What is DataVendor by HUD, and who pays for your data?
DataVendor is a marketplace run by HUD (Human Union Data, Inc., a San Francisco startup) where teams license private codebases, RL tasksets with their environments, and raw data to AI labs. It opened a public beta on 22 Oct 2026. Vendors set their own prices; DataVendor keeps 20% of earnings, charged only when you withdraw to a bank. Selling is self-serve, while buyers join a waitlist. No completed sale prices or buyer names are public yet.
AI labs are short of one thing that scraping can't supply: private, realistic work. That means codebases with years of history, tasks an agent can be graded on, and records of how real businesses run. DataVendor is HUD's attempt to turn that into a market.
I read the full DataVendor documentation as someone who might sell on it. This is what it says, and what it leaves out. One naming note before we start: this HUD is Human Union Data, Inc., an AI agent evaluation company, not the U.S. Department of Housing and Urban Development. Search results mix the two up.
What DataVendor is
DataVendor is a marketplace for private AI training and evaluation data. Vendors list assets they already hold. Buyers, who are mostly AI labs, either buy a listing or post a request for something the catalog doesn't have.
It runs on HUD's own evaluation platform. The tasksets and environments sold there are the same objects HUD uses to run agent evaluations, so a buyer can run what they bought straight away.
HUD is a Y Combinator Winter 2025 company. It announced a $16M Series A led by Standard Capital in June 2026, according to a third-party review that tracks data-licensing platforms.
What you can sell
The docs name four kinds of supply:
| Asset | What it is | Notes |
|---|---|---|
| Codebases | Private repositories with their Git history | They want READMEs, setup steps, architecture notes, issue and PR history, tests |
| Tasksets | Sets of task templates built on the HUD platform | Priced per task |
| RL environments | Runnable software environments that tasks execute against | Sold only together with a taskset that runs on them |
| Raw data | A ZIP of documents, logs, examples or enterprise records | Up to 64 GB per file |
The page on what you can sell doesn't mention agent trajectories, chat logs or human feedback data. If that's what you hold, it would fall under raw data or need to be rebuilt as a taskset, and the docs suggest booking a call to classify it.
For every asset type, the instruction is the same: remove secrets, credentials, customer information "and anything you do not have the right to share."
How a sale works
- You create an organization with a work email and choose "Sell data". You become a Tier 1 vendor immediately, with no application review. You accept four agreements: a mutual NDA, a platform access agreement, a data processing addendum and a data sharing agreement.
- You build a listing from assets in your inventory and set prices per task, per environment, per repository or per bundle, plus a sample price.
- Quality checks run. The required checks are free. They scan the listing text for credentials and personal data, check that environments build, and check that codebases arrive with readable Git history. Optional paid checks earn a "DV quality certified" badge, and certified listings rank higher.
- Buyers evaluate without seeing raw files. They see metadata, a redacted file tree, quality evidence and taskset previews. They can buy the whole listing, buy one sample (one random environment plus one task), or make an offer.
- Delivery copies the assets into the buyer's team, so nothing links the two teams. Buy-now purchases and samples stay anonymous to the vendor. An offer reveals the buyer's organization so you can choose whether to accept.
Buyers can also post an opportunity, a request with requirements and a deadline. Only Tier 2 vendors can respond, and DataVendor decides which vendors see each one.
A sale licenses your assets for commercial use rather than transferring ownership. HUD's own seller guide says each sale grants a non-exclusive licence, so you can sell the same asset again.
What it costs, and when you get paid
| Item | Amount |
|---|---|
| Platform fee | 20% of your earnings, charged only when you withdraw to a bank |
| Converting earnings to HUD credits | 1:1, no fee, irreversible |
| Buyer transaction fee | Separate, shown to the buyer at checkout |
| Minimum withdrawal | $1,000 before fees |
| Settlement | Earnings become available 14 days after payment |
| Withdrawal review | Typically within 3 business days, then Stripe's payout schedule |
| Lost card dispute | The sale is clawed back, plus a $15 fee |
| Refunds | None. All sales are final |
All of this is from the fees and limits page, which notes the settings may change. During the beta, DataVendor can also pause purchases, invoicing and payouts temporarily.
What a listing might sell for
This is the weakest part of the public record.
- HUD's own comparison page (June 2026) says codebase pricing "starts at a $5,000 baseline per codebase and increases with grading."
- The DataVendor homepage shows example listings priced from $68K to $220K. They are not labelled as sold.
- No completed sale prices are published, and no buyers are named.
So treat any number as a vendor's asking price, not a market price.
The catches a seller should know
- The required scan covers the listing text. The docs don't say it scans your files. Scrubbing secrets and personal data out of the assets is your job, and HUD says it won't accept repos with unscrubbed personal data.
- Folder names show in the redacted file tree. A project or customer name in a directory path is visible to every buyer browsing.
- Public repositories can't be estimated. DataVendor's docs say public repositories can't be run through the codebase estimator.
- You can't browse much at first. Vendor-only teams see at most six listings per asset type until they have two live listings of their own.
- Licence terms sit in agreements behind sign-in. What a buyer may do with your code, and for how long, depends on the agreement behind the listing.
- Coverage is thin. At the time of writing, I found no independent press coverage. Most of what's online repeats HUD's own pages.
How it compares
HUD's comparison page lists the main alternatives. It ranks itself first, so read it as marketing:
- Turing's Project Lazarus: $10K–$100K for legacy code, up to $15K per modern repository.
- SimpleClosure Asset Hub: for companies shutting down. It's free within their dissolution package and deals close in one to two weeks.
- Direct deals with a lab: no platform fee, but you handle diligence, negotiation and transfer yourself, with no public price benchmark.
FAQ
Is DataVendor the same as HUD the US housing department?
No. DataVendor is run by Human Union Data, Inc., a San Francisco AI evaluation startup. It has no connection to the U.S. Department of Housing and Urban Development.
How much does DataVendor charge sellers?
It keeps 20% of your earnings, but only when you withdraw them to a bank. Converting earnings to HUD platform credits is free. The minimum withdrawal is $1,000.
Can anyone sell on DataVendor?
Vendor sign-up is self-serve with a work email, and you start as a Tier 1 vendor. Buyers have to join a waitlist and are let in in batches.
Do I give up ownership of my code or data?
Not by default. A sale licenses the assets for commercial use, and HUD says the licence is non-exclusive, so you can sell the same asset again. The exact rights depend on the agreement behind each listing.
Can buyers see my data before they pay?
Not the raw files. They see metadata, a redacted file tree, quality evidence and taskset previews, and they can buy a single sample first.
Does DataVendor scan my files for secrets?
The required check scans the listing text for credentials and personal data. The documentation doesn't say it scans the files, so scrub them yourself before uploading.
Can I sell agent transcripts or human feedback data?
The documentation doesn't name them. They would most likely go in as raw data or be rebuilt as HUD tasksets, and DataVendor suggests booking a call to classify unusual assets.
I build RL environments and tasks for coding, terminal, computer-use and GPU agents. If you're a lab looking for that kind of supply, see what I can provide.
I help teams ship agents, RAG and model evaluations that hold up in production.