Engineered around every layer of the stack
NVIDIA
ASUS
DeepSeek
Qwen
Anthropic
OpenAIMost GPU clouds resell someone else's datacentre. TAIDA designs the power, cooling, compute and software layers together, so the cost of every token is decided by engineering rather than by a landlord's margin.
Density and utilisation are the two levers. Liquid cooling fits more GPUs in a rack, scheduling keeps them busy, and the saving is passed through in the per-hour and per-token rate rather than absorbed as margin.
Direct-to-chip liquid cooling is designed in from day one rather than retrofitted. It supports the rack densities Blackwell Ultra needs and removes the airflow overhead that inflates power draw in conventional halls.
Every layer answers to the same engineering team. When a job underperforms there is no hand-off between a hosting provider, a hardware vendor and a software partner: one team owns the root cause from the chip to the endpoint.
The AI factory is built from a repeatable pod that follows the NVIDIA Reference Architecture: a fixed number of racks, a known power envelope, a known cooling loop, validated fabric and storage. Capacity is added by deploying more of the same unit, so lead times and performance are predictable and the next GPU generation slots into the same footprint.
Every layer TAIDA operates, and what sits inside it.
DeepSeek
Qwen
Kimi
Nemotron
Image, video, speech* In development.
Two node classes, one silicon generation. Both run NVIDIA Blackwell Ultra; they differ in who they are for and how much of the factory sits behind them.
Commercial, server-grade nodes for training and high-throughput inference, deployed in liquid-cooled pods.
| GPUs per node | 8× NVIDIA B300 |
|---|---|
| GPU memory | TBC GB HBM3e per GPU |
| Interconnect | NVLink · NVSwitch |
Desk-side Grace Blackwell nodes for start-ups and research teams to prototype and test on production silicon.
| Configuration | Grace CPU + Blackwell Ultra GPU |
|---|---|
| Unified memory | TBC GB coherent CPU–GPU |
| Form factor | Desk-side / rack-mountable |
Where TAIDA capacity lives today. Explore the facility layer by layer.
Two fully independent power paths, each sized for the entire load. Maintenance or failure on one path leaves the hall running on the other with no reduction in capacity.
Every cooling stage carries one unit more than the load needs. A chiller, pump or air handler can drop out for service without the hall warming up.
Facility certifications held by the AIMS Kuala Lumpur 2 site. Suitable for regulated financial-services and public-sector workloads in Malaysia and Singapore.
The output of the AI factory, delivered as a service. Run leading open-weight text, reasoning, vision, image, video and speech models, plus your own, through one API on TAIDA compute.
Tell us the workload and we will size the cluster, the terms and the timeline. Replies within one working day.
