Colab is cheap per hour when you can get the GPU. On the most recent dated measurement we found, a T4 on Colab Pro works out to roughly 12 cents an hour. Colab stops being the cheap option when your units expire unused, when the GPU you are paying for is not available, or when a job outlives a session. At that point a rented GPU with a fixed hourly price is the better buy, and the live table further down shows what one costs today.
When Colab stops being the cheap option
Colab's paid plans sell compute units, not hours. Google's signup page says the amount of access a unit buys "depends on the configuration of your VM and use of specialized accelerators like TPUs or GPUs". Google publishes no rate table. Its FAQ says only that limits and hardware "vary over time". The live burn rate appears in the resource panel of a connected runtime, and nowhere else.
So any dollars per hour figure for Colab is an estimate built on someone's measurement. Here is one, with the arithmetic shown.
Price per unit. The US signup page listed Colab Pro at $9.99 a month for 100 units as retrieved on 21 September 2026. That is $9.99 / 100 = about $0.10 per unit. Pay As You Go costs the same per unit. Colab Pro+ was $49.99 for 600 units, or about $0.083 per unit.
Units per hour. Chris McCormick, in a table he updated in March 2026, measured a T4 at 1.19 units per hour "through experimentation" from southern California. He adds that the rate "does seem to fluctuate a little".
Estimate. 1.19 units per hour x $0.10 per unit = about $0.12 per hour for a T4 on Colab Pro. This is our estimate from his measurement, not a Google price.
Now the same sum with an older measurement. varlog.info measured a T4 at 1.58 units per hour on 1 October 2024. That gives about $0.16 per hour. Measured rates for the L4 and A100 fell over the same period too, so treat any figure as a snapshot.
Here is McCormick's March 2026 table turned into estimates at the Pro price per unit:
| GPU on Colab | Measured units per hour (March 2026) | Estimated cost per hour at $0.10 per unit | Hours from 100 units |
|---|---|---|---|
| T4 | 1.19 | about $0.12 | about 84 |
| L4 | 1.71 | about $0.17 | about 58 |
| A100 40 GB | 5.40 | about $0.54 | about 18.5 |
| A100 80 GB | 7.52 | about $0.75 | about 13 |
Compare those estimates with the live rental prices below. Colab may well win on the raw hourly number. It loses in three other ways.
- Units expire. The signup page says units "expire after 90 days" and are "not refundable and not transferable". A subscription you use for three hours a month costs far more per hour than the table suggests.
- The GPU is not guaranteed. Google's paid terms say limits may be adjusted "at the sole discretion of Google". An open GitHub issue dated 6 August 2026 reports the H100 and G4 as unavailable, with no Google reply on the thread. An earlier issue is titled "Can never connect to 'H100' 'G4' or 'A100' GPU's". These are user reports, not measurements, but a GPU you cannot connect to has no hourly price at all.
- Sessions end. Free notebooks "can run for at most 12 hours". Pro+ supports "continuous code execution for up to 24 hours if you have sufficient compute units". "Runtimes will time out if you are idle." Google does not publish the idle timeout. The often quoted 90 minutes is not in its FAQ.
When your balance hits zero, Google's FAQ says you "revert to the free-of-charge tier policies and restrictions" until you buy more.
Google Colab pricing and plans
These are US prices from Colab's signup page on 21 September 2026. They were read from the page's embedded data, and prices differ by country, so check the signup page where you live.
| Plan | Price | Compute units | What Google says you get |
|---|---|---|---|
| Colab (free) | $0 | None | "Very limited access to GPUs", sessions of at most 12 hours, no reliable background execution |
| Pay As You Go | $9.99 for 100 units, $49.99 for 500 units | One-time purchase | Units expire after 90 days |
| Colab Pro | $9.99 per month | 100 per month | Faster GPUs, high memory machines, expanded Gemini quota |
| Colab Pro+ | $49.99 per month | 600 per month | Everything in Pro, priority access to premium GPUs, background execution for up to 24 hours |
| Google AI Pro | Sold as a Google AI plan | 200 | Units join the same balance as a Colab subscription |
| Google AI Ultra | Two tiers, $100 and $200 per month as of 19 May 2026 | 1,000 or 2,000 "depending on your plan" | Longer background execution and premium GPU access |
| Colab Enterprise | Billed per hour as Google Cloud infrastructure | No units | A Google Cloud product with its own price list |
Three notes on that table.
Pro+ changed. Older articles say Pro+ includes 500 units. The current page text reads: "An additional 500 compute units per month, totaling 600 compute units per month". Confirm it in a logged-in browser if the extra units are your reason to subscribe.
Google AI plans now include Colab units. Google's FAQ says an eligible Google AI plan "also unlocks premium Colab benefits", though not during a free trial. The pages retrieved did not say which Ultra tier gets 1,000 units and which gets 2,000.
Colab Enterprise is Google's own hourly price. It drops the unit system. On 21 September 2026 Google's Colab Enterprise price list for Iowa showed a T4 at $0.42 per hour and an H100 at about $11.76 per hour, each including a management fee, with CPU and memory billed on top. That T4 figure is about three and a half times our Colab Pro estimate above. Capacity without Colab's usage limits costs more, even from Google.
What the free tier gives
Less than most guides claim to know. Google's FAQ says GPU types "vary over time" and that Colab "does not publish these limits". Users widely report a T4 as the free Google Colab GPU, but Google names no GPU model, weekly quota or idle timeout. The signup page warns that for free users "execution will be interrupted when user interaction ceases". Without a positive unit balance, Google's FAQ also disallows SSH shells, remote desktops and using a web UI in place of the notebook.
Kaggle, the closest free alternative
Kaggle Notebooks work like Colab and are more open about their limits. Kaggle's notebook documentation, read on 21 September 2026, states: "12 hours execution time for CPU and GPU notebook sessions and 9 hours for TPU notebook sessions." GPU time comes from a weekly GPU quota that Kaggle sets and shows in your account settings (commonly reported as around 30 hours). Kaggle's docs give no number, and the quota floats.
The accelerator menu changes, so check it when you open a notebook. T4 and P100 class GPUs are what Kaggle has offered.
One more link between the two services: the same documentation says Colab Pro and Pro+ subscribers who link their account get 15 and 30 extra GPU hours per week on Kaggle respectively. Those hours do not consume Colab compute units, and Kaggle may change or end the offer.
Older lists of free GPUs also recommend Amazon SageMaker Studio Lab. AWS documentation now says it "is no longer open to new customers". Our guide to free cloud GPUs and credits covers the rest of the field.
What these GPUs rent for by the hour, with no unit system
This table shows the cheapest current price to rent each GPU on demand across the providers we track, and how many providers have it in stock now. You pick the GPU. The price is per hour and is fixed when you launch. There are no units.
The T4, L4 and A100 are the cards Colab users know. The RTX 4090 is here because it is the common step up from a T4 for people who leave Colab. It has 24 GB of memory against 16 GB on the T4.
The T4 itself is a Turing card from 2018. The spec table shows how far the newer cards have moved on.
| Spec | T4 | L4 | RTX 4090 | A100 | H100 |
|---|---|---|---|---|---|
| VRAM | 16 GB | 24 GB | 24 GB | 40 to 80 GB | 80 to 94 GB |
| Memory bandwidth | 320 GB/s | 300 GB/s | 1,008 GB/s | 2,039 GB/s | 3,350 GB/s |
| FP16 (dense) | 65 TFLOPS | 121 TFLOPS | 165.2 TFLOPS | 312 TFLOPS | 989 TFLOPS |
| FP16 (with sparsity) | Not published | 242 TFLOPS | 330.4 TFLOPS | 624 TFLOPS | 1,979 TFLOPS |
| Architecture | Turing | Ada Lovelace | Ada Lovelace | Ampere | Hopper |
| Launch year | 2018 | 2023 | 2022 | 2020 | 2022 |
Check who is behind the cheapest row before you rent. The cheapest offer may come from a marketplace where an independent host owns the machine. Vast.ai's own documentation says "individual hosts may have less formal security measures". That is fine for a public dataset and a hobby model. It is not fine for client data. The providers page shows each provider's prices and stock.
To see what steady use would cost, this table turns the live hourly price into daily and monthly figures at 730 hours.
Few Colab users run a GPU all month. Multiply the hourly price by your own hours and set the result against a Colab subscription.
What you give up
Zero setup. Colab opens in a browser tab with Python, CUDA and the common libraries already installed. A rented GPU is a Linux machine. You choose an image, wait for it to boot, and install what is missing.
Google Drive. Colab notebooks live in Drive and a notebook can mount Drive as a folder. On a rented GPU you move data yourself and keep code in git. Some providers charge to move data out, and our data egress reference lists them.
Billing that stops on its own. Colab times out an idle runtime for you. A rented instance bills until you stop it. Storage you keep is usually billed after that too. Set a reminder or a shutdown script on your first day.
The Gemini quota that comes with the paid plans.
What you gain: a GPU model you chose, sessions with no time limit, SSH, background jobs, and a price you can read before you start.
Getting a notebook onto a rented GPU
There are two general routes, and both work at most providers.
The first is a notebook template. Many providers offer an image with JupyterLab and PyTorch installed that opens at a URL after boot. It is the closest thing to Colab.
The second is Jupyter over SSH, which works on any Linux GPU instance:
# on the rented machine
pip install jupyterlab
jupyter lab --no-browser --port 8888
# on your laptop
ssh -L 8888:localhost:8888 user@your-instance-ip
Then open http://localhost:8888 and paste the token that Jupyter printed.
Either way, keep datasets and checkpoints on a persistent volume, push code to a repository, and stop the instance when you finish. If you are not sure how much memory your model needs, the LLM VRAM calculator and our guide to the best GPU for AI work will narrow the choice.
Decision rule, by GPU hours per week
The hour figures come from the March 2026 burn rates above, spread over a month of about 4.3 weeks. They are estimates and will move when Google changes the rates.
- A few hours a week, a T4 class GPU is enough. Stay on free Colab or use Kaggle. Kaggle shows your remaining weekly quota in your account, which makes it the more predictable of the two.
- Up to about 19 hours a week on a T4, interactive work. Colab Pro covers it on those rates. Nothing rented will be simpler.
- More than about 4 hours a week on an A100. Colab Pro's 100 units run out. Pro+ stretches to roughly 26 hours a week of A100 40 GB time on the same estimate. Set $49.99 against your monthly hours times the live price, currently $0.68 per GPU-hour on LeaderGPU (checked ). If the numbers are close, the deciding question is whether Colab gives you the A100 when you ask for it.
- Any job that must run unattended for more than 24 hours, or needs one specific GPU every time. Rent. No Colab consumer plan promises either. Google's FAQ itself points to Colab Enterprise and the GCP Marketplace for "guaranteed resources".
- You need more than one GPU, or an H100 you can count on. Rent. Start with the live table above, and read A100 vs H100 first if you are unsure which.
If you sit between two cases, the practical Colab Pro alternative is both: keep the $9.99 plan for quick experiments and rent by the hour for the long runs.
Sources
Source pages were retrieved with automated research tools on the dates shown and each figure was traced back to its source before publishing. Rental prices on this page are not typed: they are read live from GPUperhour's own data.
All accessed 21 September 2026 unless a date is given.
- Google Colab signup page: plan names, US prices, unit allotments, unit expiry, free tier wording
- Google Colab FAQ: session limits, unpublished limits, Google AI plan benefits, zero balance behaviour, disallowed uses, guaranteed resources
- Colab paid terms of service: limits adjustable at Google's discretion
- Google One help, Google AI Pro benefits and Google AI Ultra benefits: Colab units per plan
- Google, Google AI subscriptions (19 May 2026): Ultra tiers
- Colab Enterprise pricing: hourly accelerator prices, Iowa
- Chris McCormick, Colab GPUs: features and pricing (table updated March 2026): measured units per hour
- varlog.info, Colab credit consumption comparison (measured 1 October 2024)
- Colab GitHub issues 6076 (6 August 2026) and 6014: premium GPU availability reports
- Kaggle Notebooks documentation: session limits, extra GPU hours for linked Colab Pro and Pro+ accounts
- AWS, SageMaker Studio Lab availability change
- Vast.ai security FAQ