HUD Opens DataVendor So Companies Can Sell Private AI Data

DataVendor, a marketplace where companies and researchers list private data, code and AI training environments for sale to AI labs, opened a public beta on 22 October 2026, the company said in a post on X. The site uses the sign-in of HUD, a San Francisco startup from Y Combinator’s Winter 2025 batch, whose own homepage sends users to DataVendor, “where AI labs buy training and evaluation data.”

“We spent the last two years building the best solution for people buying and selling data, whether you’re a researcher in a lab or a company trying to sell data,” the post says. It adds: “Data is the fundamental bottleneck for AI progress, and DataVendor will be the infra that solves it.” Both are the company’s claims.

Business Pill · DATA LABELING

A short Business Engineer explainer on data labeling, the human work behind an AI model. It teaches the general idea only and says nothing about DataVendor or HUD.

What DataVendor Says It Sells

DataVendor says it matches valuable data, code and technical expertise with AI companies that need training and evaluation supply. Its homepage puts the pitch in one line about the rest of the world’s data: “The rest is private. That’s the supply AI labs couldn’t reach.” Its supply-types page lists five kinds of supply a vendor can bring:

• RL environments. Runnable environments where AI agents are trained and evaluated, priced per environment and run on managed Harbor or HUD infrastructure, the page says.
• Tasksets. Reusable task templates for benchmarking, priced per task.
• Codebases. GitHub repositories, delivered as a self-hosted download.
• Zip bundles. Archived files of any mix, such as datasets, documents or media.
• Operational data. CRM exports, SaaS data, SOPs and process documents, listed as coming soon; the page says the HubSpot and Salesforce connectors for it are not live yet.

Source: datavendor.ai homepage, read 23 October 2026. These are example cards shown under ‘Codebases’; the page does not say whether they are live listings or completed sales.

What Was Posted on X

DataVendor’s own post announcing the public beta on 22 October 2026. It says the team spent two years building a place to buy and sell data and calls data the fundamental bottleneck for AI progress, a claim made by the company. It matters because it is the company’s first-party announcement of the launch this piece reports.

How a Listing Works

Vendors set one unit price per kind of asset, per task, per environment or per repo, multiplied by quantity, and checkout runs through Stripe, the supply-types page says. Before buying a codebase, a buyer sees a redacted file tree: “folder names are real, file names are placeholders.” Buyers take a whole listing or a sample the platform chooses, and “partial asset purchases are not supported.”

A vendor that is not ready to list can file a data proposal instead: what it could deliver on short notice if buyers show interest. The page says proposals never appear in the public catalogue and that HUD uses them for buyer matching and supply analytics. A separate estimator puts a value range on supply a company already holds, such as a GitHub codebase. We did not find a stated platform fee or take rate on the pages we read.

Who Is Behind It

HUD’s Y Combinator page lists three founders: Jay Ram, chief executive; Lorenss Martinsons, co-CEO; and Parth Patel, chief technology officer. It gives a founding year of 2025, a team size of 24 and San Francisco as the location, and says “Over 50 businesses use HUD to build RL environments, sell them to AI labs, or train their own models from them.” The page states the mission as: “Our mission is to enable a generation of data entrepreneurs.”

HUD’s homepage describes its main product as a platform for building RL environments, with a Python SDK. When we read its homepage on 23 October, its live counters showed 4,831,034 task runs, 13,764 environments built and 22,297,355 inference calls. DataVendor’s terms and privacy pages carry the HUD logo and a HUD sign-in.

Connecting the Coverage

Our Spirit Airlines piece reported on 17 September that Mercor bid $7.5 million, Google $10 million and Micro1 $12.5 million for the collapsed airline’s data cache, which included about 30 million lines of code. Our expert-pay piece reported on 17 October, citing Bloomberg, that recruiters were offering civil rights attorneys up to $400 an hour to help train AI systems. Our harness-scaling piece reported on 23 October a study in which agents trained across 18 harness setups did better than agents trained through one. What’s new: DataVendor puts codebases, tasksets and training environments on one public price list that any company can sell into.

The Structural Read

As we read it, DataVendor treats a company’s internal code and records as inventory: the supply-types page lists codebases, zip bundles and, as coming soon, CRM exports, SaaS data and process documents as things a vendor can sell.

As we read it, the buyer side is AI labs looking for training and evaluation supply, and the seller side is any company or researcher that holds it, which is the matching HUD describes in its own pages.

As we read it, the redacted file tree is the trust mechanism: a buyer sees real folder names and placeholder file names before paying, the supply-types page says, and buys the whole listing or a sample the platform chooses.

Three Implications

FOR COMPANIES HOLDING PRIVATE CODE AND RECORDS DataVendor’s estimator puts a value range on supply a company already holds, such as a GitHub codebase, and can pre-fill a listing, the supply-types page says.

FOR AI LABS BUYING DATA Listings can bundle several asset kinds, for example RL environments plus tasksets plus a repo, and are bought whole or as a sample, the page says.

FOR DATA-LABOUR PLATFORMS Our earlier pieces reported bids of up to $12.5 million for Spirit Airlines’ data cache and specialist pay of up to $400 an hour for AI training work; DataVendor adds a self-serve listing route for the data itself.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Four Intelligence Moats.

The framework lists four moats. “The Corpus Moat” is “built by pretraining, eroding as the public web is exhausted and synthetic data closes gaps.” “The Verifier Moat” is “built by RLVR on domain-specific reward signals, building into a new asset class for reasoning-heavy verticals.”

As we read it, DataVendor is trying to price the inputs to those moats: tasksets and RL environments are what reward-signal training runs on, and private codebases and records are data the public web does not contain. The framework was published in July 2026 and does not discuss DataVendor.

What Is Not Established

The prices above are example cards on DataVendor’s homepage, not reported sales, and the company has not published sales volumes or named buyers on the pages we read. We did not find a funding announcement for HUD in two web searches; job-board listings mention a raise, but we could not confirm it from the company or a news outlet, so we do not report a figure.

The post says the team spent two years building, while Y Combinator lists HUD as founded in 2025. The site blocks plain requests, so we read it through a crawler on 23 October. We did not contact the company. Nothing here is investment advice.

The Bottom Line

HUD has opened DataVendor to the public as a store where companies list private code, task sets and agent training environments for AI labs to buy, at prices the sellers set. Its homepage example cards price codebases between $68k and $220k.