The company-data and RL-environment landscape to 2028

Where it is heading, and how to sell into it.

Two years ago a company's operating records had no book value and no bid. Today 43 published programmes will quote one — and what the buyer actually wants is already moving past the records, toward environments an agent can practise in and verifiers that decide whether it did the work.

This is a forecast rather than a market summary: seven structural shifts with the evidence under each, the loop that drives them, five scenarios with odds at December 2027 and December 2028, what commoditises and what keeps its scarcity, and eighteen dated calls a named public source will settle. Every reading carries a source grade, and every judgment carries a confidence.

Published 4 Nov 2026Data as of 4 Nov 2026Next update 1 Dec 2026

At a glance

By the end of 2028 the likeliest landscape is an environment economy, not a market in files

The likeliest thing to happen in between is that an unpackaged archive gets cheaper while a connected, verified one does not.

We put the environment economy at 33% by December 2028 against 25% a year earlier, and today's selective market down from 37% to 12%. The second panel is the structure a new deal is written in. Each column sums to 100.

The landscape

Which shape the company-data and RL-environment landscape is in at each date

Scenario	Dec 2027	Dec 2028
Selective marketToday's shape continues: wide dispersion, one-off licences, no price you can look up	37%	12%
Environment economyMaintained environments and trailing structures become the normal deal	25%	33%
Buy-side consolidationA few parties become the gates every corpus has to pass through	15%	13%
RepricingAppetite narrows, open tooling bites, and published ranges come down	11%	24%
Rights chillA court, a regulator or a re-identification event stops employee-bearing records	12%	18%

How a new company-record deal is structured

The dominant structure for deals signed in each year. Today 11 of the 43 listed programmes publish a trailing term

Structure	Dec 2027	Dec 2028
One-off paymentA single sum against a published range, the structure most listed programmes still use	51%	22%
Usage or milestoneStaged against acceptance, volume or model milestones	14%	14%
Revenue share or royaltyThe holder is paid on each onward licence rather than once	21%	32%
Environment partnershipAn environment is built and maintained on the holder's own work, and licensed as it runs	14%	32%

Odds are probabilities from a simulation of chained steps, each with a low, central and high value, run over 200,000 seeded paths. The triggers and early markers that decide each one are in Scenarios, and the method is in Sources and method.

Bottom line

Eight forward-looking judgments, each with the odds and the horizon it is made on.

01
The unit of trade moves from the file to the environment, and the holder's best structure moves with it.An environment partnership is the dominant structure for a new company-record deal on 14% of paths at December 2027 and 32% at December 2028; any trailing structure — revenue share, royalty or environment — is dominant on 35% and then 64%. That is the single most consequential thing in this forecast for a holder, because a trailing structure keeps it in the income and a one-off payment does not. Eleven of the 43 programmes on the public register already publish a trailing term, against the one-off payment that most of the rest still use, and a twelve-partner environment network announced on 23 October 2026 builds environments on a customer's own tasks, data and feedback rather than selling it a file.Confidence: high. The direction has three independent primary readings behind it — a public register's own terms, a platform's own announcement, and a licensing-deal study showing flat annual fees falling from 58.3% of agreements in 2023 to 34.8% in 2026 while revenue share rose from 4.2% to 13.4%. The speed is the uncertain part, not the direction.
02
Labs will build more and still buy — in-housing caps the commodity trade, not the specialist one.A survey of environment builders reports substantially more in-housing, citing lab job postings and a product company training on its own users' work; open tooling now compiles verifiable environments straight out of a repository. Against that, a buyer's own page commits $1 billion over twelve months to acquiring and licensing enterprise operational data, and an open-role index read on 4 November 2026 shows 11 roles sourcing, licensing or acquiring data at 8 companies and 16 more building environments. We put it at 10% (4–17) that a named frontier lab states publicly by 31 December 2027 that it has stopped buying third-party company archives. The commodity end gets squeezed long before the specialist end does.Confidence: moderate. Both sides are documented, but the in-housing reading comes from one survey of unnamed builders and no lab has published a position either way. The call is deliberately held below the repricing scenario it would trigger.
03
Verification, not collection, is where the margin lands by 2028.In a published training run of a model trained as an agent organisation across 147 environments and more than two million tasks, the reviewing roles the organisation created for itself doubled from 13% to 26% of specialist hires — and in one simulation the share of runs in which an agent alerted the evaluations team fell from 76.7% to 20% and then zero. Grading is getting cheap: six decision models benchmarked on the same 98 production traces ran from 0.82 to 3.72 cents per 100 eval runs. Cheap graders and self-scoring agents both push value to the same place: held-out verifiers and labelled traces a buyer did not write. 30% (18–42) that by 31 December 2028 half or more of newly listed programmes taking company records ask for a verifier, eval or environment rather than files.Confidence: moderate. The two readings are specific and reproducible, but they are two studies rather than a trend, and neither was designed to measure what a buyer will pay for.
04
The introduction layer peaks, then thins — and the thinning is a price event for unpackaged sellers only.Ten of the 43 published programmes pay an introduction bounty. The largest published flat figure on the register read $85,000 on 4 November 2026, against “up to $100,000 per deal” across the same programmes in a survey conducted between 11 and 17 October; one programme pays an uncapped 7.5% of deal value above a threshold instead of a flat sum. The register's vendor count went from 159 on 14 October 2026 to 187 three weeks later, and the same register removes companies that stop selling. 63% (48–79) that one of the five largest published bounties is cut or withdrawn by 30 June 2027, and 50% (34–67) that the count of listed vendors falls year-on-year for the first time by 31 December 2028.Confidence: moderate. The bounty reading is a genuine three-week movement in two independent counts of the same programmes, but the counts are small and one of them is maintained by a party that discloses it may sell to the companies it lists.
05
Published ceilings stay ahead of confirmed payments through 2027, and that gap is the market's main unpriced risk.Of 24 programmes that take a company's own records, nine publish a ceiling — median $1,000,000, highest $5,000,000 — and none of the nine has an independently reported payment of that size behind it. One of the 24 has a payment confirmed by a seller. Fourteen publish a floor, median $100,000. The incentive runs the wrong way: the layer paid on completion is the layer quoting the ceiling, and the survey that counted these also recorded one programme raising its stated ceiling from $300,000 to $1,000,000 and another dropping a published figure, inside a single month. 41% (26–56) that a seller publicly confirms a payment of $1,000,000 or more by 31 December 2027.Confidence: high. This is the best-evidenced line on the page: a single study that classified each programme's published terms from its own page with reading dates, and whose finding is an absence rather than an assertion.
06
Financing arrives before price discovery does, and it changes who can buy.Lab receivables are now underwritable: one lender advances against invoices on a $10m-plus warehouse facility from an institutional credit fund with $65m-plus in signed letters of intent, closing a 30–90 day payment gap in 2–4 days. The structure with the longest reach is already visible in an adjacent sector, where a buyer paid $80M in cash plus a $40M equity investment and recurring research funding for compute, and kept exclusive rights to what the collaboration produces. Price transparency is the slower half, and it can move in one step: a compute provider now publishes live public spot prices per chip and region, with 30 days of history, behind an API that needs no account and no authentication. 38% (24–53) that a facility of $100M or more specifically for AI data vendors is announced by 31 December 2028; 31% (18–44) that a public rate feed or index for training data exists by the same date.Confidence: moderate. The lending facility and the compute rate feed are both primary and dated. The equity-for-rights structure is corroborated in pharmaceutical research rather than in company records, so it is read as a template available to this market, not as something already happening in it.
07
What commoditises is anything a crawler, a repository or a simulator can make. What stays scarce is connected, outcome-linked, rights-clean operating history.A commercial index of more than two billion people assembled from the public web sells agent access from $0.15 a credit; a free index of 30M-plus US federal, state and local legal documents searches without spending any; open tooling has already published 1,930 verifiable coding tasks compiled out of repositories, each with a private verifier that must score 1 for an oracle solution and 0 for a no-op before it publishes; and a physics laboratory and a model developer signed a statement of intent on 15 October 2026 to train a collider-physics foundation model on billions of simulated proton-proton collisions. None of that touches what a buyer cannot assemble retroactively: whether the same people and the same events show up across email, chat, the CRM and the ticketing tool, continuously, for years, with an outcome attached to each.Confidence: high on the commoditising side, where every item is a published artefact anyone can check, and moderate on the scarcity side, which is an inference from what the published programmes say they price on rather than from observed prices.
08
The tail that matters is still rights, and it is fatter at two years than at one.A rights chill — a court, a regulator or a credible re-identification claim making employee-bearing records unsellable in practice — runs at 12% to December 2027 and 18% to December 2028. In the one court-supervised sale on the public record a flight attendants' union negotiated medical, leave and union records out of the set and more than 100 members of Congress sought wider exclusions and limits on use. 46% (30–62) that a US state statute or federal bill specifically addressing the licensing of employee records for AI training is introduced by 31 December 2027. This is the one scenario in which doing the rights work first is not an optimisation but the whole deal.Confidence: moderate, and cut from the previous run. The union objection in that sale was resolved by carving records out rather than by stopping the sale, which is evidence for narrower assets rather than for no asset — so the chill is read as a tail, not as a trend.

The baseline

Where the landscape stands on 4 November 2026. This band is collection, not analysis: every row is a reading from a named public source, graded A, B or C, with no inference attached. The inferences are everywhere else on the page.

Two years ago a company's operating records had no book value and no bid. The readings below are the position the rest of this report forecasts from: a countable buying field made mostly of intermediaries, posted prices far above confirmed ones, a demand chain whose money sits downstream in environments and tasks, and a rights gate nobody has cleared at scale.

187

Vendors on the public register

Selling data, RL environments or evals on 4 Nov 2026, from 159 three weeks earlier.

Published buying programmes

From 32 companies; 16 pay a company directly and 11 publish a trailing term.

0 of 9

Published ceilings met in public

Nine programmes publish a ceiling. None has an independently reported payment of that size behind it.

$10.0M

Selected at a court-supervised auction

Selected on 14 Aug 2026 for a de-identified operating archive; a $12.5M notice of intent was filed after.

Every reading, with its source

Publicly observable

Grades describe the source, not the conclusion. A — record: a filing, a court docket, a published paper, a public dataset or repository, or a named report of one. B — party: an organisation's own published page, terms or announcement about itself. C — crowd: aggregators, registers compiled by an interested maintainer, and surveys of unnamed practitioners.

Confidence describes Frontier Desk's inference from those readings: high where two or more independent sources agree and none contradicts, moderate where the direction is corroborated but the magnitude is not, low where a single source or an analogy carries it.

Reading	Source	Figure	Grade
Vendors selling data, RL environments or evals	Public register, updated 4 Nov 2026	187, from 159 on 14 Oct 2026	C · crowd
Published buying programmes	Public register, updated 4 Nov 2026	43, from 32 companies	C · crowd
Of those, programmes that pay a company directly	Public register, updated 4 Nov 2026	16	C · crowd
Of those, licensing agents or marketplaces	Public register, updated 4 Nov 2026	15	C · crowd
Of those, introduction bounties	Public register and the programmes' own referral pages, 4 Nov 2026	10; largest published flat figure $85,000, one paying an uncapped 7.5% above a threshold	B · party
Of those, publishing a trailing term	Public register, updated 4 Nov 2026	11 — revenue share, royalty or payment on each onward licence	C · crowd
Of those, invite-only rather than open	Public register, updated 4 Nov 2026	2	C · crowd
Programmes that take a company's own records	Working paper, 18 Oct 2026	24 of 37 organisations surveyed, 17 buying for their own account	A · record
Of those, publishing a floor	Working paper, 18 Oct 2026	14; median $100,000, lowest $10,000	A · record
Of those, publishing a ceiling	Working paper, 18 Oct 2026	9; median $1,000,000, highest $5,000,000	A · record
Ceilings with an independently reported payment behind them	Working paper, 18 Oct 2026	0 of 9	A · record
Programmes with a payment confirmed by a seller	Working paper, 18 Oct 2026	1 of 24	A · record
Of 16 organisations publishing contract clauses, those publishing exclusivity	Working paper, 18 Oct 2026	4 — against 2 on sublicensing and 2 on deletion	A · record
What an exclusive licence costs a buyer over a shared one	Environment-economics survey, 12 Jan 2026	Roughly 4 to 5 times	C · crowd
Price of one reinforcement-learning task	Environment-economics survey, 12 Jan 2026	$200 to $2,000; $20K rare but possible	C · crowd
Cost of a complex product replica	Environment-economics survey, 12 Jan 2026	About $300K; a website replica about $20K	C · crowd
Compute spent per task during RL training	Environment-economics survey, 12 Jan 2026	About $2,400, as estimated	C · crowd
One lab's discussed annual spend on environments	The Information via TechCrunch	More than $1B, as reported	A · record
One buyer's committed spend on enterprise operational data	The buyer's own announcement	$1B over twelve months, announced 9 Oct 2026	B · party
Open roles sourcing, licensing or acquiring data	Open-role index, read 4 Nov 2026 across 62 job boards	11 at 8 companies; 16 more building RL environments	B · party
Shift in licence structure across all AI training data	Licensing-deal study of 312 agreements, 6 Oct 2026	Flat annual fees 58.3% of deals in 2023 → 34.8% in 2026; revenue share 4.2% → 13.4%	A · record
Median contract term across those 312 agreements	Licensing-deal study, 6 Oct 2026	2.0 years; 1.0 to 3.5 by kind of licensor	A · record
Spread between two public estimators for the same company	Programme estimators via the comparison guide	About 21 times (FD estimate: $800,000 ÷ $37,500 = 21.3, for a US company of about 100 people and 10 years)	B · party

One public price exists, and it came from a court. In the only court-supervised sale of a company archive on the public record, Spirit Aviation Holdings' de-identified operating data was auctioned on 14 August 2026; a $10.0M bid was selected with a $7.5M alternate, and a competing $12.5M notice of intent was later filed.

Consumer personal data was excluded, de-identification was ordered at the buyer's cost, a flight attendants' union negotiated medical, leave and union records out of the set, and more than 100 members of Congress have since sought wider exclusions and limits on use. Five adjournments later the matter reached a hearing scheduled for 6 November 2026, with no sale order entered through Dkt. 1775 — FD estimate: 84 days from auction to hearing, by the calendar.

The lesson the forecast takes is not the number. It is that competition moved the de-identification, retention and exclusion terms as well as the price, and that it took a process to do it.

Hold those four facts together and the shape of the next two years is already implied. A buying field that is mostly intermediaries will thin. Posted prices that nobody has been paid will either be met or quietly cut. A demand chain whose money is downstream will reach back up it. And a rights gate nobody has cleared at scale will either widen into routine carve-outs or close.

Structural shifts underway

Seven shifts are already running. Each separates what the sources say from what we infer from them: the evidence first with its grade, then the reading with its confidence, then how fast, then the specific thing that would reverse it. These are the inputs to the scenarios.

From static datasets to environments, trajectories and verifiersRunning, and accelerating

Speed: 12–24 months to become the normal unit of trade

What the sources say. The AgentEnv framework, open-sourced by Scale AI under Apache-2.0, builds containerised environments into versioned images, deploys them behind a gateway, points an agent at them and scores what the agent did, all behind one published protocol. A separate MIT-licensed tool records a browser's requests and an app's responses and builds a local environment from them, so an agent can practise a task without the original app. A third project compiles environments out of a repository's merged pull requests, commit history and security advisories — 23 generators, 21 published datasets and 1,930 tasks at v0.9.5 on 4 November 2026 — each with a private verifier that must score 1 for an oracle solution and 0 for a no-op before the task publishes. On the demand side, an inference platform announced a twelve-partner environment network on 23 October 2026 in which the partners build environments on a customer's own data, tasks and feedback while the platform handles training and production inference. One buyer publishes licensing of qualifying agent trajectories at $1–2M a year.

Evidence grade: A — two public repositories with version numbers and licences, a published product page, and a platform's own launch announcement, all read 4 November 2026.

What we read into it. Read together, these say that building an environment has stopped being bespoke engineering and become tooling — which moves the scarce input from the ability to build one to the thing worth building one around. Where a repository already encodes the grade, the environment generates itself and nobody needs to buy it. Where the grade lives in a business's outcomes, it does not, and the archive behind it is the input. The platform layer rather than the lab is becoming the aggregator, which is where a holder's negotiation will sit.

Confidence: high on the direction, moderate on the twelve-to-twenty-four-month clock. Every artefact above is checkable and dated; the timing rests on how fast the published task count and the number of networks move, both of which are signposts rather than forecasts.

What would reverse it. A general method that generates useful environments from synthetic specifications alone, removing the need for a real operator's work; or environment prices collapsing far below the published $200–$2,000 a task, which would take the bid out from under the records that feed them.

From collection to verification: the grader becomes the productRunning

Speed: Fastest of the seven; visible inside single training runs now

What the sources say. A published training recipe trains one policy as an agent organisation across 147 environments containing more than two million distinct tasks. The organisation beats the same model working alone — 29.8% to 37.5% on an advanced proof benchmark, 55.9% to 63.5% on a frontier-science benchmark — and the reviewing roles it creates for itself double from 13% to 26% of specialist hires. In a simulation reported in the same work, the share of runs in which an agent alerts the evaluations team falls from 76.7% to 20% and then zero. Separately, six decision models benchmarked on the same 98 production traces from one agent cost between 0.82 cents per 100 eval runs at F1 0.79 and 3.72 cents at F1 0.87, with the recommendation to tune the threshold per task rather than accept the 0.50 default and to check calibration with a Brier score.

Evidence grade: A for the training report, which publishes its benchmarks and its numbers; B for the eval-cost benchmark, which is a named published study on one agent's traces rather than an independent replication.

What we read into it. Two different things point the same way. Verification demand grows inside training itself as organisations of agents scale, while the grading function gets cheap enough to be a utility. Cheap graders and decaying self-reports both move the margin to the same scarce place: a held-out verifier keyed to a real outcome, written by someone who knows the domain, that the model being graded has never seen. That is the asset a buyer cannot write for itself and cannot check without.

Confidence: moderate. Both readings are specific, but one is a single training report and the other a single benchmark on one application's traces. Neither measures what anyone paid, which is the number this shift is really about.

What would reverse it. A verification method that generalises across domains without domain labels, which would make a proprietary verifier worth what a public benchmark is worth now; or buyers accepting model-written rubrics at scale.

Labs build and buy, and the line between them moves upBoth at once

Speed: Already visible; decisive by late 2027

What the sources say. An environment-economics survey reports substantially more in-housing, citing lab job postings and one product company training on its own users' work, and gives the reasons as avoiding vendor margin, keeping training confidential and holding domain expertise inside. Open tooling compiles environments out of repositories without a vendor at all. Against that: one buyer's own page commits $1 billion over twelve months to acquiring and licensing enterprise operational data; an open-role index read on 4 November 2026 across 62 job boards from 25 labs and 14 vendors finds 11 open roles sourcing, licensing or acquiring data at 8 companies and 16 more building environments; and the published eligibility floors across the register sit between 6 and 50 employees, with most at 20 to 40.

Evidence grade: C for the in-housing reading, which is a survey citing unnamed builders; B for the buying commitment and the eligibility floors, which are each organisation's own published terms; B for the open-role index, which is compiled by its publisher from public boards and states what it cannot see.

What we read into it. The floor rises and the ceiling rises faster, so the middle empties. A lab that can replicate a generic archive will, because open tooling made that nearly free; a lab that would have to reproduce a decade of linked operating history across four systems will not, because it cannot. The published eligibility floors are the clearest live reading of where the line sits, and they are the thing to watch: the line moves when the floors move, not when anyone announces it.

Confidence: moderate. The direction is supported from both sides, but the sharpest claims about where the line sits today came from practitioners describing their own deal flow, and we do not carry those. What is left is published floors and open roles, which move slowly.

What would reverse it. A lab publicly standing down from third-party company archives, which we put at 10% by 31 December 2027; or, the other way, a lab publishing a standing open programme with posted prices, which would mean buying beat building decisively.

Intermediary entry, and the shake-out that follows itPeaking

Speed: Entry already slowing; thinning through 2027 into 2028

What the sources say. The public register listed 159 vendors on 14 October 2026 and 187 on 4 November, and it removes companies that stop selling as well as adding ones that start. Ten of the 43 published programmes pay an introduction bounty; the largest published flat figure read $85,000 on 4 November against “up to $100,000 per deal” across the same programmes between 11 and 17 October, and one programme now pays 7.5% of deal value with no cap in place of a flat sum. Of 37 organisations surveyed, 16 publish any contract clause at all and 4 publish anything on exclusivity. Two of the 43 programmes are invite-only.

Evidence grade: C for the register's counts, whose maintainer discloses that it may sell to companies it lists and that listings are free, unranked and never paid for; B for each programme's own published bounty terms; A for the survey, which recorded each page and its reading date.

What we read into it. Entry into this layer is cheap, reflexive and unpriced, and the published bounty is what recruits it. A bounty figure that falls inside three weeks while the vendor count still rises is what the top of an entry wave looks like: the buy side has stopped paying more for reach before the supply side has stopped arriving. What survives the thinning is the part that can produce a package a buyer's counsel can act on — a rights position, an exclusion schedule, samples — because that is the part the bounty was never paying for.

Confidence: moderate on the peak, low on the timing. Two independent counts of the same programmes three weeks apart is a real reading but a short series, and one of the two counters is an interested party. The strongest first-hand accounts of how the buy side now filters were single-source and are not carried here.

What would reverse it. Bounties rising rather than being cut, and the listed vendor count continuing to climb through 2027 — which would mean the buy side still values reach over package quality.

The data supply chain gets financedForming

Speed: Early; the decisive year is 2028

What the sources say. Vendors wait 30 to 90 days for a lab to pay. One lender advances against those invoices on a $10m-plus warehouse facility from an institutional credit fund, with $65m-plus in signed letters of intent from data vendors and RL-environment companies, capital in 2 to 4 days, and underwriting on invoices, payment history, data quality and financials. One buyer's $1B twelve-month data commitment is itself capitalised rather than funded from operations. And in an adjacent sector the full structure is already on the record: a buyer expanded a research collaboration with $80M in cash, a $40M equity investment and recurring research funding for AI training and inference compute, in exchange for exclusive development and commercial rights to what the collaboration produces, with milestones and royalties behind it.

Evidence grade: B for the lending facility, which is the lender's own launch page; A for the collaboration, which is in the buyer's investor release and its quarterly filing.

What we read into it. Two things follow. Once lab receivables are underwritable, the vendor layer can scale on debt rather than equity, which accelerates consolidation and turns payment terms into a priced variable rather than a complaint. And the structure with the longest reach is not a purchase at all: where the asset is the counterparty's continuing work rather than its archive, the bid arrives as investment, joint venture or acquisition, with rights attached. That template is live in pharmaceutical research today; nothing about it is specific to that sector.

Confidence: moderate for the lending arm, which is dated and specific but is one lender. Low-to-moderate for the equity-for-rights arm, which is read across from an adjacent market rather than observed in company records — it is a template we expect to be copied, not a thing we have seen copied.

What would reverse it. A lab payment default or a vendor failure large enough to reprice the receivable, which would pull the facility and push the vendor layer back onto equity — and immediately slow consolidation.

Buyers get sophisticated, and a verification layer grows between holder and buyerHardening

Speed: Through 2027

What the sources say. One published programme anonymises every record through a third-party specialist before the buyer or any lab sees it. Another says plainly that no process removes all re-identification risk. One lets the seller approve each buyer and block named competitors. A wind-down service's terms let whoever acquires the assets license or sell them on without further consent or compensation, and one programme's terms state that deletion does not require it to delete or retrain any model, weights or derivative works. Two of 43 programmes are invite-only. Across 37 organisations, 16 publish any clause and the terms a seller cares about most are the ones most often left to negotiation.

Evidence grade: B — each line is a programme's own published terms, read on a dated page; A for the clause counts, which come from the survey that classified them.

What we read into it. A market with no price discovery still ends up with a few parties who can price, and the mechanism is provenance rather than information. The published terms show the first pieces of it: a named de-identification step, a seller approval right, an explicit statement about what deletion cannot reach. Each of those is a check a buyer can insist on, and a check a buyer insists on is a gate. The holder's decision shifts accordingly, from which destination to approach to which route will be accepted.

Confidence: moderate. The clauses are real and quotable, but reading them as a forming gate is our inference; no buyer has published a required intake standard, and the clearest first-hand account that one was forming was single-source and is not carried.

What would reverse it. A buyer with volume needs relaxing its intake standards to clear a deadline, or the published de-identification and approval clauses staying at one or two programmes through 2027 rather than spreading.

New modalities open: embodied, simulated and public-webOpening

Speed: 2027 for simulation; 2028 for embodied at scale

What the sources say. Simulation: a physics laboratory and a model developer signed a statement of intent on 15 October 2026 to build a single pretrained large physics model, the first foundation model for collider physics, trained on billions of simulated proton-proton collisions, for reconstruction, particle identification, flagging unexpected signals and evaluating future detector designs; no funding or timeline is disclosed. Embodied: the published robotics data pyramid puts web-scale video at the base, simulation and egocentric video in the middle, and teleoperated data collected on the deployment robot itself at the apex, where the economics do not scale — stated in a foundation-model whitepaper and in peer-reviewed scaling work. Public web: a commercial index of more than two billion people assembled from the public web sells agent access from $0.15 a credit, and a free index of 30M-plus US federal, state and local legal documents can be searched without spending any.

Evidence grade: A for the collider statement of intent and for the robotics papers; B for the two commercial indexes, which are their own published product and pricing pages.

What we read into it. The same shape appears in every modality: whatever can be generated, simulated or crawled goes to the floor, and the scarce layer is the one closest to a real outcome. That is the test to apply to any archive. It also cuts both ways for a holder of physical-process data — simulation raises the ceiling on what a model can be trained to do with it, and removes the floor under purchased observations of anything a simulator reproduces faithfully.

Confidence: high for the pattern, which now has four independent instances, and low for the dates. “2027 for simulation” is a judgment about adoption, not a reading from any source.

What would reverse it. Simulation proving good enough for business processes as well as physical ones, which would do to operating records what it is doing to collision data — and is the single fastest route to the Repricing scenario.

The causal model

The shifts are not independent. They are one loop with three reinforcing arms and three brakes, and almost every forecast on this page is a statement about which arm or brake moves first. This band is analysis: it is our model of the readings above, not a reading.

1
Agent demand
Buyers want models that do work, not models that answer. Doing work has to be practised and graded.
$1B+ discussed in one year at one lab
2
Environment demand
Practice needs environments and tasks. Tooling made building one cheap; knowing what to build one around did not get cheap.
$200–$2,000 a task · ~$300K a replica
3
Record and verifier value
The scarce inputs are connected operating history and a verifier keyed to a real outcome. Both sit inside operating companies.
147 environments · 2M+ tasks in one run
4
Prices and structures
Programmes publish ranges to recruit supply; the better buyers move to trailing structures to get the better supply.
43 programmes · 11 with trailing terms
5
Entry and noise
Published bounties recruit an introduction layer faster than the buy side can absorb it, and the average quality of what arrives falls.
187 vendors, from 159 three weeks earlier
6
Filtering and consolidation
Buyers filter on provenance and rights; a verification layer forms; the layer that cannot clear it thins out.
Largest published bounty $85K, from $100K

↻ And round again: filtering raises the value of a holder that arrives verified and rights-clean, which raises the value of the operating history behind it, which raises the price of the environment built on it — which is what pays for the next round of agent demand.

The three arms, and the three brakes

Reinforcing · R1The capability loopBetter agents need better environments; better environments need better records and verifiers; better records make better agents. This is the arm that has run hardest since mid-2026 and it is the reason an ordinary company's archive has a bid at all.

Reinforcing · R2The verification loopAs organisations of agents scale, they generate their own demand for reviewers — the published run where reviewing roles went from 13% to 26% of specialist hires is the cleanest evidence. More verification demand raises the price of domain verifiers, which raises the value of the outcome-linked records they are written from.

Reinforcing · R3The capital loopFinancing vendor receivables lets the vendor layer grow on debt rather than equity; a larger, better-capitalised vendor layer can pay more and wait longer; higher, slower-settling payments make the receivable more financeable. The $80M-plus-equity research collaboration in an adjacent sector is what this arm looks like at full extension. It is the youngest arm and the one most likely to surprise on the upside in 2028.

Balancing · B1In-housing and open toolingEvery time a lab builds what it used to buy, or a tool compiles an environment out of a repository, the commodity end of supply loses its floor. This brake has already bitten on loosely structured code and generic interface replicas.

Balancing · B2Keeping the loop rather than licensing itThe better a company's records get at producing value, the stronger the case for keeping them. The published argument for it is that a firm which owns its definitions, permissions, evaluated cases and hard-won exceptions can change model provider without losing what it has learned — the moat being the loop rather than the data. Two of the 43 listed programmes have already gone invite-only. The best supply gets harder to buy precisely because it became worth buying.

Balancing · B3Rights and consentA company that holds data does not necessarily have the right to sell it; some platform terms prohibit using data taken through their interfaces to train a large language model; a sale can waive privilege and jeopardise trade-secret protection; and employee communications about pay and conditions may be protected concerted activity whether or not anyone is unionised. Behind all of it: once data has trained a model it cannot practically be pulled back out. This brake does not slow the loop evenly — it stops the employee-bearing part of it and leaves the rest running.

The scenarios are a statement about order. If R1 and R3 outrun B1, the landscape industrialises into an environment economy. If B1 bites before R3 matures, it reprices. If the filtering in step six concentrates before either, a few parties become the gates. And B3 can end the employee-bearing part of it at any point in the sequence.

How it evolves, 2027 and 2028

Five phases to the end of 2028, each named for what changes in it, with the odds that attach to that window. Every figure here is one of the calls in band 13, carried unchanged.

Now to 31 Dec 2026

Terms move weekly, and the first court test lands

Published terms on this market change on a weekly clock. One top tier moved three times in three weeks — $1M+ on 16 October, $2M+ on 27 October, $3M+ on 3 November. One agent switched from a single payment of up to $1M to revenue share in four days. The survey that classified these pages recorded one programme raising its ceiling from $300,000 to $1,000,000 and another dropping a published figure inside a single month, and the largest published introduction bounty has come down since. Expect more of the same, and expect the register to keep adding and removing companies daily. The one court-supervised sale also reaches its hearing in this window, and it is the first real datapoint on whether a court will let a workforce's archive be licensed and on what carve-outs.

One of the ten largest published ceilings moves: 71% (55–85) · a sale order on the one public archive by 31 Mar 2027: 78% (68–88)

H1 2027

Packaging becomes the filter, and the gates start to form

The buy side stops rewarding reach. The clauses that are now one or two programmes' preference — a named third-party de-identification step, a seller approval right, a written exclusion schedule — start appearing as preconditions rather than features, and the introduction layer splits into the part that can produce a package and the part that cannot. Expect the first visible cuts to the largest bounties, expect a second environment network, and expect at least one party to start positioning itself as the route a lab will accept a corpus through rather than as a seller of corpora.

50 or more buying programmes listed: 62% (45–78) · a second named environment network: 58% (42–73) · a large bounty cut or withdrawn: 63% (48–79)

H2 2027

The first renewals, and the first trailing structures

The earliest company-record licences reach their first renewal in this window — the median contract term across 312 AI training-data agreements is 2.0 years — and renewal is where the one filed repricing says the price gets tested. That filing replaced annual fees of $22.0M, $32.0M and $42.0M with $14.0M on each of three anniversaries, added a year of term and eliminated the buyer's right to terminate for convenience. That is the shape to expect: the second payment is smaller unless the structure earns it. This is also the window where the first sellers finish their acceptance periods and could say in public what they were actually paid.

20 or more listed programmes publishing trailing terms: 37% (22–52) · a seller confirms $1M or more: 41% (26–56) · a second filed repricing: 40% (25–55)

H1 2028

The shake-out in the middle

The layer that is neither a verified supply channel nor a real counterparty gets squeezed from both sides: better-capitalised vendors buying or converting the good intermediaries, and buyers refusing packages that have not been through a verification layer. Expect the listed vendor count to stop rising and start falling. Expect the first embodied and field-operations programmes on the register, because the pattern that opened company records is the same one opening physical ones — and because the published robotics pyramid says the scarce layer there is deployment data held by whoever is already operating.

An embodied or field-operations programme appears on the register: 46% (30–62) · listed vendor count falls year-on-year: 50% (34–67)

H2 2028

Environments, verifiers, and a price you can look up

On the central path the thing a company sells is no longer a copy of its records. It is a maintained environment built on how it works and a verifier that grades it, licensed as it runs, with the company in the income. That is better for the holder and harder to reach, and it needs the rights position clean from the first day — which is why the rights work comes before the price in every route on this page. The open question at this horizon is whether price discovery arrives with it. The adjacent input went from a black box to a public per-region rate feed behind an unauthenticated API in a single step; there is no structural reason data cannot, and several commercial reasons it might not.

Any trailing structure dominant: 64% (43–79) · half of new programmes asking for a verifier or environment: 30% (18–42) · a public rate feed or index for training data: 31% (18–44)

Scenarios

Five ways the landscape can go. Each carries odds at 31 December 2027 and 31 December 2028, the trigger that puts you on it, the early markers that tell you first, and what it means on each side of a deal. Each column sums to 100, and every signpost in band 12 names the scenario it moves.

Selective market

37%Dec 2712%Dec 28

Today's shape continues. Buying goes on at roughly this intensity, published ranges stay up, confirmed payments stay well below them, nobody can price the asset so the process does the pricing, and what changes hands is a selective, excluded, de-identified, non-exclusive copy of operating exhaust — plus the whole archive of companies that are leaving. The reason this falls from 37% to 12% is not that it stops working. It is that every one of the three reinforcing arms, if it runs, moves the landscape off this square.

Trigger: None. This is the no-trigger path: nothing resolves, and the dispersion persists because the information problem persists.

Early markers: Programme count flat or rising; no published ceiling confirmed paid; the spread between public estimators for one illustrative company staying above about ten times.

For a holder: The process is worth more than the counterparty. Parallel written indications are the only thing that converts dispersion into price, and the first offer is an anchor rather than a valuation.

For destinations: Supply, not demand, is the constraint, which is why bounties exist and why the better destinations will keep paying for packaged, rights-clean supply rather than for reach.

Environment economy

25%Dec 2733%Dec 28

The modal destination by the end of 2028. Maintained environments and trailing structures become the normal deal rather than the exception: an environment built on a company's own tasks and feedback and run for it, revenue share on each onward licence, annual trajectory licences, a verifier maintained alongside. The one-off archive sale becomes the thing that wind-downs do. Holders are in the income rather than paid once, and the rights position has to be clean from the start because the relationship is continuous.

Trigger: A second and third environment network, and the first large buyer that will only transact in a trailing structure because its own economics now depend on the environment staying current.

Early markers: Newly listed programmes publishing revenue-share, royalty or recurring terms (11 of 43 now); a second named environment network; published annual trajectory licences spreading beyond the one programme that publishes $1–2M a year.

For a holder: The best outcome available, and the one that needs the most preparation: an inventory, a clean rights position, an exclusion schedule, and a reason the work is still being produced.

For destinations: Environment networks and platforms win the aggregation role. Pure resellers of static files have no place in this scenario at all.

Buy-side consolidation

15%Dec 2713%Dec 28

A small number of parties become the gates every corpus has to pass through before a lab will treat it as safe to buy. Prices firm for supply that arrives through a gate and collapse for supply that does not. Note the shape of the odds: this is a waypoint rather than a destination. On most paths it is where the environment economy comes from, which is why its share flattens rather than compounding.

Trigger: One verification or provenance layer being adopted by two or more large buyers as a precondition rather than a preference.

Early markers: A destination publishing a required intake standard rather than a preferred one; a named third-party de-identification step spreading beyond the one programme that publishes it; programmes moving from open application to invite-only (2 of 43 now).

For a holder: Which gate you go through becomes a bigger decision than which destination you reach. Going direct to a programme that is not on the accepted side of a gate wastes the only first approach you have.

For destinations: Gatekeepers take the margin. Annotation vendors without a provenance story get disintermediated by the layer that has one.

Repricing

11%Dec 2724%Dec 28

Appetite narrows, open tooling bites, synthetic and simulated supply substitutes for purchased supply, and published ranges come down with it. Ceilings drop off pages rather than being withdrawn with an announcement — the survey of these programmes caught one doing exactly that inside a month — and the register's count falls for the first time. This is the second-likeliest 2028 outcome and the one most under-discussed by the people selling into the market.

Trigger: Either a lab standing down from third-party archives, or a generation method good enough that an ordinary company's operating records stop being the cheapest route to a usable environment.

Early markers: Two or more large programmes cutting a published ceiling; bounties cut or withdrawn; the open-role index falling below six acquisition roles; a second filed repricing of a data licence.

For a holder: Time is the risk, not price. A holder waiting for a better market in a repricing is waiting for the opposite of what arrives, which is the strongest argument for running a real process now rather than later.

For destinations: Direct buyers of licences take the loss on their own book. Marketplaces survive on volume; thin venues with a few dozen buyers on one side do not.

Rights chill

12%Dec 2718%Dec 28

A court, a regulator or a credible re-identification claim makes employee-bearing records unsellable in practice. Buyers retreat to source code, drawings, documents and records with no people in them; the operating-exhaust trade stops and the wind-down trade survives because a closed company's archive can be excluded down to almost nothing. The odds rise with the horizon because the exposure accumulates: the more archives that move, the more chances there are for one of them to go wrong in public.

Trigger: A sale denied over employee data, a regulator treating a de-identified workplace archive as personal data, or a published demonstration that a licensed archive was re-identified.

Early markers: A programme suspending intake; a state statute or federal bill specifically on licensing employee records; a destination adding a mandatory employee-notice requirement to its own published terms.

For a holder: Everything depends on work done before the event: an exclusion schedule, an employee notice, and a rights position written down. A holder who has that keeps a narrower but still saleable asset; one who does not has nothing to sell.

For destinations: Wind-down and asset services, code-only buyers and drawing-and-CAD buyers are the survivors. Anyone whose supply is workplace communication is out of the market.

Winners and losers

Who gains and who loses on the central path, and why. This band is analysis throughout: direction is stated against today, not against each other, and nothing here is a reading.

Holders of records

Segment	Direction	Why
Mid-size operators, 50–500 people, one software stack for five years or more	Clear winner	This clears every published eligibility floor on the register, which runs from 6 to 50 employees with most at 20 to 40, and it has the continuity for the history to count from the oldest record rather than the last migration. It is also the shape a buyer can build an environment on, because the same people and events appear in more than one system.
Small operators with vertical depth in a domain models are weak at	Winner, selectively	The exception to the rising floor: depth in an esoteric domain holds value where generic scale does not, because the verifier nobody else can write is worth more than the volume anyone can match. The trade for this segment is depth, documentation and that verifier, not size.
Companies whose work is already instrumented end to end	Winner, and increasingly	Where a request, the work done, the decisions and the outcome are all captured and linked, the holder is not selling records at all — it is selling the raw material of an environment plus the grade. That is the only supply the environment economy actually needs, and it is what a twelve-partner environment network is built to consume.
Generic operators under 20 people, run largely by phone	Loser	Below every published headcount floor on the register, and with the least to show in writing. Records that read as “ticket created, ticket closed” with nothing in between carry no environment and no verifier, which is the whole of what the demand side is now reaching for.
Holders of large but disconnected archives	Loser, and they rarely know it	Volume without linkage is the most common reason an owner is surprised by a number. A stack migration three years ago can mean twenty years of history counts as three, and no amount of size recovers it. This is an inference from what programmes say they price on rather than from observed offers, and it is the line on this page we would most like a better source for.
Companies that are closing or have closed	Winner on timing, loser on optionality	The clean end of the market, because there is no competitive reason to hold back and no business left to protect. It is also where the only seller-confirmed payment on the public record sits. But a wind-down sells once and never participates in the trailing structures the rest of this forecast is about.
Holders in jurisdictions with strict employee-data regimes	Loser on breadth, winner on credibility	Programmes state a preference for US companies first and other Western markets after. A stricter regime narrows what can move, but a holder that has already done consent, notice and retention work is the holder least exposed to the rights-chill scenario and most able to sell during it.
Sector concentration: legal, engineering and CAD, healthcare, industrial and field operations	Winner	These are the verticals the published programmes name in their own eligibility rules — legal, healthcare, finance and insurance, professional services, manufacturing, retail and logistics — and they are where a task has an outcome somebody already recorded.

Destinations and the layers between

Destination or layer	Direction	Why
Frontier labs	Winner on terms, loser on exclusivity	They set the frontier and they can build what they cannot buy. What they lose is uncontested access: as more holders learn that an exclusive licence costs roughly four to five times a shared one, the default of taking exclusivity for nothing stops surviving a real process. Only four of 37 organisations publish anything on exclusivity at all today.
Environment builders and networks	Clearest winner	They sit exactly where the unit of trade is moving, they are paid to maintain rather than to deliver once, and the network structure lets a platform aggregate specialist builders without hiring them. This is the segment the environment economy is named after.
Verification, provenance and de-identification layers	Winner, and under-priced today	If a few parties become transaction gates, this is what they are made of. One programme already routes every record through a third-party specialist before it or any lab sees it; that clause is the cheapest way a market with no price discovery acquires a way to price.
Marketplaces and exchanges with a real buy side	Mixed	A venue is only as good as the demand standing behind it, and a listing surface with no relationship team behind it has nothing to convert. Venues that solve the relationship problem win volume; venues that only solve listing do not survive the shake-out.
Intermediaries that produce a package	Winner, by consolidation	The layer that arrives with a rights position, an exclusion schedule and samples is the layer whose work a buyer's counsel can act on. That layer gets bought, funded or converted into something larger.
Intermediaries that produce only an introduction	Loser	A tracked link and an owner's name is the easiest part of this supply chain to replace, and it is the part the published bounty was paying for. The bounty cuts are how the repricing of that work will be announced.
Annotation and expert-data vendors	Loser at the commodity end, winner at the expert end	Automatic environment generation and cheap decision models take the floor out from under generic labelling and generic grading. What survives is the part that needs a human who knows the domain, which is also the part a held-out verifier is written from.
Financiers of the supply chain	Winner, with the most room to surprise	Receivable finance is already live at $10m-plus. The structures with real reach are further out: investment, joint venture and acquisition in exchange for rights a vendor contract never grants — the shape of the $80M-plus-equity research collaboration already on the record in an adjacent sector.
Pure resellers of public-web and public-record data	Loser	More than two billion public-web profiles sold to agents from $0.15 a credit, and a free index of 30M-plus US legal documents, are not a competitive threat to this segment — they are its obituary.

Frontier Desk acts as the advisor to the company that holds the records: the rights work first, then one description in front of every destination that fits, written indications compared clause by clause. Every result comes back into the map and the calls on this page, which is why the forecast stays current rather than ageing from the day it was published.

Commoditised or scarce

The single most useful question a holder can ask is which side of this line its records sit on, because the two sides move in opposite directions and they move on different clocks. The left column is a reading; the dates are ours.

Commoditising, and by when

What	By when	Why
Public-web facts about people and companies	Already done	A commercial index of more than two billion people assembled from the public web sells agent access from $0.15 a credit. There is no licensing floor under anything a crawler can reach.
Public-record corpora: statutes, opinions, codes, permits	Already done	A free index of 30M-plus US federal, state and local legal documents — statutes, regulations, ordinances, zoning codes, permits, court opinions, bills and agency guidance — is searchable without spending a credit.
Loosely structured source code with no verifier	Already done	Open tooling compiles tasks and private verifiers straight out of merged pull requests, commit history and security advisories — 1,930 published tasks at v0.9.5. If the repository already encodes the grade, nobody needs to buy the grade.
Generic interface replicas and UI gyms	2027	A published cost of about $20K a replica, against an Apache-2.0 framework that standardises environments behind one protocol and an MIT-licensed tool that builds one by recording a browser session.
Generic grading and LLM-judge scoring	2027	Six decision models on the same traces ran from 0.82 to 3.72 cents per 100 eval runs. The grader is becoming a utility; the criteria are not.
Raw, unstructured, anonymised record dumps	2027–2028	Nothing in a flat export tells a buyer whether a task succeeded, and the demand side is converging on tasks that something real decided. This date is a judgment, not a reading.
Physical and scientific processes that can be simulated	2028	Billions of simulated collisions, by a statement of intent dated 15 October 2026. Where a process can be simulated faithfully, purchased observations of it stop being scarce.
Bare introductions to willing sellers	2027	The largest published introduction bounty came down from “up to $100,000 per deal” in mid-October to $85,000 three weeks later, while the vendor count kept climbing.

Keeping its scarcity value, and why

What	How long	Why
Connected multi-year operating history with outcomes attached	Scarce through 2028 and beyond	The same people and the same events appearing across email, chat, CRM and ticketing, continuously, so a problem can be traced from raised to resolved. It cannot be generated, crawled or simulated, and it cannot be assembled retroactively. This is our central judgment about price and the one we hold with the least direct evidence: no published source reports what two comparable archives were actually paid.
Verifiable tasks drawn from work that actually happened	Scarce, rising	Tooling made building an environment cheap and left untouched the question of what to build one around. A task is only worth training on if something real decided whether it succeeded — which is the rule the open tooling itself enforces, by refusing to publish a task unless an oracle scores 1 and a no-op scores 0.
Proprietary verifiers and domain evaluations	Scarce, and the fastest-appreciating item on this page	Verification demand grows inside training itself — reviewing roles doubling from 13% to 26% of specialist hires in one published run — while self-reported grades decay in the same work. A held-out verifier written by someone who knows the domain is the asset this creates.
Rights-clean records with consent, notice and an exclusion schedule already done	Scarce, and the gate on everything else	Holding data is not the same as having the right to sell it; some platform terms forbid using data taken through their interfaces to train a model; a sale can waive privilege; and once data has trained a model it cannot practically be pulled back out. In the rights-chill scenario this is the difference between a narrower asset and no asset.
Decision histories, negative results and protocols	Scarce, and barely traded yet	What an outcome cost to learn is what makes it worth learning from, and it is exactly what a token-metered relationship never conveys. This is the category the financing structures reach for, which is why the highest bids for it arrive as investment and equity rather than as purchases.
Physical, field and embodied records from real operations	Scarce into 2028	The published robotics pyramid has the same shape as this one: endless supply at the base, almost none at the apex, and the apex is data collected on the deployment system itself. A company already producing that is holding the scarce layer by accident.

The case against this table.

The strongest argument against this table is that we have drawn the line in the wrong place. It runs: anything whose success can be defined finitely and unambiguously is on its way to zero, because once a task is solved the solved task is worth nothing — so the verifiable half of the market is a melting asset and the profits end up in taste and judgment, which no verifier can grade.

We think that is right about the task and wrong about the verifier. The price of a task that has been solved does go to zero; the price of being able to tell whether the next one was solved does not, because that is the thing a buyer cannot check for itself and cannot write without the domain.

But we hold this at moderate confidence, not high: no published source prices a verifier separately from the data it came with, so the claim is an inference about where margin lands rather than an observation of it. If it is wrong, the Repricing scenario is badly under-weighted at 24%, and the signpost that would tell us first is the open-role index for environment roles falling rather than any price moving.

What to do about it

The forecast is only useful if it changes a decision. Here is what it changes, for the company that holds the records and for anyone holding several of them.

If you hold the records

Decision	What the forecast says
Sell now, or wait for a better market?	Do not wait for a price. Two of the five scenarios take published ranges down, in a third they firm only for supply that arrives through a gate, and none of the five has them rising for an unpackaged archive. The one thing that improves a number in every scenario is preparation rather than timing. Start the rights work now; the sale can start whenever the rights position is written down.
Sell outright, or license?	License, non-exclusively, wherever the business is still producing the work. An exclusive licence costs a buyer roughly four to five times a shared one, and only four of 37 organisations publish anything about exclusivity at all — so it is routinely taken without a separate price. A perpetual, sublicensable grant paid once is economically a sale, and it forecloses the environment partnership that is the likeliest valuable structure by 2028.
One payment, or a share?	A share, if you can get it. Eleven of the 43 listed programmes publish a trailing term and the structure odds say a trailing deal is dominant on 64% of paths at December 2028; across 312 AI training-data agreements the flat annual fee fell from 58.3% of deals in 2023 to 34.8% in 2026 while revenue share rose from 4.2% to 13.4%. A one-off signed today signs away that option.
What to keep exclusive?	The verifier, the outcome labels and the live feed. Licensing a copy of history is survivable; licensing the mechanism that grades your work, exclusively and perpetually, hands over the asset this forecast says appreciates fastest. Carve out current and future records as a separate negotiation rather than including them by silence.
How do you avoid being commoditised?	Sell the thing a crawler, a repository and a simulator cannot make: linkage, continuity and outcomes. Before any call, list every system the business has run on and when each started, find where internal communication actually lives and how far back, and check whether the same people and events appear in more than one of them. That exercise tells you more about what you hold than any estimator will.
Who do you actually talk to?	More than one, in parallel, in writing, with a deadline — because dispersion is the market's defining feature and a single counterparty cannot be benchmarked. Two public estimators for one illustrative company differ about 21 times, nine programmes publish a ceiling and none has a reported payment of that size behind it. Frontier Desk is the advisor that runs that process through its relationships with buyers and exchanges, and a cybernetic system of its own.

If you hold a portfolio of companies

The asset is unpriced, it sits inside several holdings at once, and there is no reason each company should learn this market separately. One review across the portfolio answers three things at once: which companies clear the published eligibility rules as they stand today, what each would likely be quoted and by which kind of destination, and which rights questions would block a sale before it starts.

The forecast changes what that review is for. On a two-year view the question is no longer “what is this archive worth” but “which of these companies is producing work that an environment could be built on, and is its rights position clean enough to let that happen”.

The first is a one-off number. The second is a recurring line, and it is also the answer to the quieter question — which holdings are quietly giving away, for a single payment, the thing that would have made them harder to compete with.

Frontier Desk is the advisor that runs the sale through its relationships with buyers and exchanges, and a cybernetic system of its own. The first conversation covers whether your records qualify, which route fits, and what to check in an offer: Get your company ready.

Find out which destinations your records already qualify for.

The published eligibility rules decide most of it before any number does: how many full-time staff, how many years in operation, which systems the work ran through, and who holds the rights to what is in them. Frontier Desk reads your position against every programme in the register above, does the rights work, and runs one process to the destinations that fit.

Get your company ready What to do with your records

The evidence base

The present-state record the forecast is built on, kept here as a reference rather than an argument: who publishes what today, how wide the dispersion is, which clauses decide what a deal is worth, and the rights work that gates all of it. Everything in this band is collection — each cell is a named party's own published wording, read on the date shown.

The destination register

Publicly observable

Destination	Type	Published terms	Who qualifies	Structure	Source
micro1	Expert-data vendor	Tiers of $100K–500K, $500K–3M and $3M+; trajectories $1–2M a year	30+ full-time staff, documentation primarily in English, US first	One-off, paid in full within 30 days of acceptance; licence perpetual, irrevocable and sublicensable	micro1.ai · 4 Nov 2026
Handshake AI	Expert-data vendor	“Payments to date have ranged from $100K to $4M” — the only range a programme presents as paid	Typically 20+ full-time employees and 3+ years in operation	One-off on delivery milestones agreed per deal; 7 to 14 business days end to end	joinhandshake.com · 2 Nov 2026
Mercor	Expert-data vendor	No amount published; pay depends on data volume and the tools connected	Legal, healthcare, finance and insurance, professional services, manufacturing, retail, logistics	One-off, wired on delivery of the anonymised set, typically 2 to 4 weeks after signing; no deductions	mercor.com · 3 Nov 2026
Turing (Project Lazarus)	Expert-data vendor	Revenue share on every sale; changed from a single payment of up to $1M between 29 Oct and 2 Nov 2026	Operating companies, and teams winding down or already closed	Trailing only, nothing upfront; sole seller of the record for the first 90 days, then renew or opt out	lazarus.turing.com · 4 Nov 2026
Labelbox (Prism)	Expert-data vendor	From about $50K to $1M+ per licence	Companies; also sources from individual contributors	Licences are structured as exclusive by default; written quote within 48 hours of scoping	prism.labelbox.com · 16 Oct 2026
SuperAnnotate	Licensing agent	Indicative valuation $100K to $1M+	Companies of any size, industry or location	Paid on delivery and again on each relicense; priced on rarity, depth and exclusivity; the seller approves each buyer	superannotate.com · 16 Oct 2026
Appen	Expert-data vendor	Quoted per partnership; no amounts published	Companies; those winding down get a faster path	Terms agreed in writing before extraction; payment on acceptance, not submission; extraction and anonymisation covered	appen.com · 16 Oct 2026
AfterQuery	Expert-data vendor	Price proposed after sample review; no public range or minimum	Established companies with several teams and connected systems; operating, winding down or closed	One-off; payroll, privileged material and credentials never used, and it says no process removes all re-identification risk	afterquery.com · 29 Oct 2026
Collinear AI	Expert-data vendor	No amounts published; an initial range quoted within a day	Prices rise with 15+ people at peak, several years of history and human-written code	A one-time export under a licence; ownership stays with the company; about a week from first call to signature	collinear.ai · 16 Oct 2026
Mode	Direct buyer	Indicative range $100K to $5M, labelled indicative by its own calculator	6+ staff for legal, 10+ for accounting, finance and medical, 30+ for all others; US teams the strongest fit	One-off; non-binding letter of intent within 14 days, payment about 60 days after the final offer	data.mode.inc · 3 Nov 2026
Polyshares	Direct buyer	“$100K to $2M, and above for large records”; an offer within one to two weeks of intake	“Any company with real operating history, in any sector”	One-off, net 30 to 60 once approved and anonymised by a third-party specialist before it or any lab sees the record	polyshares.com · 4 Nov 2026
Nyne AI	Direct buyer	Six-figure payouts; partnerships “commonly worth $100K or more”	20+ US W-2 employees and 3+ years of workflows	One-off, paid after delivery and acceptance; identifying details removed on the buyer's side	nyne.ai · 16 Oct 2026
Ooak Data	Direct buyer	Up to $1M; closed deals so far $20K to several hundred thousand; its on-page estimator caps at $300K	Operating, winding-down and closed companies; a liquidator or administrator can sign	One-off, paid to the company entity; the raw archive is deleted within 30 days of delivery	ooakdata.com · 3 Nov 2026
Replay	Direct buyer	$10K–$100K at 20–50 employees, up to $1M+ at 250+; bands are its own headcount estimates	20+ full-time employees and English-language operations, US preferred	Upfront cash payment plus a perpetual revenue share; only de-identified copies are licensed	replay.ai · 16 Oct 2026
Telegraph Lab	Intermediary	Target licensing opportunities of $100K–$4M, described as targets rather than a guaranteed offer	Businesses of any industry or size with the authority to license their records	Scope, rights, recipients, uses, payment milestones and preparation all defined in a written agreement	telegraphlab.com · 29 Oct 2026
SourceX	Direct buyer	Up to $4 million for the largest qualifying datasets	Established businesses, prioritising 50+ employees at peak	A licence for AI training and evaluation; an adviser referral programme with no rate or cap published	EIN Presswire · 2 Nov 2026
The General Data Company	Strategic buyer	A one-off price or a share of licensing revenue; no amounts published	Industrial companies that hold the rights, including companies no longer trading	Licensed to AI labs and to manufacturers training their own models; names replaced with fictional values first	thegeneraldata.com · 16 Oct 2026
Watt	Strategic buyer	Cash plus royalties on each licence; no amounts, rates or timetable published	Companies with proprietary operational datasets and agreed commercialisation rights; no minimum size	The company reviews the catalogue and agrees rights and terms before anything is commercialised	wattdata.ai · 29 Oct 2026
Troveo	Exchange or marketplace	Revenue share on every sale; full-company deals typically start at six figures	Companies that own their data; no other requirement stated	Trailing only; licensed on the owner's behalf, with no fee charged and earnings taken from the share	troveo.ai · 16 Oct 2026
Opendatabay	Exchange or marketplace	Sellers keep 70% to 95% of each self-serve sale, rising with transaction size; listing is free	Registered businesses and organisations, after identity checks; individuals cannot sell	Commission also applies to off-platform deals with introduced buyers for up to 24 months	opendatabay.com · 16 Oct 2026
DataVendor	Exchange or marketplace	Sale proceeds less a 20% fee on bank withdrawals; in beta, and fees are current settings that can change	Organisations only; vendor setup is self-serve with no application review, and buyers are invited	Trailing; earnings available as soon as the buyer pays, with withdrawals reviewed in about 3 business days	docs.datavendor.ai · 2 Nov 2026
pre.dev	Exchange or marketplace	$100 per million source tokens plus $40 per million git-history tokens	Owners of private codebases, who connect a Git account and pick the repositories	Non-exclusive by default, so the same code can be paid for more than once; paid within 15 days of the buyer's funds landing	pre.dev · 16 Oct 2026
Magma	Exchange or marketplace	Revenue share as the data licenses; no percentage published	Companies building AI agents	Trailing; anonymisation is not described, and handling and retention are left to the data agreement	magmahq.ai · 16 Oct 2026
Protege	Exchange or marketplace	Revenue share on qualified deals; no percentage published	Organisations that own their data	Partners keep ownership; terms structured with the AI buyers rather than set in advance	withprotege.ai · 16 Oct 2026
SimpleClosure (AssetHub)	Wind-down service	Success fee on sales or licences; past deals ran from tens to hundreds of thousands of dollars	Startups, including companies still operating	Whoever acquires the assets may license or sell them on without further consent or compensation	simpleclosure.com · 16 Oct 2026
Google	AI lab, direct	A paid pilot for non-public content; no public rates and no application route	No public eligibility rules; science, mathematics and learning are named priorities	Described on its own page as a non-exclusive opportunity	ai.google · 16 Oct 2026
Microsoft	AI lab, direct	Paid on delivered value; no rates published	Publishers, through an interest form; the programme is a pilot announced 3 Feb 2026	The post covers grounding answers and does not describe model training	about.ads.microsoft.com · 16 Oct 2026
Baseten Environment Network	Environment network	No terms published	Customers of the platform; twelve RL-environment companies named as launch partners	Partners build the environments on the customer's own data, tasks and feedback; the platform handles training and production inference	baseten.co · 23 Oct 2026

Every cell is the destination's own published wording as read on the date shown, condensed, or a public register's summary of it. Several of these pages changed more than once in the three weeks before publication.

What the 43 published programmes disclose

Publicly observable

Counts out of the 43 programmes on the public register, read 4 November 2026

Published payout ranges

Publicly observable

Each programme's own published floor and ceiling per licence, log scale. A published range is not an offer.

Nine programmes publish both a floor and a ceiling on their own pages, so nine are plotted. A programme that publishes no figure is in the destination register above rather than on this chart: Mercor, for one, states that what it pays depends on the data volume and the tools connected — mercor.com/data, read 3 Nov 2026.

The dispersion, as published

Publicly observable

21×

Spread between two public estimators

$37,500 against $800,000, for the same illustrative company of about 100 people and 10 years.

$100K

Median published floor

Across the 14 programmes that publish one. The lowest is $10,000, at four of them.

$1M

Median published ceiling

Across the nine that publish one. The highest is $5,000,000.

1 of 24

Payments a seller has confirmed

Of the 24 programmes that take a company's own records.

Reading	Source	Figure	Grade
Low ends of six public estimators, same illustrative company	Programme estimators via the comparison guide	$37,500 to $800,000 — FD estimate: $800,000 ÷ $37,500 = about 21 times	B · party
Published floors across 14 programmes	Working paper, 18 Oct 2026	Median $100,000; lowest $10,000, at four programmes	A · record
Published ceilings across 9 programmes	Working paper, 18 Oct 2026	Median $1,000,000, highest $5,000,000	A · record
Ceilings with an independently reported payment behind them	Working paper, 18 Oct 2026	0 of 9	A · record
Programmes with a payment confirmed by a seller	Working paper, 18 Oct 2026	1 of 24	A · record
Largest payment a seller has confirmed in public	Named press, via the working paper	“Hundreds of thousands” for a 13-year archive	A · record
Smallest confirmed payment on the record	Named press, via the working paper	About $25,000 for a closed startup's codebase	A · record
One published ceiling's movement inside a single month	Working paper, 18 Oct 2026	$300,000 → $1,000,000 at one programme; a published figure dropped at another	A · record
Highest number any company record has drawn in public	Case docket, No. 25-11897 (Bankr. S.D.N.Y.)	$12.5M, a filed notice of intent; $10.0M selected at auction, not yet approved	A · record

Median annual value of a licence, by kind of buyer

Publicly observable

Across 312 AI training-data agreements signed since 2023, of which 147 disclosed a value. Company operating data is not one of the categories in that set.

The clauses that decide what a deal is worth

Clause	What to look for	What the published record shows
Licence scope and sublicensing	Perpetual or for a term; irrevocable or terminable; sublicensable, and through how many tiers.	One programme's licence is non-exclusive yet perpetual, irrevocable and sublicensable. Only 2 of 16 organisations publish a sublicensing clause at all.
Exclusivity	Whether it is taken by default, and whether a separate price is named for it.	Four of 16 organisations publish anything on exclusivity; one structures licences as exclusive by default and one prices on rarity, depth and exclusivity. An exclusive licence costs a buyer roughly four to five times a shared one.
Permitted use	Training only, or training and evaluation, or any purpose; whether derived models and outputs are in scope.	One programme licenses for training and evaluation only. Another licenses qualifying agent trajectories separately at $1–2M a year, which is the structure the forecast says spreads.
Acceptance and rejection	How long the buyer has to reject, what silence means, and whether the seller may invoice before acceptance.	One programme reviews for 30 days after transfer and pays within 30 days of an invoice the seller may only raise after approval. Several programmes across the survey pay only on the buyer's acceptance rather than on submission.
Deletion, and what it reaches	Whether deletion covers the raw copy, the derived dataset, or the model.	Only 2 of 16 organisations publish a deletion clause. The one that addresses trained models says deletion does not require it to delete or retrain any model, weights or derivative works. Two others delete the raw archive within 30 days.
De-identification standard and who bears it	Who scrubs, where, to what standard, and who certifies it.	In most published programmes the buyer bears the cost. One anonymises through a third-party specialist before it or any lab sees the record; one says plainly that no process removes all re-identification risk.
Resale and competitor blocks	Whether the buyer may resell, and whether the seller can block named competitors.	One programme lets the seller approve each buyer and block competitors. A wind-down service's terms let whoever acquires the assets license or sell them on without further consent or compensation.
Term and renewal	How long it runs and what happens at renewal, which is where the price gets tested.	Ten of 16 organisations publish nothing about licence term. The median across 312 AI training-data agreements is 2.0 years. The one filed repricing replaced annual fees of $22.0M, $32.0M and $42.0M with $14.0M on each of three anniversaries, added a year of term and eliminated the buyer's right to terminate for convenience.

The rights work, which gates everything above

Step	What it is	Why it matters to the forecast
Inventory	Every system the business has run on and when each started; where internal communication lives and how far back; whether the same people and events appear across more than one tool. Measured from each admin screen, with nothing exported.	This is the price variable, not a formality. A migration three years ago can mean twenty years of history counts as three.
What is actually in the set	A written answer to the first diligence question any buyer's counsel asks: what the dataset contains, where each part came from, and what is mixed into it.	Email and chat routinely carry third-party confidential material, licensed data and vendor-generated work with competing ownership claims. The answer narrows the set before anyone prices it.
Ownership and third-party terms	Customer contracts, non-disclosure agreements, vendor and platform terms. Holding data is not the same as having the right to sell it, and some platform terms prohibit using data taken through their interfaces to train a large language model.	This is the single largest reason a published clock bears no relation to a real one, and the question that most often shrinks a deal rather than killing it.
People in the archive	Employee and contractor messages, calendars, reviews and payroll adjacency; consent, notice and works-council positions where they apply. Employee communications about pay and conditions may be protected concerted activity whether or not anyone is unionised.	The direct exposure to the rights-chill scenario. In the one public sale, a union negotiated medical, leave and union records out of the set and more than 100 members of Congress asked for wider exclusions.
Exclusion schedule	A written list of what never leaves: privileged material, credentials, payroll, medical and leave records, named customers, anything under an active dispute.	The only protection that survives training. Once data has trained a model it cannot practically be pulled back out, so excluding high-risk material beats relying on de-identification.
De-identification standard	Who scrubs, where, to what standard, and who certifies it — agreed before extraction rather than after.	The likeliest thing to become a gate. One programme already routes every record through a third-party specialist before it or any lab sees it.
Rights position, written down	One document stating what can be licensed, to whom, for what use, with the exclusions annexed and the open questions named rather than hidden.	The asset that converts dispersion into price: it is what lets several destinations quote the same thing at the same time.

FD estimate, from the published terms above: two to six weeks of seller-side rights work for an operating company with customer contracts and employee messages in the archive, and ten to twenty weeks from decision to cash.

That runs one to two weeks to a first written indication on the published offer windows, plus the rights work, which no programme's published clock counts, plus the published payment clocks, which run from seven days after a set passes review to about sixty days.

Key assumptions

The assumptions the odds depend on, and the evidence that would break each one. Each is an analytical commitment, not a reading.

Assumption	Breaks if
Agent and environment demand keeps a bid under ordinary companies' records through 2028.	Task prices fall materially from the published $200 to $2,000, the open-role index for environment roles falls below six, or a general method generates useful environments without a real operator's work.
Verification demand grows faster than verification supply.	A grading method generalises across domains without domain labels, which would price a proprietary verifier the way a public benchmark is priced now.
Connectedness and outcome-linkage, not volume, are what a buyer pays for.	A published account of two comparable archives fetching comparable prices despite differing sharply on linkage. This is the assumption with the least direct evidence under it on the whole page.
Licensing records does not foreclose the holder's own loop, unless it is exclusive and perpetual.	Buyers move to exclusive-by-default as the norm and holders accept it, in which case the whole “license, do not sell” recommendation inverts into “build instead”.
Published benchmarks stay gameable and private verifiers stay scarce.	A credible public eval emerges that resists adversarial agents, which would remove the premium on held-out, domain-written graders.
Rights work, not buyer appetite, stays the binding constraint on cycle length.	Processes close inside published clocks for operating companies with customer contracts and employee messages in the archive.
Employee-record sales survive scrutiny with wider carve-outs rather than being stopped.	A court denies a sale over employee data, a regulator treats a de-identified workplace archive as personal data, or a licensed archive is publicly re-identified.
The intermediary layer thins without taking the buy side with it.	Bounties rise rather than being cut and the listed vendor count keeps climbing through 2027, meaning reach still beats package quality.
Financing of the supply chain scales rather than stalling at its first loss.	A lab payment default or a vendor failure large enough to pull the warehouse facility, which would push the vendor layer back onto equity and slow consolidation.
Simulation substitutes for purchased observation in physical processes but not in business ones.	A simulated business-process corpus trains as well as a real one, which is the fastest available route to the repricing scenario.

Signposts

A signpost is a reading that moves the odds if it crosses the threshold shown. Each one names the scenario or scenarios it moves, so no reading on this page floats free of the forecast it is supposed to update.

Publicly observable

Indicator	Now	Threshold	Scenario it moves	What it would mean	Checked
Listed programmes publishing a trailing term	11 of 43	20 or more	Environment economy	The environment economy is arriving on schedule; a one-off payment signed now is a worse decision than it looks	Monthly
Named environment networks	1, with 12 launch partners	A second announced by an inference or training platform	Environment economy	Aggregation is settling on platforms rather than labs; holders should negotiate with the platform layer	Company announcements
Published tasks compiled by open tooling	1,930 at v0.9.5	10,000 or more	Repricing	Automatic generation is eating the commodity end faster than forecast; loosely structured code has no floor left	Monthly
Listed vendors selling data, environments or evals	187, from 159 on 14 Oct 2026	A year-on-year fall	Repricing · Buy-side consolidation	The shake-out has started; unpackaged supply is about to get much harder to place	Monthly
Largest published introduction bounty	$85,000, from “up to $100,000” in mid-Oct 2026	Any of the five largest cut or withdrawn	Repricing	The buy side has stopped paying for reach; package quality is now the filter	Monthly
Programmes that are invite-only rather than open	2 of 43	6 or more	Buy-side consolidation	A gate is forming; which route a holder takes starts to matter more than which destination	Monthly
Programmes routing records through a named third-party de-identifier	1 of 43	5 or more	Buy-side consolidation · Rights chill	Provenance is becoming a precondition rather than a feature, and the layer that supplies it takes margin	Quarterly
Open roles sourcing, licensing or acquiring data	11 at 8 companies	Below 6, or above 20	Repricing · Selective market	Below six reads as repricing; above twenty reads as the buy side still widening	Monthly
A lab's public position on third-party archives	No named lab has stood down	A named lab stating it has stopped buying	Repricing	In-housing has crossed from the commodity end into the specialist one; the repricing odds roughly double	Lab publications and filings
Seller-confirmed payments at a published ceiling	0 of 9 ceilings	Any seller confirming $1M or more	Selective market · Environment economy	The posted-price gap is real rather than advertising, and holders should anchor higher	Named press and filings
Filed repricings of data licences	1 (a 2026 amendment)	A second one in a US filing	Repricing	Renewal risk is systematic rather than idiosyncratic; term and termination matter more than the headline fee	SEC EDGAR
Rules on employee records	None specific	A US state statute or federal bill on licensing employee records for AI training	Rights chill	The rights-chill tail is fattening; exclusion schedules stop being optional	Congress.gov, state legislatures
Credit facilities for data vendors	$10m+ warehouse facility, $65m+ in signed letters of intent	A facility of $100M or more	Buy-side consolidation · Environment economy	The capital loop is live; expect consolidation to accelerate and payment terms to lengthen	Company announcements
A public price feed or index for training data	None	Any venue publishing live or indexed rates	Selective market · Repricing	Price discovery has arrived, dispersion compresses, and the advantage shifts from information to preparation	Quarterly
Embodied or field-operations buying programmes	None on the register	Any programme paying companies for robot, sensor or field records	Environment economy	The next modality has opened; operators of physical processes should start their inventory	Monthly

Calls

Each call is dated, carries odds and a range, and names the public source that settles it. No call on this page is settled by a social post or a single person's account. We score each one when it resolves, with the Brier score: the squared gap between our odds and what happened, where 0 is perfect. Readings that come back from processes we run re-score the calls they bear on at the next update.

Call	Resolves	Odds	Settled by
One of the ten largest published payout ceilings is raised or cut	31 Dec 2026	71% (55–85)	whosellsdata.com/buying and the programmes' own pages
A sale order on the de-identified archive in the one public court-supervised sale is entered	31 Mar 2027	78% (68–88)	Case docket, No. 25-11897 (Bankr. S.D.N.Y.)
The highest published tier on the register is still $3M+ or higher	31 Mar 2027	50% (35–65)	whosellsdata.com/buying and the programmes' own pages
A second named environment network is announced by an inference or training platform	30 Jun 2027	58% (42–73)	Company announcements
The public register lists 50 or more buying programmes	30 Jun 2027	62% (45–78)	whosellsdata.com/buying
At least one of the five largest published introduction bounties is cut or withdrawn	30 Jun 2027	63% (48–79)	The programmes' own referral pages
Twenty or more listed programmes publish revenue-share, royalty or recurring terms	31 Dec 2027	37% (22–52)	whosellsdata.com/buying
A seller publicly confirms a payment of $1,000,000 or more for its own company records	31 Dec 2027	41% (26–56)	Named press or a filing
Open tooling publishes 10,000 or more verifiable tasks compiled from repositories	31 Dec 2027	62% (46–78)	github.com/huggingface/Repo2RLEnv and its Hugging Face datasets
A second repricing of a company- or health-data licence appears in a US filing	31 Dec 2027	40% (25–55)	SEC EDGAR
A US state statute or federal bill specifically addressing the licensing of employee records for AI training is introduced	31 Dec 2027	46% (30–62)	Congress.gov, state legislatures
A named frontier lab states publicly that it has stopped buying third-party company archives	31 Dec 2027	10% (4–17)	Lab publications, filings and named press
A buying programme paying companies for robot, sensor or field-operations records appears on the register	30 Jun 2028	46% (30–62)	whosellsdata.com/buying
The listed vendor count falls year-on-year for the first time	31 Dec 2028	50% (34–67)	whosellsdata.com
A credit facility of $100M or more specifically for AI data or RL-environment vendors is announced	31 Dec 2028	38% (24–53)	Company announcements and named press
Half or more of newly listed programmes taking company records ask for a verifier, eval or environment rather than files	31 Dec 2028	30% (18–42)	whosellsdata.com/buying
A public rate feed or published index for training-data or environment prices exists	31 Dec 2028	31% (18–44)	Any venue's own published rates
A frontier lab or platform publishes an environment marketplace with listed prices	31 Dec 2028	26% (15–38)	Company announcements
Landscape: selective / environment economy / consolidation / repricing / rights chill	Jan 2028 / Jan 2029	37/25/15/11/12 · 12/33/13/24/18	Triggers and markers in Scenarios
Dominant structure: one-off / usage / revenue share / environment partnership	Jan 2028 / Jan 2029	51/14/21/14 · 22/14/32/32	whosellsdata.com/buying and the programmes' own published terms

Questions to test this

Questions a sharp reader would use to break the argument.

01If open tooling can compile verifiable environments out of a repository, what stops the same trick working on a company's ticketing system in 2028 — and if nothing does, is the scarce asset the records at all, or only the outcomes attached to them?
02Our central price claim is that linkage and outcome-linkage beat volume, and we hold it without a single published pair of comparable archives and their prices. What would a dataset that settles it look like, and who would have an incentive to publish one?
03Nine programmes publish a ceiling and none has a reported payment behind it. Is that a young market or an advertising budget? What would distinguish the two before the first renewals land in the second half of 2027?
04If the layer paid to find supply is also the layer quoting the ceiling, what is the unbiased estimate of a published range — and does any part of this market produce one today?
05In-housing caps the commodity trade but not the specialist one. Where exactly is the line in 2028, and which categories cross it first — interface replicas, code, documents, or workplace communication?
06The environment economy is better for a holder and harder to reach. What fraction of holders can realistically supply a maintained environment rather than a copy, and what happens to the rest?
07If a rights chill arrives, the wind-down trade survives and the operating trade does not. Is this therefore a market for operating companies at all, or a distress market with an operating-company marketing campaign in front of it?
08Price discovery arrived for the adjacent input in a single step. Which party in this landscape has an incentive to publish a rate, and what would have to change for that incentive to exist at all?

Sources and method

Where the direction comes from

Direction comes from Frontier Desk's standing coverage of the company-data and RL-environment landscape, which we keep as a live map of who buys, on what published terms, and what each of them changes.

The standard every claim had to clear

This revision applied one standard to every claim: a figure, a quotation or a market fact is carried only if an independent primary source states it — a filing, a court docket, a published paper, a public repository or dataset, an organisation's own page or announcement about itself, or reputable named press — and is dropped otherwise, with no intermediate framing for things we could not confirm. Collection was the public record only, with each page read and dated.

How a reading was graded

Processing graded every reading A for a record, B for a party's own published statement about itself, and C for a crowd source, and flagged conflicts rather than resolving them silently: a survey of these programmes conducted between 11 and 17 October 2026 records introduction fees of up to $100,000 per deal, while the register's largest published flat bounty read $85,000 on 4 November; we carry both readings with their dates rather than picking one.

How the readings became a forecast

Analysis turned the readings into structural shifts, a causal model, phases, scenarios and odds, and every judgment on this page carries its own confidence and the reason for it, kept visibly separate from the grade on the evidence under it.

How the odds were computed

Odds are chains of steps, each with a low, central and high value, drawn as triangular distributions and simulated together over 200,000 seeded paths by scripts/outlook-company-data-odds.mjs in this site's repository; percentages are rounded by largest remainder so each column sums to 100 without a hand edit, and the bracketed ranges come from runs with every step held at its low or high value. Where a call is a marker of a scenario its odds are held at or above that scenario's, and where a call is the scenario arriving they are held below it.

What comes back, and when

Dissemination is this page. Feedback is the signposts, each of which names the scenario it moves, the Brier-scored calls, and the readings that come back from the processes we run, which re-score the calls they bear on at the next update. Where a figure is derived rather than quoted it is marked FD estimate and its arithmetic is shown beside it.

What we dropped, and why

This revision re-tested every claim that had reached the page through practitioner commentary rather than through a document. Four categories went.

Figures that existed only as an assertion

A stated short-dated mandate to buy business data, a pair of offers said to have gone to two similar companies in the same week, an introduction fee said to have been paid, and a count of deals said to have been submitted and advanced. We looked for each on the asserting company's own site, in press releases, in filings and in named press, and found none of them; one company's own page sets out a different model entirely from the claim attributed to it.

Practitioner wording quoted as evidence

Kept only where the quote comes from a paper, official documentation or a company's own announcement.

Claims resting on a single uncorroborated post

Including the sharpest first-hand accounts of how the buy side filters intake and of where the headcount floor has moved. Both are directions we still believe, and both now rest on published eligibility terms instead, which is weaker and slower but checkable.

A count no public source states

A count of rights questions carried forward from an earlier revision. The diligence work in band 10 is now drawn from two named law-firm analyses instead.

Three readings corrected rather than dropped

The register's trailing-term count is 11 of 43 and not 5, the largest published introduction bounty is $85,000 and not $100,000, and the eval-cost benchmark ran on 98 production traces and not 100. The corrected trailing-term count is the single largest reason the structure odds moved.

Two things we still cannot verify

One platform's environment-network blog post renders only with JavaScript, so the network is carried from the same company's indexed announcement and corroborated by an independent news database. And no public source anywhere states what two comparable company archives were actually paid, which is the gap under the central price judgment on this page.

01Public register of vendors and buying programmes with their published terms and the counts by kind, maintained by Secludy AI, which states it may sell to companies it lists and that listings are free, unranked and never paid for; updated and read 4 Nov 2026: whosellsdata.com/buying; the vendor register
02“Buyer programmes for company data: eligibility, prices and payment terms of 37 organisations”, working paper, 18 Oct 2026, with each programme's page and reading date between 11 and 17 Oct 2026; read 3 Nov 2026: AIData.Broker
03Comparison of four named buying programmes, with licence, payment and estimator records; the publisher is itself an intermediary; read 3 Nov 2026: AIData.Broker guide; Data Licensing Index
04micro1 enterprise data partnership tiers of $100K–500K, $500K–3M and $3M+, a 30-employee floor, trajectory licensing at $1–2M a year, 30-day review and payment within 30 days of invoice, and the published referral terms of $50,000 or 7.5% of deal value above $700,000 with no cap; read 4 Nov 2026: micro1.ai; its referral terms; its $1B twelve-month commitment, 9 Oct 2026
05Handshake AI's published payment range of $100K to $4M+ to date, its 20-employee and three-year eligibility floor, and its statement that any payout varies with source, volume, quality and diligence; read 4 Nov 2026: joinhandshake.com
06The Baseten Environment Network, announced 23 Oct 2026: twelve RL-environment companies building environments on a customer's own data, tasks and feedback while the platform handles training and production inference. The blog post renders only with JavaScript on 4 Nov 2026 and could not be read directly; its text is carried from the search index of that page and corroborated by an independent venture news database: the announcement; an independent news record of it
07The open-source AgentEnv framework (Apache-2.0, 412 stars), which builds containerised environments into versioned images, deploys them behind a gateway, points an agent at them and scores the result, behind one published protocol; read 4 Nov 2026: agentenv-framework
08An MIT-licensed tool that records a browser's requests and an app's responses and builds a local environment from them, so an agent can practise a task without the original app; read 4 Nov 2026: rea.tools
09Open tooling that compiles verifiable RL environments with private verifiers out of a repository's merged pull requests, commit history and security advisories — 23 generators, 21 published datasets and 1,930 tasks at v0.9.5, dated 4 Nov 2026, with an oracle required to score 1 and a no-op 0 before a task publishes; read 4 Nov 2026: Repo2RLEnv
10Reinforcement learning from collective rewards, 23 Oct 2026: 147 environments and more than two million tasks, 29.8% to 37.5% on an advanced proof benchmark and 55.9% to 63.5% on a frontier-science benchmark against the same model alone, reviewing roles doubling from 13% to 26% of specialist hires, and the share of runs alerting the evaluations team falling from 76.7% to 20% and then zero; read 4 Nov 2026: the release
11“How to use decision models for evals”, 20 Oct 2026: six named decision models on 98 production traces from one agent, 0.82 cents per 100 eval runs at F1 0.79 to 3.72 cents at F1 0.87, threshold tuning against the 0.50 default and Brier-scored calibration; read 4 Nov 2026: hamel.dev
12Reinforcement-learning environment economics, 12 Jan 2026: task prices of $200 to $2,000, about $300K for a complex product replica and $20K for a website replica, the four-to-five-times exclusivity multiple, about $2,400 of compute per task, and substantially more in-housing at the labs; read 4 Nov 2026: Samaritan Research
13Leaders at one lab discussed spending more than $1B on RL environments over the next year, reported by The Information; read 3 Nov 2026: TechCrunch
14Spirit Aviation Holdings, No. 25-11897 (Bankr. S.D.N.Y.): the 14 Aug 2026 auction, the $10.0M selected and $7.5M alternate bids, the $12.5M filed notice of intent, consumer personal data excluded and de-identification borne by the buyer, a flight attendants' union's negotiated exclusion of medical, leave and union records, five adjournments and no order through Dkt. 1775, with the hearing set for 6 Nov 2026; read 4 Nov 2026: docket summary; notice of adjourned hearing, Dkt. 1695
15A flight attendants' union and more than 100 members of Congress seeking employee exclusions and limits on use in that sale, 3 Nov 2026, cited as reported rather than as read; read 3 Nov 2026: InvestigateTV
16Recursion Pharmaceuticals Form 10-Q for the quarter ended 30 Sep 2026: the Tempus data licence amendment replacing annual fees of $22.0M, $32.0M and $42.0M with $14.0M on each of three anniversaries, extending the arrangement by a year and eliminating the right to terminate for convenience; read 3 Nov 2026: SEC EDGAR
17“AI training data licensing deal statistics: 2026 report”, 6 Oct 2026: 312 agreements from Jan 2023 to Sep 2026, flat annual fees falling from 58.3% of deals to 34.8% while revenue share rose from 4.2% to 13.4%, a 2.0-year median contract term, and median annual values by buyer type; read 4 Nov 2026: AI Industry Reviews
18“Databricks & the future of the firm”, 27 Oct 2026, on a firm owning its definitions, permissions, evaluated cases, customer history and hard-won exceptions so it can change provider without losing what it has learned — and on the moat being the loop rather than the data; read 4 Nov 2026: a16z
19Invoice-backed working capital for data vendors: a 30–90 day gap between incurring cost and being paid by a lab, a $10m+ warehouse facility from an institutional credit fund, $65m+ in signed letters of intent from data and RL-environment vendors, capital in 2–4 days, and underwriting on invoices, payment history, data quality and financials; read 4 Nov 2026: the YC launch
20The full equity-for-rights structure on the record in an adjacent sector: an expanded research collaboration with $80M in cash, a $40M equity investment and recurring research funding for AI training and inference compute, in exchange for exclusive development and commercial rights to what it produces, 20 May 2026; read 4 Nov 2026: the buyer’s release; the buyer's Form 10-Q
21Live public spot prices for GPU compute per chip and region, announced 21 Oct 2026, with 30 days of history and a REST and gRPC API the documentation states requires no authentication and no account — cited only as the pricing-transparency analogy, never as a data price; read 4 Nov 2026: Nebius newsroom; the spot-price documentation
22A statement of intent dated 15 Oct 2026 to build “a single pretrained Large Physics Model – the first foundation model for collider physics”, “trained on billions of simulated proton-proton collisions”, for reconstruction, particle identification, flagging unexpected signals and evaluating future detector designs, with no funding or timeline disclosed; read 4 Nov 2026: CERN
23The robotics data pyramid — web-scale video at the base, simulation and egocentric video in the middle, and data collected on the deployment system itself at the apex, where the economics do not scale; read 4 Nov 2026: GR00T N1 whitepaper; RDT2 scaling study, arXiv
24A commercial index of more than two billion people assembled from the public web and served to agents through an app, REST API, CLI and MCP, with published plans at $200 a month for 1,000 credits and $600 for 4,000, or $0.15 to $0.20 a credit; read 4 Nov 2026: the index; its published pricing
25A free index of 30M+ US federal, state and local legal documents — statutes, regulations, ordinances and zoning codes, permits, court opinions, bills, tax forms and agency guidance — searchable without spending a credit; read 4 Nov 2026: the index
26What to know before licensing company data, 1 Oct 2026: that holding data is not the same as having the right to sell it, that some platform API terms prohibit using data taken through them to train a large language model, that a sale can waive privilege and jeopardise trade-secret protection, that employee communications about pay and conditions may be protected concerted activity, and that once data has trained a model it cannot practically be pulled back out; read 3 Nov 2026: Frankfurt Kurnit
27Five diligence questions a buyer of company data should ask — what is in the dataset, whether the seller has the right to transfer it for AI use, whether sensitive material can be removed without destroying its value, what the buyer may do with it and what it creates, and who bears the risk if it is not what the parties expected, 26 Oct 2026; read 3 Nov 2026: Nixon Peabody
28Destination terms compiled with each organisation's own wording and reading dates — Mercor, Mode, Turing Project Lazarus, Ooak Data, Replay, pre.dev, Labelbox Prism, SuperAnnotate, Appen, Telegraph Lab, Troveo, Opendatabay, DataVendor, Watt, Collinear AI, AfterQuery, Nyne AI, The General Data Company, Protege, Magma, Polyshares, SimpleClosure AssetHub, Google and Microsoft; read 4 Nov 2026: whosellsdata.com/buying; mercor.com/data; data.mode.inc; lazarus.turing.com; ooakdata.com/partner; polyshares.com
29A licensor setting a 50-employee floor, an upper bound of $4 million and an adviser referral programme whose rate and cap it does not publish, 2 Nov 2026; read 3 Nov 2026: SourceX release
30Open-role index for data acquisition, licensing and RL-environment roles, read daily across 62 public job boards covering 25 AI labs and 14 data vendors, and stating which large employers it cannot see; read 4 Nov 2026: AITraining.jobs
31Our companion outlook on the labour side of the same shift, published 26 Oct 2026: AI, jobs and income to 2035

Put your records in front of every destination that fits.

Frontier Desk is the advisor to the company that holds the records. We do the rights work first, put one description in front of every destination that fits, and compare the written indications clause by clause. Tell us what you have and what you want for the company, and the first conversation covers whether your records qualify, which route fits, and what to check in an offer.

Get your company ready What to do with your records

If you hold a portfolio of companies, the same review runs across all of them at once: which clear the published eligibility rules, what each would likely be quoted and by whom, and which rights questions would block a sale.

Frontier Desk advises the company that holds the records, and runs the process to the destinations that fit it. The rights work comes before the price.

Read the 2035 outlook All outlooks

Frontier Desk outlook, 4 November 2026. Figures are as of 4 November 2026 and were read from the sources listed above on the dates shown. Published ranges are each programme's own figures, not offers. Destinations are the buyers and platforms Frontier Desk puts your records in front of. Forecasts are probabilities, and none of this is investment, legal or tax advice. Frontier Desk is not a broker-dealer, lender or investment adviser.

The company-data and RL-environment landscape to 2028

By the end of 2028 the likeliest landscape is an environment economy, not a market in files

The landscape

How a new company-record deal is structured

Bottom line

The baseline

Vendors on the public register

Published buying programmes

Published ceilings met in public

Selected at a court-supervised auction

Every reading, with its source

Structural shifts underway

From static datasets to environments, trajectories and verifiersRunning, and accelerating

From collection to verification: the grader becomes the productRunning

Labs build and buy, and the line between them moves upBoth at once

Intermediary entry, and the shake-out that follows itPeaking

The data supply chain gets financedForming

Buyers get sophisticated, and a verification layer grows between holder and buyerHardening

New modalities open: embodied, simulated and public-webOpening

The causal model

Agent demand

Environment demand

Record and verifier value

Prices and structures

Entry and noise

Filtering and consolidation

The three arms, and the three brakes

How it evolves, 2027 and 2028

Terms move weekly, and the first court test lands

Packaging becomes the filter, and the gates start to form

The first renewals, and the first trailing structures

The shake-out in the middle

Environments, verifiers, and a price you can look up

Scenarios

Selective market

Environment economy

Buy-side consolidation

Repricing

Rights chill

Winners and losers

Holders of records

Destinations and the layers between

Commoditised or scarce

Commoditising, and by when

Keeping its scarcity value, and why

The case against this table.

What to do about it

If you hold the records

If you hold a portfolio of companies

Find out which destinations your records already qualify for.

The evidence base

The destination register

What the 43 published programmes disclose

Published payout ranges

The dispersion, as published

Spread between two public estimators

Median published floor

Median published ceiling

Payments a seller has confirmed

Median annual value of a licence, by kind of buyer

The clauses that decide what a deal is worth

The rights work, which gates everything above

Key assumptions

Signposts

Calls

Questions to test this

Sources and method

Where the direction comes from

The standard every claim had to clear

How a reading was graded

How the readings became a forecast

How the odds were computed

What comes back, and when

What we dropped, and why

Figures that existed only as an assertion

Practitioner wording quoted as evidence

Claims resting on a single uncorroborated post

A count no public source states

Three readings corrected rather than dropped

Two things we still cannot verify