ALQUA CLOUD

Train and serve models in the UAE.

Managed Kubernetes, Slurm and model serving on dedicated GPUs. We run the drivers, the fabric and the scheduler; Proof tells you when a job is wasting time. Data stays in the UAE.

Get Access See Pricing

SITESDXB-1 Dubai · AUH-1 Abu Dhabi

NETWORKInfiniBand or Ethernet, up to 800 Gb/s

STORAGENVMe, parallel file system, S3

BILLINGPer second, no egress fees

01WAYS TO RUN

Five ways to run.

We operate the stack under each of them: drivers, NCCL, the fabric, the scheduler, node repair and upgrades, with engineers on call in Dubai. You bring the model and the data.

02PRICING

Prices per GPU-hour.

GPU

{{ nodesL }}{{ nodes }} × 8

03NETWORK

InfiniBand or Ethernet.

Every GPU gets its own link: 400 Gb/s on H100 to B200, 800 Gb/s on B300. Rails run to a non-blocking fat tree with no oversubscription.

RAIL-OPTIMISED FAT TREE · NON-BLOCKINGINFINIBAND / ETHERNET

	INFINIBAND	ETHERNET
SWITCHING	Quantum-2 NDR · Quantum-X800 XDR on B300	NVIDIA Spectrum-X, 800 GbE ports
PER GPU	400 Gb/s NDR · 800 Gb/s XDR on B300	400 Gb/s · 800 Gb/s on B300
TRANSPORT	Native RDMA, in-network reduction	RoCEv2 with adaptive routing
BEST FOR	Large pre-training, tightly coupled jobs	Fine-tuning, inference fleets, multi-tenant clusters
CONNECTS TO	ALQUA storage and gateway	Your VPC and on-premise network
MAX CLUSTER	2,048 GPUs	1,024 GPUs

04STORAGE AND API

Local NVMeUp to 30 TB per node for datasets and checkpoints that need to be close to the GPU.

Parallel file systemShared across the cluster on the same fabric, sized for synchronized checkpoint writes.

S3-compatible object storageFor datasets and model weights. Stored in the UAE, no egress fees.

INFERENCE · PYTHON

from openai import OpenAI

client = OpenAI(
  base_url="https://api.alqua.ai/v1",
  api_key=os.environ["ALQUA_KEY"],
)
client.chat.completions.create(
  model="falcon-h1-34b-instruct",
  messages=[{"role": "user", "content": "مرحبا"}],
)

◆ Falcon-H1◆ JaisLlama 4Qwen3DeepSeekMistralgpt-oss