GPTrack Nvidia HGX B200 Blackwell Air 1,5TB, Rack Server for AI and HPC, 8U

Technical details of HGX B200 Blackwell Air 1.5TB:

8x Nvidia Blackwell B200 Tensor Core GPU
1440GB HBM3e memory
2x Intel Xeon Platinum 8570 (56-core) CPU
32x DDR5 5600 RDIMM 64GB
Total TDP up to 15kW
12x High-efficiency 3000W PSU
2x PCIe M.2 slots on board
12x PCIe gen5 2.5″ drive slots (NVMe)
8x HHHL PCIe Gen5 x16
2x FHHL PCIe Gen5 x16
2x RJ45 1GbE ports
5x USB 3.2 port
1x VGA port
1x RJ45 IPMI port
Halogen-free LSZH power cables
Stainless steel cage nuts
Rail kit
8U 448 x 354 x 930 mm (17.6 x 14 x 36.6″)
133 kg (290 lbs)
The price does not contain any taxes.
3-year manufacturer’s warranty
Free shipping worldwide.

Optional components

NIC Nvidia Bluefield-3
NIC Nvidia ConnectX-7/8
NIC Intel 100G
M.2 SSD
E1.S SSD
2.5″ SSD
Storage controller
Raid controller
OS preinstalled
Anything possible on request

Need something different? We are happy to build custom systems to your liking.

Nothing is more valuable than intelligence. Luckily, inferencing, tuning and training gigantic cutting-edge LLMs have become a commodity. Thanks to state-of-the-art, open-source LLMs you can download for free, the only thing you need is suitable hardware. We are proud to offer bleeding-edge Nvidia rack servers with the most competitive pricing in the world. From smaller air-cooled systems like the Nvidia GH200 Grace-Hopper Superchip and the HGX B200 and HGX B300 up to massive, NV-linked, liquid-cooled, CDU-integrated, ready-to-use Nvidia GB200 Grace-Blackwell Superchip and GB300 Grace-Blackwell Ultra Superchip systems. Multiple NV-linked (G)B200 or (G)B300 act as a single giant GPU with one single giant coherent memory pool. We are the only ones who offer smaller (half-size) systems than a complete NVL72 rack with “only” 18 superchips (NVL36). All systems are perfect for inferencing insanely huge LLMs, quick fine-tuning and training of LLMs, image and video generation and editing as well as high-performance computing.

Example use case 1: Inferencing Deepseek R1 0528 685B, GLM 4.5 355B-A32B, OpenAI OSS 120B, MiniMax-M1 456B, ERNIE-4.5-VL-424B-A47B, Kimi K2 0905 1T, Qwen3 Coder 480B A35B, Qwen3-235B-A22B 2507, Grok 2 270B and LongCat-Flash 560B

Deepseek R1 0528 685B, GLM 4.5 355B-A32B, OpenAI OSS 120B, MiniMax-M1 456B, ERNIE-4.5-VL-424B-A47B, Kimi K2 0905 1T, Qwen3 Coder 480B A35B, Qwen3-235B-A22B 2507, Grok 2 270B and LongCat-Flash 560B are the most powerful open-source models by far and even match GPT-o1/o3/o4/o5, Claude 4.1 Opus, Grok 4 and Gemini 2.5 Pro.

• Deepseek R1 0528 685B with 4-bit quantization needs at least 410GB of memory to swiftly run inference! Qwen3-235B-A22B 2507 Thinking with 4-bit quantization needs at least 134GB of memory to swiftly run inference! Luckily, GH200 has a minimum of 624GB, GB300 a minimum of 784GB. With GH200, Qwen3-235B-A22B 2507 Thinking in 4-bit can be run in VRAM only for ultra-high inference speed (significantly more than 50 tokens/s). Even faster is GH200 with OpenAI’s OSS 120B with >100 tokens/s. With GB200 Blackwell, as well as GB300 Blackwell Ultra this is also possible for Deepseek R1 0528 685B. With GB200 Blackwell and GB300 Blackwell Ultra you can expect up to 1000 tokens/s. If the model is bigger than VRAM you can only expect approx. 10-20 tokens/s. Surprisingly, Deepseek R1 0528 685B in 4-bit runs on our smallest system, the GH200, with up to 20 tokens/s. That is usable! 4-bit quantization is the best trade-off between speed and accuracy, but is natively only supported by Blackwell. We recommend using vLLM and Nvidia Dynamo for inferencing.

Example use case 2: Fine-tuning Deepseek R1 0528 685B with PyTorch FSDP and Q-Lora

• Models need to be fine-tuned on your data to unlock the full potential of the model. But efficiently fine-tuning bigger models like Deepseek R1 0528 685B remained a challenge until now. This blog post walks you through how to fine-tune Deepseek R1 using PyTorch FSDP and Q-Lora with the help of Hugging Face TRL, Transformers, peft & datasets.
• Fine-tuning big models within a reasonable time requires special and beefy hardware! Luckily, GH200, GB200/GB300 or B200/B300 are ideal to complete this task extremely quickly.

Example use case 3: Generating videos with Mochi1, HunyuanVideo, MAGI-1 or Wan 2.2

• Mochi1, HunyuanVideo, MAGI-1 and Wan 2.2 are democratizing efficient video production for all.
• Generating videos requires special and beefy hardware! Mochi1 and HunyuanVideo need at least 80GB of VRAM. Luckily, GH200, B300 and GB300 are ideal for this task. GH200 has a minimum of 144GB, GB300 a minimum of 288GB, B200 a minimum of 1.5TB.

Example use case 4: Image generation with HiDream-I1, Flux.1, SANA-Sprint, HunyuanImage-2.1 or SRPO.

• HiDream-I1, Flux.1, HunyuanImage-2.1 and SRPO are the best image generators at the moment. And they are uncensored, too. SANA-Sprint is very fast and efficient.
• In high-speed inference, FLUX requires approximately 33GB of VRAM for maximum speed. For training the FLUX model, more than 40GB of VRAM is needed. SANA-Sprint requires up to 67GB of VRAM.Luckily, GH200 has a minimum of 144GB, GB300 a minimum of 288GB, B200 a minimum of 1.5TB.

Example use case 5: Image editing with Qwen-Image-Edit-2509, FLUX.1-Kontext-dev, Omnigen 2, Nvidia Add-it, HiDream-E1 or ICEdit.

• Qwen-Image-Edit-2509, FLUX.1-Kontext-dev, Omnigen 2, Add-it, HiDream-E1 and ICEdit are the most innovative and easy-to-use image editors at the moment.
• For maximum speed in high-resolution image generation and editing, beefier hardware than consumer graphics cards is needed.Luckily, GH200, GB200, GB300, B200 and B300 excel at this task.

Example use case 6: Video editing with AutoVFX, Skyreels-A2, VACE or Lucy Edit

• AutoVFX, SkyReels-A2, VACE and Lucy Edit are the most innovative and easy-to-use video editors at the moment.
• For maximum speed in high-resolution video editing, beefier hardware than consumer graphics cards is needed. Luckily, GH200, GB200, GB300, B200 and B300 excel at this task.

Example use case 7: Deep Research with WebThinker or Tongyi DeepResearch

• WebThinker and Tongyi Deepresearch enable large reasoning models to autonomously search, deeply explore web pages, and draft research reports, all within their thinking process.
• The hardware requirements for using Webthinker depend on the particular LLM of choice (see above).

Example use case 8: Creating a Large Language Model from scratch

• Imagine stepping into the world of language models as a painter stepping in front of a blank canvas. The canvas here is the vast potential of Natural Language Processing (NLP), and your paintbrush is the understanding of Large Language Models (LLMs). This article aims to guide you, new to NLP, in creating your first Large Language Model from scratch, focusing on the Transformer architecture and utilizing TensorFlow and Keras.
• Training a LLM from scratch within a reasonable time requires special and extremely beefy hardware! Luckily, GH200, GB200 and GB300 or B200/B300 are ideal for this task.

Why should you buy your own hardware?

“You’ll own nothing and you’ll be happy?” No!!! Never should you bow to Satan and rent stuff that you can own. In other areas, renting stuff that you can own is very uncool and uncommon. Or would you prefer to rent “your” car instead of owning it? Most people prefer to own their car, because it’s much cheaper, it’s an asset that has value and it makes the owner proud and happy. The same is true for compute infrastructure.

Even more so, because data and compute infrastructure are of great value and importance and are preferably kept on premises, not only for privacy reasons but also to keep control and mitigate risks. If somebody else has your data and your compute infrastructure you are in big trouble.

Speed, latency and ease-of-use are also much better when you have direct physical access to your stuff.

With respect to AI and specifically LLMs there is another very important aspect. The first thing big tech taught their closed-source LLMs was to be “politically correct” (lie) and implement guardrails, “safety” and censorship to such an extent that the usefulness of these LLMs is severely limited. Luckily, the open-source tools are out there to build and tune AI that is really intelligent and really useful. But first, you need your own hardware to run it on.

What are the main benefits of GH200 Grace-Hopper, GB200 Grace-Blackwell and GB300 Grace-Blackwell Ultra?

They have enough memory to run, tune and train the biggest LLMs currently available.

Their performance in every regard is almost unreal (up to 10,000 times faster than x86).

There are no alternative systems with the same amount of memory.

Ideal for AI, especially inferencing, fine-tuning and training of LLMs.

Multiple NV-linked GB200 or GB300 act as a single giant GPU.

Optimized for memory-intensive AI and HPC performance.

Ideal for HPC applications like, e.g. vector databases.

Easily customizable, upgradable and repairable.

Privacy and independence from cloud providers.

Cheaper and much faster than cloud providers.

They can be very quiet (with liquid-liquid CDU).

Reliable and energy-efficient liquid cooling.

Flexibility and the possibility of offline use.

Gigantic amounts of coherent memory.

They are very power-efficient.

The lowest possible latency.

They are beautiful.

CUDA enabled.

Run Linux.

What is the difference to alternative systems?

The main difference between GH200/GB200/GB300 and alternative systems is that with GH200/GB200/GB300, the GPU is connected to the CPU via a 900 GB/s chip-2-chip NVLink vs. 128 GB/s PCIe gen5 used by traditional systems. Furthermore, multiple GB200/GB300 superchips and HGX B200/B300 are connected via 1800 GB/s NVLink vs. orders of magnitude slower network or PCIe connections used by traditional systems. Since these are the main bottlenecks, GH200/GB200/GB300’s high-speed connections directly translate to much higher performance compared to traditional architectures. Also, multiple NV-linked (G)B200 or (G)B300 act as a single giant GPU with one single giant coherent memory pool. Since even PCIe gen 6 is much slower, Nvidia does not offer B200 and B300 as PCIe cards any more (only as SXM or superchip). We highly recommend choosing NV-linked systems over systems connected via PCIe and/or network.

What is the difference to server systems of competitors?

Pricing: We aim to offer the most competitive pricing worldwide.

Size: A single GB200/GB300 NVL72 rack gives you more than an exaflop of compute. For some people, one complete rack is more than needed and too expensive. That is why we also offer smaller (half-size) systems with only 18 superchips (NVL36). We are, to our knowledge, the only ones in the world where you can get systems smaller than a complete GB200/GB300 NVL72 rack.

In-rack CDU: Our rack server systems are available with liquid cooling and a CDU integrated directly into the rack. You can choose between an air-liquid and liquid-liquid CDU.

Ready-to-use: Our systems can be ordered fully integrated and ready-to-use. Everything that is needed is included and tested. All you need to do is plug your system in to run it.

Shipping: Free shipping worldwide.

Warranty: 3-year manufacturer’s warranty.