A training data marketplace is a platform on which data providers list datasets and model developers license them, under rules set by the platform's operator [2, 4]. The operator reviews each listing, supplies a standard contract, bills the buyer and delivers access as files, an API or a read-only share [6, 7, 8, 17]. This explainer follows a listing from publication to the end of its term, then covers intermediary rules and crawler licensing. It is research, never legal advice.
What a training data marketplace does
Markus Spiekermann's 2019 paper defines a data marketplace as "a digital platform on which data products are traded" and says such a platform must act like a neutral intermediary, open to anyone, or at least to many registered customers, who wants to sell data products [2]. A 2023 review of data acquisition for machine learning names the parties: data providers who sell, data acquirers who need data for their own machine learning applications, and sometimes data brokers between the two [4].
The operator earns from the flow: a commission per transaction, or fees for membership, listings, storage or data services [2], or time-based subscriptions, revenue shares and one-off connection fees [3]. AWS publishes its schedule: 3% of the pre-tax total contract value on public offers of data products, and 3%, 2% or 1.5% on private offers as contract value rises, with 1.5% on renewals [13].
A 2022 survey listed 180 data trading entities, studied 97 of them in depth and counted 2,015 data providers selling in a sample of nine marketplaces [3]. The 2023 review called the market "ad-hoc" and said that none of the existing data marketplaces met all four properties it set out for an ideal one: budget awareness, price transparency, multiple providers and useful information shared with buyers before purchase [4].
Why buyers cannot value data in advance
Kenneth Arrow set out the problem for information in 1962: "its value for the purchaser is not known until he has the information, but then he has in effect acquired it without cost" [1]. A 2019 paper on market design adds that data is freely replicable, that its value depends on its correlation with other data, and that the usefulness of training data is hard to verify without first applying it to a prediction task [5]. Providers are often unwilling to share a dataset before purchase, and buyers are unwilling to share their use case or models, for confidentiality [4]. The 2023 review calls the result "the biggest challenge in the marketplace": how to value data with limited information [4].
Marketplaces answer in a few ways. The common practice is to share only metadata or a few samples before purchase [4]. AWS lets a provider attach up to 10 samples of up to 50 MB to each data set, which subscribers can view and download before they subscribe [9]. Snowflake listings can offer a limited trial of 1 to 90 days, with a subset of the data or the full product for a short period [16]. Other designs move the computation to the data: AWS Clean Rooms ML lets two or more parties run machine learning models on their data without sharing it with each other [15].
How listings and prices work
A listing is the record a buyer evaluates. On AWS it holds a name, short and long descriptions, data samples, a logo and support contacts, and AWS reviews each product against its guidelines before it enters the catalog [6]. Snowflake requires dataset listings on its Marketplace to describe update frequency, geographical scope, region availability and the completeness of fields, and their documentation link to give more information about the product, such as the sources and methodology used to compile it [19].
A listing of federal contract records, for example, needs to name the reporting systems behind them, which FPDS, SAM.gov and USAspending: where contract data comes from traces. A listing of trades and quotes needs to explain fields whose rules changed on set dates, as TAQ data explained shows. MLCommons' Croissant format, released in March 2024 as an extension of schema.org, makes such descriptions readable by machines [24].
Prices are often private. In the 2022 survey, 25% of data marketplaces gave no clear public information on how their pricing works, and 57% of the entities that did publish price information supported fixed prices [3]. The 2023 review found that in most markets prices are not advertised and buyers have to contact sales [4]. On AWS, a public offer sets prices for durations of 1 to 36 months, in U.S. dollars only, paid upfront, and a provider can extend a custom offer to one buyer's account [8]. Snowflake supports usage-based plans, such as a fixed price per query billed in arrears with a maximum total charge per month, and subscriptions billed upfront [18].
The license that comes with a listing
AWS offers a standard Data Subscription Agreement as the default, and a provider may edit it or upload its own [8]. The July 2022 template grants a nonexclusive, worldwide, nontransferable license to receive, retain, use and modify the data and to create "Derived Data" [14]. It bars distributing the data, or any material subset of it, to third parties, and bars sublicensing [14]. It lists "models" among the items that count as Derived Data, has the subscriber warrant that it owns the Derived Data apart from the provider's data inside it, grants the provider no rights in that Derived Data, and requires the subscriber to remove the data within 90 days after the term ends and, if the provider instructs, to destroy other copies [14]. That is one template, and each provider can change it [8].
Snowflake's policies say each listing should state its costs, use restrictions and license grants, and require providers to hold the rights to share or sell their products [19]. What to check in those terms before signing is the subject of buying training data.
Delivery: copies, shares and revocation
Data often reaches buyers as a bulk download or through an API [3], and the cloud platforms add access in place. On AWS, a file product arrives as a copy the subscriber is entitled to, while S3 data access lets subscribers read the same objects the provider maintains [7]. Snowflake's Secure Data Sharing copies no data between accounts, shared objects are read-only, and consumers pay only for the compute they use to query them [17]. Google's BigQuery sharing creates a read-only linked dataset, and a publisher can switch off its copy, export and snapshot functions [22]. The open Delta Sharing protocol gives a recipient pre-signed file links or temporary cloud credentials [21].
An AWS provider can revoke a subscriber's access to a revision at any time, typically for compliance reasons, after which subscribers cannot export or query it [10]. In this explainer's reading, revocation stops access through the platform, while a copy already exported falls under the agreement, and a training pipeline that needs files outside the platform needs a listing whose terms and controls allow export [10, 14, 22].
Marketplaces, intermediaries and data brokers
The EU's Data Governance Act, which applies from 24 September 2023, names data marketplaces among its examples of data intermediation services [25]. Such a service goes through a notification procedure, may not use the intermediated data for other purposes, must operate through a separate legal person and must offer fair, transparent and non-discriminatory access, prices included [25]. A service that aggregates or enriches data and licenses the result, without connecting data holders with data users, falls outside the definition [25]. The European Commission's register listed 40 providers in 10 member states on 8 November 2026, by this explainer's count [28]. On 19 November 2025 the Commission proposed turning the mandatory regime into a voluntary one under the Data Act [26], and the proposal was awaiting a committee decision in Parliament on 8 November 2026 [27].
California defines a data broker as a business that knowingly collects and sells to third parties the personal information of consumers with whom it has no direct relationship [29]. Since 1 August 2026, data brokers must access the state's deletion mechanism at least once every 45 days and process consumer deletion requests [29]. The questions a buyer asks a vendor about it are set out in due diligence questions to ask a data vendor.
Licensing content for crawlers
Some sites license access to crawlers directly. Under Cloudflare's pay per crawl, a crawler either presents payment intent in its request headers or receives an HTTP 402 Payment Required response with the price, and Cloudflare acts as merchant of record [31]. In January 2026 Cloudflare acquired Human Native, a data marketplace that connects creators with developers [32]. Really Simple Licensing (RSL) 1.0, published in December 2025, lets a site publish machine-readable terms, with usage tokens such as ai-train and payment types that include a one-time purchase, a subscription and a payment per crawl [33]. An IETF Internet-Draft of September 2026 defines a vocabulary for usage preferences, and says agreements that explicitly permit a use likely apply despite them [34]. Licensing also runs outside these channels: Shutterstock reported increased demand for access to its metadata for model training in its 2025 annual report [35].
Rillor's data marketplace offers ready-made datasets to license for model training and research, with access set in a written data agreement and delivery by authenticated API, scheduled export or agreed files.
Read the full report
The full report sets out each mechanism with its source, tabulates delivery on four platforms and AWS's fees, and lists all 35 sources.