01
Exclusive nodes
A whole GPU node is allocated to your workspace for the contracted term — not a slice of a shared host.
Platform · Dedicated GPU
Dedicated GPU nodes, exclusively yours
Whole-node H100 and L40S clusters inside Malaysia with predictable performance and your own change windows.
REF · SC-2026-GPUD
Mode · Whole node Tenancy · Single-tenant GPU · H100 / L40S Residency · Malaysia
Platform · Dedicated GPU
The whole node, with nothing noisy beside it
Dedicated capacity removes the two variables that make shared GPU clouds hard to reason about: a neighbour changing your effective throughput, and a provider changing your maintenance window. You get the node, the interconnect and the schedule.
For long training runs and steady inference, shared capacity is the wrong shape. Take the whole node, keep the interconnect, and operate it under your own change policy.
- Whole-node allocation with no oversubscription or noisy neighbours.
- NVLink / NVSwitch available on H100 node shapes for distributed work.
- Private networking and storage attached to your workspace.
- Change windows and maintenance coordination agreed in the contract.
Capabilities
What dedicated capacity includes
The operational and commercial surface around an exclusive GPU node.
02
Cluster interconnect
H100 node shapes include NVLink / NVSwitch for multi-GPU training and tensor-parallel inference.
03
Private networking
Attach the cluster to your private network, with no public data-plane exposure by default.
04
Your change windows
Maintenance and upgrades are scheduled with you, under the change policy agreed in the contract.
05
Sovereign storage
High-throughput volumes and S3-compatible object storage remain in-country with encryption at rest.
06
Audit and export
Access, provisioning and maintenance events are logged immutably and exportable to your SIEM.
Hardware
Dedicated node profiles
Reference profiles. Final node specification, power envelope and availability are confirmed in the statement of work.
H100 8-GPU node
SXM · NVLink- GPU memory
- 8 × 80GB HBM3
- Aggregate memory
- 640GB
- Interconnect
- NVLink / NVSwitch
- Workload
- Training · large inference
Distributed fine-tuning, multi-node training and high-throughput model serving.
L40S 8-GPU node
PCIe · 48GB- GPU memory
- 8 × 48GB GDDR6 ECC
- Aggregate memory
- 384GB
- Interconnect
- PCIe Gen4
- Workload
- Inference · document AI · media
Production inference, RAG and visual pipelines that need predictable throughput.
Availability is confirmed per order.
Process
How a dedicated cluster is delivered
Capacity, configuration and acceptance are agreed in writing before the term starts.
- 01
Size
Model the workload against node profiles and agree the GPU count, storage and network requirements.
- 02
Contract
The statement of work records the node specification, term, change policy, support model and acceptance criteria.
- 03
Deploy
The cluster is provisioned in-country, connected to your network and validated against the acceptance criteria.
- 04
Operate
Run production workloads with your change windows, audit exports and the escalation path named in the contract.
Use cases
When dedicated capacity is the right fit
Patterns where predictable throughput matters more than elastic scale.
Long training runs
Keep multi-day jobs on the same interconnect without competing for shared capacity.
Steady production inference
Serve a predictable request load with stable latency and no neighbour interference.
Large model serving
Hold a large model and its KV cache on a node that is not shared with another tenant.
Regulated workloads
Use single-tenant capacity for data classes that rule out shared infrastructure.
Controls
Sovereignty controls
Single-tenant controls applied to the dedicated capacity model.
In-country deployment
Nodes are deployed inside Malaysia and supported from Malaysia; any cross-border access requires written authorisation.
Single-tenant boundary
The node, its interconnect and its storage are assigned to one customer workspace.
Immutable audit trail
Administrative, access and maintenance events are recorded and exportable.
PDPA 2010 alignment
Access, retention and deletion procedures map to Akta 709 obligations.
FAQ
Dedicated GPU questions
Is this bare metal?
Dedicated nodes are allocated exclusively to your workspace for the contracted term. Whether a specific profile is delivered as bare metal or through a thin virtualization layer is confirmed in the statement of work.
How long is the commitment?
Terms are quote-based and can be structured around a project, a quarter or a longer operating period. The commitment length, capacity floor and renewal terms are recorded in the contract.
Can I burst beyond the reserved capacity?
Subject to availability, additional on-demand instances can be added alongside the dedicated nodes. Burst capacity is not guaranteed unless it is written into the contract.
Do you publish an SLA?
Support and availability commitments are set per contract. We do not publish a universal uptime percentage because the obligation depends on the node profile, redundancy and support tier you select.
Platform · Dedicated GPU
Reserve a dedicated cluster
Share the model, throughput target and term. We will return a node specification, availability window and itemised quote.
Quotes are itemised by configuration, term and support. No fixed list price is published.