Some links in this guide are affiliate links. If you sign up through them we may earn a small commission, at no extra cost to you. It never affects what we recommend.
Start with the cheapest thing that works
Think of free compute as a ladder. Each rung gives you more hardware, but asks for more in return: an account, then proof that you are a student, then a written proposal, and sometimes a supervisor's letter. Most student projects never need to go past the second rung, and the people who get the larger research allocations usually show they have already outgrown the free tiers.
| Rung | Examples | What you need | Good for |
|---|---|---|---|
| 1. Free notebooks | Google Colab, Kaggle Notebooks, Lightning AI free tier | An account | Coursework, tutorials, fine-tuning small models |
| 2. Student cloud offers | Azure for Students, GitHub Student Developer Pack | Student status (school email or verification) | Hosting demos, APIs and small apps |
| 3. Research credit programmes | Google TPU Research Cloud (TRC), AWS Cloud Credit for Research, Google Cloud research credits, Lambda Research Grant | A short research proposal | Thesis and paper experiments |
| 4. National allocations | NAIRR Pilot: Advancing AI Research Resources, Isambard-AI and Zenith (AIRR) Gateway route, EuroHPC AI Factories | A proposal, usually through your supervisor or institution | Large training runs |
Free notebooks: Colab, Kaggle and Lightning AI
Google Colab. Google says Colab is free of charge to use and gives access to computing resources including GPUs and TPUs [source]. The catch is that the free tier has no fixed allowance. Colab says resources are not guaranteed and not unlimited, that usage limits sometimes fluctuate, and that it does not publish those limits because they change over time [source]. It also says the free tier prioritises people who are actively programming in a notebook, so long unattended jobs are the first thing to be cut off [source]. Save checkpoints to Google Drive and expect to be disconnected.
Kaggle Notebooks. Kaggle lets you add a free NVIDIA Tesla P100 GPU to a notebook, and also lists a two-GPU T4 option and TPU sessions [source]. Kaggle warns that free GPU availability is limited, so you may wait in a queue at busy times [source]. Its technical specifications give 12 hours of execution time for CPU and GPU sessions and 9 hours for TPU sessions, so plan runs that save their progress before the limit [source]. Kaggle also has many public datasets and notebooks you can fork, which makes it a quick way to get started.
Lightning AI. Lightning AI's pricing page offers up to 50 free GPU hours on its free tier, one free Studio that runs around the clock but must be restarted every four hours, and no credit card [source]. After that it is pay as you go.
All three are fine for learning and for fine-tuning small open models. None of them is a good place for a multi-day training run.
Student cloud offers: Azure and the GitHub Student Developer Pack
Azure for Students is for full-time university students. Microsoft says you sign up with your school email, need no credit card, get a $100 credit plus free amounts of popular services, and can renew every year while you are a student [source]. The GitHub pack describes the main offer as for students aged 18 and over, with a separate Azure for Students Starter for ages 13 to 17 [source]. Azure credit can pay for GPU virtual machines, but $100 does not last long on GPUs. It is better spent hosting a model behind an API or a demo app.
The GitHub Student Developer Pack bundles offers from many companies for verified students. It includes GitHub Copilot Student, which GitHub says comes with unlimited code completions and an allowance of AI credits, with limited chat and agent use [source]. Other offers in the pack include Heroku credit of $13 a month for 24 months and $50 of MongoDB Atlas credit [source]. These are useful for shipping a project. They are not training compute.
Hugging Face Spaces is another way to host a demo. Its documentation says you can request a free hardware upgrade if you are building a demo for a side project, and that you can apply for a community GPU grant for a Space from the Space's settings [source]. It is also listed as Hugging Face Community GPU Grants (for Spaces). Spaces on free hardware are suspended after a period without use [source].
Research credits: when you have a real project
When you have a defined research question, the bigger programmes become realistic. They want to know what you will run and what you will publish.
- Google TPU Research Cloud (TRC) gives accepted researchers Cloud TPUs at no charge and says it sends invitations on a rolling basis [source]. Participants are expected to share their work through publications, open-source code, blog posts or similar, and to give Google feedback [source]. The TPUs are free but other Google Cloud services are not, so you will pay small amounts for the virtual machines and storage that drive them [source].
- AWS Cloud Credit for Research is open to graduate, postgraduate and PhD students at accredited research institutions. AWS caps student awards at $5,000 [source].
- Google Cloud research credits are limited to faculty, PhD students and postdoctoral researchers at accredited institutions in approved countries [source]. Undergraduates and master's students are not eligible on their own.
- Lambda Research Grant: Lambda offers qualifying researchers up to $5,000 in cloud credits [source].
- Model API credits are covered by OpenAI Researcher Access Program (Up to $1,000 in API credits, valid 12 months) and Anthropic AI for Science program (Up to $20,000 in API credits over 6 months). These pay for calls to hosted models, not GPUs.
Our guide on how to write a compute-credits application covers what to put in the proposal. The short version: name the model sizes, estimate the hours, and say what you will release.
National allocations: NAIRR, Isambard-AI and EuroHPC
Public research computing is the largest free source of GPUs, but it usually goes through a supervisor.
The US NAIRR Pilot: Advancing AI Research Resources accepts proposals from US-based researchers and educators at US institutions. Graduate students can apply with a support letter from a faculty advisor [source]. Awards last 12 months, and the call stays open until the pilot ends or the resources are committed [source]. NAIRR also says some resources, such as pre-trained models, datasets and platforms, can be used without a proposal [source].
In the UK, the Isambard-AI and Zenith (AIRR) Gateway route offers Up to 10,000 GPU hours on the Isambard-AI and Zenith supercomputers, used within three months; no funding. In Europe, EuroHPC AI Factories Playground Access is the lightest-touch way into EuroHPC's AI Factories. We compare these in free GPU and supercomputer access in the UK and Europe.
Making free compute go further
- Checkpoint everything. Colab and Kaggle sessions end, by design. Write a checkpoint every few minutes to Drive or a Kaggle dataset, so a cut-off costs you minutes rather than a day.
- Use smaller models first. Debug the training loop on a tiny model and a small slice of data, then scale up. Most wasted hours are spent on bugs that a small run would have caught.
- Stack, do not stretch. Prototype on a free notebook, run the real experiment on research credits, and host the demo on student cloud. Each tool is good at one job.
- Watch expiry dates. Credits expire. Apply when you are ready to use them, not months before.
- Ask your department. Many universities have their own GPU clusters or cloud agreements. They are often easier to get than any external programme.
If you do end up paying, renting a GPU by the hour from a specialist cloud such as Runpod or Vast.ai is often cheaper than the big clouds for short jobs. We compare prices in the cheapest GPU clouds for AI research. Both also run startup credit programmes, Runpod Startup Program and Vast.ai Startup Program (GPU credit matching), if your project turns into a company.
Beyond compute
Free compute pairs well with funded programmes that also give you time. Many AI safety programmes include a compute budget on top of a stipend, and some competitions give out credits as prizes. See AI funding for students for fellowships, scholarships and competitions, and the student's guide to AI competitions for contests where you can test what you have built.
