What AI labs want from a codebase
Public code is already in most training sets. Private, production code is scarce. It shows how real systems handle edge cases, scale and change over time. Turing's Project Lazarus acquires defunct startups' private repos, design docs, testing logs and infrastructure manifests (Turing).
- Full git history with meaningful commit messages
- Tests and test results
- Infrastructure and deployment configs
- Design docs and architecture notes
- Linked issue tickets (see selling Jira data)
A repo with years of commits shows how a system grew, broke and got fixed. A single snapshot of the final code shows far less.
Codebase price points
| Source | Figure | Notes |
|---|---|---|
| MIT Technology Review | ~$25K for one crypto startup's codebase | Workspace data pending at ~$15K |
| Forbes | $10K to $100K per SimpleClosure deal | ~100 deals in the prior year |
| HUD/DataVendor | $5K baseline per codebase, seller keeps 80% | Company's own claim |
| Turing Project Lazarus | No public prices | Buys codebases and operational data |
SpaceX's AI unit held informal talks about buying data from failed startups. No deal was reported (Bloomberg).
What raises a codebase's value
HUD says it grades codebases on five factors (HUD, company claim):
| Factor | What it means |
|---|---|
| Production history | The code ran in production with real users |
| Completeness | All services, configs and history are present |
| Runnability | A buyer can build and run it |
| Rights | You clearly own it |
| Domain specificity | It covers a niche field such as fintech or logistics |
Sunset adds that Jira tickets tied to code commits are worth more than standalone docs (Forbes).
Ownership and chain of title
- Contractor code. Confirm that every contractor and early founder assigned IP to the company in writing.
- Client code. Code written for clients under contract may belong to them.
- Third-party code. Vendored libraries and open-source components carry their own license terms. Identify them.
- Proof. SimpleClosure runs chain-of-title checks before sale (01net). Expect any serious buyer to ask for the same.
Tax treatment (sale vs. license) depends on deal terms, so consult a tax advisor. This is general information, not legal or tax advice.
How the codebase market works
Most codebase sales so far run through startup shutdown services. SimpleClosure says it has supported 350+ companies and reports AI buyers up 9x versus 2025 (01net). It has completed 150+ deals (MIT Technology Review). Sunset is a competing service (Forbes).
Some buyers acquire code directly. Turing's Project Lazarus, launched in December 2025, buys defunct startups' codebases along with their operational data (Turing). Platforms like HUD/DataVendor list codebases for buyers and take a cut. HUD says sellers keep 80% (HUD, company claim).
Code often sells for less than people expect. The public figures above sit mostly in the five-figure range. Bundling code with tickets, design docs and workspace data generally makes a stronger package than code alone.
How to prepare a repository
- Run a secrets scan across the full git history, not only the latest commit.
- Remove customer data from fixtures, seeds, logs and test databases.
- Rotate or revoke every credential that ever appeared in the repo.
- Write a short README on how to build and run it.
- List third-party and open-source components.
- Gather IP assignment agreements.
- Package related tickets, design docs and test logs.
See how to anonymize company data.
Operating vs. shutting-down companies
Codebase sales so far mostly come from shut-down startups. An operating company is usually unwilling to license its live product code. It may still license retired products or internal tools it no longer uses. If you are winding down, see the startup shutdown data checklist.
Find out what your repo may be worth: get a 60-second estimate.
Frequently asked questions
Can I sell my startup's source code to AI companies?
Yes, if you clearly own it. Shutdown services and buyers like Turing's Project Lazarus acquire codebases, often with tickets, docs and test logs.
How much is a codebase worth to AI labs?
Public examples are modest: one crypto startup's codebase sold for about $25K. SimpleClosure deals have run about $10K to $100K each.
What do buyers check before buying code?
Ownership and chain of title, whether the code ran in production, completeness, whether it builds and runs, and how specialized it is.