Kaggle

What is Kaggle?

Kaggle is Google's community for machine-learning practice: prize competitions, public datasets, micro-courses, and hosted notebooks. There is no Kaggle Premium.

A free account is enough to browse and run CPU notebooks. GPU and TPU unlock after phone verification.

Kaggle's own notebook docs still describe adding a free NVIDIA Tesla P100 (or, when offered, T4 x2) from the editor Settings pane; community and Google program pages in 2026 commonly quote about 30 GPU hours per week and a smaller TPU v3-8 pot, with sessions that time out (often cited around 12 hours, shorter on some competitions). Quotas reset weekly.

CPU notebooks are not on that weekly GPU meter. Busy periods put you in a queue.

Linking Colab Pro / Pro+ has at times added extra Kaggle GPU hours as a promotion — treat that as optional and check the current banner. BigQuery or AutoML calls from a notebook can bill your GCP account; that is not Kaggle charging you for the notebook itself.

A concrete walkthrough: spend one weekly GPU pot on a real baseline, not idle Settings

Pick a Getting Started competition, write a CPU notebook first, and only flip Accelerator to GPU when the training cell is ready. A P100 hour spent on pip installs and EDA is an hour you do not get back until next week.

Checkpoint to the working directory so a 12-hour (or shorter contest) timeout is a restart, not a loss. If you need more than ~30 GPU hours, that is a signal to rent a small cloud GPU or use Colab Pro extras if the promo is live — not to open five Kaggle accounts on one phone (Kaggle treats that as abuse).

For serving the trained weights, export and host elsewhere; Kaggle is the lab, not the production URL.

Key features

• Prize competitions with public leaderboards and shared winning notebooks
• Large public dataset library you can attach to a notebook in one click
• Hosted notebooks with a free P100 (and often T4 x2); TPU v3-8 as a separate quota
• About 30 GPU hours/week after phone verify, per 2026 program write-ups; CPU is uncapped by that weekly pot
• Micro-courses and Discuss forums; competition hardware can differ (e.g. RTX 6000 reserved for specific contests)
• No credit card for notebooks — GCP products you call from a notebook are a different bill

How to get started

1. Create a free Kaggle account and verify a phone number if you want an accelerator
2. Start with a Getting Started competition or a dataset you already understand
3. Open a notebook, attach the data, then turn on GPU only when training actually needs it
4. Save checkpoints; a session timeout is not a billing event, but you will lose unsaved RAM state
5. Read the top public notebooks after you have a first submission — that is the real course

Kaggle pricing

Kaggle itself is $0. September 2026: notebooks need no card.

GPU/TPU need phone verification; expect on the order of 30 GPU hours/week (P100 or T4 x2) and a separate TPU quota, resetting weekly. Session length is capped.

Competition machines can be different and sometimes internet-disabled. Colab Pro extras, when offered, are a Google promo, not a Kaggle SKU.

BigQuery/AutoML from a notebook can charge GCP.

Our take: Use Kaggle when the job is practice or a cheap weekly GPU. Do not treat 30 hours as production hosting. If you need a model up 24/7, that is Replicate, Hugging Face Endpoints, or a VM — Kaggle will time you out.

Prices verified 2026-09. Plans change often — confirm on the official site above before you buy.

Use cases

• Learning ML on messy, real competition data instead of toy CSVs
• A free weekly GPU pot for fine-tunes that do not justify a cloud invoice
• Portfolio rankings and public notebooks hiring managers can click
• Reproducing a paper's baseline on a public dataset before you rent an H100

Strengths and tradeoffs

Kaggle alternatives

• Hugging Face: The Hub is for publishing models and Spaces, not prize leaderboards.
• fast.ai: fast.ai is a structured, code-first course; Kaggle is open-ended practice.
• DeepLearning.AI: Video specializations with certificates, not a competition arena.
• Replicate: Pay-per-second hosted inference when you outgrow a 12-hour notebook.

Stay on Kaggle to learn and experiment. Move weights to Hugging Face when you want to share a model.

Pay Replicate or an Endpoint when something must stay online. Take DeepLearning.AI or fast.ai when you want a curriculum instead of a leaderboard.

FAQ

Is Kaggle free to use?

Kaggle's plans: Free; phone-verified GPU/TPU quotas, no Kaggle subscription. Pricing and free-tier limits change over time, so check the official site above for the latest details.

What is Kaggle used for?

Free competitions, datasets, and notebooks — GPU hours are a weekly quota, not a paid plan.

Who is Kaggle best for?

Kaggle is a good fit for learning ML on messy, real competition data instead of toy CSVs, or a free weekly GPU pot for fine-tunes that do not justify a cloud invoice.

What are the alternatives to Kaggle?

Commonly compared alternatives include Hugging Face, fast.ai, and DeepLearning.AI, see the comparison above for how they differ.

How much does Kaggle cost?

Kaggle is free. There is no site-wide Premium. GPU/TPU are weekly quotas after phone verification, commonly described as about 30 GPU hours/week. You only pay if you call a billed GCP service from inside a notebook.

What's the best Kaggle alternative?

Hugging Face is the most commonly recommended swap: the Hub is for publishing models and Spaces, not prize leaderboards. Which one is actually best depends on which of those tradeoffs matters more for what you're doing.

Does Kaggle charge for GPU notebooks?

No. Kaggle does not sell a notebook subscription. After you verify a phone number you get a weekly accelerator quota (commonly about 30 GPU hours). You can hit a queue. Calling BigQuery or AutoML from that notebook can still bill Google Cloud.

Is Kaggle enough to host a production model?

No. Sessions time out and GPUs are a shared weekly pot. Export the model and host it on an always-on endpoint (Replicate, Hugging Face, your own VM) if users need a URL that stays up.