Skip to content

Deal economics

AI data buying mandate vs minimum commitment: a six-rung demand test

By SourceX Editorial · Reviewed by Noah Loul ·

11 sections

Short answer

An AI data buying mandate authorizes a buyer to purchase; a minimum commitment obliges it to pay. Only the second is demand an investor can count before orders arrive. Read the contract for a stated floor, its term and the buyer's termination rights, then discount everything above the floor by buyer concentration. U.S. federal contracts draw the same line between a required minimum and a ceiling (FAR 16.504).

License terms

AI data buying mandate

This page covers AI data buying mandate vs minimum commitment: a six-rung demand test.Illustrative

Key takeaways

  • A mandate is permission to buy up to a ceiling; a minimum commitment is an obligation to pay for a floor. Demand estimates should start from the floor.
  • Six rungs separate interest from enforceable spend: expressed interest, evaluation scope, allocated budget, framework mandate, signed order and minimum commitment.
  • Filings show the gap: Innodata describes services subject to customer requirements and, in many cases, terminable on 30 to 90 days' notice, while one customer produced about 58% of its 2025 revenues.
  • Commitment and concentration compound: a floor from the only buyer is worth less than the same floor across independent buyers, and buyers on one budget cycle count as one.
  • Uncommitted demand is an option held by the buyer, not nothing; value it on order history and conversion by dated cohort, never on the mandate's headline size.

Does an AI data buying mandate commit the buyer to a minimum purchase?

No. A buying mandate authorizes a buyer to purchase: it names who may buy, what, from which suppliers and up to what amount. A minimum commitment obliges the buyer to pay for a stated quantity or spend over a stated term, whether or not it uses all of it. Only the second is demand an investor can count before orders arrive.

The two are easy to confuse because both arrive as signed paper. A master services agreement, a preferred-supplier listing, a framework award and an approved internal data budget can each be called a mandate, and each sets the rules for future orders. None obliges the buyer to place one unless a clause says so: a minimum annual commitment, a take-or-pay obligation or a shortfall payment.

Public procurement law draws the line explicitly. Under FAR 16.504, a U.S. federal indefinite-quantity contract must state a minimum and a maximum quantity, and the minimum must be more than nominal but should not exceed what the government is fairly certain to order. The regulation allows this contract type when the government cannot predetermine quantities above a specified minimum and it is inadvisable to commit to more than that minimum. The maximum is a planning number; the minimum is the commitment. An AI data contract deserves the same reading.

So the decision behind an ai data buying mandate minimum commitment question is narrow and answerable: how much of a data company's reported demand is a floor buyers must pay, and how much is an option they may never exercise. Contract language, order history and buyer concentration answer it. Announcements do not.

Why do headline mandates overstate AI data demand?

Headlines report the ceiling or the expected run rate, while revenue arrives order by order and can stop on notice. Three sets of public filings show the mechanics from different sides.

Innodata, a listed data engineering company, describes its services in its 2024 annual report as typically subject to customer requirements and, in many cases, terminable on 30 to 90 days' notice. Its 2025 annual report says one customer generated about 58% of total revenues in 2025 and 48% in 2024, with no other customer at 10% or more. The filing language describes demand set by the customer's requirements, not a purchase floor.

Appen shows what happens when that kind of relationship ends. It told the ASX in January 2024 that Google, as part of a strategic review, would terminate its global inbound services contract, with all projects ceasing by 19 March 2024. Appen's 12 February 2024 update then announced measures to achieve $13.5 million in annualized cost savings. The supplier had built capacity around the buyer; the filing describes no obligation on the buyer to keep using it.

Committed value behaves differently, but it is not permanent. According to Reddit's S-1, filed in 2024, the data licensing arrangements it entered in January 2024 totaled $203.0 million in aggregate contract value across two- to three-year terms, and it expected to recognize at least $66.4 million of revenue from them in 2024. That is a disclosed floor with a schedule. Reddit's annual report for 2024 then puts substantially all of its licensing contract value with two partners and cautions that renewal is not assured, or may come on less favorable terms. Every floor ends with its term.

The commitment ladder: six rungs from interest to enforceable spend

Place every claimed buyer relationship on the highest rung its documents support, and count only rungs five and six as demand before delivery. The ladder merges two questions investors usually ask separately: how strong is the buyer's intent, and what could the seller enforce if the buyer changed its mind.

Rungs one to three are evidence of need. Interest says a buyer has a problem, an evaluation scope says it is measuring one, and an allocated budget says the need survived an internal funding decision, which the buyer can still reverse. Rung four sets terms without volume. Rung five binds the buyer for one order, subject to acceptance. Only rung six binds it beyond the order in hand, and only until the term ends.

The counting rule is deliberately strict. It does not say lower rungs are worthless; it says they are options held by the buyer, and options need their own evidence, such as conversion by dated cohort, before they enter a demand estimate.

The commitment ladder: six rungs from interest to enforceable spend
RungWhat the paper saysWhat the seller can enforceEvidence to requestHow to count it
1. Expressed interestOpen call, inbound inquiry or a published list of data types soughtNothingA dated written request naming a task, data type and ownerPipeline signal only; not demand
2. Evaluation scopeSample request, evaluation license, pilot statement of work or benchmark planThe evaluation terms, if signedSigned evaluation terms, test plan, decision owner and decision dateCount the pilot itself; treat conversion as unproven
3. Allocated budgetAn approved budget line or requisition for a named data needNothing; the buyer can move the budgetBudget owner's confirmation of amount band, period and approval dateA probability input, never revenue
4. Framework mandateMSA, preferred-supplier status, framework award, an up-to or estimated valueTerms for future orders, but no volumeQuantity, termination and exclusivity clauses of the agreementOption value, discounted by notice period and concentration
5. Signed orderPurchase order, order form or statement of work with price and scopePayment for that order, subject to acceptanceThe order, acceptance criteria, cancellation terms and invoicesContracted backlog for that order only
6. Minimum commitmentMinimum annual commitment, take-or-pay or shortfall payment clauseThe floor amount over the term, used or notFloor amount, term, true-up mechanics, termination and renewal termsCommitted demand until the term ends, then renewal risk
A six-step staircase from expressed interest to minimum commitment, with a dashed line before the signed-order step marking where buyer options end and countable demand begins, and what the seller can enforce at the foot of each step.
The commitment ladder. Editorial framework: steps are ordered by what the seller can enforce, not by measured data.

What is the difference between data interest and a funded mandate?

Interest names a need; a funded mandate names a budget owner, an amount band, a period, a scope and an approval that survived a funding decision. To evaluate ai data buying mandate evidence, sort every claimed signal into one of four bins, exploratory interest, evaluation scope, allocated budget or enforceable commitment, and require a dated document for each.

At a glance

What is the difference between data interest and a funded mandate

Public statements from model developers sit at the bottom rungs by design. OpenAI's Data Partnerships page invites organizations to share large-scale datasets that reflect human society and are not already easily accessible online, either for an open-source archive or as private datasets for its proprietary models, and it excludes sensitive personal information and data belonging to third parties. It describes the data OpenAI wants and how to partner, which makes it a statement of interest, not a purchase commitment to any one supplier.

Evaluation demand is real spending, but it is scoped and specific. OpenAI's GDPval benchmark, published in 2025, covers 44 occupations across 9 sectors with 1,320 specialized tasks, 220 of them in an open-sourced gold set. Anthropic's engineering post on agent evals recommends starting with 20 to 50 simple tasks taken from real failures. Evaluation buyers need expert-made tasks with clear success criteria, counted in hundreds or thousands, so an evaluation scope shows a measurement problem. It does not show that a training purchase will follow.

Classify first, then test each bin with a question only a budget owner can answer. Champions and researchers can describe interest; they rarely control the line item, so their answers belong in rung one or two however senior they are.

  • Exploratory interest: what task would this data change, who owns that task, and what date is on the written request?
  • Evaluation scope: what result on the test triggers a purchase decision, who makes it, and by when?
  • Allocated budget: which line, what amount band, approved by whom and on what date, and can it be moved to another use?
  • Enforceable commitment: which clause obliges payment if the buyer takes less, for how long, and what ends it?

Which contract clauses move a deal up or down the ladder?

Quantity, termination and acceptance clauses set the rung, whatever the agreement is called. A document titled data license agreement can sit on rung four, and a two-page order form can sit on rung six, so read these clauses before building any revenue model.

  • Up-to, estimated, anticipated or not-to-exceed values: a ceiling. Count none of it until orders are placed.
  • No minimum purchase, or a non-exclusive grant with no volume clause: rung four. The buyer can route the same work to another supplier.
  • Termination for convenience: caps the life of uncommitted volume at the notice period. Innodata's 2024 annual report cites 30 to 90 days for many engagements.
  • Subject to acceptance: an order pays only for accepted scope, so a signed order can shrink at delivery.
  • Minimum annual commitment, take-or-pay or shortfall payment: rung six. Confirm the floor, the measurement period and whether unused value carries forward or lapses.
  • Minimum guarantee against royalties, with a true-up: rung six for the guaranteed amount only; the royalty above it is option value.
  • Renewal at the buyer's option: the floor ends at term end. Reddit's annual report for 2024 names this renewal risk for its licensing arrangements.
  • Exclusivity, most-favored pricing or volume rebates: change what a fixed floor is worth to the seller even when the floor itself does not move.

The buyer concentration and commitment matrix

Commitment and concentration compound each other: a floor from the only buyer is worth less than the same floor spread across independent buyers, and a mandate from a dominant buyer is the weakest evidence of all. The matrix gives a decision for each combination. Its concentration bands use the 10% line found in public-company customer disclosures, such as Innodata's, plus a majority band; the decisions in the cells are editorial judgments, not measured results.

Measure concentration on qualified demand, meaning signed orders plus committed floors over the trailing four quarters, not on inbound interest. Record each buyer's highest rung with the clause that puts it there, then read across.

The buyer concentration and commitment matrix
Highest rung heldNo buyer at 10% or more of qualified demandTop buyer at 10% to 50%Top buyer above 50%
Rungs 1-3: interest, evaluation or budgetPipeline. Track conversion by dated cohort and count none as demandPipeline with a dependency. Ask what replaces this buyer's interest if it fadesA thesis, not demand. Do not underwrite on it
Rung 4: framework mandateAn option portfolio. Value it on order history, discounted by notice periodsA concentrated option. Stress-test losing the top buyer within its notice periodDependence without a floor. Require a path to rung six or to new buyers
Rung 5: signed ordersContracted backlog. Count orders and check acceptance and cancellation ratesBacklog with renewal exposure. Treat the next order as uncertainOne customer's backlog. Count it, but value the company as if the next order may not come
Rung 6: minimum commitmentCommitted, diversified demand. The strongest position; check term lengthsCommitted but concentrated. Map each floor's end date and renewal termsA contract, not a market. Real until term end; renewal is a single event

Eight public documents coded onto the ladder, and the case that cuts against it

In a small set of public primary documents, only Reddit disclosed a committed contract value with a schedule; Innodata and Appen described relationships set by customer requirements and notice, and the developer documents sat at interest and evaluation. The set is original reporting for this article, coded on 2026-10-26.

Sampling frame: primary documents from 2023 to 2026 published by companies that sell data or data services to AI developers, or by AI developers describing the data they seek, where the text states a commitment, a termination or concentration term, or a stated data need. Sources were SEC filings, ASX announcements and developer pages. Excluded: press-reported deal values not confirmed by a party, equity investments presented as demand, private-company statements without a filing, and any SourceX record, since none is published. Every document is public; no interviews or private contracts were used.

The case that cuts against the headline is Innodata. Its filings describe demand set by customer requirements, yet its largest customer's share rose from 48% in 2024 to about 58% in 2025 while total revenues grew from $170.5 million to $251.7 million, according to its 2025 annual report. An investor who counted only disclosed floors would have found little to count, and the demand compounded anyway. The lesson is not that mandates are worthless. Uncommitted demand is an option held by the buyer, and its value has to be shown by order history and concentration, not assumed from the mandate's size.

What the set cannot show: it is small, chosen because the documents disclose terms, and limited to listed companies and published pages. Contract terms behind each relationship are mostly private, so a rung is the highest the public language supports, not the true contractual position.

Eight public documents coded onto the ladder, and the case that cuts against it
DocumentLanguage that sets the rungRung supportedWhat it does not show
Reddit S-1 (2024)Aggregate contract value of $203.0 million, terms of two to three years, at least $66.4 million expected in 20246: disclosed committed value with a schedulePer-partner amounts and renewal terms
Reddit annual report for 2024Substantially all licensing contract value from two partners; may not be renewed6, with concentration and renewal riskWhat happens at term end
Innodata annual report for 2024Services subject to customer requirements; many terminable on 30 to 90 days' notice4: framework language, no disclosed floorTerms of any single contract
Innodata annual report for 2025One customer about 58% of 2025 revenues and 48% of 2024 revenuesConcentration on rung 4 or 5 demandThat customer's contract terms
Appen ASX update, February 2024Google to terminate its contract; all projects to cease by 19 March 20244: a framework that ended within weeks of noticeWhether any notice period or floor applied
OpenAI Data Partnerships pageSeeks large-scale datasets not already easily accessible online1: open interestVolumes, budgets or prices for any supplier
OpenAI GDPval (2025) and Anthropic's agent evals post1,320 expert tasks; 20 to 50 tasks from real failures as a start2: evaluation scopeAny purchase that follows an evaluation

Could a prestigious mandate be worth more than its floor?

Sometimes, which is exactly why it needs its own evidence. A mandate from a demanding buyer can validate quality, shorten the next buyer's diligence and teach a supplier what labs accept. Those benefits are real, but they are not revenue, and the same mandate creates dependence without a floor.

At a glance

Could a prestigious mandate be worth more than its floor

Dependence shows up in cost and in leverage. A supplier hires, trains and builds tooling to one buyer's specifications; when the work stops, much of the cost stays until it is cut, as Appen's cost program after the Google termination illustrates. Leverage moves too: a supplier that cannot replace the volume within the notice period negotiates every renewal from the weaker side.

The alternative explanation for demand itself also deserves a test. Capability progress and benchmark results show where buyers' needs are moving, but they are not purchase orders; GDPval measures model performance on tasks, not spending on data. Sequoia Capital's 2023 essay AI's $200B Question estimated that a $50 billion GPU revenue run rate implied about $100 billion of data center spending and, with a 50% margin for end users, $200 billion of revenue needed, leaving a $125 billion or larger hole for each year of capital spending at that level. Investment ahead of revenue can reverse, so read announcements as rung one until a dated document moves them up.

Four tests separate a real prestige premium from a story:

  • Second buyer: did a different buyer sign an order within a stated period after the first mandate, and did the first mandate figure in its diligence?
  • Replacement: could the supplier replace the top buyer's volume within that buyer's notice period?
  • Conversion: what share of evaluation scopes reached a signed order, by dated cohort, with open cases shown separately?
  • Price: at renewal, did the price hold, fall or rise, and with what change in scope?

What could make this commitment framework wrong?

The framework is wrong if contract language stops predicting cash, and six developments could make it so. Name them before the upside; each comes with the public evidence an investor can track.

  • Contracts are mostly private. Public filings such as Innodata's and Reddit's describe terms in aggregate, so the ladder is only as good as the documents a company will share under confidentiality.
  • Floors can be renegotiated. A buyer with leverage can trade a lower floor for a longer term or a broader scope, so a stated minimum is a starting position, not a fixed outcome.
  • Committed value expires. Reddit's annual report for 2024 warns that its licensing arrangements could end without renewal, or be renewed on less favorable terms.
  • Uncommitted demand can outperform. Innodata's revenues grew from $170.5 million in 2024 to $251.7 million in 2025 with demand its filings describe as set by customer requirements; a rule that zeroes mandates would have missed it.
  • Budgets can fall together. If infrastructure spending outruns end-user revenue, as Sequoia Capital's 2023 analysis warned, buyers may cut data budgets at once, and spreading demand across labs protects less than the matrix assumes.
  • Evaluation may not lead to training purchases. Agent evaluation can start with 20 to 50 tasks, per Anthropic's guidance, so evaluation demand can stay small even when it is widespread.

Questions to ask any AI data company about its mandates, SourceX included

Ask for the clause, the order and the date behind every demand claim, and apply the same ladder to every company, SourceX included.

SourceX is one application of this question. It helps businesses license selected operating records to AI developers as datasets, and handles sourcing and qualification, rights review, contracting, preparation, delivery and payment along the way. It holds no stock of datasets for AI buyers; sourcing begins when a buyer uses the request form to set out scope, volume, language, format, timing and licensing needs. By this article's ladder, a request like that is rung one or two evidence: a reason to source, not a commitment to pay. On the supply side, no data is delivered until the company that owns the records has approved the scope and terms and signed the agreement.

SourceX has not published buyer, commitment or concentration metrics, and nothing here describes its demand. It shares company and commercial metrics with investors directly and is raising from accredited investors under Rule 506(c); investor materials are available on request through the investors page. A useful first diligence request is this article's matrix filled in: each buyer relationship on a rung, with the clause and date that put it there.

Checklist · 6 items

  1. For each buyer counted as demand, what is the highest rung, and which clause puts it there?
  2. What share of the last four quarters' qualified demand came from the top buyer, and which other buyers share its budget cycle?
  3. Which commitments carry a floor, how long is each term, and when does each renewal fall?
  4. What share of evaluation scopes converted to signed orders, by dated cohort, with open cases reported?
  5. If the top buyer gave notice today, how long is the notice period, and what capacity would sit idle?
  6. Which announcement or benchmark in the pitch has not yet produced a dated order?

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

Methods note: how the ladder and matrix were built, and what was not measured

The commitment ladder, the concentration and commitment matrix, the clause checklist and the document coding are editorial frameworks built for this article on 2026-10-26. Rungs are ordered by what a seller could enforce if a buyer changed course, with FAR 16.504's separation of a required minimum from a stated maximum as the reference model. The evidence set is the eight public documents in the coding table, each coded by one writer to the highest rung its language supports, without a second coder.

The interview design behind the funded-mandate test is published here as a method; its results are not yet published, and SourceX has not run the study. Population: people who control AI data budgets at model developers and enterprises, recruited independently of any supplier's customer list, excluding champions without budget authority. Variables: the rung reached, the document behind it, its date, an amount band rather than an amount, the approval path and the termination notice. Cohorts: quarter of first written request. Denominator: every signal that entered rung one in the cohort, with open cases kept as censored rather than dropped.

Not measured: the share of AI data revenue under minimum commitments in any market, conversion rates between rungs, typical floor sizes or notice periods across contracts, and any SourceX buyer or commitment data. The 10% and majority concentration bands are editorial thresholds anchored to public disclosure practice, not industry standards.

Frequently asked questions

Is a master services agreement a commitment to buy AI data?

Not by itself. A master services agreement sets terms for future orders, such as pricing schedules, security, rights and liability. It obliges the buyer to buy only if it contains a minimum purchase, take-or-pay or shortfall clause; otherwise each order form or purchase order is the binding commitment, for its own scope.

What is the difference between a minimum commitment and a minimum guarantee in a data license?

Both create a floor. A minimum commitment usually obliges the buyer to buy or pay for a stated volume or spend over a period, while a minimum guarantee usually sets the least the licensor receives against royalties, often with a true-up. Read the clause rather than the label: the amount, the period and whether unused value carries forward decide what the floor is worth.

How should investors count an evaluation license or a paid pilot?

Count the pilot itself if it is paid, and treat conversion to a training purchase as unproven until a dated order exists. Evaluation needs can be small: Anthropic's guidance says 20 to 50 tasks drawn from real failures is a great start for agent evals. Ask for conversion rates by dated cohort, with open cases shown.

Why does buyer concentration change the value of a minimum commitment?

A floor protects revenue only until its term ends, and a supplier with one dominant buyer negotiates the renewal with little leverage. Reddit's annual report for 2024 attributes substantially all of its licensing contract value to two partners and warns the arrangements may not be renewed. The same floor spread across independent buyers carries less renewal risk.

Does SourceX disclose its buyer commitments?

No. Buyer, commitment and concentration metrics for SourceX are unpublished, and its company and commercial metrics are provided to investors directly. Through sourcex.si/investors an investor can request materials and ask which rung of the ladder each buyer signal reaches; any offering is made solely to verified accredited investors.

Sources

  • FAR 16.504 requires a U.S. federal indefinite-quantity contract to state minimum and maximum quantities; the minimum must be more than nominal but should not exceed what the government is fairly certain to order, and the type fits when quantities above a minimum cannot be predetermined. Source
  • Innodata's annual report for 2024 says its services are typically subject to customer requirements and, in many cases, terminable on 30 to 90 days' notice. Source
  • Innodata's annual report for 2025 says one customer generated about 58% of total revenues in 2025 and 48% in 2024, no other customer reached 10%, and total revenues were $251.7 million in 2025 and $170.5 million in 2024. Source
  • Appen's 12 February 2024 ASX update restates that Google would terminate its global inbound services contract as part of a strategic review, ceasing all projects by 19 March 2024, and announces measures to achieve $13.5 million in annualized cost savings. Source
  • Reddit's S-1 disclosed data licensing arrangements entered in January 2024 with an aggregate contract value of $203.0 million and terms of two to three years, with at least $66.4 million expected to be recognized in 2024. Source
  • Reddit's annual report for 2024 says substantially all contract value associated with its licensing revenue comes from two partners and that the arrangements may not be renewed or may be renewed on less favorable terms. Source
  • OpenAI's Data Partnerships page seeks large-scale datasets that reflect human society and are not already easily accessible online, for an open-source archive or private datasets for proprietary models, and excludes sensitive personal and third-party information. Source
  • OpenAI's GDPval benchmark covers 44 occupations across 9 sectors with 1,320 specialized tasks, 220 of them in an open-sourced gold set. Source
  • Anthropic's engineering post Demystifying evals for AI agents says 20 to 50 simple tasks drawn from real failures is a great start for agent evals. Source
  • Sequoia Capital's essay AI's $200B Question estimated that a $50 billion GPU revenue run rate implied about $100 billion of data center spending and $200 billion of revenue needed with a 50% end-user margin, a $125 billion or larger hole per year of capital spending at that level. 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