What it costs to fine-tune a model on an 8-GPU B300 node

When someone asks us what a B300 node costs, the honest answer is another question: what are you going to run on it? The rate is the easy part. The number that decides your bill is how long your job keeps the machine busy. So rather than quote a headline price and leave you to guess, here is the actual arithmetic — with worked examples you can drop your own hours into.

The only formula you need

Cost of a job = GPU-hours × the per-GPU-hour rate. That is it. A FlexiCloud B300 node is eight SXM GPUs, so one hour on the node is eight GPU-hours. At the floor rate of $4.50 per GPU-hour, a node-hour is 8 × $4.50 = $36. Keep the node for a day and that is $864; keep it for a week, $6,048. Everything below is just that sum with different hour counts.

One caveat before the examples: $4.50 is the floor — the rate on the longest commitments. A six-month minimum term sits higher, and the price comes down as the term lengthens. So treat these figures as the best case, and the real number as "this, or a little more, depending on your term." You can see the current rate on the pricing page.

Three worked examples

These are illustrative. Real GPU-hours depend on your model size, dataset, number of epochs, batch size and method (a LoRA adapter is far cheaper than a full fine-tune). But the shape is right, and the maths is honest.

• A short experiment or small LoRA fine-tune — say the job keeps the node busy for 6 hours. That is 48 GPU-hours. At the floor rate: ~$216.
• A serious fine-tune — a larger model, several epochs, roughly two days on the node (48 hours). That is 384 GPU-hours: ~$1,728.
• A long training run — two solid weeks on the node (336 hours). That is 2,688 GPU-hours: ~$12,096.

Notice what is not in those numbers: no per-token meter, no per-request charge, no surprise egress bill for pulling your own checkpoints out. You are renting the machine, not paying by the sip. Once the node is yours for the term, running it flat out costs the same as leaving it idle — which is exactly why utilisation is the thing to optimise.

How the commitment changes the bill

Because the rate falls as the term lengthens, the same job gets cheaper the longer you reserve the node. That sounds like a catch, but it is the opposite: dedicated bare metal is reserved for you, and a longer reservation is what lets us price it down. If you know you have months of training and inference ahead, a longer term is simply a lower effective cost per run. If you are not sure yet, start at the six-month minimum and we will size it with you.

When a node is the wrong tool

We would rather tell you this now than after you have signed. If your workload is small, spiky or occasional — a bit of inference here, an experiment there — a pay-per-token API will almost certainly cost you less, because you only pay for the seconds you use. A dedicated node earns its keep when you can keep it busy: sustained training, high-throughput inference, or work that has to stay on hardware and in a region you control. Below that line, rent nothing and use an API.

Getting a real number for your job

The examples above get you to an order of magnitude. To get to a figure you can put in a budget, we need to estimate the GPU-hours together — and that means a short conversation about the model, the method and the deadline, not a checkout page. Tell us what you are training and we will come back with a GPU-hour estimate, a firm rate for your term, and an honest read on whether a node is even the right call. Start on the B300 page, or talk to an engineer.