Model versioning, training artifacts versioning, training code versioning, training and test data versioning, providing developer environment on a gpu machine, serving a model and some more tasks are very costly for a company to implement on their own using open source tools.
A manager platform that does all that certainly is valuable and I will recommend it to any new ML team.
If these use cases don't click with you I suggest thinking through working with ML models long term on a team with a big customer base. What if you introduce new data and your model does better but then later does worse than the original? what if you tested your model and it had great performance but after 3 months in production it sucks, now you want to go back to your test data at the time and see why. what if you personal machine does not have gpus? what if you need a custom dev environment? what if different customers need your model at different versions of your training data.
I think it is rightfully a nascent niche.
I would divide the main challenges for AI startups as follows: 1. Support Enterprise Features. (Security, IAM, VPC-SC, Ecripyion) 2. If providing Compute Resources do not make that your main source of income (i.e. DeepNote, Saturn Cloud) which that may not scale. 3. Data integration. (BigQuery, S3, GCS, etc)
Databricks is one of the ones that have integrated with each of the clouds and provide this E2E workflow nicely, in addition they have seen the nascent Analytics market and invest on it. My concern with some startups: Cohere.AI similar to OpenAI GPT3, is that some are only solving some part of the ML workflow: (Seldom, OctoML, etc.) they may get some customers now, but will be hard to scale, and probably best destiny is getting acquire by major players.