Kaggle Notebooks: free weekly GPU quota, including 2x T4

GoogleAbout 30 GPU hours a week, free

Cost: Free. No card needed for the GPU quota.
How long: Ongoing. The quota resets weekly.
Last verified: Nov 9, 2026 · kaggle.com

What you get

A weekly GPU quota that Kaggle says is "30 hours or sometimes higher depending on demand and resources". It resets every week.
A choice of accelerator per notebook: 1x NVIDIA Tesla P100 (4 CPU cores, 29 GB RAM) or 2x NVIDIA Tesla T4 ("T4 x2", 4 CPU cores, 29 GB RAM).
A free TPU v3-8 option as well, also limited by availability.
A ready-made Python/R environment with GPU libraries installed, 20 GB of saved output space (/kaggle/working) and versioned notebooks.

Limits

Each session runs for at most 12 hours on CPU or GPU, and 9 hours on TPU.
Free GPUs and TPUs are shared. At busy times you may wait in a queue.
Interactive sessions stop when idle, and idle time still counts against your quota, so stop sessions you aren't using.
The quota is per account. Kaggle doesn't publish a fixed number above 30 hours.

Who qualifies

Anyone with a Kaggle account.
Some Kaggle features need a phone-verified account (Kaggle's docs say so for LLM API quotas, for example). Kaggle may ask you to verify a phone number.
Optional: Colab Pro and Pro+ subscribers who link their accounts get 15 or 30 extra GPU hours a week on Kaggle. Kaggle calls this an experimental, limited-time promotion that can end without notice.

Claim it with an AI agent

A browser or computer-use AI agent, for example Grok Bot or Meta's Muse, can do most of the clicking and setup. You keep the steps tied to your identity, money and legal consent.

The agent can do

Open kaggle.com and the Notebooks docs, and explain the P100 vs. 2x T4 choice for your workload.
Once you're signed in, create a new notebook, set Settings > Accelerator to "GPU T4 x2" (or P100), and paste in the starter code you asked for.
Check the GPU with a first cell (for example, nvidia-smi) and show you how much of the weekly quota is left.
Write the training or inference code, save versions, and stop the session when it's done so quota isn't wasted.
If you prefer, push notebook versions with Kaggle's official API instead of keeping a browser session open.

You do yourself

Create the Kaggle account and accept Kaggle's Terms of Use yourself.
Do phone verification yourself if Kaggle asks for it. Never hand a verification code to a third party.
Sign in to your own account yourself (password, 2FA codes).
Accept any competition rules or model licenses yourself before the agent attaches that data or model.

Example prompt: "Open the official Kaggle page (kaggle.com) and check that this offer still matches what it says there. Once I've signed in, set it up for me. Stop and hand control back to me at any sign-in, phone or ID check, card field, captcha or terms acceptance."

Kaggle can suspend accounts that abuse GPU resources. Use one account and the quota it gets.

Sources

Checked Nov 9, 2026 on the official pages above. Offers change, and the official page always wins. All deals