Free Cloud GPUs in 2026: What Works and What Closed

Which free GPU options still work in September 2026, which closed, who qualifies for credits, and what the cheapest paid GPU costs once the quota runs out.

By Faiz Ahmed
13 min read

As of 21 September 2026, free GPU time still comes mainly from four places: Google Colab's free tier, Kaggle Notebooks, Hugging Face ZeroGPU for short inference demos, and small monthly credits from serverless vendors such as Modal. Every one of them gives you an older or shared GPU under a hard quota. When the quota runs out, the cheapest paid step is an hourly rental of a T4, RTX 3090 or RTX 4090, and the live table further down shows what that costs today.

Everything here was read from the vendors' own pages on 21 September 2026 unless a row says otherwise. Free tiers change without notice. Treat the date as part of each fact.

Free GPU options at a glance

OptionWhat you getThe catchStatusAs of
Google Colab, free tierA hosted notebook with "very limited access to GPUs". Sessions run "at most 12 hours". Google does not name the GPU. Users report a T4.No published quota. Idle runtimes time out. SSH, remote desktops and web UIs are not allowed without a paid balance.Open21 Sep 2026
Kaggle NotebooksT4 and P100 class GPUs are what Kaggle has offered. "12 hours execution time for CPU and GPU notebook sessions". A weekly GPU quota that Kaggle sets and shows in your account settings (commonly reported as around 30 hours).Kaggle's docs give no quota number, and the quota floats. The accelerator menu changes.Open, check the menu and quota in your account21 Sep 2026
Hugging Face ZeroGPUA shared RTX PRO 6000 Blackwell. The default size is half a GPU with 48 GB. A free account gets 5 minutes of GPU time a day.Gradio Spaces only. 60 second default GPU function. No torch.compile. It serves demos. It does not train models.Open21 Sep 2026
Lightning AI, free plan"1 free active Studio (4 hour restarts)" and persistent storage with a 50 GB limit."Free Studios become paid after 4 hours". We could not confirm the monthly free credit amount.Open, details unconfirmed21 Sep 2026
Paperspace Gradient, Free planA "Free GPU (M4000)" machine for notebooks in private workspaces, with 5 GB of storage.Auto-shutdown after 6 hours at most. We did not test whether free capacity is obtainable.Documented, page last verified 13 Jul 202621 Sep 2026
Modal, Starter plan"$30 / month free compute" on a serverless platform.It is code-first serverless, not a notebook. Usage past the credit is billed.Open21 Sep 2026
Amazon SageMaker Studio LabExisting users: GPU sessions of "up to 4 hours at a time with a limit of 4 hours in a 24-hour period"."No longer open to new customers."Closed to new users21 Sep 2026
Colab Pro for EducationA free year of Colab Pro for US students and faculty, launched 18 July 2025."No longer available for new signups at this time."Closed to new users21 Sep 2026
GitHub Student Developer Pack"$100 in Azure credit" for students aged 18 and over. Camber's student plan with "5 GPU hours".No DigitalOcean offer appeared on the pack page as retrieved.Changed21 Sep 2026
Hyperscaler free trialsGoogle Cloud: $300 over 90 days. AWS: up to $200 over 6 months. Azure: $200 for up to 30 days.Google's trial cannot add GPUs. AWS's free plan covers "select services only".Open, mostly not for GPUs21 Sep 2026

Three rows need a note on how sure we are.

Colab. Google's FAQ says "The types of GPUs and TPUs that are available in Colab vary over time" and "Colab does not publish these limits". The T4, and every weekly hour figure you see quoted elsewhere, comes from users and not from Google. The 12 hour session cap is Google's own wording, qualified by "depending on availability and your usage patterns". Google also warns free users not to rely on background execution: "execution will be interrupted when user interaction ceases, and the VM will be deleted soon after that."

Kaggle. The session limits are from Kaggle's notebook documentation, read on 21 September 2026: 12 hours for CPU and GPU sessions and 9 hours for TPU sessions. The size of the weekly GPU quota is not in Kaggle's docs. It is shown in your account settings and it floats, so the figure of around 30 hours is what users commonly report, not a Kaggle statement. 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, without spending Colab compute units, and that Kaggle may change or end this.

Lightning AI. The pricing page is a client-rendered app. The free plan strings quoted above were read from the page's JavaScript on 21 September 2026. Third-party sites disagree with each other about the monthly credit, so we do not quote one.

What the hardware is

The GPUs behind free tiers are old. The T4 launched in 2018 with 16 GB of memory. The Quadro M4000 on Paperspace's free plan launched in 2015 with 8 GB. For comparison, a rented RTX 4090 has 24 GB and 2.5 times the T4's dense FP16 throughput (165.2 against 65 TFLOPS, both dense figures from our spec table).

SpecT4Quadro M4000RTX 3090RTX 4090RTX PRO 6000
VRAM16 GB8 GB24 GB24 GB96 GB
Memory bandwidth320 GB/s192 GB/s936 GB/s1,008 GB/s1,792 GB/s
FP16 (dense)65 TFLOPSNot published71 TFLOPS165.2 TFLOPS503.8 TFLOPS
FP16 (with sparsity)Not publishedNot published142 TFLOPS330.4 TFLOPS1,007.6 TFLOPS
Launch year20182015202020222025
ArchitectureTuringMaxwellAmpereAda LovelaceBlackwell
Figures from the vendor datasheets: T4, Quadro M4000, RTX 3090, RTX 4090, RTX PRO 6000, checked 13 Sep 2026. With-sparsity figures assume 2:4 structured sparsity and are twice the dense figure, so compare dense with dense. "Not published" means the vendor gives no figure.

ZeroGPU is the exception on age and the extreme case on quota. Hugging Face's documentation says it "dynamically allocates and releases NVIDIA RTX Pro 6000 Blackwell GPUs as needed". The default large size is half a GPU with 48 GB. The xlarge size is the full 96 GB GPU and burns quota twice as fast. A free account gets 5 minutes of GPU time per day and an unauthenticated visitor gets 2. The quota "resets exactly 24 hours after your first GPU usage". PRO costs $9 a month, raises the quota to 40 minutes, and lets you buy more "at the rate of $1 per 10 minutes of GPU time". That is a way to host a demo. It is not a way to fine-tune a model.

What closed by September 2026

Amazon SageMaker Studio Lab. AWS's documentation states: "Amazon SageMaker Studio Lab is no longer open to new customers. Existing customers can continue to use the service as normal." AWS gives no closure date and says "we do not plan to introduce new features". The replacement it suggests is paid SageMaker Studio, which needs an AWS account.

Colab Pro for Education. Google launched a free year of Colab Pro for US students and faculty on 18 July 2025. The Colab FAQ now says these subscriptions "are no longer available for new signups at this time." Verified users keep the benefit until their year ends.

The DigitalOcean credit in the GitHub Student Developer Pack. Older guides cite a DigitalOcean credit. No DigitalOcean offer appeared on the pack page as retrieved on 21 September 2026. The compute offers we did find were Azure ("$100 in Azure credit"), Heroku, Codespaces, a Deepnote plan with no GPU, and Camber's student plan with "40 CPU hours, 5 GPU hours, 50 GB of storage".

Credits: who qualifies

Sign-up credits from the big three clouds

These are real money and mostly useless for GPUs. Google Cloud's free trial gives "a $300 Welcome credit to spend over 90 days", but during the trial you cannot "Add GPUs to your VM instances", and "GPUs and TPUs are not included in the Free Tier offer." AWS gives "$100 in credits immediately" and up to $100 more for exploring services, "up to $200 over 6 months", on a free plan "Limited to select services only". Azure gives a "$200 credit to explore Azure for up to 30 days" and asks for a phone number and a non-prepaid card. Azure's page does not say whether a new subscription gets any GPU quota, and we do not know.

Student programmes

Azure for Students gives "$100 credit to use on Azure services within 12 months" with "No credit card required", for full-time university students. Its page lists only burstable CPU machines as free services and does not state GPU eligibility. Modal's academic grant is the largest GPU-relevant offer we found: "Graduate students, labs, and researchers can get up to $10k free compute credits". You apply for it. It is not automatic.

Startup programmes

ProgrammeAmount as publishedConditions as published
Google for Startups Cloud Program"$2,000 to build your MVP". Seed to Series A: "$200,000 in cloud credits (or up to $350,000 for AI-first startups)"Tier depends on funding stage
AWS ActivateFounders: "$1,000 USD in Activate Credits", with some participants qualifying for up to $5,000. Portfolio: "Up to $200,000 USD""Pre-series B, founded in the last 10 years, AWS Account on Paid Tier Plan"
Microsoft for Startups"up to $150,000 in credits" across eligible Azure servicesTiers and eligibility are not shown on the landing page
NVIDIA Inception"free cloud credits from NVIDIA and partners", no amount stated, plus "preferred pricing on select NVIDIA hardware and software"Incorporated, less than 10 years old, at least one developer, a working website. No fees and no equity
Koyeb startup programme"up to $30k in credits"Apply through Koyeb
Modal"Early-stage startups can get free compute credits", amount not statedApply through Modal

All rows were read on 21 September 2026. Inception is not a cash-credit programme in itself. Its credits come through partners, and NVIDIA publishes no figure.

Provider credits

Modal's Starter plan includes "$30 / month free compute". Baseten says new accounts "come with credits" and gives no amount. Replicate says "You can run select models on Replicate for free, but after a bit you'll be asked to set up billing." Beam's and Cerebrium's pricing pages list $0 plans billed on usage and state no free credit amount. RunPod's referral programme pays a new user who loads at least $10 "a random weighted bonus between $5 and $500", or a fixed $5 in Europe, and requires Google sign-in. Our sister article on serverless GPU pricing lists what these platforms charge once the credit is gone.

Converting credits to GPU-hours

A credit is only worth the GPU time it buys. The conversion is one division:

GPU-hours = credit amount divided by the hourly price of the GPU you will use.

Two rules keep it accurate. First, take the hourly price from a live source on the day, such as the table below or the provider pages. Do not use a number typed into an article, including this one. Second, use the price at the vendor where the credit can be spent. A cloud credit buys GPU-hours at that cloud's rate, which is often well above the cheapest rate on the market. Cerebrium's pricing page, for example, says AWS and GCP cloud credits cannot be applied to its platform. Check the expiry as well: the Google Cloud trial credit lasts 90 days and the Azure one 30.

When free costs more than paid

Free tiers cost time. Sessions end at 12 hours on Colab and Kaggle. Colab may interrupt you when you stop interacting. Kaggle's quota is weekly, ZeroGPU's is minutes per day, and Studio Lab's is 4 hours in 24 for the users who still have it. Work that outgrows these limits turns into checkpoint juggling, account switching (which Colab's terms forbid) and waiting for a reset. Here is what the cheapest real GPUs cost right now.

GPUCheapest $/GPU-hrProviderProviders in stock
Tesla T4none in stock
RTX 3090$0.29LeaderGPU2
RTX 4090$0.75Vast.ai2
L4$0.49RunPod2
A100$0.68LeaderGPU8
Cheapest in-stock on-demand price per GPU-hour, from providers with live stock tracking. Latest stock observation: .

And the same prices as a daily and a monthly bill if you left the GPU running all the time, which you would not.

GPU$/GPU-hrPer day (×24)Per month (×730)Provider
Tesla T4none in stock
RTX 4090$0.75$18.00$548Vast.ai
Cheapest in-stock on-demand price per GPU-hour, for one GPU running the whole time. Per day = hourly price × 24 hours. Per month = hourly price × 730 hours (365 days × 24 ÷ 12), rounded to the nearest dollar. Storage, data transfer and tax are not included. Latest stock observation: .

Read the hourly column, then multiply by the hours your job needs. Do that sum with today's prices before you spend another evening on quota. For a job of a few hours on a consumer card, a short paid rental often beats a week of fitting the same job into quota windows on a 16 GB T4. A rented RTX 3090 or L4 gives you 24 GB and no weekly quota. If you are unsure how much memory your model needs, the LLM VRAM calculator and our guide to the best GPU for AI will narrow it down. The current cheapest T4 we track is no in-stock on-demand offer right now.

The decision rule

Students. Learn on Colab free and Kaggle. Use Kaggle when you need a quota you can see. Use Colab for quick notebooks. Do not plan on Colab Pro for Education or Studio Lab, because both are closed to new users. When a course project outgrows a 16 GB T4, rent a 24 GB card by the hour. Graduate students should apply for Modal's academic grant.

Hobbyists. Use ZeroGPU to host a demo and Kaggle to experiment. The moment you want SSH, a web UI such as a Stable Diffusion front end, or a run longer than 12 hours, stop fighting the free tiers. Colab's FAQ says free users "commonly experience runtime terminations" when they bypass the notebook UI. Rent an RTX 4090 or RTX 3090 for the hours you need and shut it down after.

Startups. Apply to every programme you qualify for on day one: Google for Startups, AWS Activate, Microsoft for Startups and NVIDIA Inception. Inception has no fees and takes no equity. Then run the conversion above with the live price at the cloud where the credit is spendable, and compare it with the cheapest rate for the same GPU across all providers. Credits expire. Build on them, but keep your stack portable so you can move to the cheapest paid GPU when they do.

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.

Frequently asked questions

Is there a free cloud GPU in 2026?

Yes. Google Colab's free tier, Kaggle Notebooks and Hugging Face ZeroGPU all gave free GPU time as retrieved on 21 September 2026. Each one runs on an older or shared GPU under a quota the vendor can change.

Which GPU does Kaggle give you for free?

T4 and P100 class GPUs are what Kaggle has offered. The accelerator menu changes, so open a notebook and check it, along with the weekly quota shown in your account settings.

Is there a free GPU for students?

Less than there used to be. Colab Pro for Education is closed to new signups, and the GitHub Student Developer Pack page showed no DigitalOcean offer as retrieved. What remains is Azure credit for students, a small Camber GPU allowance, and Modal's academic grant for graduate students and labs.

Can I use the Google Cloud free trial credit for GPUs?

No. Google's documentation says you cannot add GPUs to VM instances during the free trial, and that GPUs and TPUs are not part of the Free Tier offer.

How do I convert GPU credits into GPU-hours?

Divide the credit amount by the hourly price of the GPU at the vendor where the credit can be spent. Read the price from a live table on the day, because a typed price goes stale.

Can I still sign up for Amazon SageMaker Studio Lab?

No. AWS's documentation says Studio Lab is no longer open to new customers. Existing customers can keep using it, and AWS does not plan new features.

Related Posts