How the market works

Three kinds of parties. Buyers are labs, AI startups, and enterprises training or evaluating models. Vendors produce or license data as their business. Data owners generate data as a by-product of something else, a hospital or a publisher for instance, and sell it directly or through a broker.

Inside a lab, the research team that has a gap usually holds the budget. The buyer is a person who needs one specific thing their model cannot do yet. The bigger labs now have data teams and formal purchasing, which makes the process slower and the questions sharper. And often the party you sign with is a vendor buying for one lab, with a contract that lets them pass your data to that customer.

Every deal is private. There is no exchange and no list price. The same dataset can close at very different numbers depending on who is asking, how badly they need it, and whether the seller knows the going rate.

Data is bought in bulk. Nobody pays for one image or one document. They pay for a package, sized in items, hours, or gigabytes, under a license that says what they may do with it. The total is the number that matters, and it is what this site records.

What buyers pay for

In rough order of how much each one moves the price.

Rights you can prove

A clear chain of title is worth more than any other attribute. Scraped data with a shaky rights story sells at a steep discount or not at all. First-party data with consent sells at a premium.

Realism

Records of real work, captured as it happened, now beat datasets built to a spec. Labs have learned that invented tasks, however hard, do not predict what their models must do for customers.

Who made it

Data written or checked by licensed professionals, senior practitioners, or researchers costs many times what crowd work costs. The largest driver for text and code.

Scarcity

Medical imaging, legal work product, and internal company records are hard to get. Web text is not. Finance, healthcare, and law are where the money is going now.

Just beyond the model

Buyers want tasks their models fail today and could pass with training. Anything far below that bar is a commodity. Anything far above it cannot be learned from yet.

Verifiability

Data whose correctness can be checked, like code with tests or tasks with a known answer, is worth more because it can train and grade a model directly. This is why coding data was the first market to mature.

Exclusivity

Being the only buyer roughly doubles the price. Buyers ask for it because data every lab can buy gives none of them an edge. Exclusivity for a fixed term is the usual compromise.

Labeling

Annotations, expert review, rubric grades, and written rationales each add cost on top of the raw data.

Size

Bigger packages cost more in total, but far less than proportionally. Ten times the data does not fetch ten times the price.

Freshness

Recent data is worth more than old data, and a feed that keeps coming is worth more than a one-time export.

Reading the benchmarks

Every figure on this site comes from three or more deals in a segment. The headline is the median, the middle deal, so one unusually large or small contract does not drag it. Charts show the middle half, from the 25th to the 75th percentile. Vendor pages also show the full spread.

Amounts are rounded before they are stored. The step is about a tenth of the amount on a 1, 2, 5 series, so 172,000 becomes $160k to $180k and 1.7 million becomes $1.6M to $1.8M. Medians and ranges are computed on those rounded values and rounded again, so no one can read a single contract back out of a benchmark.

Almost every contract makes the price and the parties confidential. That is why this site shows only aggregates and never names a source.

A segment with fewer than three deals shows nothing. Adding a deal to a thin segment is the fastest way to make it useful, including to you.

Buyer due diligence

What a careful buyer asks before signing. Each answer becomes a warranty in the contract, backed by your indemnity, so bring the evidence and not just the answer.

• Where did each source come from, and under what consent or contract? Can you show it?
• Is any of it scraped? From where, and does the source permit it?
• How was personal data removed, and who checked the free text?
• Is there third-party content inside? Stock media, quoted articles, user uploads?
• Is any of it AI-generated, and is that disclosed?
• Is there health, biometric, payment, or children's data inside?
• Who else has licensed it, and on what terms?
• How was it labeled, by whom, with what agreement rate?
• What quality checks ran on it, and can you show the pipeline rather than describe it?
• Can we see a representative sample under an evaluation agreement first?
• What is the all-in price for the quantity we actually need?

Preparing data to sell

1. Inventory the rights. For every source, write down who created it, what consent or contract covers it, and whether onward licensing is allowed. Keep the consent records themselves. Contracts require you to produce them on request and hold them for years. If you cannot answer, the deal stalls.
2. Remove what you cannot sell. Customer and employee personal data, anything a third party gave you in confidence, and anything a contract restricts. De-identify what remains and document the method.
3. Write a datasheet. One page: what is in it, how much, how and when it was collected, what checks it passed, known gaps. Buyers use it to decide whether to take the call. It also becomes the manifest the contract points to, so it is what you are bound to deliver.
4. Cut a sample. A representative slice, small enough to share under a simple evaluation agreement. Most deals are won or lost here.
5. Set a floor. Use the benchmarks here to decide the lowest total you would accept before the first conversation.
6. Choose delivery. Parquet or JSONL for text and tables, DICOM or standard image formats for imaging, with a manifest. For sensitive data, offer access inside a controlled environment instead of a download. The contract will set a delivery deadline, and missing it lets the buyer walk without paying.

Negotiating

• Lead with the gap you fill. The buyer wants one specific capability their model lacks. Show that your data closes it and price follows.
• Anchor on comparables. Say what similar packages have gone for. That is what the benchmarks are for.
• Unbundle. Price the raw data, the labels, exclusivity, and the refresh feed as separate lines, and narrow exclusivity to a field or a term. A buyer who wants one piece pays for one. A buyer who wants everything sees the discount.
• Prefer a shorter term with renewal over a perpetual license at the same price.
• Expect an uncapped indemnity for rights claims and a cap on everything else. Push on the cap, and only warrant what you can prove.
• If the buyer is a vendor, limit sublicensing to the one end customer they name.
• For custom builds, agree acceptance criteria before price. Rejected items are the usual dispute.
• You get paid once the buyer accepts the data. Negotiate a short review window and make silence count as acceptance.
• A declined quote is information. Note the number and the reason, and add it here.