Executive summary:
The scale of MBZUAI's mission demands infrastructure that does not slow research cycles or constrain how models are designed. MBZUAI needed sovereign high-performance compute and a heterogeneous environment where workloads would shift across accelerators as research evolved.
With Core42's AI Cloud, MBZUAI trains frontier models at scale, accelerates research timelines, and operates on a sovereign foundation for long-term national capability.
The Challenge:
Scaling frontier research places unique pressure on infrastructure. Training frontier models like Jais and K2 requires massive computational scale, with thousands of GPUs working in parallel through distributed training frameworks. That scale depends on high-performance, low-latency networks, where even milliseconds of delay can compound across training runs that span days or weeks.
Scale alone was not sufficient. Different research workloads have different optimal compute profiles. Frontier model training demands NVIDIA clusters, large exploratory workloads run efficiently on AMD infrastructure, and specialized architectures such as Cerebras support both large-scale training and high-throughput inference. Research directions shift quickly, and workloads had to move across accelerators without reengineering.
Sovereignty was equally critical. Models designed for national use cases, from Arabic language intelligence through Jais to climate intelligence through Jais Climate, need to be trained, stored, and managed within a secure environment that meets strict strict regulatory, data protection and national control requirements.
Finally, the infrastructure had to keep up with the speed of research. Breakthrough ideas need immediate access to performance, scale and sovereignty so they can be tested, trained and advanced without delay. MBZUAI also needed consistency across the entire research pipeline, from initial experiments through full-scale training and production inference. Infrastructure that performs well in training but creates friction at inference can slow the path from promising research to real-world deployment.
Approach:
AI Cloud: Building the Engine of Discovery
The Core42 AI Cloud provided MBZUAI with a sovereign, heterogeneous compute fabric built for large-scale scientific workloads. The platform combines multi-accelerator infrastructure, high-performance networking and unified operations to support rapid experimentation, large-scale training and deployment within a single environment.
Multi-accelerator Infrastructure deployed:
Unified Platform Layer
All workloads are submitted through a single control panel that abstracts the complexity of a multi- accelerator environment. Scheduling, routing, and resource allocation are managed automatically, allowing researchers to focus on model development rather than infrastructure orchestration.
High-Performance Foundations
The network fabric is optimized for high-performance, low-latency interconnects optimized for gradient synchronization and allow thousands of GPUs to coordinate efficiently during model training. Storage systems use memory-mapping, asynchronous writes and structured formats that prevent I/O bottlenecks. Together, these capabilities reduce time-to-train and maintain stable performance as workloads scale.
Result:
Research Velocity Unleashed via Core42 AI Cloud
With the Core42 AI Cloud, MBZUAI has trained and deployed multiple frontier models including Jais, K2, and Jais Climate. Each represents months of intensive computation delivered in timeframes that match the rapid pace of AI research. Infrastructure bottlenecks have been reduced, giving researchers access to compute when they need it. As a result, training cycles are shorter, experimentation can move faster and publication timelines can accelerate.
Models Built for Real Use Cases
Jais, was developed as one of the world's highest-quality Arabic large language models, serving 274 million Arabic speakers.
Jais Climate is the world’s first bilingual LLM dedicated to climate intelligence. Fine-tuned with 1.4 million climate-related instructions, it was launched during COP28 UAE to support climate research and decision-making.
K2 matches frontier model performance at a fraction of the size, enabling faster inference and reducing inference cost.
Specialized Portfolio: MBZUAI also launched a diverse portfolio of models across different research directions. These include regional language models such as Nile-Chat for Egyptian Arabic, Atlas-Chat for Moroccan Darija, SHERKALA for Kazakh and Turkish languages, and NANDA for Hindi, alongside Vicuna as a lightweight chatbot, FM For Bio – GET for gene expression prediction, and World Model for embodied reasoning and physical-world simulation.
Global Recognition
Access to sovereign, high-performance infrastructure strengthens MBZUAI’s ability to compete with leading AI research institutions globally. It also supports the university’s ability to attract faculty and students who want to work on advanced models using the same class of systems deployed by major AI laboratories worldwide.
For researchers, the value is clear: infrastructure becomes an enabler of ambition rather than a limitation on it.
Operational Excellence
Core42 AI Cloud’s heterogeneous environment gives MBZUAI greater efficiency by allowing each workload to run on the accelerator best suited to its performance and cost profile. AMD infrastructure can support large exploratory research workloads, while NVIDIA H100 clusters can be reserved for frontier scale training.
At MBZUAI's deployment scale, these operational efficiencies can support more research activity, enable greater experimentation and reduce the trade-offs between ambition, cost and capacity. Flexible capacity scales dynamically with research demand, reducing both overprovisioning waste and underprovisioning bottlenecks.
Complete sovereign control also ensures that research breakthroughs to remain within UAE and MBZUAI ownership, strengthening long-term national AI capability.
Takeaways: