • The rights reach the new records, and the next batch can be made. Records created after an earlier deal may sit under newer customer contracts, notices or vendor terms, so each fresh batch needs its own rights check, and the seller needs a repeatable way to produce the next one for less than its increment.

A rights-normalized worksheet for fresh dataset pricing

The worksheet turns the five proofs into one number: the premium a buyer could rationally pay above the historical alternative, and the margin left after the seller's refresh cost. Inputs are in units where 100 is the rights-normalized value of the best historical alternative, so no money amount or SourceX figure appears.

The rule is a product of the proofs. The supportable premium equals the increment measured at delivery, times the share of it kept on average over the use window given its half-life, times the share no cheaper route could deliver, times the share of new records the rights cover. Subtracting the cost of producing one fresh batch gives the net freshness margin, and a case pays only if that margin is above zero. A script computed the five illustrative cases below:

A rights-normalized worksheet for fresh dataset pricing
Illustrative caseIncrement; half-life; window; substitutable; rights-covered; refresh costSupportable premium (units)Margin after refresh costVerdict
Fast-changing task: answers tied to current prices and policies40; 6 months; 12 months; 25%; 90%; 814.6+6.6Pays, but far below a quote of 160
Slow-changing workflow: procedures that rarely move10; 36 months; 12 months; 50%; 90%; 84.0-4.0Price at the historical baseline
Facts a buyer can retrieve at answer time40; 3 months; 12 months; 80%; 90%; 82.4-5.6Retrieval beats a premium
Fast-changing task, 60 percent of new records outside the rights40; 6 months; 12 months; 25%; 40%; 86.5-1.5Fix the rights before pricing recency
Evaluation items kept unseen after a model's cutoff30; 9 months; 6 months; 10%; 100%; 1221.6+9.6Pays while the items stay unseen

One asking price, taken apart

Applied to the first case, a quoted price of 160 units against a historical baseline of 100 leaves most of the quote unproven. The increment measured at delivery is 40 units; decay over a 12-month window with a 6-month half-life removes 18.4, substitutes remove 5.4 more and the rights gap 1.6, so the supportable price is 114.6.

That leaves 45.4 units of the quote with no evidence behind it. The seller's refresh cost of 8 units still fits inside the 14.6-unit premium, so this is a real freshness business, only a much smaller one than the quote implies. The chart shows each step, and every value in it is illustrative.

Whichever deduction is largest tells a diligence team where to dig. Here decay dominates, so the questions concern the buyer's retraining cadence and the half-life evidence. In the retrieval case the substitute line dominates instead, and no amount of recency evidence would rescue the quote.

Waterfall chart in illustrative units: a historical baseline of 100 plus a 40-unit increment at delivery, minus 18.4 for decay, 5.4 for substitutes and 1.6 for the rights gap, reaching a supportable price of 114.6 against an asking price of 160, with 45.4 unproven.
One asking price taken apart. Illustrative: all values are example units from this article's worksheet, not measured data.

Is a quoted freshness surcharge evidence of what recency is worth?

No. A surcharge on a quote, a list price or a public payout range shows what a seller hopes for, not what a buyer paid after testing substitutes. Statistical agencies treat the difference as basic: the U.S. Bureau of Labor Statistics asks businesses reporting to the Producer Price Index for realistic transaction prices that reflect all discounts, premiums, rebates and allowances, rather than list or book prices, and calls the use of list prices "the exception, rather than the rule."

Freshness quotes deserve the same treatment. The strongest case for taking them at face value is that buyers in fast-moving tasks do pay for recency, as delayed-data pricing shows. That holds where records age in minutes and the delayed copy is the substitute. For business records the substitute is usually a larger historical archive, public data or a retrieval step, and a quote never tested against those is a hypothesis.

The test is easy to state: ask what a buyer actually paid for the same scope delivered older, or what the same buyer pays for the next batch once the first has aged. Until one of those exists on paper, the worksheet's supportable price, not the quote, belongs in the model.

What could make a freshness thesis wrong

Five developments could shrink or erase a recency premium. Put them on the table before modeling the upside, and ask which one a company is most exposed to:

  • Retrieval absorbs the premium. If buyers keep facts current by looking them up at answer time instead of training on them, recency moves from the dataset to the lookup, as the retrieval case in the worksheet shows.
  • Diffs replace snapshots. The TemporalWiki result suggests that when only the changes matter, a buyer can update on deltas for a fraction of the compute, which caps what a full fresh snapshot can command.
  • Fresh turns out not to be new. If a fresh batch mostly repeats earlier content, as the Dated Data authors found with old copies inside new crawl dumps, the novelty share sets value, not the delivery date.
  • Rights lag the records. New records arrive under new contracts and notices; a seller that cannot re-clear each batch has a stale-rights problem, not a fresh-data business.
  • Refresh costs climb. Extraction, review and preparation recur with every batch, and if they rise above the increment the flow turns negative even while demand for recency is real.

Questions for any fresh-data seller, SourceX included

Ask for the evidence behind each proof, defined your way, and hold SourceX to the same list. On its published terms, licenses are exclusive for AI training, the term and any alternatives are set in each deal's Dataset Deal Confirmation, and the term covers only the selected data, with future data and unselected records not included. A supplier returning with more data reuses its Master agreement and signs a new Confirmation, so whether newer records earn more than historical value is settled deal by deal rather than carried over from an earlier one.

SourceX does not publish how it prices offers, and none of its deal, volume or commercial figures are public; the company keeps those for investors rather than articles. Because the company's raise relies on Rule 506(c), only verified accredited investors can take part; the investors page takes requests for materials, and any question below is a fair one to send with it.

  • On which buyer task was decay measured, against which historical alternative, and on a holdout from which period?
  • What share of the last fresh batch was genuinely new once duplicates, revisions and continuations of older cases were set aside?
  • Which cheaper routes, such as more historical volume, public data or retrieval at answer time, were tested against the fresh records, and what did each recover?
  • What half-life does the evidence support, and how often does the buyer retrain?
  • Were the new records rights-checked as a batch of their own, and what did that check exclude?
  • What did the last fresh batch cost to produce, and how does that compare with the increment it delivered?

Methods note: sources, worksheet rules and limits

The three-price split, the five proofs and the worksheet were written for this article on 2026-10-28 as editorial tools. Public evidence: Nasdaq's US equities and options data policies, the EU Open Data Directive (2019/1024), Luu and colleagues' NAACL 2022 study of temporal misalignment, the TemporalWiki benchmark paper (EMNLP 2022), Cheng and colleagues' Dated Data paper (2024) and the BLS Handbook of Methods chapter on Producer Price Index concepts, all checked through search excerpts on the publishers' own sites and paraphrased unless quoted, plus SourceX's how-it-works and investors pages, read directly.

Worksheet rules, as the script computes them: 100 units is the rights-normalized value of the best historical alternative; the increment at age t is the delivery increment times 0.5 raised to t over the half-life; the average share kept over a window W is the half-life divided by the product of W and ln 2, multiplied by one minus 0.5 raised to W over the half-life; the supportable premium is the increment times that share, times the non-substitutable share, times the rights-covered share; the margin is the premium minus one batch's refresh cost, and a case pays only when the margin is above zero. Values are rounded once to one decimal, and the chart's steps are differences between the rounded values. Every input is illustrative.

Not measured: decay, substitution, novelty or refresh cost for any real dataset, the prices buyers actually pay for recency in AI data deals, and anything about SourceX's deals. Those need transaction records nobody has published. The worksheet shows how the evidence should combine once a seller produces it.

Frequently asked questions

Is fresh data always worth more than historical data?

No. Recency adds value only when the buyer's task changes over time and no cheaper route closes the gap. In this article's illustrative worksheet, a slow-changing workflow supports a premium of 4.0 units on a 100-unit baseline, less than the 8 units a fresh batch costs to produce.

How do you measure a dataset's freshness premium?

Build on the best historical alternative and on recent records, score both on a holdout from the latest period, and translate the gap into value for the buyer's task. Then discount it for decay over the use window, for cheaper substitutes and for new records the rights do not cover.

Why does freshness decay, and how fast?

Each batch starts aging the day it is delivered, so the advantage it gave a model shrinks as the world moves on. The speed depends on the task: quotes and prices age in minutes or months, while procedures can hold for years. A half-life estimated from the buyer's own retraining history is better evidence than any general rule.

Do AI buyers pay extra for real-time data?

Some markets price real-time access separately, as Nasdaq's delayed-data policy shows, but most business records used for AI training are not real-time products. For them the question is whether recent records beat a larger historical archive or a retrieval step on the buyer's task, which only a test can answer.

What should an investor ask a company selling fresh data?

Ask which task decay was measured on, what share of the last batch was genuinely new, which substitutes were tested, what half-life the evidence supports, whether the new records were rights-checked as their own batch and what the batch cost to produce.

Sources

  • Nasdaq's data policies let distributors delay certain information, recognize delay intervals including 15 minutes, 24 hours and midnight local time, and say delayed usage of an eligible product delayed under the policy may carry no charge. Source
  • The EU Open Data Directive defines dynamic data by volatility or rapid obsolescence, says its economic value depends on immediate availability and regular updates, requires release immediately after collection (Article 5(5)) and allows delay where that exceeds a body's capacities (Article 5(6)). Source
  • Luu et al. (NAACL 2022) found across eight tasks that temporal misalignment degrades end-task performance and that continued pretraining on recent text helps far less than finetuning on labeled data from the target period (paraphrase). Source
  • TemporalWiki (Jang et al., EMNLP 2022) reports that continual learning on Wikipedia snapshot differences reaches similar or better perplexity than training on the full snapshot at 12 times less computational cost (paraphrase). Source
  • Cheng et al. (2024) find that effective knowledge cutoffs often differ from reported cutoffs, citing old data in new Common Crawl dumps and deduplication complications with semantic and near-duplicates. Source
  • The BLS Handbook of Methods says the PPI seeks realistic transaction prices including discounts, premiums, rebates and allowances rather than list or book prices, with list prices the exception rather than the rule. Source
  • SourceX's how-it-works page says licenses are exclusive for AI training, the term and alternatives are set in each Dataset Deal Confirmation, the term covers only the selected data with future data and unselected records not included, and returning suppliers sign a new Confirmation. Source
  • SourceX's investors page describes it as licensing business data directly from companies and onward to AI developers, and offers company and commercial metrics to investors on request. 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