Google Colab is still the first place most people go for a free GPU, but the rules have shifted a lot. Google doesn’t publish the free quota, scheduled notebooks were switched off in March 2026, the free student year closed to new sign-ups, and since September 2026 Google AI subscriptions include Colab compute units. So how much free GPU time do you actually get, what does a paid plan buy, and when is Kaggle or a cheap cloud GPU better? Here’s the full picture for November 2026, with the numbers I’d use to plan a project.
- Colab Free: usually an NVIDIA T4 (16 GB) when available, sessions up to 12 hours, no published quota. Heavy users get cut off for a day or more. Web UIs like ComfyUI or Gradio tunnels need a paid balance.
- Paid: Pay As You Go $9.99 per 100 compute units, Colab Pro $9.99/month (100 CU), Pro+ $49.99/month (600 CU, background runs up to 24 h).
- What 100 CU buys (measured): about 80-90 h of T4, 60 h of L4, 18 h of A100 40 GB, 11 h of G4 (RTX PRO 6000 96 GB), 5-6 h of H100.
- Google AI plans now include CU: AI Plus 50, AI Pro 200, Ultra 1,000 or 2,000 a month (from 22 Sept 2026).
- Best free alternative: Kaggle (about 30 GPU hours a week, T4 x2 or P100, 12 h sessions).
What Colab Free Gives You
Google describes free GPU access as “very limited” and deliberately doesn’t publish a fixed quota. In practice:
- GPU: an NVIDIA T4 with 16 GB VRAM when one is free. At busy times you may only get a CPU runtime.
- Session length: up to 12 hours, often less. Idle notebooks time out much sooner, and you’ll see an “Are you still there?” prompt.
- Quota: dynamic. If you use the GPU a lot, Colab blocks GPU runtimes for a while (commonly a day or more) with a “You cannot currently connect to a GPU due to usage limits” message.
- Restrictions: SSH, remote desktops and hosting web UIs (for example a ComfyUI or Stable Diffusion interface through a tunnel) require a paid compute balance.
- Storage: the runtime disk is wiped when the session ends. Save to Google Drive.
A T4 is an older Turing card: fine for learning PyTorch, small LLMs, LoRA fine-tunes of small models and SD 1.5/SDXL-class image models, but slow for big modern image and video models, and it has no BF16 support.
Colab Plans and Prices
| Plan | Price (US) | Compute units | Notable extras |
|---|---|---|---|
| Colab Free | $0 | None | T4 when available, no web UIs |
| Pay As You Go | $9.99 / 100 CU or $49.99 / 500 CU | One-off | Access to faster GPUs while balance lasts |
| Colab Pro | $9.99 / month | 100 / month | Faster GPUs, more RAM, terminal |
| Colab Pro+ | $49.99 / month | 600 / month | Background execution up to 24 hours with the browser closed |
| Colab Enterprise | Google Cloud hourly rates | No CU | Vertex AI, team controls |
Compute units expire 90 days after purchase. The free year of Colab Pro for verified US students and faculty is closed to new sign-ups; existing holders keep it until their year ends.
How Far Compute Units Go
Google doesn’t publish burn rates either, but independent measurements from March and July 2026 give a reliable guide:
| Accelerator | VRAM | Hours per 100 CU (approx.) | Good for |
|---|---|---|---|
| T4 | 16 GB | 80-90 | Learning, small models, notebooks that run for hours |
| L4 | 24 GB | 60 | Best value: BF16, FP8, image models, 7B-14B LLM fine-tunes |
| A100 40 GB | 40 GB | 18 | Bigger fine-tunes, faster training |
| A100 80 GB | 80 GB | 13-15 | 30B-class models, long context |
| G4 (RTX PRO 6000) | 96 GB | 11 | Video models and large image pipelines |
| H100 | 80 GB | 5-6 | Short, heavy training runs |
| TPU v5e-1 | 16 GB HBM | 33 | JAX and TPU-ready code |
| TPU v6e-1 | 32 GB HBM | 25 | JAX and TPU-ready code |
My rule of thumb: on Colab Pro, the L4 gives the best balance of speed and hours for AI work. The H100 is quick but burns a month’s units in an afternoon.
Colab Units in Google AI Plans
Since 22 September 2026, Google AI subscriptions include Colab compute units that share the same balance as a Colab plan:
- Google AI Plus: 50 CU a month
- Google AI Pro ($19.99): 200 CU a month, double Colab Pro
- Google AI Ultra: 1,000 or 2,000 CU a month depending on tier, plus background execution
If you already pay for Gemini through AI Pro, you probably don’t need a separate Colab Pro subscription. My breakdown of those plans is in Google AI Plus vs Pro vs Ultra.
How to Get a GPU in Colab (Step by Step)
- Open colab.research.google.com and create a new notebook.
- Click Runtime > Change runtime type.
- Under Hardware accelerator, choose T4 GPU (free) or L4, A100, G4 or H100 (paid balance).
- Click Save, then Connect at the top right.
- Check the GPU with a code cell:
!nvidia-smi. You should see the card name and 15-16 GB of memory for a T4. - Mount Drive so your work survives:
from google.colab import drive; drive.mount('/content/drive'). - When finished, use Runtime > Disconnect and delete runtime so you stop using quota or units.
On paid plans, the resources panel (click the RAM/Disk indicator) shows your remaining units and the current burn rate per hour.
How to Make Free GPU Time Last
- Prepare on CPU first. Write and debug code on a CPU runtime, then switch to a GPU only for the heavy cells.
- Disconnect when done. An idle GPU runtime still counts against your quota.
- Cache downloads on Drive. Downloading a 10 GB model every session wastes GPU minutes.
- Use mixed precision. FP16 on a T4 roughly doubles throughput over FP32 for many models.
- Use quantised models. 4-bit GGUF or bitsandbytes models fit and run far better in 16 GB.
- Save checkpoints often. Free sessions can end without warning.
- Don’t run keep-alive scripts. They break Colab’s rules and can get GPU access restricted.
Can You Run Ollama, ComfyUI or Fine-Tuning?
| Workload | Colab Free (T4) | Colab paid (L4/A100) |
|---|---|---|
| Ollama / llama.cpp from a notebook cell | Yes, 7B-14B Q4 models | Yes, up to 70B Q4 on A100 80 GB |
| ComfyUI or other web UIs via tunnel | Not allowed on free | Allowed with a paid balance |
| LoRA fine-tune (Unsloth, PEFT) | Small models (up to about 8B in 4-bit) | Up to 30B-class depending on VRAM |
| Image generation scripts (diffusers) | SDXL-class, slow on newer large models | Modern image models comfortably |
| Video generation | Impractical | Possible on A100 80 GB or G4 |
For running a UI like ComfyUI for hours, a rented cloud GPU is usually cheaper and simpler than Colab. See the cheapest cloud GPU for AI for current per-hour rates.
Free and Cheap Alternatives Compared
| Service | Free GPU | Limits | Best for |
|---|---|---|---|
| Kaggle Notebooks | T4 (x2) or P100 | About 30 h/week (shown in account settings), 12 h sessions, phone verification | Most free GPU hours with a clear quota |
| Lightning AI | Free Studio | One active Studio, 4-hour restarts, 50 GB storage | VS Code-style environment |
| Modal | $30/month credit | Code-first serverless, billed beyond credit | Developers deploying scripts and APIs |
| Hugging Face ZeroGPU | Shared RTX PRO 6000 | 5 min/day free, 40 min/day with PRO ($9/month), Gradio Spaces only | Demos |
| Paperspace Gradient | Free M4000 in some workspaces | 6-hour auto-shutdown, 5 GB storage | Basic notebooks |
| SageMaker Studio Lab | – | Closed to new users | – |
Which Option Should You Use?
- Learning or a class project: Colab Free, with Kaggle as a backup when Colab blocks you.
- Regular weekend AI projects: Colab Pro on L4, or AI Pro if you also want Gemini.
- Long training runs: Pro+ for background execution, or a rented GPU billed by the second.
- ComfyUI, Ollama servers or anything with a web UI: a cloud GPU host or your own GPU. If you’re shopping, my best GPU for local LLMs guide covers prices.
Troubleshooting Colab GPU Errors
“You cannot currently connect to a GPU due to usage limits”
You’ve hit the hidden free quota. Wait 12-24 hours (sometimes longer), use Kaggle meanwhile, or buy a Pay As You Go pack.
“CUDA out of memory”
The model doesn’t fit in 16 GB. Use a 4-bit model, a smaller batch size, gradient checkpointing, or restart the runtime to clear memory left by earlier cells.
Runtime disconnected after a few minutes
Idle timeout. Keep the tab active while cells run, avoid closing the laptop lid, and save outputs to Drive.
!nvidia-smi says command not found
You’re on a CPU runtime. Change runtime type to a GPU and reconnect.
My units are gone faster than expected
An idle A100 or H100 runtime still burns units. Disconnect and delete the runtime when you stop.
Frequently Asked Questions
How many hours of free GPU does Google Colab give?
Google doesn’t publish a number. Free users usually get a T4 for sessions of up to 12 hours, but heavy use triggers a temporary block, often for a day or more.
What GPU does Colab Free use?
Usually an NVIDIA T4 with 16 GB VRAM, when available. Faster GPUs like the L4, A100, G4 and H100 need compute units.
How many hours do 100 compute units last?
Roughly 80-90 hours on a T4, 60 on an L4, 18 on an A100 40 GB, 11 on a G4 and 5-6 on an H100, based on independent 2026 measurements.
Is Colab Pro worth it in 2026?
Yes if you use an L4 a few evenings a week. If you already pay for Google AI Pro, it includes 200 units a month, so a separate Colab Pro plan is usually unnecessary.
Can I run ComfyUI on Google Colab for free?
No. Colab doesn’t allow web UIs through tunnels on the free tier. You need a paid compute balance, or a cheap cloud GPU instead.
Is Kaggle better than Colab for free GPUs?
Kaggle gives a clearer weekly quota (commonly about 30 hours) and offers two T4s, so for regular free use it is often better. Colab is easier to share and integrates with Google Drive.
Do Colab compute units expire?
Yes, purchased compute units expire 90 days after purchase.
Final Thoughts
Colab Free is still a great way to learn, but don’t plan a serious project around it: the quota is hidden and changes with demand. For steady work, the best value right now is either Google AI Pro (200 units plus Gemini) or Colab Pro on an L4. Keep Kaggle bookmarked for when Colab says no, and move to a rented GPU when you need a web UI or long runs.
Related reading: how much VRAM you need for AI, how to check GPU VRAM and usage, and decentralized GPU networks if you want even cheaper hourly rates.