What matters for private AI
Starting point: EMD Lab Studio, EMD Lab AI or the QNAP AI server
Get a quote for this
Your AI. Your data.
Private AI, running on your own hardware.
Find the right AI model, and the machine to run it, for your business. Answer 7 quick questions to get a tailored recommendation.
What should your AI do? Pick one to start.
- Private AI assistant
- Chat with company documents
- Coding
- Image generation
- Customer-service chatbot
- Scanning documents
- Transcription and translation
- Fine-tuning on our own data
Already have hardware? See which models it runs
- 25 yearsdistributing PC hardware in Indonesia
- About 500active resellers
- Brands we partner withAMD, ASUS, ASRock and Supermicro
What matters for your work
Build-to-order workstations and servers from EMD Lab, configured for video, 3D, engineering and AI work, and delivered across Indonesia.
Pick your main work to see what matters most
Brands we build with
ASUS ExpertCenter Pro ET900N G3
A deskside AI supercomputer on NVIDIA DGX Station architecture: 748 GB of memory that the processor and the graphics chip share, in one tower, for teams that run the largest open models on their own premises.
- Model
- ASUS ExpertCenter Pro ET900N G3
- Superchip
- NVIDIA GB300 Grace Blackwell Ultra: a 72-core Arm processor and a Blackwell Ultra graphics chip in one package
- Memory
- 748 GB: 252 GB of fast graphics memory (HBM3e) plus 496 GB of system memory (LPDDR5X)
- AI performance
- Up to 20 petaFLOPS (ASUS preliminary figure)
- Price
- Rp 3.499.000.000, contact us for details
Office AI server: QNAP QAI-h1290FX
Ask your company files anything. Nothing leaves the office.
Your team asks in Bahasa Indonesia or English and gets the answer with the file it came from, searching only folders they are already allowed to open. One box is your file server and your AI.
- Office AI
- Teams of up to about 30 people (RTX PRO 4500 Blackwell, 32 GB)
- Company AI
- About 30 to 100 people, or 120B-class models (RTX PRO 6000 Blackwell Max-Q, 96 GB)
- Included
- Models and apps installed, users and folder permissions set up, and a written test report
EMD Lab ProArt PC
Two creator PCs.
One ProArt ecosystem.
Wood, glass and ProArt parts, neatly built and fully tested by EMD Lab, for work that depends on accurate colour and predictable render times.
Price on request for the ProArt PC 401 and PC 602
- Ryzen 9 9950X16 cores, 32 threads
- RTX 50 SeriesProArt 5070 Ti, 5080, 5090
- 2 × USB440 Gbps per port
- 10 GbE + WiFi 7Studio networking
- Tested before shippingCPU, GPU and memory stress tests
DaVinci Resolve
Adobe Premiere Pro
Adobe After Effects
Blender
Cinema 4D and Redshift
Unreal Engine
Autodesk Revit
AutoCAD
Rhino
Lumion and Enscape
Local LLMs
PyTorch
ComfyUI
MATLAB
Python data science
Virtualization
Expected performance
Two graphics cards with similar specs can perform very differently in the same application. These ranges show how the parts we build with compare in real software.
What each machine runs
Estimated speed for one person, with 4-bit (compressed) models and 32K context (about 32,000 tokens of conversation and documents kept in mind). A token is a short piece of a word; for scale, people read at roughly 5 tokens per second. “Fits” means the whole model fits in the graphics card’s memory (VRAM); “partly fits” and “does not fit” mean some or most of it runs from much slower system memory. Calculated by the EMD Lab AI selector, not measured; lab measurements will replace these.
| Model | Good for | Studio, RTX 5090 (32 GB) | AI, 2 × RTX PRO 6000 Max-Q (192 GB) |
|---|---|---|---|
| gpt-oss-20b | Fast general assistant | Fits, about 200 tokens/s | Fits, about 200 tokens/s |
| Qwen3.8 27B | Office assistant, document questions, coding help | Fits, about 65 tokens/s | Fits, about 65 tokens/s |
| Llama 3.3 70B | Large general model | Does not fit, about 2 tokens/s | Fits, about 25 tokens/s |
| gpt-oss-120b | Stronger reasoning for many users | Partly fits, about 20 tokens/s | Fits, about 175 tokens/s |
Why choose EMD Lab?
The parts are the same everywhere. The difference is who checks them, who builds and tests the machine, and who supports you after delivery.
Tested before it ships
Every system is assembled to the approved specification, then stress-tested. The results are recorded and travel with the machine, so problems are found here, not on your desk.
Parts from the source
We partner with AMD, ASUS, ASRock and Supermicro and build from our own stock: genuine parts, not parallel imports. Warranty claims go through us instead of a chain of shops.
Help before and after you buy
Our teams in Jakarta, Bandung and Yogyakarta help you choose before you buy. After delivery we support you online, and we ship across Indonesia with the specification and the test report.
Specialists who ask first
We start from your software, your projects and your budget, and recommend a specification with the reason behind each part. No upsell to parts your work does not use.
Guides
Practical advice on choosing hardware for your work.
AIHow much GPU memory do local LLMs need?
New
ComparisonGuideCloud AI or your own box?
Where your documents go, what it costs, and what each does better.
HardwareToolWhich hardware runs which AI?
Pick a card or AI box and see which models fit.
Clients we have supplied
Companies, universities and public institutions across Indonesia.
Tell us about your work
Send us the software you use and what slows you down. A specialist will reply with a recommended specification and a quote.
What happens after you message us
VisitJakarta (Mangga Dua), Bandung, Yogyakarta