[OutOfHere]

For the 99% of use cases, you don't actually need AWS. I would stick to open source software/services, efficient code and algorithms in the fastest languages, cheaper smaller clouds, AI APIs, and even an in-house cluster for GPUs when needed. Keep iterating. If you're still running say Java with its enormous memory requirements, you are however stuck with the big clouds.

[ceinewydd]

There’s a middle ground:

For cloud-like provisioning experiences and paying by the hour you can use DigitalOcean, Vultr, or alternatives.

Deploying GPUs? Something like CoreWeave may be optimal.

Want inexpensive costs for dedicated hosts? Hetzner, OVH.

There are many, many advantages to using AWS but equally many alternatives that are viable and which aren’t “doing it yourself” in a hardware and network sense.

None of this is to say DIY is a terrible choice either and everyone’s got an opportunity to consider time (initial setup; ongoing maint), costs, and also expertise and ability to do a good job.

[OutOfHere]

Linode is a good small cloud too in my experience. Personally, I would steer clear of Hetzner and OVH, but that still leaves various low-cost small clouds on the table. In my experience, Hetzner disables accounts without warning if you run at 100% for too long. OVH is good for alpha software that is okay with a week-long unannounced downtime and full data loss.

Due to inflation, the cost equation is rapidly changing to where lowering cloud expenses is vital.

[debian3]

I have been using OVH since 2005. Still waiting to loose my data.

Uptime wise I had 2 outages last December with one of my server with them (electrical issue). Prior to that I don’t even remember, maybe in 2018 or something. Thankfully I had no server in Strasbourg, I saw pictures of their containers datacenter there and decided to choose Roubaix.

OVH is great if you need a lot of bandwidth. Servers price are no longer as competitive as before.

[OutOfHere]

With OVH the issue is more about prolonged multi-day downtimes; this is unsuitable for most production applications. I had given OVH a fair chance.

[triblemaster]

You want to hire a 120k/year person to maintain 10 GPUs?

[OutOfHere]

Why do they need to be "maintained"? Assuming they're set up with good cooling and dust-free air, they just sit there and work. They can perhaps be slightly underclocked and undervolted too at say 85% to run for a longer number of years. Same for the memory clock and power draw of the GPU. If any replacement was needed, it's a small part of someone's job.