Sovereign Infrastructure for Frontier AI Research

How Core42's sovereign, multi-accelerator AI cloud helped MBZUAI train frontier models while keeping strategic data and models in the UAE.

THE CHALLENGE

MBZUAI needed infrastructure that could support frontier training, exploratory research and inference as priorities evolved. It also needed accelerator flexibility without rebuilding the operating environment for each project, while keeping nationally significant models and data under UAE jurisdiction. Compute bottlenecks, fragmented tooling or reliance on a single architecture would slow research and weaken sovereign control.

THE CORE42 SOLUTION 

Core42 delivered a sovereign, heterogeneous AI cloud that combines large-scale compute, unified orchestration and in-country control in one research environment:

→  Multi-accelerator compute — NVIDIA DGX SuperPOD, AMD MI210 GPUs and specialised inference architectures.

→  Unified orchestration — managed Kubernetes and Slurm coordinate workloads across clusters.

→  High-speed data fabric — ultra-fast storage with InfiniBand and Ethernet networking.

→  Sovereign governance — UAE hosting, access management, model governance and unified security controls.

PLATFORM CAPABILITIES

DEEP DIVE
 

MBZUAI's research portfolio spans multilingual foundation models, climate intelligence and advanced reasoning. Each programme creates a different infrastructure profile: frontier training needs dense GPU clusters and fast storage; exploratory research benefits from flexible accelerator choice; and inference requires efficient, reliable deployment. At the same time, models and data with national strategic value must remain under UAE jurisdiction. The challenge was therefore to give researchers freedom to move between workloads without fragmenting the operating environment or compromising sovereign control.

To achieve this, Core42 built a sovereign, heterogeneous AI cloud hosted in the UAE. NVIDIA DGX SuperPOD supports frontier training, AMD MI210 GPUs provide capacity for exploratory workloads and specialised architectures serve inference. Core42 unified these resources through managed Kubernetes and Slurm, supported by high-speed InfiniBand and Ethernet networking and ultra-fast storage. Access management, model governance and integrated security controls extend across the environment, so research teams can select the right accelerator for each project while working through one consistent operating model.

The resulting platform has supported more than five frontier models across the Jais and K2 families. Jais advances high-quality Arabic AI for a language community of 274 million speakers. Jais Climate, fine-tuned on 1.4 million climate-related instructions, became the first bilingual large language model dedicated to climate intelligence. K2 Think combines 32 billion parameters with advanced reasoning performance that surpasses flagship models 20 times larger. Together, these programmes show how sovereign infrastructure can accelerate frontier research: compute, orchestration, security and governance operate as one system, giving MBZUAI the scale to move from experimentation to nationally significant AI without moving strategic data or models outside the UAE.

OUTCOMES AT GLANCE