11 November 2026 · VRAMfarm team

Google Colab vs a rented GPU: disconnects, compute units and the real cost per hour

“Runtime disconnected.” If you have seen that message three hours into a training run, with the last checkpoint an hour old, you already know Colab’s weak spot. Colab is a very good deal for short notebooks on a 16 GB T4, and it is free to start. It gets expensive in a less visible way when your jobs run long, need more than 16–24 GB of memory, or must survive you closing the laptop. Then a rented GPU, billed per minute from €0.30 an hour for a 24 GB RTX 3090, usually gives you more finished work per euro.

How Colab works in November 2026

Colab’s plans change often, so here is the model in words rather than a price list. Check Google’s Colab pricing page for current numbers.

• Free. You get a GPU when one is available, usually an NVIDIA T4 with 16 GB. Usage limits are dynamic and not published. A notebook runs for at most 12 hours, and an idle runtime is disconnected much sooner. Some days you get no GPU at all.
• Colab Pro. A monthly subscription that comes with a pool of compute units (about 100 a month for about €10–11 in November 2026). Units buy access to faster GPUs, such as the L4 (24 GB) and the A100, while they last, plus more RAM.
• Colab Pro+. More units per month and background execution: the notebook keeps running for up to 24 hours after you close the browser, as long as you have units.
• Pay As You Go. Buy units without a subscription. They do not last for ever, so a pack you buy in spring may be gone by autumn.
• Google AI plans. Google has also started bundling Colab benefits into some of its Google AI subscriptions. What each tier includes changes, so check it before you count on premium GPUs or background execution.

What matters most: units burn per hour while a runtime is connected, and the rate depends on the GPU. A T4 uses only a couple of units an hour; an A100 uses several times more, so 100 units last dozens of hours on a T4 and somewhere around a dozen or fewer on an A100. Units also burn while the GPU sits idle and you read documentation with the notebook open.

Disconnects, limits and what they cost you

A Colab session ends for several reasons, and only some of them are under your control:

• you were idle for too long (no interaction with the notebook);
• the session hit its maximum length (12 hours, or 24 with background execution on Pro+);
• you hit a usage limit, or your compute units ran out mid-run;
• on plans without background execution, you closed the tab or your laptop went to sleep;
• Google needed the GPU for someone else.

When the runtime ends, its disk is wiped: the packages you installed, the models you downloaded, and any output you did not copy out. The usual fix inside Colab is to mount Google Drive and write checkpoints there every few hundred steps. It works, but Drive is slow with many small files, and the free 15 GB fills quickly with model weights.

On a rented server the rules are different:

• It runs until you stop it (or your balance runs out). Close the laptop and the job keeps going. Over SSH or in the Jupyter terminal, start long jobs in tmux or with nohup so they do not depend on your connection.
• There is no session length. Billing is per minute from your prepaid balance, with a 15-minute minimum per launch, and the price per minute is fixed when you launch.
• Idle protection exists, but it is predictable. After 2 hours with no SSH, Jupyter or GPU load you get an email; one hour later the server saves /workspace and stops. A training run keeps the GPU busy, so it does not count as idle. If your job runs on the CPU only, for example a long download, switch the auto-stop off in Settings.
• If the host machine fails, billing stops from the moment it went missing and you are notified.

Which GPU and how much memory you get

With Colab you request a GPU type and get what is free. With a rented GPU you choose the exact card, and you see the machine’s country, driver and CUDA version before launch.

Memory is often the deciding factor. The T4 has 16 GB and is a 2018 design: slow at FP16 by today’s standards and without bfloat16, which many current training scripts expect. It handles SD 1.5, small LLMs in 4-bit and course notebooks. It struggles with Flux at full quality, cannot fit Wan 2.2 14B video (30 GB or more), and makes fine-tuning anything above a few billion parameters painful.

Typical rented options by memory:

• 24 GB: RTX 3090 from €0.30, RTX 4090 from €0.59. Several times faster than a T4 for image generation and most training.
• 32 GB: RTX 5090 from €0.76.
• 48 GB: RTX A6000 from €0.72, for 70B models in 4-bit and Wan 2.2 14B.
• 80 GB: A100 from €1.46, H100 from €3.90.

Not sure how much your model needs? The VRAM picker gives the number and the cards that fit.

Price per useful hour

A fair comparison counts the price per useful hour: time the GPU spends on your job, as opposed to setup, waiting or work you lose.

In raw numbers, Colab is cheap. If your notebook fits a T4, the free tier costs nothing, and paid units make a T4 hour cost cents. Even an A100 hour on Colab comes out at about a euro or a little more (November 2026, depending on the unit price and the A100’s unit rate), which is not more than renting an A100 by the hour. If that is all you need, stay on Colab.

What changes the sum is overhead:

• Setup every session. Each new Colab runtime starts empty: pip install, then downloading the model again. With a 10–20 GB model that can be 10–20 minutes of burning units before the first step.
• Connected but idle. Units go while the notebook is open and you think, read or fix a bug.
• Lost runs. A disconnect at 90 % of a run without a fresh checkpoint means you pay for the whole run twice.
• The wrong card. A job that takes 3 hours on a T4 often finishes in about an hour on a 24 GB RTX 4090, which has roughly three times the memory bandwidth and much faster tensor cores.

A small example. A LoRA training run that needs 2 hours on an RTX 4090 costs €1.18, plus the minutes you spend installing packages; the launch itself is not billed. You start it, close the laptop and come back to the result. Saving the folder with the dataset and the trained LoRA costs €0.03 per GB per month, so the next session starts where this one ended. The same run on a free T4 is free if it finishes, but it takes several times longer and must fit into 16 GB and a single session.

When Colab is enough

You do not need a rented GPU for:

• learning: PyTorch tutorials, course notebooks, fast.ai-style lessons;
• small notebooks: models that fit in 16 GB and runs under an hour or two;
• sharing: a notebook a colleague opens with one click;
• occasional experiments where a lost session costs you nothing but a rerun.

Kaggle notebooks are another free option with a weekly GPU quota. For many students, the two together cover everything.

When to move to a rented GPU

Move when one of these happens more than once:

• your compute units run out before the month does;
• jobs run longer than a few hours, or must run overnight without your browser;
• the model needs more than 16–24 GB: 70B LLMs, Flux at full quality, Wan 2.2 14B video, fine-tuning above 7B;
• you want to use a web interface such as ComfyUI or Open WebUI, or SSH and VS Code, which Colab restricts, at least on the free tier;
• you need a specific CUDA version, or a known environment that is still there tomorrow;
• you want to know which country the machine is in before you upload data.

Moving a notebook is quick:

1. Launch an empty PyTorch + Jupyter server on the card you need. It is usually reachable in 2–5 minutes.
2. Open Jupyter from My servers and upload your .ipynb.
3. Delete the from google.colab import drive and drive.mount(...) cells, and change /content/... paths to /workspace/....
4. Run it. !pip install and !nvidia-smi work the same way.
5. When you stop, choose Save and stop to keep /workspace for the next session, or download your results first.

For SSH access from your own terminal or VS Code, add a public key once, as described in the SSH key guide. Billing details, including the minimum and traffic, are on the pricing page.