Built by EMARQUE
AI Work 100
- Max GPU memory per card
- 96 GB
- RTX PRO Blackwell GPUs
- Up to 2
- Max system memory
- 192 GB
The price depends on the GPU model, whether you take one card or two, and the memory and storage you choose.

Key features
One or two RTX PRO Blackwell GPUs
Choose the RTX PRO 4500 with 32 GB, the RTX PRO 5000 with 48 GB or the RTX PRO 6000 with 96 GB, each with ECC GDDR7. A 7B model runs at 16-bit precision on any of them, and a 4-bit 70B model fits on one RTX PRO 5000 or 6000.
Room for the work around the model
Sixteen Ryzen 9 cores, up to 192 GB of DDR5 and up to 8 TB of Gen5 NVMe, so data preparation, the working dataset and the experiment cache all stay on the machine.
Fits under a desk
An air-cooled tower that runs from a standard 240 V wall socket. It does not need a rack or any change to the room.
Assembled and tested by us in Malaysia
Built and tested here, with the runtime you name installed before delivery. Warranty and parts are handled in Malaysia.
Is the AI Work 100 right for you?
Researchers, developers and small teams who want an on-prem AI workstation at the desk rather than a rackmount server, assembled and supported by us in Malaysia.
Choose it when
- You are running 7B to 70B open-weight models for yourself or a small team.
- One or two RTX PRO Blackwell cards give you the GPU memory you need: 32, 48 or 96 GB on a card, up to 192 GB across two.
- The machine has to live in an office, on a standard wall socket.
Choose something else when
- A whole department will be served from it. The AI PRO 500 is the step up.
- You need four GPUs, or ECC registered memory, which this consumer platform does not have.
- The workload is bound by bulk storage rather than by GPU memory.
When you outgrow it
Move to the AI PRO 500 when you need four GPUs, ECC memory, or a machine that lives in a rack. Move to the AI Server when one machine has to serve a whole organisation.
Use cases
- Private assistant
A model that answers from your own files
Serve an open-weight chat model from the workstation, so the prompts and the documents stay on it.
- Fine-tuning
LoRA and QLoRA on your own data
Train an adapter on a domain dataset overnight, then evaluate it on the same machine the next morning.
- Before a server
Prove the workload first
Build the retrieval pipeline and measure it here, then move to the AI PRO 500 when more people need it.
Specifications
The final parts list is confirmed in your quote.

- GPU
- 1 or 2 × NVIDIA RTX PRO 4500, 5000 or 6000 Blackwell · 32, 48 or 96 GB GDDR7 with ECC per card (up to 192 GB across two)
- GPU interconnect
- PCIe Gen5 ×16. No NVLink: the two cards do not share memory
- CPU
- AMD Ryzen 9 9950X, 16 cores / 32 threads
- Memory
- Up to 192 GB DDR5 across 4 DIMM slots, non-ECC
- Primary storage
- Up to 8 TB Gen5 NVMe (RAID 0 or 1)
- Bulk storage
- Optional SATA SSD or HDD for datasets
- Cooling
- Air cooling, sized to the GPU model and count
- Networking
- 2.5 GbE on board · 10 GbE optional add-in card
- Form factor
- Tower workstation, standard 240 V wall socket
- OS
- Ubuntu 24.04 LTS or Windows 11 Pro
- Warranty
- 3-year EMARQUE system warranty · 90-day exchange · lifetime labour
Questions about the AI Work 100
Can it run a 70B model?
Quantised to 4 bits, yes. A 4-bit 70B model needs roughly 40 GB, so it fits on one RTX PRO 5000 or RTX PRO 6000. On the RTX PRO 4500 it takes two cards, and because they do not pool their memory the model is split across them over PCIe. At 16-bit precision a 70B model needs about 140 GB, which takes two RTX PRO 6000 cards; we confirm the card edition and power supply in your quote. If you need more than two cards, look at the AI PRO 500.Which GPU should I choose?
Choose by the models you plan to run. The RTX PRO 4500 (32 GB) runs 7B and 13B models at 16-bit precision. The RTX PRO 5000 (48 GB) holds a 4-bit 70B model on one card. The RTX PRO 6000 (96 GB) leaves room for longer context and larger models. Tell us the models and we will recommend one in the quote.One GPU or two?
One card runs 7B and 13B models at 16-bit precision. A second card lets a larger model be split across both, or keeps one model loaded while a second job uses the other card. Adding a card later means labour and retesting, so decide at the quote if you expect to need two.The system memory is not ECC. Does that matter?
This build uses non-ECC DDR5; the memory on the RTX PRO cards has ECC. For serving models and short fine-tuning runs that is normal. If you need ECC registered memory, more PCIe lanes and four GPUs, the AI PRO 500 on Threadripper PRO is the step up.What is the warranty?
Three years from us, covering assembly, configuration and thermal performance, with a 90-day exchange on failed parts and lifetime labour. The component makers' warranties apply on top, and we handle every claim in Malaysia.
Get a quote for the AI Work 100.
Tell us what it needs to run and who will use it. We reply with a written quote in MYR.
