Full Precision · FP8
$96/hr $0.0267/sec · $70,080/mo at 24/7
Max quality for frontier dense models. Frontier dense models, 1M-ctx serving.
VRAM pool
288GB ×8 = 2.3TB
Position
vs H100 8×: ~2× tokens/$ on MoE
Compute die
NVIDIA B300 288GB — FP8 tensor cores, tuned for frontier dense models, 1m-ctx serving.
HBM3e memory
2,304 GB of HBM3e across the 8 GPUs.
NVLink 5 interconnect
NVLink 5 across the 8 GPUs — one memory domain.
Per-second metering
$96/hr → 0.0267/sec. Pause the deployment and the meter hits $0.
Scale-to-zero
Autoscale + region pins (GLOBAL / US / EU / APAC) on the control plane.
Memory budget
2,304 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
$96/hr is $0.0267/sec while running and $0 while parked; park/wake freezes and restores state instead of paying for idle.
Best fit
2,304 GB across eight B300 288 GB cards on NVLink 5 at $96/hr, $12 per GPU-hour. The catalog's note for this row is max quality for frontier dense models, and it is the FP8 row of the 8-card tier: at one byte per parameter a 1T-parameter dense checkpoint is ~1 TB of weights, leaving about 1.3 TB of the 2,304 GB pool for KV cache and batch (rule of thumb).
Trade-off
gb300-16x-fp8 is $184/hr — about 1.9× the hourly — for 16 GB300 GPUs and 4,608 GB, still FP8 and slightly cheaper per GPU-hour: $11.50 against $12. The step up doubles the cards and the memory pool, and it is the only catalog row that adds a Grace CPU memory domain over NVLink-C2C.
$96/hr on runaii.cloud, metered per second: $0.0267/sec while the VM runs and $0 while parked. A month of 24/7 uptime would be about $70,080, and park/wake means busy-fraction pricing in practice.
Yes. Per-second metering with park/wake: idle freezes the VM state into a signed artifact and stops the meter the same second; the next request wakes it byte-exact
2,304 GB of VRAM. Pooled HBM keeps frontier dense models and 1T-class checkpoints resident (rule of thumb: about one byte per parameter at FP8).
2,304 GB — eight 288 GB B300 cards on NVLink 5, $96/hr, $12 per GPU-hour.
The catalog lists the 8-card B300 row at FP8 and the 4-card B300 row at NVFP4; FP8 is the higher-precision of the two catalogued modes and this row carries the max-quality note for frontier dense models.
At FP8 and one byte per parameter, a 1T-parameter checkpoint is ~1 TB of weights with about 1.3 TB of the 2,304 GB pool left for KV cache and batching (rule of thumb).
RTX PRO 6000 slices
from $0.30/hr · 12-48GB
Blackwell slices for light jobs
1× NVIDIA T4 16GB
$0.44/hr · FP16
Starter
1× NVIDIA L4 24GB
$0.87/hr · FP8
Efficient
2× NVIDIA L4 24GB
$1.8/hr · FP8
Efficient ×2
4× NVIDIA L4 24GB
$3.6/hr · FP8
Efficient ×4
8× NVIDIA L4 24GB
$7.2/hr · FP8
Efficient ×8
1× NVIDIA RTX PRO 6000 Blackwell 96GB
$2.4/hr · FP8
Workstation
2× NVIDIA RTX PRO 6000 Blackwell 96GB
$4.8/hr · FP8
Workstation ×2
4× NVIDIA RTX PRO 6000 Blackwell 96GB
$9.6/hr · FP8
Workstation ×4
8× NVIDIA RTX PRO 6000 Blackwell 96GB
$19.2/hr · FP8
Workstation ×8
8× NVIDIA B200 180GB
$80/hr · NVFP4
Throughput
4× NVIDIA B300 288GB
$48/hr · NVFP4
Minimal
4× NVIDIA B200 180GB
$40/hr · NVFP4
Efficient
16× NVIDIA GB300
$184/hr · FP8
Frontier Max