guide · valuation
How to value a dead startup's source code, models, and data
Buying source code from a defunct startup is a salvage transaction: you're paying for time saved, not future cash flows. This guide walks through how to price a codebase, a trained model, and a scraped dataset when the company behind them is gone — and how to buy a startup without inheriting cap-table drama.
$ cost_to_rebuild
Start with what it would cost to rebuild the codebase from scratch at market engineering rates — then discount 60–90% for the salvage nature. If a repo would take a team 12 months and $600k to reproduce, a fair salvage bid usually lands between $30k and $120k.
$ trained_model_premium
Trained ML models with proprietary datasets carry a premium because they can't be reproduced without the original data. Value them on training cost + data acquisition cost, then discount 40–60% for as-is risk.
$ scraped_data_replaceability
Data assets are worth what it costs to re-scrape them, plus the compliance risk of the scrape method. Public data = low value. Proprietary integrations or years of longitudinal data = high value.
$ buy_source_code_checklist
Before you buy source code, verify: readable commit history, working local build, third-party license inventory, absence of embedded secrets, and a clean IP chain-of-title in the APA.
Frequently asked
What's a fair price to buy source code from a failed startup?
Most defunct SaaS codebases trade for 5–20% of the total capital raised, adjusted for how production-ready the code is and whether trained models or proprietary data are included.
Is it better to buy a startup or just buy the source code?
Asset-only deals (source code + models + data) are almost always faster and cleaner than acquiring the legal entity. You skip the cap table, the debt, and the outstanding contracts — you take the assets and go.
How do you value a trained ML model with no company behind it?
Add the compute cost of training runs to the cost of acquiring the training data, then discount 40–60% because you don't get the team that built it. If the dataset is unreproducible, discount less.
Ready to poke around the graveyard?
Browse the graveyard