Bare-metal GPU clusters
Dedicated accelerated nodes on high-bandwidth, low-latency fabrics — no virtualisation overhead.
Compute, on demand.
Our platform turns raw infrastructure into a seamless experience. Provision a single node or a thousand-GPU cluster, schedule training runs and serve models in production — with transparent metering and enterprise-grade isolation.
ClusterHKG-01 · Training
Healthy
GPU utilisation87.4%
Active jobs246 queued
Hall power1.82 MWPUE 1.24
Throughput · tokens / s2.41M
Fleet map · 256 GPUs
llm-pretrain-70bRunning512 GPU
vision-ft-a12Running64 GPU
inference-api-prodServing32 GPU
rag-embed-batchQueued16 GPU
Interface shown for illustration.
(01)Capabilities
Everything between
the GPU and your model.
One control plane for training, fine-tuning and serving — on infrastructure we design and operate ourselves.
Managed Kubernetes & Slurm
Bring your containers or batch jobs; we handle orchestration, scheduling and health checks.
Inference endpoints
Serve models behind autoscaling, low-latency APIs with usage-based pricing.
High-performance storage
Parallel file and object storage placed next to your GPUs, so data loading never starves training.
Observability & metering
Real-time telemetry, per-job cost attribution and transparent invoicing.
Enterprise isolation
Private tenancy, dedicated networks and Hong Kong data-residency options.
GridPort · Compute distribution platform
Need tokens,
not GPUs?
GridPort delivers compute as open-model inference: one OpenAI-compatible API, billed per token, with routing, metering and failover handled for you.
(02)Developer experience
From login to
first job
in minutes.
Work the way your team already works. Launch clusters from the console, script them from the command line, or drive everything through the API.
Console, CLI & API
Everything you can click, you can automate.
Bring your stack
Your containers, frameworks and schedulers run unchanged.
Cost you can trace
Utilisation, throughput and spend, broken down job by job.
# Reserve a 512-GPU training cluster in Hong Kong
$ hkc cluster create train-01 \
--region ap-hongkong-1 \
--nodes 64 --gpus-per-node 8 \
--fabric infiniband --scheduler slurm
✓ train-01 ready · 512 GPUs · 4m 12s
$ hkc job submit --cluster train-01 \
--image registry/llm-pretrain:latest \
--gpus 512 -- python train.py --config 70b.yaml
✓ llm-pretrain-70b running
$ hkc job logs llm-pretrain-70b --follow
step 1200 · loss 2.184 · 2.41M tok/s# Serve a model behind an autoscaling endpoint
from hkcompute import Client
hkc = Client()
endpoint = hkc.endpoints.create(
name="inference-api-prod",
model="registry/my-model:v3",
region="ap-hongkong-1",
gpus=4,
autoscale=(2, 32),
)
print(endpoint.url)
# https://inference-api-prod.ap-hongkong-1.…# Your existing manifests run unchanged
apiVersion: batch/v1
kind: Job
metadata:
name: vision-ft-a12
spec:
template:
spec:
containers:
- name: trainer
image: registry/vision-ft:latest
resources:
limits:
gpu: 8
restartPolicy: NeverCommands and SDK shown for illustration.
(03)Security & tenancy
Isolated
by design.
Built for teams whose models and data are their most valuable assets — and for the regulators who oversee them.
Private tenancy
Single-tenant clusters with no shared compute, storage or management plane.
Dedicated networks
Private connectivity and isolated fabrics, with direct interconnects to your own sites and clouds.
Data residency
Options to keep data, models and logs within Hong Kong at every stage.
Access control & audit
Role-based access, single sign-on and audit trails for every action.