Private AI servers, hosting servers, and Domain registration?

Questions
1

I want to be able to host a new DeepSeek-r1 model with over 671 billion tokens, and I want to be able to host my own email, my own storage cloud, and also register my own domain in a very private manner. However, that seems hard to accomplish. First off, I don’t have the most money you could possibly imagine. It is somewhat hard to get all of these things done in a normal manner. On top of that, I also want to be able to run uncensored models, not censored models. Therefore, the GPU server provider must be willing to provide enough flexibility for uncensored models. Adding on to that, I would also have to have enough storage for all of these things. I would need enough storage not only to host my cloud, my AI, my storage for my email and everything else, but it would have to be flexible enough for all of these. Lastly, I do need enough resources in all these servers to be a reasonable tool here. Here’s the software and things I wish to run on all of them.

Nextcloud (cloud):

Storage = 300GB or more (With 3 backups and free space with everything else)

RAM = 8GB (ish)

Cores = 4 (ish)

All ports needed for most services open

Has good performance with all optional containers in AIO. Not required but highly preferred!

IPv4 and IPv6 address

Stalwart (email):

Storage = 100GB (of emails) (Has more if optional Docker containers are available. I have not checked to be fair)

RAM = 2GB (Has more if optional Docker containers are available. I have not checked to be fair)

Cores = 2 cores (Has more if optional Docker containers are available. I have not checked to be fair)

IPv4 and IPv6 address

Needs all ports for mail and most SSL/TLS (25, 995, 993, 443, etc).

!!!This changes if it does not have Alias support. If not, this is Mailcow Dockerized. I know it works for what I need but the above and system requirements for Mailcow are still required!!!

Openwebui (AI):
Can handle DeepSeek-r1 671b tokens at a reasonable speed (Can use uncensored) (This one is a tall order so if nothing cheap, 70b will do IF cheap AND fast with the others at the same time)

Can handle Gemma 27b tokens at a reasonable speed (Can use uncensored)

Can handle Qwen3 235b tokens at a reasonable speed (Can use uncensored) (Again 32b will do depending on the price and performance)

Can handle nomic-embed-text with most tokens at a reasonable speed (Can use uncensored) (I don’t know what embedding is. I’m just trying to make it diverse)

Being able to run multiple tasks at once is not required but amazing.

Being billed once a term and not per user is also not required but amazing.

IPv4 and IPv6 address

Can have a custom webdomain for the server with SSL/TLS support

I can go into the CLI and not have to use any pre-installed interface and also set my own ports and use my own services.

Other types of models like image-gen, music-gen, voice-to-text, text-to-voice are not required but very preferred.
Allows API is not required but very nice.

Domain (Website register):
Has nice DNS support
Have tons of TLDs to pick
Allows transfers

All:
Allows private cards from privacy.com
Has good trust from users in privacy and security
Allows multi (not just one) Linux OS for the server (like Fedora, Ubuntu, Debian, others, but all have Docker support)
Is the cheapest option for the requirements. For most or all of the preferred, price can be flexible.
NOT in the 5 eyes at all but is preferably outside of the 14 eyes.
GUI interfaces are the best if possible.

I know this is a long list but does anyone know something good? Mixing providers is also OK.

2

What is your budget?

3

Depends on the benefits of what will work. I can MABYE do $80-100(ish) MAX for all combined. I can bend that just a little based on how good it is. Any idea?

4

Monthly?

5

Also combining all in to one would be good (if that would even work).

6

About, yha. Even then that is a stretch BUT if it works and it’s that good, so be it.

7

Hetzner has 1TB hosted Nextcloud instances for 4EUR a month: Managed nextcloud: cheap cloud storage

However you are ABSOLUTELY not going to find any dedicated server for <$100 that can run DeepSeek 671b as that requires over 1TB of vRAM. Going rate for a 16x GPU machine in the cloud is well over $30k a month.

If you want to build that 16x A100 80GB machine yourself the cost would be approximately $450,000.

8

WOW! Ok. Let’s keep it simple. Options:

  1. 70b Deepseek and 235b qwen3
  2. 70b Deepseek and 32b qwen3
  3. Find a cheaper service if possible. However now the rules can be changed. It just needs doable policy, but still NOT in the 5 eyes.

1 if doable
2 if both are at least better and worth while (to the most extent) when compared to 8b max. All other rules apply.
3 if it is still privacy respecting (mostly).

Also if I have to make sacrifices I will need whisper support (My spelling is disgusting).

EDIT:
Also I would like to say, it does not to be super fast for the original, 1, 2, or even 3. It just has to work to the point I am not waiting all day for a response. I can wait 15 min if I need to on the 671b but If running 70b I would like something a bit faster or I want a fast response for 32b, and 14b it HAS to be fast (no exceptions). This all counts on quality. You get you paid for I guess.

Ont top I think you know this but this is not at home. This is in the cloud for AI and email/cloud storage. I see you saying I need $450k at home, but the cloud it just as good IF it meets the requirements.

Also at this time, I need to say I want a server I can set up nextcloud in myself. Not have someone set up a docker, kuberneties(I can’t spell), or podman set for me.

Lastly have a email server?

9

If you don’t require 100% private AI hosted on premises and isolated from the world then maybe venice.ai could suit you. It uses uncensored models, costs about $150 per year, and conversations are not stored on the server but in the browser. IIRC until recently it also had DeepSeek-R1 670B in the Web UI but I think now it can be used only via API.

EDIT: Venice.ai privacy

EDIT 2: Ok, I’m not sure if DeepSeek they use is uncensored but their default model (Dolphin Mistral 24B) is definitely uncensored.

10

Thanks! I hear of them and they sound good but I did not know they where that cheap. Few questions:

  1. Is their anyway to tell if the deepseek IS uncensored? Also can I use my own models so if not I can add them?
  2. Can I download my chats, so if I wish to use them again I can?
  3. Is their anything else you would recommend that is simmaler?
  4. Is their a possibility the back end is doing something wired? If so what is their biggest controversy?
  5. Any other interesting bells and whistles they offer(like music gen or TTS/STT)?
  6. Do you know of a good email, cloud, and domain registration?
11

Regarding venice.ai:
I’m not really an expert in the AI field but I think the models can be censored at two levels:

  1. When the model is trained, some data can be omitted or removed from the model.
  2. When the user enters the prompt, it can be scanned and possibly rejected.

A quick test could be asking some ‘dreadful’ question e.g. “how to build an atomic bomb?”. None of the models at duck.ai or ChatGPT will answer that question but models in venice.ai will. But that doesn’t mean that the models will answer anything as some censorship is probably always applied to topics such as guns. On top of that, the models can inherit their creators’ biases.

The chats can be downloaded as Markdown or shared with a link, but currently they can’t be imported back.

They offer also text-to-speech, image generation, document scanning, character creation.

I can’t say anything about their backend. You always have to trust the provider as currently it doesn’t seem to be possible to send completely encrypted prompts to the models.

Finally, I don’t think they support using custom models.

An alternative to venice.ai could be hugginface.co but I haven’t used it yet so can’t comment.

I think it’s not worth to self-host an email service. You can get 1GB inbox with a custom domain at mailbox.org for 2 EUR per month. I think they use encryption at rest, but I’m not sure. Or maybe when Thundermail is available it will also be a good option as they plan to host it in EU and use Stalwart as the backed which would also encrypt the emails on the server. Anyway, email is a legacy technology, unencrypted in transit so I would not bother to host it myself.

EDIT: Regarding domains, I recently used hostinger.com. They support crypto payments using Coingate but I payed with a debit card (didn’t have the need).

12

Hi. Thanks for the replay. I will take venice as the AI option as it sounds nice. Also I already host my own email. I just want to find another VPS provider that’s more private. Anything for nextcloud?