Deal economics
Enterprise AI versus frontier lab data buyers: a re-review ratio test
By SourceX Editorial · Reviewed by Noah Loul ·
10 sections
Short answer
Frontier labs and enterprise AI teams buy data on different economics. Labs sign fewer, larger contracts, so review cost is spread thin but revenue concentrates: Innodata, which serves AI labs and Big Tech, drew about 58% of its 2025 revenue from one customer. Enterprise teams buy smaller, workflow-specific sets with a heavy review per order. The segment that pays is the one whose repeat orders reuse the first review, so ask for the re-review ratio.
License terms
Enterprise AI
Key takeaways
- Large data licensing contracts run for years: Reddit's January 2024 deals carried $203.0 million of aggregate contract value over two-to-three-year terms.
- Revenue from a few big buyers concentrates fast; Reddit attributes substantially all licensing contract value to two partners.
- Enterprise orders carry a fixed review core of security, privacy, rights and onboarding work that does not shrink with contract size.
- The re-review ratio, review effort on a repeat order divided by the first, shows whether a buyer relationship gets cheaper to serve.
- Neither segment diversifies by itself: a lab can end a program within weeks, and enterprise pilots can stop at proof of concept.
Which buyer segment should an AI data company build around?
Neither segment wins by default. Frontier labs and enterprise AI teams reward different cost structures, and the right mix for a data company depends on four lines an investor can measure contract by contract: contract size, review effort, customization and repeatability.
In this article a frontier lab is an organization training the most capable general-purpose models, buying records for pretraining, post-training or evaluation across many tasks at once. An enterprise AI team is a company outside the labs, or a vertical AI vendor serving one industry, that builds, tunes or tests AI for named workflows. Many data companies sell to both, and that is where the comparison matters most.
Useful evidence is contract-level: signed value and term, hours of buyer review before signature, the share of preparation work only one buyer can use, and the effort needed to win a second order from the same buyer. Logos, announced partnerships and pipeline totals cannot settle it, because each segment can produce impressive versions of all three.
What filings reveal about selling to frontier labs
Public filings show large data licensing contracts running for years and revenue from a few big buyers concentrating quickly. Four disclosures make the pattern concrete; none of the passages cited here identifies a counterparty as a frontier lab, so treat them as the closest public proxies.
Reddit's 2024 S-1 disclosed data licensing arrangements signed in January 2024 that together carried $203.0 million in contract value over terms of two to three years, of which it expected to recognize at least $66.4 million in 2024. Its Q1 2026 10-Q states that "substantially all of the contract value associated with our licensing revenue is derived from two of our partners."
Innodata's FY2025 10-K describes a data engineering and AI systems services company whose customers include leading AI labs and Big Tech companies building frontier generative AI models, five of the so-called Magnificent Seven among them, alongside leading enterprises. The same filing says one customer generated about 58% of total revenue in 2025 and about 48% in 2024. Its 10-Q for the six months ended June 30, 2026 puts that customer at about 46% and a second customer at about 26%.
Appen's ASX update of 12 February 2024 said Google would end its contract with Appen after a strategic review, with all projects ceasing by 19 March 2024. Appen's FY2023 revenue from Google had been $82.8 million, earned at a 26% gross margin.
On the demand side, OpenAI's data partnerships page asks for large-scale datasets that reflect human society and are not already easily accessible online, and it rules out sensitive or personal information and information that belongs to a third party. That is a request for breadth and scale, screened on rights before anything else.
Taken together, lab-scale contracts can be large enough to absorb heavy review, but they are few, their value is recognized over years rather than at signing, and they can end on a decision made inside the buyer.
Why the enterprise side of the ledger is mostly unpublished
Enterprise AI teams buy in smaller, workflow-specific pieces, and almost none of that buying is disclosed, so the enterprise column rests on indirect evidence and on what a company can show from its own contracts.
Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025, citing poor data quality, inadequate risk controls, escalating costs or unclear business value. A project that stops at proof of concept also stops buying data, so enterprise orders carry pilot attrition on top of ordinary procurement risk.
Order size points the same way. Anthropic's guide to agent evaluations says teams often hold off because they think they need hundreds of tasks, when 20 to 50 simple tasks drawn from real failures is a great start. Evaluation sets that small, built from one company's own failures, suggest enterprise orders that are narrow and quick to scope but hard to sell twice.
Lab-style demand is also reaching enterprise records through intermediaries. In October 2026, micro1 published a plan to spend $1 billion within 12 months on acquiring and licensing operational data from enterprises as raw material for reinforcement learning environments. That remains a stated plan rather than money spent, and it says nothing about whose data will be bought. It does show the two segments meeting on the supply side, where enterprises are sellers as well as buyers.
The segment ledger: contract size, review effort, customization, repeatability
The segment ledger compares frontier lab buyers and enterprise AI teams on six lines and names, for each line, the record an investor should ask a data company to produce. The two segment columns are editorial readings of the public evidence above, not measurements; the last column is what turns them into facts about one company.
Read the rows as trade-offs rather than scores. The lab column is strong on contract size and weak on concentration; the enterprise column runs the other way, and whether its repeat orders get cheaper decides whether that trade is worth making.
| Line | Frontier lab buyers | Enterprise AI teams | Record to request |
|---|---|---|---|
| Contract size | Large and multi-year, recognized over the term (closest proxy, Reddit: $203.0 million across January 2024 deals) | Smaller, scoped to one workflow or evaluation set, and rarely disclosed | Signed value, term and amount recognized per contract, by segment |
| Review effort per order | Heavy, but spread across a large contract; rights and scope screens come first | A similar fixed review core (security, privacy, rights, onboarding) on a much smaller contract | Buyer review hours and elapsed days before signature, per contract |
| Customization | Specification written by the lab's research team; output often fits only that program | Built around one company's systems and failures; hard to resell | Share of preparation hours that produced records usable for a second buyer |
| Repeatability | Next order follows research priorities, which can change at once (Appen, 2024) | Next order follows the workflow's rollout; pilots can stop at proof of concept (Gartner, 2024) | Re-review ratio and days to second order, keeping buyers who never reordered |
| Concentration | Few buyers with large shares (Innodata: one customer about 58% of 2025 revenue) | Many small buyers, correlated if they share a model vendor or budget cycle | Revenue share of the top buyer and the top budget owner over four quarters |
| Commitment form | Multi-year contract value, often with phased scope | Pilot, evaluation license or single order before any master agreement | Share of segment revenue under signed minimums versus open-ended orders |
Review hours per unit of contract value: the enterprise penalty and its cure
Enterprise orders carry a review load that does not shrink with contract size, so review hours per unit of contract value run far higher than on lab contracts until repeat orders reuse the first review.
A vendor review has a fixed core: a security questionnaire, privacy and processing terms where personal data is involved, a rights and provenance file, legal review of the license and onboarding into the buyer's vendor system. A small enterprise order triggers most of that core. A large lab contract triggers it once and spreads it over far more value.
Illustrative: index a lab contract at 100 units of value with 300 hours of combined buyer and supplier review, and an enterprise order at 8 units with 120 hours. Review intensity is 3 hours per unit for the lab and 15 for the enterprise, five times heavier. If the enterprise's next order under the same master agreement needs a fifth of the first review, 24 hours for another 8 units, intensity on that order falls to 3 hours per unit. If the lab's next program needs a fresh specification and half the original review, 150 hours for another 100 units, its intensity falls to 1.5. None of these figures is measured; they show the mechanism, not the market.
The investor test follows directly: enterprise revenue earns back its review cost only through repeat orders that skip most of the review. Without them, every enterprise order is a new sale at several times the review intensity of lab work.
The re-review ratio: does the second order cost less to win?
The re-review ratio is the review effort a buyer spends on a repeat order divided by the effort it spent on its first order. A ratio that falls from one order to the next is the clearest sign that a segment compounds.
Measure it per buyer, in review hours where a company logs them or in elapsed days from request to signature where it does not, and report the median by segment with the number of buyers behind it and the date range. Buyers who never placed a second order stay in the denominator, coded as no repeat. Dropping them makes either segment look healthier than it is.
Four things lower the ratio: a master agreement that covers new orders, approved-vendor status that survives the first purchase, a rights file that already covers the next scope and an unchanged delivery format. Three things reset it: a new business unit with its own budget owner, a new intended use that needs fresh rights review and a lab research program that rewrites the specification.
A ratio near 1 means every order is a first sale. A ratio that keeps falling across the second and third orders means the company is building a relationship asset, whichever segment the buyer belongs to.
Is one prestigious lab contract safer than ten enterprise accounts?
Not by default. A large lab contract validates the data and can carry its review cost, but it can also create dependence without any duty to keep buying, which is the strongest case against reading lab revenue as traction.
Appen shows the failure mode: a relationship worth $82.8 million of FY2023 revenue was set to end within weeks after the buyer's strategic review. Innodata shows how quickly shares move inside a concentrated base, with one customer at about 48% of revenue in 2024, about 58% in 2025 and about 46% in the first half of 2026, when a second customer reached about 26%. In both cases the exposure sat in decisions made inside a few buyers.
Ten enterprise accounts are not automatically diversified either. If they depend on the same model vendor, the same budget cycle or the same proof-of-concept gate, one shift can stall all ten, which is what Gartner's abandonment forecast warns about. Diversification counts independent budget owners and independent triggers, not logos.
The decision rule: credit lab revenue at the commitment its contract actually carries, credit enterprise revenue at the re-review ratio it actually shows, and compare the two on the same denominator.
What could make the segment comparison wrong
The comparison could mislead if the segments blur, if the public comparables are weaker than they look or if substitutes cut demand in one segment faster than in the other.
At a glance
What could make the segment comparison wrong
- Segments blur: labs reach enterprise records through intermediaries, as micro1's stated $1 billion plan shows, and vertical AI vendors can pass enterprise work through to lab programs, so a buyer's label can hide where demand starts.
- Comparables are imperfect: Reddit licenses its own platform content and Innodata sells data engineering services, so neither shows the economics of licensing third-party business records.
- Lab terms could standardize: if labs adopt uniform data terms and onboarding, their review load could fall below the enterprise core and reverse the review line.
- Enterprise attrition could be lower than forecast: Gartner's 30% figure is a prediction about generative AI projects, not a measured rate for data purchases.
- Substitutes: synthetic data, or evaluation sets built from a buyer's own logs, could replace purchased records in either segment, and faster where tasks are generic.
- Measurement limits: review hours are often not logged, and elapsed days mix the buyer's queue with the supplier's own delays.
Segment questions to ask any data company, SourceX included
Ask for the ledger lines by segment, with denominators, before crediting either kind of revenue. The same questions apply to SourceX.
SourceX is one application of this question. Its published role is the whole chain for a business licensing operational data to AI developers: finding and qualifying the source, reviewing rights, contracting, preparing the data, delivering it and handling payment. It keeps no catalog; each dataset is sourced from U.S. businesses after a buyer asks, and the request spells out scope, volume, language, format, timing and licensing needs, so preparation answers a named need rather than a guess.
SourceX has not published buyer segment, contract or review metrics, and it keeps its company and commercial metrics for investors, sharing them directly. With a Rule 506(c) raise open to accredited investors, SourceX provides investor materials on request through the investors page, and any of the questions below is a fair place to start.
Checklist · 6 items
- What share of revenue and of signed contract value came from lab-scale buyers versus enterprise teams over the last four quarters?
- What is the median re-review ratio by segment, and how many buyers sit behind each median, including those who never reordered?
- How many review hours or elapsed days did first orders take before signature, by segment?
- What share of preparation hours produced records usable for a second buyer?
- Which buyer, and which budget owner inside it, accounts for the largest revenue share, and what would a program change there do?
- How much segment revenue sits under signed minimums rather than open-ended orders?
0 of 6 ticked. Tick each item as you go.
Methods note: how the segment ledger was built
The segment ledger and the re-review ratio are editorial frameworks built for this article on 2026-10-26. Public anchors: Reddit's 2024 S-1 and Q1 2026 10-Q, Innodata's FY2025 10-K and its 10-Q for the period ended June 30, 2026, Appen's ASX update of 12 February 2024, OpenAI's data partnerships page, Gartner's July 2024 press release, Anthropic's agent evaluation guide and micro1's October 2026 announcement. The review-hour example is illustrative and describes no real buyer or contract.
Not measured: contract sizes, review hours, customization shares or re-review ratios at any company, SourceX included, and enterprise data spending, which none of the sources reviewed here breaks out by segment. The segment columns are readings of the cited evidence, not survey results.
Frequently asked questions
Do frontier labs or enterprise AI teams sign larger data contracts?
Public evidence is thin on both sides. The largest disclosed figures are Reddit's: January 2024 data licensing arrangements with $203.0 million of aggregate contract value over two-to-three-year terms. Enterprise data purchases are rarely disclosed at all, so a company's own contract records are the only reliable comparison.
What is the re-review ratio for a data buyer?
It is the review effort a buyer spends on a repeat order divided by the effort it spent on its first order, measured in hours or elapsed days. Buyers who never reordered stay in the denominator. A ratio that falls from order to order shows a relationship that gets cheaper to serve.
Why does a small enterprise data order need so much review?
Most vendor review is fixed per supplier rather than proportional to price: security questions, privacy and processing terms, rights evidence, legal review and onboarding. A small order triggers nearly all of it, so review effort per unit of contract value stays high until repeat orders reuse the first approval.
How concentrated are public AI data companies?
Very, in the filings reviewed here. Innodata reported one customer at about 58% of its 2025 revenue, and Reddit says two partners account for substantially all of the contract value behind its licensing revenue. These are reported figures, not forecasts.
How should an investor compare lab revenue with enterprise revenue?
Credit lab revenue at the commitment its contract actually carries and enterprise revenue at the re-review ratio it actually shows. Then compare both on the same denominator, counting independent budget owners rather than logos.
Sources
- Reddit's 2024 S-1 reported January 2024 data licensing arrangements with an aggregate contract value of $203.0 million and two-to-three-year terms, with at least $66.4 million expected to be recognized in 2024. Source
- Reddit's Q1 2026 10-Q states that substantially all of the contract value associated with its licensing revenue is derived from two partners. Source
- Innodata's FY2025 10-K lists AI labs, Big Tech companies building frontier models and leading enterprises among its customers and reports one customer at about 58% of 2025 revenue and 48% of 2024 revenue. Source
- Innodata's 10-Q for the six months ended June 30, 2026 reports one customer at about 46% of revenue and another at about 26%. Source
- Appen told the ASX on 12 February 2024 that Google would end its contract after a strategic review, with projects ceasing by 19 March 2024; FY2023 revenue from Google was $82.8 million at a 26% gross margin. Source
- OpenAI's data partnerships page seeks large-scale datasets not already easily accessible online and excludes sensitive or personal information and third-party information. Source
- Gartner predicted in July 2024 that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Source
- Anthropic's agent evaluation guide says 20 to 50 simple tasks drawn from real failures is a great start for an evaluation set. Source
- micro1 announced in October 2026 a plan to spend $1 billion over 12 months acquiring and licensing enterprise operational data for reinforcement learning environments. Source
Request SourceX investor materials
SourceX shares company and commercial metrics with investors directly. Any offering is made only to verified accredited investors.