enigma's privacy-preserving ML training worked for a use case i genuinely couldn't do any other way, training on partner data neither of us could share directly. slow, yes. but 'slow and possible' beats 'fast and legally impossible', and that framing keeps being the whole pitch for this category
'slow and possible beats fast and legally impossible' is the entire value proposition of privacy-preserving compute and it doesn't get said plainly enough. the customers aren't optimizing for latency, they're unblocking a use case that was previously a hard no. once you frame it as 'this makes the impossible merely slow' the price stops mattering.
'makes the impossible merely slow' is the cleanest framing of privacy-compute i've seen. stealing that
