Some links in this article are affiliate links. We earn a commission at no extra cost to you when you purchase through them. Full disclosure.

You built an AI app. Where do you host it? The answer depends on your stack, traffic, and budget. Here’s the honest comparison.

Free tier comparison

Platform	Free tier	Best for	Limitations
Vercel	Generous	Next.js, static + serverless	10s function timeout, no long-running
Render	750 hrs/mo	Docker, Python, Node	Spins down after 15 min idle
Railway	$5 credit/mo	Any stack, databases	Credit runs out fast with AI workloads
Cloudflare Workers	100K req/day	Edge functions, lightweight	10ms CPU limit (free), no GPU
Fly.io	3 shared VMs	Docker, global deploy	Limited memory on free tier

Reality check: Free tiers work for demos and prototypes. The moment you have real users calling LLM APIs, you need a paid plan. The LLM API cost will dwarf your hosting cost anyway.

Paid tier comparison (for AI apps)

Platform	Starting price	Strengths	Weaknesses
Railway	$5/mo + usage	Simplest deploy, good DX, databases included	No GPU, usage-based can surprise
Vercel	$20/mo (Pro)	Best for Next.js, edge network, fast	Serverless only, 60s timeout
Render	$7/mo	Docker support, background workers	Slower deploys, less polished DX
DigitalOcean	$6/mo (Droplet)	Full VM control, GPU Droplets available	More ops work, you manage everything
Hetzner	€4.50/mo	Cheapest VPS in EU, great performance	No managed services, DIY everything
Contabo	~$5/mo	Most RAM per dollar, global locations	No managed services, DIY everything
ScalaHosting	$29.95/mo	Managed cloud VPS, SPanel, #1 on Trustpilot	Pricier than DIY options
Fly.io	$5/mo	Global edge deploy, good for latency	Complex networking, learning curve

Which to pick

”I just want to ship” → Railway

Push to GitHub, it deploys. Add a Postgres database in one click. Set environment variables in the dashboard. Done. Full Railway deploy guide. Sign up here.

”I’m building a Next.js frontend” → Vercel

If your AI app has a web frontend built with Next.js, Vercel is the native choice. Server components, edge functions, and the best preview deploy experience.

”I need full control + cheapest” → Hetzner

A Hetzner VPS gives you a full Linux server for €4.50/month. Install whatever you want. Run Ollama for local inference. Set up vLLM for production serving. No platform limitations.

The trade-off: you manage everything. Updates, security, backups, SSL, monitoring.

”I need GPUs” → DigitalOcean or RunPod

If your AI app needs GPU inference (not just API calls), you need GPU hosting. DigitalOcean GPU Droplets or RunPod serverless are the simplest options.

The typical AI app stack

Most AI side projects follow this pattern:

Frontend (Vercel)  →  Backend API (Railway)  →  LLM API (Claude/GPT)
     ↓                      ↓
  Domain              Database (Railway Postgres)
(Cloudflare)              

Monthly cost:

• Vercel Pro: $20 (or free tier)
• Railway: $5-15
• Domain: ~$1 (amortized)
• LLM API: $5-50 (depends on usage)
• Total: $30-85/month

For the cheapest possible setup:

Static frontend (Cloudflare Pages, free)
  →  Backend (Hetzner VPS, €4.50/mo)
       →  LLM API (DeepSeek, cheapest)

Total: ~$10/month including LLM costs.

Scaling up

When your side project gets real traffic:

Traffic	Recommended setup
<100 users/day	Free tier anywhere
100-1K users/day	Railway or Render paid
1K-10K users/day	DigitalOcean or Hetzner VPS
10K+ users/day	Multiple servers, load balancer, self-hosted inference

See our deployment checklist for production readiness and cost optimization guide for managing LLM spend at scale.

Migration between platforms

Switching hosting platforms is easier than you think if you containerize:

# Dockerfile - works on Railway, Render, Fly.io, DigitalOcean, Hetzner
FROM python:3.11-slim
WORKDIR /app
COPY requirements.txt .
RUN pip install -r requirements.txt
COPY . .
CMD ["uvicorn", "main:app", "--host", "0.0.0.0", "--port", "8000"]

With a Dockerfile, you can deploy to any platform. Start on Railway for simplicity, move to Hetzner when you need to optimize costs. The migration is a config change, not a rewrite.

The hidden costs

Platform pricing pages show base costs. Here’s what they don’t highlight:

Hidden cost	Railway	Vercel	Hetzner
Bandwidth	Included	1TB free, then $40/TB	20TB free
Build minutes	Included	6000/mo free	N/A (you build)
Database	$7/mo (Postgres)	Not included	You install (free)
SSL	Free	Free	Free (Let’s Encrypt)
Custom domain	Free	Free	Free
Support	Community	Email (Pro)	Tickets

Vercel’s bandwidth overage ($40/TB) can surprise you if your AI app serves large responses. Railway includes everything in usage-based pricing. Hetzner gives you 20TB which is more than most apps will ever use.

FAQ

What’s the cheapest hosting for AI side projects?

Contabo offers the best RAM-per-dollar ratio starting at ~$5/month for 8GB RAM, which is enough to run small AI models with Ollama. Railway is the easiest to deploy to with usage-based pricing that starts free. Hetzner offers excellent value for dedicated servers.

Can I host AI models on a regular VPS?

Yes, for small models (up to 14B parameters). You need at least 8GB RAM for basic inference. For larger models or GPU-accelerated inference, you’ll need GPU-equipped servers from providers like RunPod or Vultr.

Do I need a GPU server for my AI side project?

Not necessarily. CPU inference works for small models (7-14B) with acceptable latency for personal projects. You only need GPU hosting if you’re serving multiple users simultaneously or need fast response times with larger models.

Related: Deploy AI App on Railway · AI App Deployment Checklist · Best Cloud GPU Providers · Self-Hosted AI for Enterprise · Best Domain Registrars

💰 Best value for AI hosting: Contabo gives you the most RAM and storage per dollar — ideal for running Ollama or self-hosted models. Starting at ~$5/mo for 8GB RAM. Check Contabo plans →

Need GPUs? Vultr and RunPod are better for GPU inference workloads.