The Free Tier Ran Out Mid-Job
Colab disconnected at hour three. Kaggle hit the weekly cap on Tuesday. The free tier looked like a plan until the job needed more than a free tier offers. H
GPU Cost | 5 min read | 2026-04-07
Colab disconnected at hour three. Kaggle hit the weekly cap on Tuesday. The free tier looked like a plan until the job needed more than a free tier offers.
How it usually goes
You start on a free notebook because the workload looks small. It grows. The runtime disconnects mid-epoch, the checkpoint is gone, and you are back at zero. The real cost was not the GPU rate. It was the wasted hours.
What free tiers actually give you
- Colab free: GPU available sometimes, disconnects after 12 hours or less, no guaranteed runtime
- Kaggle free: 30 GPU hours per week, no persistent sessions, hard weekly cap
- Both: no SLA, no persistence guarantee, shared and preemptible hardware
When free tiers actually make sense
- Short experiments that finish in under two hours
- Quick inference tests where losing the session costs nothing
- Learning runs where repeating from scratch is fine
When they stop making sense
- Any training job longer than a couple of hours
- Workloads where a disconnect costs you real time to recover
- Anything with a deadline where availability matters
The real comparison
| Setup | Actual cost | Risk |
|---|---|---|
| Colab free, 6hr job | ₹0 in theory, 6hrs lost in practice | Disconnect, lose checkpoint, restart |
| RTX 4090 rental, 6hr job | ~₹600 total | Job finishes, checkpoint saved |
| Colab free, 2 failed attempts + RTX rental | 12hrs wasted + ₹600 | The free option ended up costing more |
The honest rule
Free compute is not free when a disconnect means restarting a long job. Once the workload needs more than two uninterrupted hours, the free tier is borrowing time you will pay back in restarts.
Need a run that actually finishes?
Browse stable GPU rentals with per-second billing. No weekly caps, no surprise disconnects.
Browse GPUs