[ MSP.DD ] For investors and founders · Updated
What makes an AI startup defensible?
Short answer
Three things last when models improve: data the product generates that competitors can't copy, a position inside a customer workflow that is costly to replace, and a team that has shipped production systems in the domain. Model access, prompts and being early don't last, because every competitor can buy the same model.
Why it matters
Foundation models get better and cheaper every few months. Anything that depends only on being first to a capability is temporary, and the valuation should treat it that way.
How to check
- 01List what the company owns versus rents: data, models, integrations, distribution.
- 02Check whether each new customer makes the product better for the next one.
- 03Estimate switching cost: how long would a customer need to move to a competitor?
- 04Check revenue quality: do pilots convert to paid contracts, repeatedly?
- 05Check capital efficiency with compute counted: burn relative to growth.
Red flags
- The moat is "our prompts" or "our fine-tune" with no unique data behind it.
- Pilots that don't convert to paid contracts.
- Burn that rises faster than growth once compute is counted.
Good signs
- Pilots converting to paid contracts repeatedly.
- Data that improves the product with every customer.
- Deep integrations into systems customers can't switch off.
The numbers
- CRV: "Usage should generate data competitors can't replicate." [1]
- CRV looks for pilots converting to paid contracts repeatably, and treats many repeatable expansions as stronger than a few large deals. [1]
- CRV: "Burn relative to growth reveals more than either number alone", because compute costs aren't priced into standard software benchmarks. [1]