DGX Spark on a desk
A 15 by 15 cm computer with 128 GB of unified memory that NVIDIA says runs models of up to 200 billion parameters. For developers, testing and small teams; two to four units can be joined together.
NVIDIA GPU servers for AI
We supply and set up almost the whole NVIDIA range: the DGX Spark desktop, RTX 5090 and RTX PRO 6000 cards, eight-GPU H200 and Blackwell servers, and NVL72 racks for data centres. Serving software is built for NVIDIA first, so almost every open model runs on it from the day it is released.
Why NVIDIA
vLLM, SGLang, TensorRT-LLM and llama.cpp are built and tested on NVIDIA CUDA first, and model makers usually publish their official serving instructions for NVIDIA cards. NVIDIA is the best choice when dozens of people use the system at once, or when you want to run a new model from the day it ships. Its limit is memory per card: a large model needs several cards and a more expensive server.
Which tier for which job
One or more cards in a workstation
RTX 5090 with 32 GB for models up to about 30 billion parameters, and RTX PRO 6000 with 96 GB for models of about 70 billion at 8-bit; for a department or a team.
A four- or eight-card server
Eight RTX PRO 6000 Server Edition cards with 768 GB in total, or H200 and B200 servers, for large models and dozens of concurrent users across an organisation.
Blackwell Ultra and Rubin for data centres
HGX B300 with 2.1 TB of memory in one server and liquid-cooled NVL72 racks, for a data centre with the power and cooling to run trillion-parameter models for a whole organisation.
Data-centre cards and servers
Specifications from NVIDIA's official pages, reviewed on 22 October 2026.
| Product | Memory per GPU | Bandwidth per GPU | In an eight-GPU server | What fits |
|---|---|---|---|---|
| H100 | 80 GB HBM3 | 3.35 TB/s | 640 GB | MoE models up to about 400B parameters at 8-bit |
| H200 | 141 GB HBM3e | 4.8 TB/s | 1.1 TB | Models such as DeepSeek-V4-Flash and GLM-5.2 at 8-bit |
| B200 | 180 to 192 GB HBM3e | 8 TB/s | About 1.4 TB | Models up to about one trillion parameters at 8-bit, with little room for context |
| B300 (Blackwell Ultra) | 288 GB HBM3e | 8 TB/s | 2.1 TB | Trillion-parameter models such as Kimi K2.7 Code at 8-bit |
| Vera Rubin (HGX Rubin NVL8) | 288 GB HBM4 | 19 to 22 TB/s (NVIDIA's pages give two figures) | About 2 TB | The B300 class with faster memory |
| GB300 NVL72 and Vera Rubin NVL72 racks | About 288 GB | NVLink across all 72 GPUs | About 20 TB per rack | Several trillion-parameter models at once; liquid-cooled |
| RTX PRO 6000 Blackwell Server Edition | 96 GB GDDR7 | 1.6 TB/s | 768 GB | A model of about 70B per card, in an ordinary air-cooled server |
| L40S | 48 GB GDDR6 | 864 GB/s | 384 GB | A model up to about 30B at 8-bit per card |
| A100 (40 and 80 GB) | 40 or 80 GB HBM2e | Up to 2 TB/s | Up to 640 GB | Previous generation; if you own them, we set a suitable model up on them |
The "what fits" column is our estimate, not a vendor claim: each billion parameters takes about 1 GB of memory at 8-bit and about half that at 4-bit, and room must be left for context and concurrent users.
Desktop and workstation
Specifications from NVIDIA's official pages, reviewed on 22 October 2026.
| Product | Memory | Bandwidth | Power | What fits |
|---|---|---|---|---|
| DGX Spark (GB10) | 128 or 64 GB unified | 273 GB/s | 240 W power supply | Up to 200B parameters according to NVIDIA; two units up to 400B |
| DGX Station (GB300) | 252 GB HBM3e plus 496 GB LPDDR5X | 7.1 TB/s (HBM) | 1,600 W | A model of about 200B at 8-bit in fast memory |
| RTX PRO 6000 Blackwell (Workstation and Max-Q) | 96 GB GDDR7 | 1.8 TB/s | 600 or 300 W | A model of about 70B at 8-bit |
| RTX PRO 5000 Blackwell | 48 or 72 GB GDDR7 | 1.3 TB/s | 300 W | A model of about 30B to 50B at 8-bit |
| RTX 5090 | 32 GB GDDR7 | 1.8 TB/s | 575 W | A model up to about 30B at 4-bit |
| RTX 4090 and RTX 3090 | 24 GB GDDR6X | About 1 TB/s | 350 to 450 W | A model up to about 27B at 4-bit; if you own them, we set it up on them |
| RTX A6000 | 48 GB GDDR6 | 768 GB/s | 300 W | A model of about 30B at 8-bit |
The "what fits" column is our estimate, not a vendor claim: each billion parameters takes about 1 GB of memory at 8-bit and about half that at 4-bit, and room must be left for context and concurrent users.
The software we set up on it
| Layer | What we install and tune |
|---|---|
| Serving for many users | vLLM or SGLang, both of which batch concurrent requests together |
| Light and test serving | llama.cpp or Ollama for a single user, testing and small models |
| Programming interface | An OpenAI-compatible API, so applications that call a foreign service today connect to the local model without a rewrite |
| Assistant and documents | A Persian assistant, search over internal documents with citations (RAG), and n8n workflows |
| Sign-in and monitoring | Sign-in with company accounts, logging of questions and answers, and monitoring of GPU use, request queues and response time |
From the first call to a working system
Hardware is chosen after a model has been measured on your own work, not from a table. A quote is issued only once it is clear which model has to serve how many people, and how fast.
- 01
First conversation
We go through the task, the number of users, the data, where the hardware goes and what you already own.
- 02
Measured on your work
We test several open-weight models on samples of your real work and size memory and speed from those results.
- 03
Written quote
You get the exact configuration, the proposed model, the delivery time and the support terms in writing. Prices appear only in the quote.
- 04
Installation
We install the hardware, bring up the model and connect the assistant to your systems, then test it once with international internet cut off.
- 05
Handover and support
You get documentation and training for your team; afterwards we maintain the system and update the models under contract.
How to get in touch
Fill in the form or message us on WhatsApp. The clearer these are, the more exact the quote:
- What the model has to do (answer from documents, code, summarise, process images or speech)
- How many people use it at the same time, and what response time is acceptable
- The model or hardware family you have in mind, if you already know
- Where it goes: your server room, a data centre, or hardware you already own
- How sensitive the data is, and whether it must run with no internet at all
We reply within one business day.
Zharf
Try the ZharfAI assistant right now
Ask about Persian assistants, document automation, or how to start a pilot.
Frequently asked questions about NVIDIA GPU servers
Is an RTX 5090 enough for AI?
For models up to about 30 billion parameters, yes. The RTX 5090 is a gaming card with 32 GB of memory at 1.8 TB/s; it is enough for a model such as Qwen3.8-27B (about 17 GB at 4-bit), and the official vLLM recipe for that model was verified on two RTX 5090 cards. For a larger model, or round-the-clock work in a server, the RTX PRO 6000 with 96 GB is the better fit.
H100, H200 or B200?
All three are data-centre cards and differ mainly in memory: 80 GB on the H100, 141 GB on the H200 and about 180 to 192 GB on the B200 per card. More memory means a larger model in one server and room for more concurrent users. If your model fits on H100s, you do not need a newer generation.
What is DGX Spark?
NVIDIA's small desktop computer with the GB10 chip and 128 GB of unified memory; a 64 GB version from other makers arrives in late October 2026. NVIDIA says one unit runs models of up to 200 billion parameters and that two to four can be joined. Its memory bandwidth (273 GB/s) is lower than RTX cards, so it writes answers more slowly; it suits development and small teams.
Gaming card or server card?
Gaming cards such as the RTX 5090 cost less but have less memory and are not designed for round-the-clock work in a rack. RTX PRO and data-centre cards have more memory with error correction (ECC) and cooling suited to servers. A gaming card is enough for testing and a small team; a system the whole organisation depends on needs server cards.
Do you rent GPU servers by the hour or month?
No. ZharfAI does not rent GPUs by the hour or month. There are three ways to work with us: we supply the hardware and install it on your premises, we run a dedicated machine for you and maintain it ourselves, or we set the model up on hardware you already own. If you only need a few hours of compute for a test, say so in the first conversation and we will suggest a suitable route.
How much does it cost?
We do not publish prices, because they depend on the configuration, quantity, the day's market and how it is delivered. After the first conversation and a test on your work you receive a written quote that also names the exact hardware model, the delivery time and the support terms.
Does it work without international internet?
Yes, if the model, the serving software, the assistant, user sign-in and monitoring all sit inside your network. Anything that depends on a foreign service stops during a cut, so before handover we test the system with international internet deliberately cut off.
Tell us how many people the model has to serve
From the model, the number of users and the response time you need, we propose the right card and server, and if one card is enough, we will not propose an eight-card server.