96 GB and larger GPUs you can deploy

The 96 GB and larger tier covers the three biggest single GPUs on DeployGPU: the RTX PRO 6000 at 96 GB, the H200 at 141 GB, and the B200 at 180 GB. The table below covers all three at once, with live prices.

How the three compare

Specifications from NVIDIA's RTX PRO 6000, H200 and DGX B200 pages. NVIDIA's own B200 figure is 180 GB per GPU, worth stating plainly since some listings elsewhere advertise 192 GB. B200 is also the scarcest of the three: when the B200 page shows no stock, the H200 is the next-largest card available today. The RTX PRO 6000 is also the only one of the three with no NVLink option at all: The H200 comes as SXM modules or as the NVL card with NVLink bridges, and the B200 sits on SXM boards with NVLink, covered in more depth on their own pages.

What the extra memory buys

These are weights-only sizes of published checkpoints, calculated rather than measured, shown across all three cards so the difference is easy to see in one place.

Llama 3.3 70B is the clearest example: it needs 2 GPUs on both the RTX PRO 6000 and the H200, and fits on a single B200. For most current models under about 65 GB, the RTX PRO 6000's 96 GB is already enough, and the 96 GB fit table covers more of those.

What full fine-tuning costs at this tier

Serving numbers are not fine-tuning numbers. Full-parameter fine-tuning in mixed precision costs roughly 18 bytes per parameter for weights, gradients and optimizer state, so a 70B model's full fine-tune runs past 1,200 GB and needs several of even these cards. LoRA-style methods bring that down to roughly 2 times the parameter count in gigabytes, closer to 140 GB for the same 70B model, within reach of 2 RTX PRO 6000s or a single H200. Axolotl, a common fine-tuning framework, ships example configurations for several current architectures, including gpt-oss and DeepSeek-style MoE models, which is a reasonable place to start once the memory math above says a job fits.

Multi-GPU on this tier

The RTX PRO 6000 has no NVLink, so a multi-GPU VM built on it connects over PCIe. The H200 and B200 both use NVLink, the better fit for a tensor-parallel job that spans several GPUs at once, such as the Llama 3.3 70B row above. The multi-GPU page covers how to confirm the interconnect on a specific offer before launching it.

Moving up or down a tier

For models that fit comfortably in 80 GB, the 80 GB tier costs less per hour on an A100 or H100. There is no tier above this one: for more than 8 GPUs in total, or a fixed term at a fixed price, request a cluster quote, which is also where to ask about reserved B200 capacity instead of waiting on open stock.

For H200 prices across the whole market, compare all providers on GpuPerHour. For a reference price over time, see the H200 rental index (GPUIndexes).