Right now I'm running everything on AWS, but the cost feels overkill for my current needs. Since my past professional background is in AWS, it was the default choice, but given that this project is currently funded through my own savings and I'm currently unemployed, I'm looking to minimize costs. Scalability isn't a concern as the user base is small.
Currently I'm using 3rd party LLM APIs, but in the future I may want to explore serving open-source models that I fine-tune specifically for language learning and Buddhist text exploration.
Does anyone have recommendations for budget friendly alternatives that would suit these needs? Ideally, something that could handle a future transition to self-hosted models if needed.
Also, if anyone has experience with applying spaced repetition or other memorization techniques to chatbot interfaces and long-form text retention, I'd love to hear your thoughts.
They have GPU-backed servers too, obviously theyre more expensive, and not ideal for your usecase maybe.
Modal is a good serverless alternative if you want to scale to zero and be able to handle spiky loads.
There have been plenty of stories just on HN about this. Not a place I'd choose to host anything important.
Edit: Not sure if I misread, or if you edited, but I see you addressed this, and yes in that case sounds like this would be a good option if that kind of price range is ok.
But you really have to do the calculus, as you buy their cheapest servers for at least a month at a time. And with fluctuating loads, a serverless approach might actually be better.
See also: replicate or runpod. But I like modal best; their SDK is pretty comprehensive and has a sane developer experience, replicate is too constrained and runpod instances were unavailable for me half the time...
VPS providers like Vultr and Digital Ocean are a little more expensive but a little easier in some cases. They’re still way cheaper than big cloud.
The sign up was smooth and easy.
netcup.eu has instances that are generally about the best value you'll find anywhere. They've been around a long time and good rep.
buyvm.net while not quite as cheap as a random lowend host (keep in mind there are lots of fly-by-night lowend hosts with crazy prices (buyvm is NOT one of those)), and harder to find VM vacancies there, is super solid, fantastic rep, great network and large transfer, if you need large block device attached they have up to 10TB "slabs" you can attach to your VM for $5/TB/mo (you won't beat that pricing anywhere-- I've had one attached for a year or so and no issues very stable. And I definitely use the space). And they have a fantastic discord community and documentation/help, and are fast to respond to issues in their Discord.
But in general lowendtalk.com offers section and lowendbox are your friend (but pay close attention to provider reviews/offer comments to avoid unreliable and/or fly-by-night hosts).
If you have a Python or Node.js or similar there are a wealth of inexpensive options. I personally really like Fly because the pricing is predictable (you effectively pay per container per month) and it's easy to get anything that runs in Docker to run there. https://fly.io/docs/about/pricing/ starts at $1.94/month for a 256MB container which may be enough if your backend is small and efficient.
Heroku is a bit more expensive but still a relatively cheap and very robust and proven option.
If you want to maximize the performance you can get for your money a dedicated server from Hetzner is great value, but you'll have to do a lot more management than you would with a PaaS like Fly or Heroku.
Don't spend too much time worrying about serving your own models. This is a MUCH harder problem - you need GPU allocations etc - and a whole lot more expensive. Fine-tuning models is rarely the right solution so I don't think it's worth assuming you'll be doing that. If you DO do that you can solve that hosting problem separately later on.
(Fly have a GPU product which is great but like all GPUs it's expensive compared to regular hosting.)
If you're not interested in sysadmin work... well, that's what you're paying extra for! I've never bothered but you can find discussions of open source PaaS options here on HN keword Dokku².
I see no reason not to buy or repurpose a dedicated laptop/Raspberry Pi/NUC and host everything on your own home internet connection while getting started. This helps separate concerns a bit; less secure would be a VM (probably ok) or even docker container (yikes!). Cloudflare Tunnel³ is a free endpoint vs. NAT port forwarding exposing your home IP.
I can't provide the voice of experience regarding self-hosted LLMs, but there have been a few HN discussions on building your own. You might find a deal on a near-dead GPU and can buy 'without graphics card' gaming machines capable of powering them from others reselling after pulling new GPUs.
⁰https://www.oracle.com/cloud/free
¹https://www.scaleway.com/en/stardust-instances
²https://hn.algolia.com/?query=dokku%20comments%3E0
³https://developers.cloudflare.com/cloudflare-one/connections...
Edit: replied to the wrong comment
https://aws.amazon.com/government-education/nonprofits/nonpr...
In terms of a budget friendly alternatives, definitely look into fly.io. We have a AWS setup, but run some compute for ephemeral use cases in fly.io.
> Also, if anyone has experience with applying spaced repetition or other memorization techniques to chatbot interfaces and long-form text retention, I'd love to hear your thoughts.
Have some experience with conversational experiences and LLMs. It all depends on what the modality of the interaction is. Is it something that users can have a long chat conversation with? Or is it more like the anki experience where "flashcards" get presented to you?
Easy, straight-forward no-frills hosting.
Could you share the app if/when it's available somewhere? I have a Buddhist monk friend moving to Thailand later this year who would be quite interested in something such as what you described.
Regarding spaced repetition, RemNote 1.17 an 1.18 has a bunch of AI features for summarizing PDFs and YouTube transcripts and turning that into Q&A for spaced repetition, and they also have a Chatbot interface in-app. Perhaps you could check some of what they are doing for inspiration [4].
--
1: https://developers.cloudflare.com/workers/languages/
Also replacing some of CF primitives with something else later can be viable, specially if you plan for it.
The ID verification seems to be a bit random, I didn't have any issues with it. I would suggest looking at how much you're actually spending per month, and then think about how much time and effort moving to another provider will be.
There's a very good chance you'll end up spending so much time moving this to another provider, you could have used that time to find work opportunities or do something else worthwhile. AWS credits are pretty easy to get, assuming you haven't gotten them before.
Description of the stack:
- The frontend is a TypeScript Next.js app, with Prisma and RDS Postgres for persistence, and S3 for storing files uploaded by users.
- On the backend I have a BullMQ worker for handling asynchronous tasks such as audio/video transcription, backed by Redis.
- A Telegram bot that allows quickly adding resources from a phone and for having practice conversations with AI.
- Everything is deployed via a Github Actions pipeline, with images hosted on ECR and other AWS resources deployed using Terraform.
Preferably the solution would allow me to reuse at least parts of the existing CI/CD pipeline and Terraform for managing resources. Some of the suggestions by commenters such as running this on a machine in my home, while may be optimal from a cost minimization perspective, compromise a bit too much on quality-of-life for me—not simply from a developer perspective, but I also live somewhat of a nomadic lifestyle and prefer to have few physical possessions.
I already tried applying for AWS Activate but was rejected due to not meeting their criteria. I do not have a company or non-profit entity formally set up for this project. It initially started as a research project purely for personal interest, to explore how LLMs can be used for memory augmentation (in general, not just for language learning). I’ve been a user of conventional SRS tools such as Anki for many years, and have been frustrated by certain UX limitations that hindered consistent long-term use. Initially, I built it as a tool for myself, with me as the sole target audience. Onboarding foreign monks in Thailand was a fortuitous occurrence that I had not initially anticipated.
One thing that might be worth looking at is moving your backend into a lamda. There are various shims that make this very easy for node based backends.
For low traffic websites (even for quite high traffic sites) this can be almost free.
Your main costs will then be whatever database you are using.
You can fit a lot of storage in there, and while you won’t get AWS-level availability or durability, the price, latency and IOPS will blow AWS out if the water, as long as you're using those IOPS and bandwidth within the same facility. Similarly, you can fit quite a bit of compute, but you’ll start needing to pay for more power if you need more than a couple servers worth.
At your scale, Hetzner is probably a better idea.
Given that this doesn't make any money, you might as well use local or self-hosted LLMs instead of cloud based LLMs.
AWS for this use-case (especially when you are not intending to make money) is indeed overkill.
We’re a nonprofit with apps that see ~1000s of users a month, but mostly in the US. I think we see good savings using Google Cloud Run and scaling down when there’s less traffic. You could probably set up AWS Fargate similarly. Modern app frameworks start quickly so cold starts aren’t terrible. Docker containers are portable if you outgrow that kind of environment and want to shift to dedicated VMs or the other way to serverless in future. I would also look at fly.io.
If you don't mind being offline when doing maintenance occasionally and being down on the rare occasion when the hardware fails (like once in a few years) until you get the replacement part, then look up "server housing". You can buy a computer, either rack mounted, or tower, maybe used, and you put it into a datacenter that manages bandwidth, power and cooling.
That would be probably the cheapest option but pick something close to where you live.
SST uses Pulumi so (per other comments) there are constructs for Hetzner among other budget hosts.
Hetzner (Pulumi) example outside of SST: https://timozander.de/blog/using-pulumi-with-hcloud/
Then you can get $5K ($4K actual) package by using a code from Mercury bank or Secret plan purchase ($200).
I created a service to easily deploy modern apps to VPS. Like Vercel but separating control plane and servers.
Supporting both CPU and GPU servers.
Similar to Coolify but much easier and cheaper to use.
The platform is called DollarDeploy and we will be launching soon on HN!
Explanation of the details:
For 33 years[1] I've hosted several huge websites from several servers in several peoples homes for less than 5 euro per month. As a professional datacenter builder and as ISP I've never found a cheaper option and that is not an opinion but a deeply and constantly researched measurement based on hundreds of servers. In 2025 we lowered the price to 3 Euro per month. By sharing a server with other customers, we can lower the cost to below 1 euro per month (redundant backup servers double this cost).
In short, you have these costs:
1) Energy cost per kWh per year from a supplier or the levelized cost of electricity (LCOE) of a solar/battery, windmill/battery or hydro/battery system.
2) ISP or datacenter cost per Mbit/s per year for the internet connection, fiber optic or copper leased line, satellite.
3) Colocation or closet/attic/barn space cost per year (with redundant servers you need several spaces)..
4) Server hardware cost per year based on the Total Cost of Ownership (over the lifetime).
5) Labour setting up and maintaining the software on the server (excluding content) per year.
6) Labour setting up and maintaining the content (excluding software and maintenance cost) per year.
7) Domain name registration, IP number registration, certificate cost.
8) Cooling cost[3].
The energy cost is highest of the world in my region (EU, The Netherlands) because of the inflating energy prices per kWh since November 2021. So you move the server to another home or datacenter with the lowest kWh price of the region (7.0 Euro cent from a hydro dam in rural Spain). Or you install solar panels and batteries (1.2 Euro cent per kWh LCOE).
With an €479 M4 Mac mini server in 2024 I lowered my server's power from an average 12 watt (a 2006 Mac mini using 10.4-30.6 watt and in period 1993-2006 a 20.2-31.3 watt server) to today (januari 2025) a 4.0-4.7 watt with an expected lifetime of 12-14 years. The server power includes the modem and router power. If you need a few terabytes disk space your cost per month will go up by around a euro per month.
In 2025 I expect the customer wil pay €37.21 for a year redundant availability of 10 TB at 1 Gbps for IP transit traffic per month, 16 GB DRAM unix server with 10 cores, 10 GPU cores and 16 neural engine (unlimited websites, email). If you would max out the server's 43 Trillion operations per second at max performance the power cost might go up 10x. Almost 3 euro per month for a server that can run an LLM continuously.....
You might hear of hosted servers below 3 euro. They usually involve some discount scheme to lure new customers in. With a minimum price of €7 for the cheapest domain name and certificate registration (yearly, worldwide), servers maintenance labour of €30 per hour these price quotes below 3 euro per month usually are not based on true cost.
[1] In the perod 1986-1992 we hosted email, FTP and Gopher on the university's unix servers with UUCP and TCP/IP. Since 1992 I hosted on the TCP/IP servers in my home, my friends home and several offices. Since 1999 we host in the EU, US, Canada with options for Brazil and Asia.
[2] M4 Mac Mini Cluster https://www.youtube.com/watch?v=GBR6pHZ68Ho
Anyone with experience or proof of a cheaper solution? Please contact me so we can bid on your knowledge.
[3] I heat my rooms with several servers, so the cooling and fan cost are $0, even in summer during heatwaves.
Yes, that is a fair speculation, but it is not really true: everyone can save a lot, not just me. I rent out this knowledge to my customers, so they can save tons as well.
You can test this out by actually renting a M4 Mac mini server for 3 Euro with me, still cheaper than most hosters who rent you a fraction of a server (VPS) at (for example) nearlyfreespeech.net or fly.io because with me you get a full-blown server with traffic and energy cost.
My customers are companies who want to move off the cloud to self-hosting. They pay my labour and save tens of thousands per year on their hyperscaler(like AWS) and datacenter bills.
The expert knowledge comes from 40 years of planning, simulating, building and operating several datacenters, servers and websites over decades and delaying hundreds of kilometers of optical fiber.
Every time my self-host solutions have proven to be cheaper than hyperscalers or datacenters (data available on request. As you see clearly, labour cost dominate so the trick is to eliminate labour costs almost entirely while still have enough redundant low power servers in the cheapest locations.
Anyone interested in a $3.19 per month server should contact us. We would be happy to put a redundant M4 Mac mini server in your location for many years. You get an almost free server, we get global coverage.
