Around the world, governments and industries are moving quickly to secure leadership in AI. For many countries, this momentum intersects with various pressing demographic realities: an ageing population, a shrinking workforce, and the urgent need to reinvent productivity. We cannot afford to fall behind. And the rise of agentic AI promises to accelerate this transformation.
Unlike traditional AI models, agentic AI does not just respond to queries — it reasons, plans, and takes actions across systems. For example, instead of simply answering a question on travel recommendations, an agentic system would book your flights, update your calendar, send reminders, and even adjust your itinerary based on weather or delays, all without being prompted for each step. This marks a shift from passive AI responses to proactive, collaborative systems that work alongside humans. The rise of agentic AI will require significantly more compute power, not just for single tasks or queries, but for extended workflows that involve reasoning, planning, and continuous adaptation.
As the technology for agentic AI matures and adoption expands, the world is effectively adding billions of virtual users into the compute fabric. The question for every country, including Singapore, is whether its AI infrastructure is ready to support this scale and complexity.
AI is more than GPUs
High-performance GPUs often dominate AI conversations, especially for training and running large-scale models. But CPUs are just as critical in powering AI systems behind the scenes, handling essential tasks such as data movement, memory management, thread coordination, and orchestrating GPU workloads.
In fact, many AI workloads, including language models with up to 13 billion parameters, image recognition, fraud detection, and recommendation systems, can run efficiently on CPU-only servers.
As AI models evolve into more modular architectures — such as mixture-of-experts systems popularised by DeepSeek and others — the need for smarter resource orchestration grows. CPUs must deliver high instructions per clock, fast input/output, and the ability to manage multiple concurrent tasks with precision.
Equally critical is connectivity, the “glue” that binds modern AI systems together. Advanced networking components, such as smart network interface controllers, help route data efficiently and securely between components, offloading traffic from GPUs and reducing latency. High-speed, low-latency interconnects help ensure data flows seamlessly across systems, while scalable fabric ties nodes together into distributed AI clusters.
In the age of agentic AI, heterogeneous system design becomes critical. AI infrastructure must go beyond raw compute; it must integrate CPUs, GPUs, networking, and memory in a flexible and scalable way. Systems built this way can deliver the speed, coordination, and throughput needed to support rapid, real-time interactions of billions of intelligent agents. As adoption scales, rack-level optimisation, where compute, storage, and networking are tightly co-designed, will be key to delivering performance and efficiency.
Why openness matters in the AI race
As AI systems grow more complex and distributed, the need for openness in software, hardware, and systems design becomes a strategic imperative. Closed ecosystems risk vendor lock-in, limit flexibility, and can constrain innovation at a time when adaptability is key to scaling AI.
This is why open software stacks are essential. These provide developers and researchers the freedom to build, optimise, and deploy AI models across a wide range of environments. They typically support popular frameworks, include tools for performance tuning, and offer portability across hardware — all available as open source. In the context of Singapore’s ambition to foster innovation across academia, start-ups, and industry, open AI software offers broader accessibility, faster iteration, and lower barriers to entry.
Similarly, openness at the hardware and systems level is significant. As AI compute evolves toward large-scale, heterogeneous deployments, rack-scale architecture becomes foundational. Open standards such as the Open Compute Project support modular system design, while collaborations like the Ultra Accelerator Link (UALink) aim to create high-bandwidth connections between AI accelerators across servers. Meanwhile, the Ultra Ethernet Consortium (UEC) is defining networking standards designed for AI, enabling low-latency, high-throughput data movement across distributed systems.
These initiatives allow cloud and data centre operators to build flexible, interoperable infrastructure that can scale with AI demand. For Singapore, embracing an open ecosystem can position the country to benefit from global innovation while cultivating local differentiation. It enables governments and businesses to build infrastructure that is performant and energy-efficient without being locked into proprietary limitations.
In the upcoming era defined by multi-agent AI, openness is positioned as a prerequisite for scale and long-term competitiveness.
Looking ahead in 2026
As agentic AI develops, the focus extends beyond GPUs to include CPUs, high-speed interconnects, and networking, which support the coordination of AI workloads at scale. An open ecosystem, including open software, rack-scale standards, and industry collaborations such as UALink and UEC, is also relevant to flexibility and interoperability across systems.
For Singapore, building open, heterogeneous, and scalable infrastructure like this is more than a technology choice – it is a strategic foundation for national competitiveness. As the country navigates rising automation needs and growing regional AI ambitions, future-ready AI infrastructure will be essential to unlocking sustainable growth, innovation, and resilience.
