In short,
The first target is internal customers for a consulting dev shop — i.e. me + n first partners / founders / employees (I'll begin alone but with a goal to grow to ~3 people asap).
So website, storage w. inet-facing endpoints, security all the way (encrypt etc), and basic IT stuff for dev: e.g. private GitLab + straightforward CI/CD pipe. I figure all developers need is an endpoint to push git, and access to the automation processes.
____
Longer elaboration, FWIW (maybe not much to answer the question but I'm a bit over my head on this topic so I'd rather share too much, to get debunked where needed).
The rationale to bootstrap as self-hosted is simply that it's already paid (gear is humming as I write) and I'm “99% sure I can guarantee 99% uptime” (the full setup is split in two homes 500km apart, each behind 1Gbps symmetrical business-grade fiber, each site being in High-Availability locally— from router/firewalls to storage passing by proxies and k8s-pods).
Re certs, well in a bootstrap scenario we'd worry about those later I suppose, when there's actual demand to justify the overhead cost (first clients are likely to be extremely low-profile, looking for a "break" too). Legally, clients own their code and infra at the end of the day, we would deliver a product/tool but not host it (not an MSP; we'd typically rather work with their MSP— or recommend one —to smooth out their operation, add critical value to key pain points).
In terms of actual workloads, I'm looking at custom domain-parametized ML/DL services (think: selling very specialized neural nets that solve one problem extremely well for this one company, e.g. classifying internal docs). So I need some in-house hardware anyway to experiment when developing models (it's painful and limiting to be forced to spend at the exploratory stage, for each project).
I figure my first partners or employees would either run their stuff locally or tap into my GPUs; we'd likely pay for the big guns on AWS or GC only for actual prod training iterations (those that earn us actual money when our work is 'good enough', final drafts so to speak).
It's a bit of a maverick project¹, hoping to leverage ML in profitable ways for small operations (obviously as we accumulate clients and experience, we'd get better at nailing areas of best-value optimization). It will be a rocky road during the early phase (first two years at least). The costs can skyrocket pretty fast with ML on clouds, I'm extremely wary of such fuckups, not to mention running the ongoing risk of being cancelled or shutdown (account terminated or frozen) for whatever reasons — so better have "plan B"... which historically might have begun as plan A, i.e. self-hosting.
I've also got a long-term angle of building expertise, so I've a positive bias for on-prems: if this thing works and grows enough, we'll definitely look at running the operation in-house, because clouds cost proposition for ML is too steep to scale profitably if it's your core business, or so I hear (and before one says "but Spotify is on whatever cloud", yeah, that's not their core business, they don't sell AI but music; we're the guys who sell AI to Spotify ideally). So, in ML for now, gotta get'em GPUs and TPUs to make a mark. Thus I'd rather learn all of it (myself maybe, as a company surely) sooner rather than later. It's gonna come down to this anyway, if this maverick thing works.
1: "maverick" as of 2020, but I suspect there will be a whole market of such consulting / ML shops by 2030.