Compute
13 GPU classes.
How does choosing a GPU work?
You choose a class — a family of interchangeable accelerators with the same memory and comparable throughput — rather than a specific machine. Capacity is then sourced as serverless compute across multiple providers. What you decide is memory and speed; where it comes from is not something you have to manage.
Why classes rather than machines
A class is a family of interchangeable accelerators with the same memory and comparable throughput. That is the level at which a decision actually gets made — does the model fit, how fast will it run, what does it cost per hour — and which individual card ends up serving the job changes none of those answers.
Choosing a class rather than a specific machine is also what allows capacity to be found wherever it exists rather than waiting on one queue. The cost of that abstraction is that you cannot pin a region or a host; the benefit is a shorter wait and a lower rate.
Training and serving are priced separately on every class because they are different workloads. Training is bursty and bounded — a run starts, finishes and releases capacity. Serving is resident and continuous, holding capacity available whether or not requests arrive.
| Class | VRAM | Training / hr | Serving / hr | Good for |
|---|---|---|---|---|
| RTX 3080 | 12 GB | $0.09 | $0.11 | Fine-tuning models under 2B parameters · Serving tiny models |
| RTX 4000 Ada | 20 GB | $0.09 | $0.11 | Low-latency serving of small models · Tiny fine-tunes |
| L4 | 24 GB | $0.17 | $0.20 | Power-efficient inference · Steady-state serving of models up to 8B |
| RTX 3090 | 24 GB | $0.22 | $0.26 | QLoRA on 7-8B models · LoRA on 3-4B models |
| RTX 4090 | 24 GB | $0.38 | $0.43 | LoRA on 4-8B models · Multi-adapter serving of an 8B base |
| A40 | 48 GB | $0.42 | $0.48 | LoRA on 12-14B models · Preference tuning that needs a reference model · QLoRA on 24B |
| RTX 6000 Ada | 48 GB | $0.61 | $0.71 | Fast 14B LoRA · QLoRA on 24B models · Mixture-of-experts fine-tunes |
| L40S | 48 GB | $0.78 | $0.90 | Multi-adapter serving on an 8B base · Modern 48GB training |
| A6000 | 48 GB | $0.65 | $0.75 | 14B LoRA · Multi-adapter serving · Long-context work at 48GB |
| A100 40 GB | 40 GB | $1.17 | $1.35 | 12-14B LoRA · Full fine-tunes of small models · High-bandwidth training |
| A100 80 GB | 80 GB | $1.68 | $1.94 | QLoRA on 70B models · LoRA on 32B models · Long reasoning runs |
| H100 80 GB | 80 GB | $2.59 | $2.99 | Fastest single-GPU runs · fp8 serving · Premium training and serving |
| H200 | 141 GB | $4.55 | $5.25 | Mixture-of-experts training · Very long context · Frontier-scale work |
Picking one
Two steps. First, filter by memory: the model catalogue publishes an admission threshold per model for half-precision and four-bit training, and a class below that will not be scheduled because the run would fail. Second, choose between the qualifying classes on the trade between hourly rate and throughput.
That second step is where money is won and lost, and it is not always obvious. A faster, dearer class that finishes in half the time costs less overall — which matters most when the cheaper option would require quantised training, because that is meaningfully slower per step.
Last verified 6 August 2026.
Start with the free tier
A magic link creates your account, your tenant and your first API key. No card until you ask for compute.