Are you rolling out your product as a SaaS offering, or is this something you're planning to license to people to run on-premises? If the latter, I expect that you'll get lots of requests for source, either for inspection or for escrow.
Personally, I wouldn't be scared of BigCo ripping you off. For the most part, large companies care a lot about staying on the right side of their contracts, and also it's generally really hard for a large company to out-innovate a startup. So I would be pretty surprised to see them steal your source (assuming you put in place an appropriate NDA etc.). EDIT: I'd be even more surprised to see them try to compete; the worst likely case is that they steal the source and stop paying you, not that they steal the source and get into the data analytics / ML business themselves.
However, I think that in the ML world, this "what the hell is the algorithm doing" question is a really common one, and it'd be super-worthwhile to invest in some sort of tooling to peel back the cover of the algorithm a bit. Validation of appropriate responses against future data is a real quagmire right now, to the point that some people are using ML to help find a solution, but then trying to re-implement the logic more traditionally once the ML algorithms figure out what to design for. I think there's something there, at least for a good subset of use cases.
Also, it's common for a large enterprise to require some sort of source code escrow if they do a big deal with a startup. Sounds like this is different than what they're asking for, since the source in escrow won't be available to them until the escrow conditions are triggered. Again, I wouldn't be concerned about signing escrow agreements, but I would make it a negotiation point, rather than a standard term.
Can you elaborate a little on this? The reason I ask is that my experience has been different. I have seen a lot of requests for code escrow (source is released to big company if and only if small company goes out of business) when a large company was dependent on a third-party product, but I have almost never seen a request to review source code (the one exception being selling into classified government environments where security considerations required a source code review).
I wouldn't say it's been the norm in my experience, but I've definitely been on the receiving end of requests for source (both when working at a large enterprise and at a startup). In my experience, this is much more common when selling something that will run on-prem or in process, however.
Personally, I wouldn't be scared of BigCo ripping you off. For the most part, large companies care a lot about staying on the right side of their contracts, and also it's generally really hard for a large company to out-innovate a startup. So I would be pretty surprised to see them steal your source (assuming you put in place an appropriate NDA etc.). EDIT: I'd be even more surprised to see them try to compete; the worst likely case is that they steal the source and stop paying you, not that they steal the source and get into the data analytics / ML business themselves.
However, I think that in the ML world, this "what the hell is the algorithm doing" question is a really common one, and it'd be super-worthwhile to invest in some sort of tooling to peel back the cover of the algorithm a bit. Validation of appropriate responses against future data is a real quagmire right now, to the point that some people are using ML to help find a solution, but then trying to re-implement the logic more traditionally once the ML algorithms figure out what to design for. I think there's something there, at least for a good subset of use cases.
Also, it's common for a large enterprise to require some sort of source code escrow if they do a big deal with a startup. Sounds like this is different than what they're asking for, since the source in escrow won't be available to them until the escrow conditions are triggered. Again, I wouldn't be concerned about signing escrow agreements, but I would make it a negotiation point, rather than a standard term.