Miami Mike gets a blank stare almost every time NVIDIA Brev comes up in a customer conversation, and it comes up a lot, because I use it constantly. When I need to test a use case or kick the tires on an NVIDIA blueprint, I do not open an AWS console and start requesting quota. I put credits on my Brev account, pick a GPU, and deploy. A few clicks, a few minutes, and I am inside a running environment. The strange part is that a tool this useful, owned by the most talked about company on earth, is still basically unknown outside a small circle of developers.
The short version: NVIDIA Brev is a service that rents you a preconfigured NVIDIA GPU by the hour, sourced from clouds like AWS, GCP, and Azure, with drivers, CUDA, Python, Docker, and Jupyter already installed. You pay with a credit balance, connect over SSH or VS Code, and delete the instance when you are done.
What is NVIDIA Brev
Brev started life as Brev.dev, a small startup founded in 2021 that NVIDIA acquired in 2024. The pitch was simple then and has not changed. Getting a GPU from a big cloud is a miserable experience for an individual developer. You need an account, a quota increase, the right instance family, the right driver, the right CUDA version, and the patience to debug all of it before you run a single line of your own code.
Brev sits on top of that mess. It pulls GPU capacity from multiple cloud providers, wraps it in a machine image that is actually ready for AI work, and hands you one console and one bill. You never touch the underlying cloud account. You do not even need to know which provider your instance landed on, though the console tells you. NVIDIA now runs it at brev.nvidia.com, and it has quietly become the default way NVIDIA distributes its own demos and blueprints.
How does NVIDIA Brev work
Three concepts carry the whole platform. GPU instances are the compute, cloud virtual machines with one or more NVIDIA GPUs attached. Environments are the software stack that comes preloaded, so CUDA and PyTorch version roulette stops being your problem. Launchables tie the two together with code into a single deployable package.
An instance moves through a simple lifecycle. While it runs, you are billed by the hour. Stop it and Brev releases the GPU back to the cloud provider while keeping your data, so you can pick up where you left off tomorrow without paying for idle compute. Delete it and everything is gone, permanently, along with the charges.
Brev launchables are the part worth understanding
Brev launchables are shareable links that bundle a GPU configuration, a container or setup script, and source code into one deploy button. Somebody builds a working environment once, publishes it, and anyone with the link gets the exact same thing, working, on their own account. The launchables documentation covers the builder in detail, but the console also has a searchable explore tab with featured and trending launchables, each card showing the cost per hour before you commit to anything.
This is the feature that killed reproducibility excuses for me. A tutorial that ships as a launchable cannot rot the way a README does.
Setting up NVIDIA Brev
The setup is the whole reason I keep recommending it, so here is the entire process. It took me longer to write this section than it takes to do it.
- Create an account at the Brev console with an email address
- Add credits with a card. Your balance is shared across your organization if you set one up, and you can turn on auto recharge so a long job never dies mid run because the balance hit zero
- Click Create New Instance, pick a GPU and a configuration
- Name it and hit Deploy. Boot takes a few minutes
- Connect through the browser, SSH, or VS Code and run nvidia-smi to confirm the GPU is there
That is it. No quota tickets, no VPC decisions, no IAM policy archaeology. The official quickstart walks the same path with screenshots if you want them.
The Brev CLI
The Brev CLI is where it gets good for daily use. Once installed, brev shell drops you into a terminal on your instance and brev open launches VS Code connected to it over remote SSH. The CLI manages your SSH config for you. You can create instances from a Git URL, attach setup scripts that run on boot, and stop or delete from the terminal without visiting the console. Data in the workspace directory survives a stop, which makes the stop and resume pattern cheap and painless.
Deploying NVIDIA blueprints with a few clicks
This is my actual workflow and the reason this post exists. NVIDIA publishes blueprints, which are reference workflows for things like RAG pipelines, video search and summarization, and agent safety evaluation. Nearly every blueprint page on build.nvidia.com now has a Deploy on Cloud button that is, underneath, a Brev launchable. Click it, confirm the compute, and the entire stack builds itself, containers, models, notebooks, all of it.
The video search and summarization blueprint, for example, deploys through Brev onto a two GPU RTX PRO 6000 instance on AWS with a Jupyter notebook that walks the whole setup. Standing that up manually means Docker Compose files, NGC keys, container toolkit versions, and an afternoon you will not get back. Through the launchable it is a coffee break.
For someone in my seat, presales and advisory, this changes what a customer conversation can be. I am not describing a demo from a slide. I ran it last night, it cost me a few dollars in credits, and I deleted it before bed.
NVIDIA Brev pricing and billing
Pricing is prepaid credits burned down at an hourly rate that varies by GPU, provider, and region, and the console shows the rate on every instance card before you deploy. There is no subscription. A stopped instance stops accruing compute charges, though storage charges can still apply while your data sits preserved, so the habit worth building is delete, not stop, for anything you will not return to. The console shows your balance, an estimated burn rate while things are running, and remaining hours at that rate.
The discipline is the same one I wrote about in the piece on AI inference costs. The unit price is rarely what hurts you. The forgotten instance running over a weekend is. Brev at least makes the meter visible, which is more than I can say for a hyperscaler bill that arrives as a surprise thirty days later.
NVIDIA Brev vs going straight to AWS, GCP, or Azure
The comparison question comes up immediately, usually as a suspicion that Brev must be marking something up. It brokers the same capacity, so what you are really weighing is convenience against control.
| NVIDIA Brev | Direct cloud account | |
|---|---|---|
| Time to first GPU | Minutes | Hours to days with quota requests |
| Environment | CUDA, Docker, Jupyter preinstalled | You build the image |
| Billing | Prepaid credits, hourly, one bill | Monthly invoice per provider |
| Provider choice | Brev picks across clouds for the GPU you want | You commit to one console |
| Enterprise controls | Basic teams and orgs | Full IAM, networking, compliance tooling |
For production workloads that need VPC peering, private networking, and a compliance story, you still want a real cloud account, and Brev is not pretending otherwise. For development, testing, and anything exploratory, the direct route is friction you are choosing on purpose.
Which GPU should you pick
The GPU catalog runs from a T4 with 16GB of VRAM at the entry level up through L4 and L40S in the middle, and on to A100, H100, H200, and B200 with 192GB of HBM3e at the top. The rule of thumb from NVIDIA’s own docs is that a 7 billion parameter model needs about 14GB in fp16, so size VRAM to your model and stop there. Most of my test sessions run on midrange cards. Renting a B200 to poke at a notebook is how credit balances die.
When Brev makes sense and when it does not
Use Brev when the work is bursty. Testing a blueprint before a customer meeting, fine tuning a small model over a weekend, running a workshop where twenty people need identical environments, reproducing a paper. Anything where the GPU is needed for hours, not months.
The math flips when usage becomes steady. An always on workload rented by the hour will eventually cost more than hardware you own, which is the whole argument in my on premise AI inference piece, and the deciding variable there is utilization. Brev actually acknowledges this better than most rental platforms do. Its Brev Connect feature lets you register your own machines into the same console, so my DGX Spark home lab can sit next to a rented H100 in one interface, and I burn credits only when a job outgrows the box on my desk.
I have not stress tested the team features beyond a small org, and I cannot tell you how Brev behaves at fifty seat scale, so take my read as a practitioner’s, not a procurement review.
If you have been meaning to try a blueprint, or you have a model you want to fine tune without buying anything, this is the lowest stakes way I know to do it. Put twenty dollars on an account, deploy something, delete it, and see what that does to your sense of what is possible on a Tuesday afternoon.
