01
On-demand instances
Request an H100 or L40S instance in the shape your workload needs. No reservation required to start.
Solution · AI Computing
Sovereign GPU compute, scheduled around your workload
NVIDIA H100 and L40S capacity inside Malaysia — provisioned per workload, billed per second, and never routed across a border.
REF · SC-2026-AIC
Compute · NVIDIA H100 / L40S Billing · Per second Residency · Malaysia Data plane · In-country
Solution · AI Computing
Capacity that behaves like infrastructure, not a cross-border lease
Most GPU clouds treat Malaysia as an edge location bolted onto a global control plane. Bijak Cloud is built the other way around: the compute, storage and audit boundary are all in-country, and the control plane is designed to keep them there.
Elastic GPU capacity without exporting the data plane. Choose the GPU shape, attach sovereign storage, and let the scheduler handle placement and isolation.
- H100 80GB for large-model training, fine-tuning and high-throughput inference.
- L40S 48GB for inference, document AI, graphics and video pipelines.
- Per-second billing with instant hibernation when a workload goes idle.
- Persistent volumes, private networking and audit logs stay inside Malaysian jurisdiction.
Capabilities
The compute surface
The primitives teams need to move from a notebook to a governed production workload.
02
Instant hibernation
Stop paying for compute when a job finishes or a team steps away. Hibernation preserves the instance state for a fast resume.
03
Framework-ready images
Start from maintained images with CUDA, PyTorch, vLLM, TGI and common data-science tooling preinstalled.
04
Private networking
Place instances on an isolated private network and reach them over your approved connectivity path.
05
Sovereign storage
Attach S3-compatible object storage and persistent volumes that remain in Malaysia with encryption at rest.
06
Audit-ready activity
Provisioning, access and lifecycle events are written to immutable audit logs you can export to your SIEM.
Hardware
GPU configurations
Public hardware specifications for the two GPU families currently offered. Availability is confirmed per order.
NVIDIA H100
SXM · 80GB- Memory
- 80GB HBM3
- Memory bandwidth
- 3.35 TB/s
- Interconnect
- NVLink / NVSwitch
- Shapes
- 1× · 2× · 4× · 8× GPU
Large-model fine-tuning, training runs, high-throughput inference and HPC.
NVIDIA L40S
PCIe · 48GB- Memory
- 48GB GDDR6 ECC
- Memory bandwidth
- 864 GB/s
- Interconnect
- PCIe Gen4
- Shapes
- 1× · 2× · 4× · 8× GPU
Inference, RAG, document AI, rendering, simulation and video pipelines.
Availability is confirmed per order.
Process
From request to workload
A short operating path with an explicit hand-off at each stage. Timelines are agreed in the statement of work, not assumed from a marketing page.
- 01
Scope
Share the model, dataset size, latency target and residency constraints. We map them to a GPU shape and quota.
- 02
Provision
Capacity is allocated in-country, placed on your private network and connected to the storage you designate.
- 03
Validate
Run your workload against a representative dataset, confirm throughput and capture the acceptance result.
- 04
Operate
Use per-second billing, hibernation and audit logs in production, with a named escalation path for incidents.
Use cases
Workload patterns that fit
Representative patterns — not customer stories. Each can be scoped as a pilot before a production commitment.
Fine-tuning and training
Adapt open models to Malaysian legal, financial or clinical corpora without moving data offshore.
Production inference
Serve internal or customer-facing models with predictable capacity and in-country logs.
Document intelligence
Process PDFs, scans and records through OCR, extraction and retrieval pipelines that stay onshore.
Simulation and analysis
Run GPU-accelerated analytics, rendering and risk models next to the data they depend on.
Controls
Sovereignty controls
Design goals and shipped controls that keep the data plane inside Malaysian jurisdiction.
Malaysia-only data plane
Compute, storage and backups are designed to remain in-country, and support is delivered from Malaysia. Cross-border transfer requires written authorisation.
PDPA 2010 alignment
Controls are mapped to Akta 709 obligations for personal-data handling, retention and access.
Immutable audit trail
Every provisioning and access event is recorded and exportable for internal or regulator review.
Bahasa Melayu-native
Tokenisation, retrieval and evaluation support Bahasa Melayu as a first-class language, not an afterthought.
FAQ
Questions procurement teams ask
Where does my data actually reside?
The compute, storage and audit data plane is designed to remain in Malaysia, and support is delivered from Malaysia. Any cross-border transfer or remote access from outside Malaysia requires your written authorisation and is recorded in the audit log.
How is GPU usage billed?
Compute is metered per second while an instance is running. Hibernated instances stop consuming GPU billing; attached storage is billed separately. Your quote itemises every line.
Can I reserve capacity for a long-running project?
Yes. Reserved and dedicated shapes are available alongside on-demand capacity. The commitment length, capacity floor and commercial terms are agreed in the statement of work.
What happens to model weights and checkpoints?
Model weights, checkpoints and training data remain yours. They are stored in the storage location you designate and are not reused for any other purpose without written instruction.
Solution · AI Computing
Size your first workload
Send the model, dataset and residency constraints. A solution architect will map them to a GPU shape and a quote.
Quotes are itemised by configuration, term and support. No fixed list price is published.