Skip to content

Deal economics

Proof, leverage or dependence: an AI data buyer concentration risk map

By SourceX Editorial · Reviewed by Noah Loul ·

11 sections

Short answer

AI data buyer concentration risk is usually larger than the revenue share shows. One large AI buyer can validate a startup's data, but it also gains pricing leverage and can become the whole budget. Place the relationship on a five-stage ladder, measure the top buyer's share of revenue, pipeline, preparation work and specialist capacity, and stress-test delay, cancellation and reduced scope separately. Growth that vanishes without that buyer is dependence, not traction.

Value drivers

Proof

This page covers proof, leverage or dependence: an AI data buyer concentration risk map.Illustrative

Key takeaways

  • A large AI buyer supplies validation, pricing leverage and budget dependence at once; only accepted, paid deliveries prove the first.
  • Revenue share lags. Pipeline, preparation exposure and specialist capacity show where dependence is heading, and the highest of the four sets the risk band.
  • Map the buyer's alternatives from public evidence instead of guessing its data budget from infrastructure spending.
  • Delay, cancellation and reduced scope hit different measures, so test each one separately on the company's own records.
  • Report growth excluding the top buyer beside the headline: Appen's FY24 group revenue fell 14.2 percent while adjusted revenue without Google rose 16.0 percent.

What does one large AI buyer really mean for a data startup?

It means three things at once, and an investor has to price them separately: validation, because a demanding buyer will test and pay for the data; pricing leverage, which the buyer gains as the startup comes to depend on it; and budget dependence, where one program's spending decisions set the startup's revenue. AI data buyer concentration risk is the distance between the first and the other two.

The validation is real but bounded by stage. A buyer that has accepted and paid for a delivery has told the market something about quality; a buyer that has only taken a meeting has not. Leverage and dependence grow quietly, because they appear first in the pipeline, in preparation work built to one buyer's specification and in the specialists assigned to its program, and only later in revenue.

This article separates the three with tools an investor can reuse: a five-stage relationship ladder, a Single-Buyer Exposure Map of four concentration measures, a map of the buyer's alternatives, three downside cases tested one at a time, and a revenue bridge that shows whether growth survives without the top buyer. Every example value is labeled Illustrative; public-company figures come from their own filings.

Which stage has the buyer relationship actually reached?

Place the relationship on one of five stages before counting anything: introduction, evaluation, mandate, signed purchase or paid delivery. Each stage proves something specific, and most overstated traction comes from describing an earlier stage in the language of a later one.

The public record shows how wide the gap between stages can be. OpenAI's Data Partnerships page, checked on October 26, 2026, invites organizations to offer large-scale datasets that are not already easily accessible online, either to an open-source archive or as private datasets for proprietary models, and says OpenAI is not seeking sensitive personal information or information that belongs to a third party. A startup that answered that invitation has reached the introduction stage. It has not reached evaluation, let alone a purchase.

A mandate needs particular care. A program approved to source a category of data is a budget signal, not a minimum purchase from any one supplier, and its commitment language, conditions and cancellation rights decide what it is worth. Ask for the document, not the description of it.

Which stage has the buyer relationship actually reached?
StageWhat it provesWhat it does not proveEvidence to request
IntroductionA buyer team will look at the categoryInterest from anyone who controls a budgetA dated written request that names a task, not a meeting log
EvaluationThe buyer will spend staff time testing a sampleThat a purchase follows, or at what priceThe evaluation license, the acceptance test and the named decision owner
MandateA program has approval to source a categoryA minimum purchase from this supplierThe commitment language, its conditions and who can cancel
Signed purchaseBoth sides executed terms with scope, price and payment fixedThat deliveries will be accepted or renewedThe executed contract, acceptance criteria and termination rights
Paid deliveryThe buyer accepted a delivery and paid for itThat a second order will comeThe acceptance record, invoice, cash receipt and renewal terms

Validation, pricing leverage or budget dependence: what is the buyer supplying?

Ask which of the three the relationship has produced so far, using evidence rather than the buyer's name. Validation shows up as accepted deliveries against a written test. Pricing leverage shows up as terms that move in the buyer's favor at each renewal. Budget dependence shows up when one program's decision can move the startup's revenue more than any other event.

Reddit's filings show the pattern at public-company scale. The data licensing arrangements it entered in January 2024 carried $203.0 million of aggregate contract value over terms of two to three years, its 2024 S-1 registration statement disclosed, and the same filing said one partner accounted for substantially all of that licensing contract value. The annual report for 2024 attributed substantially all of that value to two partners and warned that renewal was not assured and that any renewal could bring less favorable terms. The contracts validated the asset; the renewal language is where the leverage sits.

Appen shows what dependence looks like when it ends. On 12 February 2024 the company told the ASX that Google, as first disclosed on 22 January, would terminate its global inbound services contract as part of a strategic review, with all projects ceasing by 19 March 2024. Appen put its FY23 revenue from Google at $82.8 million, at a 26 percent gross margin after crowd expenses, and announced $13.5 million of annualized cost savings in response.

Neither company is an early-stage startup, and neither case says a buyer behaved badly. The lesson is the order of events: validation arrived first and was public, while leverage and dependence were disclosed later, as risk factors or results. Diligence has to find them earlier, in records a startup can produce on request.

The Single-Buyer Exposure Map: four measures of AI data buyer concentration risk

Measure the top buyer's share four ways: revenue, pipeline, preparation exposure and specialist capacity. Revenue shows where the company has been; the other three show where it has already committed to go. The highest measure, not revenue, sets the risk band.

Use one denominator per measure and write it down. Revenue means recognized revenue over the trailing twelve months, not bookings or total contract value. Pipeline counts opportunities at the evaluation stage or later, shown both unweighted and stage-weighted, because a weighted figure hides whatever assumptions set the weights. Preparation exposure counts hours and third-party commitments spent on work built to one buyer's specification before that buyer signed. Specialist capacity counts the share of scarce staff time, such as credentialed domain reviewers or delivery engineers, committed to one program.

Add the effective number of buyers to the revenue line: one divided by the sum of each buyer's squared revenue share. A company whose five buyers hold 45, 25, 15, 10 and 5 percent of revenue has about 3.3 effective buyers, a more honest headline than five logos.

Read the four together. When pipeline, preparation or specialist concentration runs well above revenue concentration, dependence is growing even while the revenue mix looks stable. When all four are moderate and the buyer has paid for an accepted delivery, the large buyer is mostly validation.

The Single-Buyer Exposure Map: four measures of AI data buyer concentration risk
MeasureHow to compute itWhat it catches that revenue missesIllustrative reading
Revenue concentrationTop buyer's share of trailing-twelve-month recognized revenue, plus the effective number of buyersNothing on its own: it is the lagging baseline the other three are compared against45 percent; about 3.3 effective buyers
Pipeline concentrationTop buyer's share of open opportunities at evaluation or later, unweighted and stage-weightedWhether next year's revenue depends on the same program70 percent
Preparation exposureShare of preparation hours and third-party commitments spent on one buyer's specification before it signedSunk work with no second buyer if the first one walks away80 percent
Specialist capacityShare of scarce specialist time, such as domain reviewers, allocated to one buyer's programA team that cannot pivot quickly even when a new buyer appears4 of 6 specialists

Who holds the pricing leverage? Map the buyer's alternatives, not its budget

Leverage belongs to whoever has the better alternative at negotiation time, so map the buyer's alternatives from public evidence and leave its budget alone. Large AI buyers rarely disclose data budgets, and a figure inferred from capital spending or press coverage will be wrong in ways nobody can check.

Big infrastructure numbers are the usual trap. In a September 2023 Sequoia Capital essay, AI's $200B Question, David Cahn estimated that a $50 billion run rate of Nvidia GPU revenue implied about $100 billion of data center spending and, assuming end users earn a 50 percent margin, $200 billion of lifetime revenue to pay it back, leaving what he called a $125 billion hole. Sequoia re-ran the arithmetic in June 2024 as AI's $600B Question. Those essays describe the economics of the whole buyer group; they say nothing about what any one lab will spend on a dataset.

What the public record does show is that large buyers have choices. OpenAI's GPT-4o system card, published in August 2024, describes data partnerships that give OpenAI access to non-publicly available material such as pay-walled content, archives and metadata. Each row below is an alternative that strengthens a buyer's hand when it is real, with the question that tests whether it is.

Who holds the pricing leverage? Map the buyer's alternatives, not its budget
Buyer alternativePublic evidence it existsEffect on the startup's pricing powerQuestion to ask
Other data partnersOpenAI's GPT-4o system card (August 2024) describes partnerships for pay-walled content, archives and metadataThe dataset competes with other partners for the same budget lineWhich other sources could fill this task, and how did they score on the buyer's test?
Open contributionsOpenAI's Data Partnerships page offers an open-source archive track alongside private datasetsA buyer may take data contributed openly instead of paying for private supplyIs the dataset clearly outside what is already easily accessible online?
Synthetic dataEpoch AI's 2030 outlook says synthetic data can be generated in large quantitiesLeverage rises if synthetic records pass the buyer's acceptance test for the taskHas the buyer tested a synthetic substitute on this task, and with what result?
Investing in a supplierScale AI said in June 2025 that a Meta investment valued it at over $29 billion, with Meta holding a minority stakeA buyer with its own supply relationship needs any one startup lessDoes any buyer hold equity, exclusivity or board rights in a competitor?
Ending the programAppen said Google ended its global inbound services contract in 2024 as part of a strategic reviewThe buyer can stop entirely; the startup cannot replace the budget quicklyWhat notice and termination rights does the contract give each side?

Stress-test delay, cancellation and reduced scope one at a time

Run each downside case on its own, because each hits different measures and calls for a different protection. A delay mostly costs time and cash, a cancellation removes the buyer's whole share from revenue and pipeline at once, and a reduced-scope purchase keeps the relationship while cutting price and stranding part of the preparation work.

Use the company's own records and state every assumption. The illustrative company in the table carries the exposure readings above: 45 percent of revenue, 70 percent of pipeline, 80 percent of preparation work and 4 of 6 specialists tied to one buyer. In the delay case acceptance slips two quarters; in the reduced-scope case the buyer takes half the planned scope. None of these values describes a real company.

What the company keeps decides the cancellation case. Prepared datasets the startup may license to other buyers on non-exclusive terms keep their value; work built to one buyer's private specification, or licensed exclusively to it, does not. Appen's response to the Google termination, cost cuts announced within weeks of the notice, is the shape of the cancellation case even for a company with other large customers.

Stress-test delay, cancellation and reduced scope one at a time
CaseWhat changesEffect on the four measuresWhat to modelWhat survives
Base case (Illustrative)The buyer accepts on schedule and pays on its termsNo change: 45 percent, 70 percent, 80 percent and 4 of 6Acceptance date, invoice date and cash receiptAll current work and the relationship
DelayAcceptance slips two quarters; the contract survivesRevenue deferred; pipeline slips; preparation carried; specialists idleMonths of operating cost carried before cash, and whether idle specialists can move to other buyersRevenue, once accepted, and the relationship
CancellationThe buyer ends the program after evaluation or terminates the contractRevenue and pipeline lost; preparation stranded; specialists idleRevenue and pipeline without the buyer's share, and the resale value of prepared work under non-exclusive rightsOnly work the company may license to others, and specialists it can redeploy
Reduced scopeThe buyer accepts half the planned scope and renegotiates priceRevenue and pipeline cut; preparation part stranded; specialists underusedThe stranded share of preparation and the price per accepted unit after renegotiationThe accepted slice and the relationship, on weaker terms
A Single-Buyer Exposure Map: bars show an illustrative top buyer's share of revenue, pipeline, preparation work and specialists, beside a grid of how delay, cancellation and reduced scope hit each measure.
The Single-Buyer Exposure Map with three downside cases. Illustrative: values are examples, not measured data.

When growth is concentrated in a single buyer, what counts as traction?

Only growth that survives removing the top buyer counts as broad market traction; the rest is evidence of one procurement program. Build a buyer-cohort revenue bridge that splits each period's change between the top buyer and all other buyers, and report growth excluding the top buyer next to the headline rate.

The illustrative bridge below shows why. Headline revenue rises 60 percent, a number that would stand out in any comparison of startup growth metrics, but all of it comes from one buyer's expansion. Revenue from every other buyer ends the period where it started, and the top buyer's share climbs from 30 percent to about 56 percent.

Public filings show the same split running the other way. In results released on 26 February 2025, Appen reported FY24 group operating revenue of $234.3 million, down 14.2 percent primarily because of the Google termination, and adjusted operating revenue, which leaves out the Google work, of $220.9 million, up 16.0 percent and driven mainly by generative AI projects. In its Global Services division, revenue fell 38.3 percent to $118.1 million, but excluding Google it fell 3.9 percent to $104.7 million. The headline and the ex-buyer rate tell opposite stories, and an investor needs both.

Zero bases and launch spikes do the most damage. A company that moves from one small pilot to one contract can report growth of several hundred percent from a single buyer decision. State the opening value and dates for every rate, and treat a rate computed from fewer than three paying buyers as a launch effect rather than a trend.

When growth is concentrated in a single buyer, what counts as traction?
Bridge lineTop buyerAll other buyersTotal
Opening revenue, indexed (Illustrative)3070100
New buyers0+10+10
Expansion from existing buyers+60+5+65
Contraction0-5-5
Lost buyers0-10-10
Closing revenue, indexed9070160
Growth over the period200 percent0 percent60 percent

Which growth claims survive a single-buyer check?

A growth claim survives when its metric, base, window and buyer mix are stated and reconcile to source records. Use the checklist below on any data startup, then turn it into a cohort dashboard that follows each quarter's new buyers through the quarters after.

At a glance

Which growth claims survive a single-buyer check

The dashboard needs five fields per cohort: buyers acquired, independent budget owners among them, revenue retained from the cohort in each later quarter, repeat purchases, and the top buyer's share of the cohort's revenue. Count budget owners, not logos. Two teams that draw on one approved budget are one buyer, and a lab plus a partner buying on its behalf may be one buyer too.

Then run the correlated-demand scenario. If every buyer in one segment, such as frontier model developers, cut spending in the same quarter, how much of the revenue bridge would remain? Buyers that answer to one budget cycle diversify less than their count suggests.

  • Metric: one named measure per claim, whether recognized revenue, bookings, contract value or cash, never mixed in one chart.
  • Base: the opening value and dates behind every growth rate, with rates from fewer than three paying buyers flagged as launch effects.
  • Window: like periods compared with like, and a trailing-twelve-month view shown beside any single quarter.
  • Bridge: the period's change split into new buyers, expansion, contraction and lost buyers, for the top buyer and for all others.
  • Ex-top-buyer growth: reported next to the headline rate, the way Appen reported its Global Services revenue excluding Google.
  • Budget owners: buyers counted by who approves the spend, with related teams and buying agents merged.
  • Reconciliation: each headline tied to contracts, acceptance records, invoices and cash receipts.
  • Correlation: a stated scenario for a same-quarter pullback by every buyer in the largest segment.

When concentration is the right bet, and when this map misleads

Concentration is not automatically a flaw, and the map has blind spots. Name both before acting on a score; each item below is a way this analysis could point in the wrong direction.

  • A demanding first buyer can be the best evidence a young company has. When that buyer writes the acceptance test later buyers adopt, early concentration buys a standard, and screening on revenue share alone would reject the companies where validation is real.
  • The buyer market may be concentrated by nature. Reddit, a large public platform, still attributed substantially all licensing contract value to two partners in its annual report for 2024. A startup cannot diversify beyond the number of buyers with budgets for its category, so judge it against that number.
  • Budgets can fall together. Sequoia's 2023 and 2024 essays questioned whether AI infrastructure spending would earn back its cost; if buyers cut back in the same cycle, several AI lab customers behave like one.
  • Substitutes can improve. Epoch AI's 2030 outlook expects synthetic data to be generated in large quantities; if synthetic records pass a buyer's acceptance test for a task, buyer leverage rises against every supplier of that task.
  • Ownership ties change incentives. Scale AI said in June 2025 that a Meta investment valued it at over $29 billion, with Meta holding a minority stake; a buyer's equity in one supplier can change how buyers treat the rest of the market.
  • The forward measures rely on management's own records. Pipeline stages, preparation hours and staff allocation are classified by the company, so ask for definitions, timestamps and an audit trail before trusting the map.

What independent buyer evidence would lower concentration risk at SourceX?

The same four measures, with buyers counted by independent budget owner, would show it, and SourceX has not published them. This section applies the framework to SourceX as one example; it reports no SourceX buyer, revenue or pipeline figures.

SourceX helps businesses license selected records to AI developers as datasets, and preparation sits inside its own process alongside sourcing and qualification, rights review, contracting, delivery and payment. Nothing is prepared for stock: sourcing from U.S. businesses starts with a buyer's request, which sets out scope, volume, language, format, timing and licensing needs. That design limits speculative inventory, one kind of preparation exposure, but it also means preparation starts from a particular buyer's request, so preparation exposure by buyer, and whether a prepared dataset may be licensed to others, are the first things to examine.

Its company and commercial metrics are shared with investors directly rather than published. SourceX's Rule 506(c) raise is open to accredited investors, and its materials are shared on request through the investors page; a useful request names the evidence below and the definitions to apply.

  • The number of independent budget owners that issued dated written requests in each of the last four quarters, and the largest one's share.
  • A revenue bridge split between the top buyer and all others, with growth excluding the top buyer beside the headline rate.
  • Preparation hours and third-party commitments started before a signed purchase, grouped by buyer.
  • Specialist capacity by buyer program, with the share held by the largest program.
  • For each large buyer: its stage on the ladder, its termination and renewal rights, and whether prepared datasets may be licensed to others.
  • The delay, cancellation and reduced-scope cases run on the company's own records, with every assumption stated.

How the exposure map was built, and what it leaves out

The relationship ladder, the Single-Buyer Exposure Map, the leverage map, the three downside cases, the revenue bridge and the growth-claim checklist are editorial frameworks written for this article on 2026-10-26. Public evidence comes from Reddit's 2024 S-1 and its annual report for 2024, Appen's ASX announcements of 12 February 2024 and 26 February 2025, OpenAI's Data Partnerships page and GPT-4o system card, Sequoia Capital's 2023 and 2024 essays, Scale AI's June 2025 announcement and Epoch AI's 2030 outlook. Company figures are quoted in each company's own definitions, which are not comparable with a startup's bookings.

All exposure readings, case assumptions and bridge values are Illustrative and describe no real company or deal. Not measured: buyer concentration at any private AI data company, SourceX's buyer mix or pipeline, any buyer's data budget, the share of evaluations that become purchases, and how concentration affects price. Measuring any of them would take company records that nobody has published.

Frequently asked questions

How much buyer concentration is too much for an AI data startup?

There is no universal threshold. Compute the top buyer's share of revenue, pipeline, preparation work and specialist capacity, and let the highest of the four set the risk band. Then compare the result with the number of buyers that have budgets for the category, since a startup cannot diversify beyond them.

Is a frontier AI lab customer a good sign for a startup?

It is a good sign only up to the stage the relationship has reached. An accepted, paid delivery is real validation; an introduction or an evaluation is interest. The same buyer also gains pricing leverage as the startup comes to depend on it, so ask for the contract, the acceptance record and the renewal terms.

How do you calculate customer concentration for a data company?

Divide the largest buyer's trailing-twelve-month recognized revenue by total revenue, then compute the effective number of buyers as one divided by the sum of squared revenue shares. Repeat the share calculation for qualified pipeline, preparation hours and specialist staff time, because revenue alone lags the commitments a company has already made.

What happens when a data supplier loses its largest AI buyer?

Appen is a public example. It told the ASX in February 2024 that Google would end its global inbound services contract, with projects ceasing by 19 March 2024, after Google contributed $82.8 million of FY23 revenue. Appen announced cost savings within weeks, and its FY24 group operating revenue fell 14.2 percent, primarily because of the termination.

Do two AI lab customers diversify buyer risk?

Only partly. Two buyers with independent budgets and different tasks reduce dependence on one decision, but buyers exposed to the same spending cycle can cut back together. Count independent budget owners, and run a scenario in which every buyer in the segment pulls back in the same quarter.

Sources

  • OpenAI's Data Partnerships page seeks large-scale datasets not already easily accessible online, through an open-source archive or private datasets for proprietary models, and says it is not seeking sensitive personal or third-party information. Source
  • OpenAI's GPT-4o system card (August 2024) describes data partnerships that give access to non-publicly available data such as pay-walled content, archives and metadata. Source
  • Sequoia Capital's September 2023 essay AI's $200B Question estimated that a $50 billion Nvidia GPU revenue run rate implied about $100 billion of data center spending and $200 billion of lifetime revenue needed, a $125 billion hole. Source
  • Sequoia Capital re-ran the GPU revenue-gap analysis in June 2024 as AI's $600B Question. Source
  • Reddit's 2024 S-1 disclosed January 2024 data licensing arrangements with an aggregate contract value of $203.0 million and two-to-three-year terms, with substantially all licensing contract value from one partner. 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
  • Appen's 12 February 2024 ASX announcement says Google would terminate its global inbound services contract as part of a strategic review, ceasing all projects by 19 March 2024; FY23 Google revenue was $82.8 million at a 26% gross margin; $13.5 million of annualized cost savings. Source
  • Appen's FY24 results (26 February 2025): group operating revenue $234.3 million, down 14.2% primarily due to the Google termination; adjusted operating revenue $220.9 million, up 16.0%; Global Services down 38.3% to $118.1 million, or down 3.9% to $104.7 million excluding Google. Source
  • Scale AI announced in June 2025 a new investment from Meta that values Scale at over $29 billion, with Meta holding a minority of Scale's equity. Source
  • Epoch AI's 2030 outlook says synthetic data can be generated in large quantities and judges a data bottleneck surmountable. 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