Google Colab pricing vs renting a GPU by the hour

Google Colab costs nothing to start, and $9.99 or $49.99 a month once you want better GPUs: Colab Pro is $9.99 a month for 100 compute units, Colab Pro+ is $49.99 a month for 600, and Pay As You Go sells 100 units for $9.99 with no subscription. Those are Google's US prices on 2026-09-28.

The honest comparison with renting a GPU by the hour is not a price per hour, because Google's pricing page does not say how many units an hour each GPU uses. It is the shape of your work. A notebook you watch for an hour or two costs least on Colab. A run that has to go unattended for many hours, needs a GPU with a known amount of memory, or needs SSH is simpler on a rented GPU, and the worked example below shows it can cost less too.

Colab, Colab Pro and Pro+ prices on 2026-09-28#

Colab has four self-serve ways to pay, and three of them come with compute units:

Three rules apply to the units that come with Pay As You Go, Pro and Pro+: units expire after 90 days, they are not refundable or transferable except where the law requires it, and when your balance reaches zero you drop back to the free tier's limits until you buy more. Google's AI plans now include Colab units as well, added to the same balance, but the Colab FAQ does not list how many each plan gives. Colab Enterprise is priced separately, as pay for what you use.

What an hour of Colab GPU costs#

A compute unit costs about $0.10 on Colab Pro and Pay As You Go, and about $0.083 on Colab Pro+ (our calculations: $9.99 / 100 = $0.0999, $49.99 / 500 = $0.09998, and $49.99 / 600 = $0.0833). What a unit buys is the part Google leaves open: the pricing page says it depends on the machine and on the GPU or TPU attached, and it lists no hourly rates. To find yours, note your balance in Colab's resource monitor, run the GPU runtime you actually use for an hour, and note the balance again.

With that rate in hand, the break-even is one division: a rented GPU's hourly price over Colab's price per unit gives the units per hour at which both cost the same. At the prices of 2026-09-27 (our calculations):

Prices checked 6 Oct 2026, 02:40 UTC

If your Colab runtime uses more units an hour than the figure in its row, the rented GPU costs less per hour. The comparison is only fair when both GPUs can run the job, so check what Colab assigned you with nvidia-smi first: Google's FAQ says the GPU types on offer vary over time.

A worked example: three overnight training runs a month#

Unattended training is where the difference is clearest. Say you run three 10-hour fine-tuning jobs a month and want each one to keep going after you close the laptop.

On Colab, only Pro+ offers that. Google says its background execution keeps code running for up to 24 hours after you close the browser, as long as you have compute units, and that users without a paid plan should not count on execution continuing in the background. So the plan is Pro+, at $49.99 a month. Its 600 units cover the 30 hours only if the runtime uses 20 units an hour or fewer (our calculation: 600 / 30). Above that you buy more units at $9.99 per 100, and Google says to expect the runtime to be terminated when the units run out.

On QuantaCloud, the same 30 hours on one RTX A6000 cost $14.40 at the 2026-09-27 price (our calculation: 30 x $0.48; today: $0.48/GPU-hr), and you pay only from launch to stop. A job that finishes after 9 hours 20 minutes costs $4.48 (our calculation: 9.333 h x $0.48), because the unused seconds come back when you stop. For this month, that is $49.99 against $14.40.

If you can keep the tab open instead, Colab Pro at $9.99 covers the same month as long as the runtime uses 3.3 units an hour or fewer (our calculation: 100 / 30). Above that rate, each further 100 units is a $9.99 Pay As You Go pack.

Money you do not spend is the last thing to compare. Colab units expire 90 days after you get them, used or not.

Where Colab is the better deal#

Colab wins whenever the session is short, you are watching it, and the job fits on the GPU it hands you. The free tier costs nothing and needs no setup, notebooks live in Google Drive and share like a Google Doc, and Colab offers TPUs, which QuantaCloud does not. For a tutorial notebook you run for an hour, or an occasional burst of GPU time, a $9.99 pack of 100 units that lasts 90 days is the simplest way to pay. Distance matters too: Google sells Colab's paid plans in 67 countries, while QuantaCloud's GPUs run only in the US, in Virginia and the Midwest, which you will feel in an interactive notebook opened from far away. What is a neocloud explains how GPU-focused clouds differ from hyperscalers.

Where renting wins#

Renting wins on run length, memory and control.

Run length comes first. A QuantaCloud instance has no idle timeout and no 12-hour ceiling: it runs until you stop it or your balance cannot cover the next hour. Checkpoint long jobs anyway: QuantaCloud has no SLA, and Colab makes no promise about resources either.

Memory comes second. You pick the exact GPU before you pay, with 48 to 141 GB per GPU and up to 8 GPUs in one VM, and the offer you launch is the GPU you get. On Colab, the paid GPUs are subject to availability, and the types on offer change over time.

Control comes third. You get SSH as ubuntu, Docker, and any port you forward over SSH. On Colab's free tier, without a positive unit balance, Google disallows SSH shells and using a web UI in place of the notebook.

The notebook workflow does not have to change. The PyTorch + Jupyter template puts JupyterLab in your browser on the rented VM, behind your QuantaCloud login.

What neither side keeps for you#

Neither side keeps your machine for you. Google says Colab VMs are deleted after sitting idle for a while and have a maximum lifetime, and stopping a QuantaCloud instance deletes its disk, with no volumes or snapshots to fall back on. The difference is where the notebook lives: Colab keeps it in Google Drive, while on a rented VM it sits on the VM's disk until you download it. Keep code in Git, and download notebooks and outputs from JupyterLab before you stop.

Moving a Colab notebook to a rented GPU#

Moving a notebook takes five steps, and only the Colab-specific lines need rewriting:

1. In Colab, download the notebook as an .ipynb file from the File menu.
2. Launch the PyTorch + Jupyter template on a GPU with enough memory. The RTX A6000, with 48 GB, is where I would start (templates docs).
3. When the instance is running, open JupyterLab with the Open Application button on the Deployments page and upload the notebook.
4. Replace the Colab-only parts: google.colab imports and drive.mount() do not exist outside Colab, so fetch data with wget or scp instead, and pull models as in downloading Hugging Face models fast.
5. When the run ends, download the notebook and outputs from JupyterLab, then stop the instance.

The first time through, how to rent a GPU for AI covers the account, credit and SSH key steps, and the pricing page has the billing rules.

My rule: stay on Colab while the work fits in a session you are watching, and move to a rented GPU the first month you would pay for Pro+ only to keep jobs running overnight.