GPU billing questions for LivePortrait AI video generation

[cryptobratru-source]

Hello!

We are planning to use Akash Network for AI video generation (LivePortrait neural network). Before starting, we need to clarify several
important points about billing and GPU availability:

1. GPU idle billing
Are we charged for GPU when computations are stopped (container is idle)? Or do we pay strictly for seconds of active work?

2. Billing precision
Is billing per-second or per-minute? If a task completes in 30 seconds, do we pay for 30 seconds or a full minute?

3. Available GPUs
What GPU configurations are available (RTX 3090, RTX 4090, A100)? What is the cost per hour for each?

4. Container management
Can we stop/start containers via API or CLI? Does stopping a container immediately stop billing?

5. Docker image support
Can we deploy custom Docker images with CUDA/PyTorch? Any restrictions on image size or dependencies?

6. Storage costs
What is the cost of storing data (model weights, datasets) when the container is stopped?

7. Payment methods
What cryptocurrencies are accepted? Can we pay with USDC instead of AKT?

Thank you!

[brewsterdrinkwater]

Apologies for the delays. Really appreciate you asking these questions. Here's how billing and deployments actually work on Akash for something like LivePortrait. The one thing to get straight first: when you lease a GPU on Akash, you're renting the whole box for as long as the lease is open. You're not paying per computation. Almost everything below follows from that.

Also, if you want real time support, the Akash Discord is full of builders.

1. GPU idle billing

Yes. Once your lease is open, you pay the same rate whether the GPU is busy or just sitting there. There's no auto scale-down and no cheaper "idle" rate. If your container finishes a job and waits around for the next one, you're still paying full price until you close the deployment.

So if your workload is bursty, like a job every now and then, leaving a lease running 24/7 to catch the occasional 30 second task is the pricey way to do it. Better to close the lease between jobs, or put something in front that opens a lease only when there's work.

2. Billing precision

It settles per block, and a block is about 6 seconds. It's basically continuous, not rounded up to whole minutes or hours, and you can close a deployment whenever you want and get the leftover escrow back.

A single 30 second task costs almost nothing in compute (roughly 5 blocks, a fraction of a cent on most cards). What actually costs you is how long the lease is open around that task: startup, image pull, loading the model into VRAM, and any idle time before you close. So budget on total lease uptime, not on how long the job runs.

3. Available GPUs

Akash runs NVIDIA GPUs, and it's a marketplace, so prices move with supply and which providers are bidding. All three you mentioned show up regularly.

Rough starting prices per GPU per hour (snapshot from September 2026):

GPU	Approx. starting price/hr
RTX 3090	~$0.15
RTX 4090	~$0.48
A100	~$1.78
H100	~$2.71

Plenty of other cards are around too (T4, L4, RTX 4080, RTX 6000/8000, RTX 5090, A6000, H200, etc). For live numbers, check the GPU pricing page or the deploy screen in Console. The prices above are a snapshot, not a fixed rate card.

You pick a card in your SDL under the gpu section, like:

gpu:
  units: 1
  attributes:
    vendor:
      nvidia:
        - model: rtx4090

You can list a few acceptable models to open up more providers to bid on you. Leave the model off and any NVIDIA GPU will do.

One more thing: if you need higher-end GPUs and aren't finding enough supply on the open marketplace, the core team can help. They have relationships with a number of data center partners and can pull more resources onto the marketplace pretty quickly when there's real demand. So if you're planning something larger and want to line up capacity ahead of time, reach out on Discord and they can help match you.

4. Container management

You can manage deployments from the CLI (akash / provider-services), the Console web UI, or the API/SDK. The usual actions are create, update, and close (e.g. akash tx deployment close).

Here's the catch though: Akash doesn't have a "pause" that keeps your spot but stops the meter. If you want billing to stop, you close the deployment. Closing ends the lease right away, stops the charges, and refunds whatever's left in escrow to your balance. To run again, you open a new deployment.

So it's close and redeploy, not stop and resume. And since closing tears the container down, anything on ephemeral disk is gone next time. Keep your weights and any data you care about on persistent storage or pull them at startup. More on that in the next two answers.

5. Docker image support

Yep, normal Docker images work fine, including custom CUDA and PyTorch builds, and private registries are fine too. You just point the image field at it. A couple of things worth knowing:

• No hard size limit that I'm aware of, but big images take longer to pull and have to fit on the provider's disk. So a bloated image means slower cold starts and fewer providers that can run it. Keep it lean if you can.
• Don't use the :latest tag. Providers cache images hard, so pin a real version tag or you might not get the build you think you're getting.
• GPU access comes from the gpu block in the SDL (above), so your CUDA runtime lines up with the card you asked for.

6. Storage costs

There are two kinds: ephemeral (dies with the container) and persistent (a named volume that survives container restarts inside the lease). Both are billed as part of the lease, the whole time it's open, including while the container is idle.

The thing to know for holding weights and datasets while you're "off": Akash doesn't have standalone object storage that keeps charging after you close a deployment. Persistent volumes live inside a lease, so if you close the lease to stop paying, that storage goes away with it. Two ways people handle it:

• Keep a small cheap lease alive (little or no GPU) that just holds a persistent volume with your weights, and spin GPU deployments up and down on their own.
• Park the weights and datasets somewhere external (a bucket, Hugging Face, IPFS, whatever) and pull them when the container starts. That way you pay nothing for storage when idle, at the cost of a slower cold start.

For LivePortrait the model isn't big, so pulling weights at startup is usually the easiest and cheapest route.

7. Payment methods

At the protocol level you can pay in a few ways:

• AKT, the native token
• USDC (Axelar USDC), if you'd rather use a stablecoin

So yes, USDC works as an alternative to AKT. One thing to note: the take rate isn't the same for both. Paying in AKT has a lower network take rate than USDC, on purpose, to nudge people toward AKT. So the same compute ends up a little more expensive paid in USDC.

And if you'd rather not deal with crypto at all, Console lets you pay by card in plain USD and does the token conversion for you behind the scenes.Please let us know if you have any more questions. When you have things up and running, would love to learn more about how you all are using the Akash Network.