Serve a small LLM
We'd pick RTX 4090 · Dedicated
An 8B model in FP16 is about 16 GB of weights. A 24 GB card serves it with room for the KV cache, at the lowest price we have.
Use cases
Start from the job.
Start from the job.
We'll match the GPU.
The right GPU is the cheapest one your work fits on. Here's what we'd rent for the jobs people bring to packet.ai, and why.
Serve a 70B LLM
We'd pick B200 · Dynamic
70B in FP8 needs about 70–75 GB before the KV cache. A B200 holds it on one card with over 100 GB to spare for long contexts and batching. Quantize to 4-bit and an A100 does it for $0.89.
Fine-tune a model
We'd pick RTX PRO 6000 · Dynamic
QLoRA on a 32B model needs roughly 35–45 GB. 96 GB of GDDR7 covers it, and full LoRA on 8B, at the lowest price per GB we sell.
Generate images & video
We'd pick RTX PRO 6000 · Dynamic
Flux.1 [dev] in BF16 with its encoders uses about 33 GB. The RTX PRO 6000 has 96 GB and Blackwell FP8. SDXL fits on an RTX 4090.
Notebooks & dev boxes
We'd pick RTX PRO 6000 · Dynamic
A workspace you leave running wants room more than speed: 96 GB for $0.66 an hour, about $482 a month if it never stops. Add Jupyter or VS Code at launch.
Train and run clusters
We'd pick B200 · Dedicated → bare metal
Prove the recipe on one Dedicated B200, then move to 8-GPU bare metal with NVLink, or a multi-node InfiniBand cluster.
Not sure what you need?
Ask an engineer Tell an engineer the model and the traffic. You'll get a straight answer, even if it's a smaller GPU.
Your next run
starts at $0.39.
Sign up, top up, launch. You'll be on a GPU in minutes.
No contractHourly or flat monthlyTalk to a human any time