DEV Community

niuniu
niuniu

Posted on

I Stopped Paying for Colab Pro and Run LLMs on My Own GPU Now — My $480/Year Setup Costs $0

Last year I spent $480 on Colab Pro because I believed the marketing: you need cloud GPUs to run serious models. Then I actually read the specs of the RTX 3060 12GB sitting in my desktop and felt stupid.

I cancelled Colab Pro, moved everything local, and after 8 months I'm not going back. Here's the honest math.

The setup that replaced Colab Pro

Component What I use Cost
Inference Ollama (ollama pull qwen2.5:7b-instruct-q4_K_M) $0
Fine-tuning Unsloth + my 3060 12GB $0
Notebooks JupyterLab locally $0
Overflow (rare) Colab free tier for T4 jobs $0

Total: $0/month vs $49.99/month. The only thing I still use Colab for is the occasional T4 session when I need CUDA memory I don't have — and the free tier covers that.

The numbers that made me switch

Colab Pro reality check (my usage, Jan–Aug 2026):

  • Compute units burned: ~1,800/month (mostly idle timeouts I forgot to kill)
  • Actual productive GPU hours: ~35/month
  • Effective cost per productive hour: $1.43
  • Times I hit the "you've used too much GPU, come back tomorrow" wall on the paid tier: 9

Nine. On a $50/month plan. That's when I started looking at what a used 3060 12GB costs: $180. Breakeven at Colab prices: under 4 months.

What actually runs on 12GB VRAM

qwen2.5:7b-instruct-q4_K_M    → 4.7GB VRAM, ~28 tok/s on 3060
qwen2.5:14b-instruct-q4_K_M   → 9.2GB VRAM, ~15 tok/s
llama3.1:8b-q4_K_M            → 5.1GB VRAM, ~31 tok/s
Fine-tune 7B with Unsloth QLoRA → fits in 11GB, ~45 min per epoch on 5k samples
Enter fullscreen mode Exit fullscreen mode
# The entire local inference API — replaces my Colab notebooks
import ollama

resp = ollama.chat(model="qwen2.5:7b-instruct-q4_K_M", messages=[
    {"role": "user", "content": "Explain gradient checkpointing in 2 sentences"}
])
print(resp["message"]["content"])
Enter fullscreen mode Exit fullscreen mode

That's it. No runtime disconnects, no "session expired", no uploading my dataset to Google's cloud.

The honest downsides

  • No A100s. If you need to train a 70B model from scratch, this doesn't work. I don't.
  • Electricity. My 3060 pulls ~170W under load. Measured: about $4/month in power for my usage.
  • Colab free tier got worse — 12-hour session caps and more aggressive idle kills since late 2025. But I only use it ~2 hours/month now, so I don't care.

The controversial part

Colab Pro is a bad deal for anyone whose workload fits on a $180 used GPU. Google is selling you convenience, and the convenience is worse than a local box for iterative work — no setup latency, no surprise disconnections, no quota anxiety.

The people who genuinely need Colab Pro are the ones training models that don't fit on consumer cards. Everyone else is paying a subscription for a machine they already own.

If you're exploring the local-AI stack (Ollama, Continue, self-hosted tooling), I put my full notes and config files here: https://ly.cyberserval.tech/iIETXiF

Are you still paying for Colab Pro, and if so — what does it do that a local 12GB card can't? I genuinely want to know if I'm missing something.

Top comments (0)