H100, H200 and B200 Dedicated GPU Systems
NVIDIA H100 (80 GB), H200 (141 GB) and B200 (180 GB) systems are not in the MassiveGRID store: sales configures and prices each one, so there is no list price on this page. Send the GPU count, memory, storage, network and term you need through contact us. If a 96 GB card covers your workload, the RTX PRO 6000 servers on the GPU dedicated servers page can be ordered online from $1,656.15 a month.
Can you rent a dedicated H100, H200 or B200 server from MassiveGRID?
Yes, by quote. H100, H200 and B200 are NVIDIA's data-center accelerators for training and large-scale inference. MassiveGRID supplies them as dedicated systems that sales configures for each customer, which is why this page states no price and no delivery time: both depend on the configuration and are given in the quote.
What can be ordered online today is a different class of card. The four servers in the store use the RTX PRO 4000 Blackwell (24 GB) and the RTX PRO 6000 Blackwell Max-Q (96 GB), singly or four to a server (384 GB). For many inference and adapter fine-tuning jobs a 96 GB card is enough, and published third-party benchmarks further down put it within reach of a single H100 on models that fit one card. This page helps you decide which route you need before you write to sales.
If you are looking for an H100 by the hour for a short experiment, a dedicated system is the wrong product: hourly GPU clouds list single H100 instances, and the market table below gives their public prices as reference points.
H100 vs H200 vs B200: what are the verified specifications?
From NVIDIA's product pages and its Multi-Instance GPU documentation, checked on 16 November 2026. Where a figure could not be verified on NVIDIA's pages, the cell says so.
| GPU | Memory | Memory bandwidth | Power | MIG instances | How to get it here |
|---|---|---|---|---|---|
| H100 SXM | 80 GB | 3.35 TB/s | Up to 700 W (configurable) | Up to 7 at 10 GB | Quote |
| H100 NVL (PCIe) | 94 GB | 3.9 TB/s | 350 to 400 W | Up to 7 at 12 GB | Quote |
| H200 SXM | 141 GB HBM3e | 4.8 TB/s | Up to 700 W | Up to 7 at 18 GB | Quote |
| H200 NVL | 141 GB | Not verified | Up to 600 W | Up to 7 at 16.5 GB | Quote |
| B200 | 180 GB HBM3e | Not verified | Per-GPU figure not verified; a DGX B200 with 8 GPUs draws about 14.3 kW | Up to 7 | Quote |
| RTX PRO 6000 Blackwell Max-Q | 96 GB GDDR7 ECC | 1,792 GB/s | 300 W | Up to 4 at 24 GB | Order online, $1,656.15 a month |
Three things stand out. Memory per card rises from 80 GB to 141 GB to 180 GB, so an eight-GPU node holds 640 GB, 1,128 GB or 1,440 GB. Memory bandwidth on the H100 and H200 is roughly two to three times that of the RTX PRO 6000, which matters most for training and for large models split across cards. And the RTX PRO 6000 Max-Q has more memory than an H100 at 300 W of board power against up to 700 W for the H100 SXM, which is the case for it in single-card inference. For the generation-to-generation detail, read H100 vs H200.
Who really needs H100-class hardware, and who does not?
Start from the model and the job, not from the card name. Sizes are from Hugging Face and the Ollama library, checked on 16 November 2026.
| Workload | What it needs | Route |
|---|---|---|
| Inference on models up to about 80 GB of weights: Llama 3.3 70B at 4-bit (43 GB) or 8-bit (75 GB), gpt-oss 120b (65 GB), Qwen3.8 27B at 16-bit (56 GB) | One 96 GB card | RTX PRO 6000 server, $1,656.15 a month, order online |
| The same with large RAG indexes or heavy preprocessing in host memory | One 96 GB card and 256 GB to 768 GB of RAM | RTX PRO 6000 Xeon server, from $2,485.05 a month, order online |
| Models between 96 GB and 384 GB: Llama 3.3 70B at fp16 (141 GB), Qwen3 235B at q8_0 (250 GB), Llama 3.1 405B at Q4_K_M (243 GB) | Four 96 GB cards | 4x RTX PRO 6000 server, $5,966.44 a month, order online |
| QLoRA fine-tuning up to 70B (Unsloth minimum: 41 GB) | One 96 GB card | Order online |
| Open models above 384 GB: DeepSeek-R1 and V3.1 (688.6 GB in FP8), Mistral Large 3 (681.5 GB), DeepSeek V4-Pro (864.7 GB) | An eight-GPU H200 or B200 node (1,128 GB or 1,440 GB) | Quote |
| Full fine-tuning of 70B-class models (LLaMA-Factory estimate: 600 GB at 16-bit) and training from scratch | Eight H100-class GPUs or more, with matching system RAM and storage | Quote |
| Very long contexts on large models: the Qwen3 235B card states about 1,000 GB of GPU memory for a 1 million token context | Eight H200 or B200 GPUs | Quote |
| Highest throughput per card for a high-traffic API | H100, H200 or B200 | Quote, after measuring |
Model authors say the same in their cards. OpenAI states that gpt-oss 120b fits "into a single 80GB GPU (like NVIDIA H100 or AMD MI300X)" and can be fine-tuned on a single H100 node. Mistral states that Mistral Large 3 runs in FP8 "on a single node of B200s or H200s". If your model is in the first four rows, an H100 is a choice about speed, not a requirement. The multi-GPU server page has the full list of what fits in 384 GB.
What should you send sales to get a quote?
A quote is faster and more accurate when the first message has the facts sales needs. Copy this list into the contact form and fill it in.
- GPU model and count. H100, H200 or B200, and how many per server and in total. If you are not sure, name the model you want to run or train and its precision.
- GPU memory target. Total vRAM for the largest job, including context: for example "DeepSeek V3.1 in FP8, 688.6 GB of weights".
- System memory and CPU. RAM per server and whether data loading is CPU-heavy.
- Storage. NVMe capacity for datasets and checkpoints. A single DeepSeek-R1 repository is 688.6 GB before any checkpoints.
- Network. Port speed, monthly traffic, and whether several servers must exchange data during training.
- Term. How many months you plan to keep the system, and the date you need it by.
- Location and compliance. If data residency matters, say which country or region you need, and ask which data center the system would be delivered in.
- Operating system. The Linux distribution you want, or your own image.
Sales answers with a configuration and a price; ask for the delivery date in the same message. MassiveGRID has been a hosting provider since 2003, is ISO 27001 and ISO 9001 certified, and has 24/7 support.
RTX PRO 6000 vs H100: what do published benchmarks show?
These are third-party results, not ours, and each used the 600 W RTX PRO 6000 Workstation Edition, not the 300 W Max-Q card in the MassiveGRID servers. Treat them as the direction of the difference, then measure your own model.
| Source and setup | RTX PRO 6000 Blackwell | H100 | Reading |
|---|---|---|---|
| StorageReview, updated 3 November 2026: vLLM, Ministral 3 14B in FP8, 32 concurrent chat streams, output tokens per second | 2,283 | 3,096 (H100 NVL 94 GB) | H100 NVL about 36% ahead |
| Same test, Ministral 3 3B | 4,155 | 5,873 | H100 NVL about 41% ahead |
| Same test, 14B, long documents (8,192 tokens in, 1,024 out) | 608 | 859 | H100 NVL about 41% ahead |
| CloudRift, 27 November 2025: vLLM, GLM-4.5-Air AWQ 4-bit, one card, 1,000 tokens in and out | 3,140 | 2,987 (H100 SXM) | About equal |
| CloudRift, same post: Qwen3-Coder-480B AWQ across four cards | Baseline | About 31% faster | H100 ahead on split models |
| CloudRift, same post: GLM-4.6 in FP8 across eight cards | Baseline | About 3 times faster | H100 well ahead at eight cards |
Where the H100 class wins: large models split across many cards, full training, and the highest throughput per card. Where the RTX PRO 6000 wins: 96 GB against 80 GB on one card, a published monthly price, and online ordering. For a single-card inference service the gap in these tests runs from none to about 40%. The RTX PRO 6000 vs H100 vs RTX 5090 comparison goes through the sources in detail.
By quote, or order a 96 GB server online
H100, H200 and B200 systems are configured by sales. The 4x RTX PRO 6000 server (384 GB) and the RTX PRO 6000 Xeon server (96 GB) have published prices.
NVIDIA H100 / H200 / B200
4x NVIDIA RTX PRO 6000
- CUDA, FP4 and Multi-Instance GPU (MIG) support
- 1 dedicated IPv4 included
NVIDIA RTX PRO 6000 Xeon
- 3,511 AI TOPS
- CUDA, FP4 and Multi-Instance GPU (MIG) support
- 1 dedicated IPv4 included
Every RTX PRO server includes: a 1 Gbps port with unlimited traffic, one dedicated IPv4 address, hosting in an ISO 27001 certified data center, and your choice of Ubuntu, Debian, AlmaLinux, Rocky Linux, CentOS Stream, openSUSE or Arch Linux, or a rescue system to install your own OS. Billing runs monthly, quarterly, semi-annually, annually, biennially or triennially, and the setup fee is charged once with the first invoice.
Options on the RTX PRO 6000 Xeon: 512 GB DDR5 ECC with 2x 1.92 TB NVMe for $994.68/mo more, 768 GB DDR5 ECC with 4x 3.84 TB NVMe for $1,989.37/mo more, and a 10 Gbps uplink with 20 TB of traffic included for $71.29/mo more.
What do H100, H200 and B200 instances cost per hour on public clouds?
These are public on-demand list prices of hourly clouds, checked on 16 November 2026. They are not MassiveGRID prices and they are not an indication of a quote. They are useful for one thing: judging what a month of continuous use costs elsewhere before you commit to a dedicated system.
| Provider, on demand | 1x H100, per hour | 730 hours | H200 and B200, per GPU-hour |
|---|---|---|---|
| Oracle Cloud, list price per GPU-hour | $10.00 | $7,300.00 | H200 $10.00; B200 $14.00 |
| Azure NC40ads_H100_v5, East US | $6.98 | $5,095.40 | Not read |
| AWS, N. Virginia | $6.88 (p5.4xlarge) | $5,022.40 | H200 $7.91 (8-GPU p5en at $63.296); B200 $14.24 (8-GPU p6-b200 at $113.9328) |
| Paperspace (as listed in its documentation) | $5.95 | $4,343.50 | Not read |
| Lambda | $3.79 PCIe, $4.89 SXM | $2,766.70 to $3,569.70 | B200 $7.99; H200 not listed |
| Google Cloud a3-highgpu-8g, Iowa, 8-GPU shape only | $11.06 per GPU ($88.49 for 8) | $64,597.70 for 8 | H200 $10.60 (8-GPU a3-ultragpu-8g at $84.806908) |
| CoreWeave, 8-GPU instances only | $6.16 per GPU ($49.24 for 8) | $35,945.20 for 8 | H200 $6.31 ($50.44 for 8); B200 $8.60 ($68.80 for 8) |
AWS lists the same single H100 at $8.60 an hour in Frankfurt and Azure at $9.08 in West Europe, and the hyperscalers' figures exclude storage and egress ($0.09 per GB on AWS, $0.087 on Azure, $0.12 per GiB on Google's Premium Tier). For a week of experiments, paying by the hour is the right way to buy an H100; a dedicated system is for hardware that runs most of the month. For reference, the orderable 96 GB server works out at $2.27 an hour over a 730-hour month and the four-card server at $2.04 per GPU-hour.
How to decide and get started
- Write down the largest job. Model, precision, context length, concurrent users, and whether it is inference, adapter fine-tuning or full training.
- Check it against 96 GB and 384 GB. Add the weights to a margin for context. If it fits, you can order a server online and skip the quote.
- Measure before you commit. vLLM ships a benchmark command,
vllm bench serve, which CloudRift used for the figures above. Run it against your model at your real concurrency. On a multi-GPU machine, serve with tensor parallelism:vllm serve <model> --tensor-parallel-size 4 curl http://localhost:8000/v1/models - If you need H100, H200 or B200, send the list above to sales. Include the benchmark result if you have one: it tells sales whether you are buying memory or throughput.
- Plan the first day. Dedicated servers arrive with a clean operating system and root access. NVIDIA's current Linux driver, 620.70.06, is installed from its apt repository with
apt install nvidia-open; the first-hour guide covers drivers, CUDA and Docker.
A dedicated system is a single machine or a set of them: redundancy and off-server backups of checkpoints are part of your design, whichever card you choose.
Servers you can order without a quote
The RTX PRO Blackwell servers and the workloads they are sized for.
Multi-GPU Server
Four RTX PRO 6000 cards in one server for $5,966.44 a month.
96 GBRTX PRO 6000 Server
One 96 GB card, more memory than an H100, from $1,656.15 a month.
LLMLLM Hosting
Open-weight models matched to 24 GB, 96 GB and 384 GB.
InferenceAI Inference Server
An OpenAI-compatible endpoint on hardware you control.
CostGPU Cloud Alternative
Break-even hours between hourly GPU clouds and a monthly server.
OverviewAI Infrastructure
How MassiveGRID approaches GPU and HPC hosting projects.
Choosing between GPU classes
NVIDIA H100 vs H200: Key Differences in Performance, Memory & AI Power
NVIDIA H100 vs H200: The NVIDIA H100 and H200 are both high-performance GPUs based on NVIDIA’s Hopper architecture, designed for AI, high-performance.
GuideRTX PRO 6000 Blackwell vs H100 vs RTX 5090
RTX PRO 6000 Blackwell vs H100 vs RTX 5090: 96 GB, 80 GB and 32 GB compared on specs, vLLM benchmarks, licence terms and rental prices on 16 November 2026.
GuideA100 vs H100 vs RTX 6000 Ada: Which GPU for Which AI Workload
A100, H100 and RTX 6000 Ada compared for AI: memory bandwidth vs compute, FP8, NVLink, hourly vs monthly cost, and which card suits each workload.
GuideBest GPU for LLM Inference in 2026: Pick by Model Size, Not by Name
Best GPU for LLM inference in 2026, ranked by use case: 24 GB runs models to 32B, 96 GB runs 70B and 120B, with a VRAM table for 30 current open models.
GuideRenting a GPU Server vs Buying: The 3-Year Cost Model
GPU rent vs buy over three years: what a purchase really includes, published rental totals, the utilisation break-even, and when each option genuinely wins.
GuideGPU Server Pricing in 2026: Dedicated Monthly Servers vs Hourly Cloud
GPU server pricing in 2026: on-demand hourly and 730-hour cost at AWS, Azure, Google Cloud, Oracle and others on 16 November 2026, with break-even hours.
Frequently Asked Questions
Tell us what you need to run
Send the GPU model, count, memory, storage, network and term, and sales will answer with a configuration and a price. If 96 GB or 384 GB is enough, order a server online.