Hardware
Your own AI server, in your own office.
Four graphics cards with 128 GB of memory on an Intel Xeon workstation. Your whole team shares it, and what you send it stays in the building. $24,450. A $1,000 deposit reserves yours.
8–16
Users at the same time, depending on the model.
32
People on one team, at typical use. Not everyone asks at the same moment.
128 GB
Graphics memory across four cards, 32 GB each.
$24,450
Complete. Intel build or Radeon build, same price.
Why run your AI on your own server
It stays in the building
The server sits on your office network. Prompts, files and answers stay on it. No outside AI company sees them or trains on them.
One server for the whole team
People and agents use it at the same time. Requests run side by side across the four cards.
No meter on every use
Rented AI bills for every request. This server does not. Your team can use it as much as the work needs.
Built and tested before it ships
We assemble it, install the software, load the models and run our speed checks. It arrives ready to answer.
Models kept current
We test new open models on the same hardware. When one is better for your work, we load it onto your server.
Works with your IT company
They keep the network, the backups and the repairs. We look after the AI software on top.
Two builds, one price
Both builds use the Intel Xeon workstation our own server runs on, with all four card slots filled. You pick the cards.
Intel build
4 x Intel Arc Pro B70
32 GB each, 128 GB in total
The cards in our own server. It runs our AI every day.
$24,450
$1,000 deposit reserves it
Radeon build
4 x AMD Radeon AI PRO R9700
32 GB each, 128 GB in total
Four AMD cards with the same 128 GB of graphics memory.
$24,450
$1,000 deposit reserves it
Speed and capacity
Planning figures for a four-card, 128 GB server, by model. Your models and the way your team works change them.
Qwen3.8-27B, 8-bit (FP8)
- Users at once
- 8
- Reading speed, tok/s
- 7,025
- Answer speed, all users, tok/s
- 813.9
- Context, tokens
- 139,264
Qwen3.8-27B, 4-bit (MXFP4)
- Users at once
- 16
- Reading speed, tok/s
- 9,831
- Answer speed, all users, tok/s
- 1,426
- Context, tokens
- 131,072
| Model | Users at once | Reading speed, tok/s | Answer speed, all users, tok/s | Context, tokens |
|---|---|---|---|---|
| Qwen3.8-27B, 8-bit (FP8) | 8 | 7,025 | 813.9 | 139,264 |
| Qwen3.8-27B, 4-bit (MXFP4) | 16 | 9,831 | 1,426 | 131,072 |
From published results for a four-card AMD Radeon AI PRO R9700 server. Answer speed is the total for all users at once, with 4K tokens of context each. The Intel build can differ.
Need more? Add a second server.
Each server adds the same capacity again: another 8 to 16 users at the same time, for another team of about 32.
- tok/s:
- tokens a second, the speed an answer comes out. A token is about three quarters of a word.
- Context:
- how much text the model reads for one request: the files, the history and the question.
Plug it in. Your team uses it.
We build it, load the models and test it before it ships. On your side it is three steps.
Power
Put it in your office or in your IT company's rack, and plug it in.
Office network
Connect it to your network. It gets one address there.
Your team
People and AI tools reach it at that address. Apps made for cloud AI services connect the same way.
Make it know your business
It runs open models when it arrives. We can also build an AI system on how your business works and run it on this server. Your know-how stays yours.
Need a specific open model? Ask in the form. We test it on your cards before we load it.
Own your AIReserve your build
$24,450
Intel build or Radeon build. Same price.
- Cards
- 4 x Intel Arc Pro B70 or 4 x AMD Radeon AI PRO R9700, 32 GB each, 128 GB in total
- Processor
- Intel Xeon 654, 18 cores, 36 threads
- Board
- ASUS Pro WS W890E-SAGE SE, with remote management
- Memory
- 128 GB DDR5-5600 ECC (4 x 32 GB), four slots free
- Storage
- 4 TB NVMe SSD
- Network
- Two 10 Gb Ethernet ports
- Software
- Linux, the AI server software and your models, installed and tested
How the $1,000 deposit works
- 01
Pay the $1,000 deposit on Stripe's secure checkout. Pick Intel or Radeon there.
- 02
An engineer emails you from [email protected] to confirm your order.
- 03
The deposit counts toward the $24,450. We then build and test your server.
Reserve now
Pay the $1,000 deposit on Stripe's secure checkout. You pick Intel or Radeon there.
Pay the $1,000 depositQuestions first?
Send the form. An engineer replies from [email protected].