16× NVIDIA GB300

Memory budget

4,608 GB of VRAM on this preset. Pooled HBM keeps frontier dense models and 1T-class checkpoints resident (rule of thumb: about one byte per parameter at FP8).

Typical loads

frontier dense-model serving at max quality, jumbo-batch inference. Fine on: torch, JAX, raw CUDA, axolotl, vLLM, sglang, ComfyUI.

Billing math

$184/hr is $0.0511/sec while running and $0 while parked; park/wake freezes and restores state instead of paying for idle.

Best fit

4,608 GB across sixteen GB300 288 GB GPUs at $184/hr, $11.50 per GPU-hour — the largest pool in the catalog, and the row the catalog notes for 1T+ parameter models. The GB300 row is also the only one that pairs the Blackwell GPUs with a Grace CPU over NVLink-C2C, one shared memory domain, rather than NVLink 5 between GPUs only.

Trade-off

There is no rung above this one in the catalog. The honest tradeoff is against b300-8x-fp8 below it: $96/hr for eight B300 cards and 2,304 GB, so gb300-16x-fp8 is about 1.9× the hourly for twice the GPUs and twice the VRAM — and $11.50 per GPU-hour against $12, the lowest per-GPU B300 price in the catalog.