Key Insights from NVIDIA AI Day Singapore 2026: AI Value Is Moving from Models to Systems

I spent most of yesterday at NVIDIA AI Day Singapore 2026, and here are some of my takeaways.

The event marked a pivotal moment, clearly indicating that AI is entering a new phase. It brought together developers, enterprises, government agencies, researchers, and technology partners from across the region. Discussions reflected a shift from experimentation to building, deploying, and scaling AI in real-world applications.

That made the day feel less like a showcase of what AI could do, and more like a window into how organisations are beginning to turn that potential into infrastructure, products, workflows and operating models.

I already knew, at least at a high level, that NVIDIA’s role in AI had expanded well beyond GPUs. But spending a day listening to NVIDIA, its customers and organisations building on its technology made the breadth of that role much clearer.

GPUs remain essential, but NVIDIA is now deeply involved in infrastructure, networking, software, models, development frameworks, inference, robotics, and the architecture required for agentic AI.

The initial sessions provided a useful framework for understanding these developments.

NVIDIA describes AI as a five-layer stack:

Energy → Chips → Infrastructure → Models → Applications

Energy underpins the system, with chips converting energy into computing power. Infrastructure integrates computing, networking, and storage. Models generate intelligence, which applications must translate into economic or societal value. NVIDIA emphasises that successful AI applications depend on every underlying layer.

It sounds straightforward, but as the sessions progressed, it became clear that the most valuable insights arose from the interactions between these layers.

The primary complexity lies in how these layers interact.

The focus is moving from individual models to integrated systems.

In recent years, business discussions have centred on comparing individual AI models by intelligence, benchmarks, context window size, or standardisation. The emergence of agentic AI is now shifting the focus from model comparisons to broader systems architecture.

NVIDIA’s session on the Nemotron open-model ecosystem explained this clearly.

A sophisticated agent now goes beyond a large language model. It functions as an application that integrates planning, retrieval, tool use, coding, testing, and multiple specialised models. The presentation noted that larger models manage planning, while smaller, specialised models execute tasks more efficiently.

This approach is evident in NVIDIA’s Nemotron range, which provides open models, weights, training data, and recipes to support developing specialised AI agents. The models vary in size and capability, enabling selection based on reasoning, efficiency, and deployment requirements.

For enterprises, this means moving from asking “What is the best AI model?” to considering systems-level questions: Which model best fits each task, what data and tools are required, where should inference occur, and what levels of latency or privacy exposure are acceptable? Organisations must also determine where human judgment is essential and how to measure success. In this context, AI architecture increasingly resembles the intentional design of an enterprise intelligence system rather than procuring a standalone software product.

Sea provided a clear example of becoming an AI-native organisation.

Sea’s presentation notably positioned infrastructure as a core component of the enterprise operating model. Sea has stated its goal to become an AI-native company and established an AI Centre of Excellence in Singapore earlier this year.

Its official structure focuses on three areas:

• These areas include foundational AI development, evaluation frameworks, and internal tooling.

• Scalable deployment focuses on translating research into production systems with measurable outcomes.

• Talent development and experimentation with new AI-enabled work methods are also priorities.

This approach is noteworthy because enterprise AI adoption cannot scale if each use case depends on a small central AI team to build solutions from scratch.

Sea is already deploying AI across Garena, Shopee, and Monee. Examples include AI agents and workflows in gaming, AI-assisted seller content and product discovery in Shopee, and fraud detection and credit risk modelling in financial services.

One example from the event clearly demonstrated the organisational impact.

At SPX Express, a business team developed an AI agent for fuel-claim review in under two weeks, moving from sample checking to comprehensive AI-assisted review.

This example illustrates two key themes.

The obvious one is automation. The more interesting one is enablement.

A team with direct operational knowledge transformed its domain expertise into a functional AI solution by leveraging shared enterprise capabilities.

Specialised AI solutions should be developed where deep technical expertise is required.

Organisations must also establish platforms, models, skills, governance, and safeguards to enable broader innovation safely and effectively.

This presents both an organisational design and a technological challenge.

HTX provided a more practical perspective on sovereign AI.

The HTX session added another perspective.

“Sovereign AI” can sometimes sound like an infrastructure conversation.

HTX’s approach demonstrates that sovereign AI involves more than infrastructure.

The Home Team’s AI strategy is built around three principles:

Sovereignty. Trust. Operational impact.

In this context, sovereignty means retaining control over sensitive operational systems, data, and AI capabilities.

HTX’s sovereign AI stack includes infrastructure, models, and applications. Its NGINE infrastructure provides secure AI compute under Home Team control, while the Phoenix family of models is being developed specifically for Singapore’s context and Home Team requirements.

Phoenix is a notable example of localisation. HTX’s latest Phoenix models are designed for Singapore’s laws, culture, terminology, and languages. Phoenix-VL 1.5 Medium, developed with Mistral AI, adds multimodal capabilities while remaining focused on secure operational use. HTX explicitly positions these systems as supporting human judgement rather than replacing it.

This distinction is important. The location of compute resources does not define local AI alone.

It also depends on whether a system understands the language, institutional context, policies, users, and operating environment in which it is deployed.

HTX is integrating its infrastructure and model capabilities with applications.

Its AI Suite includes tools such as Teammate for secure generative AI use by Home Team officers, as well as other workflow-specific products. HTX’s stated objective is not AI for its own sake, but improved mission outcomes, citizen experience, and officer productivity.

This distinction is also valuable for commercial organisations.

Capability alone is not the goal; operational impact is.

Customisation starts with evaluation, not training

I also attended a deeper session on NVIDIA’s Nemotron open models.

Organisations now face an expanding spectrum of options to adapt AI to their specific needs, ranging from lighter-touch methods like prompt engineering and retrieval-augmented generation (RAG) to deep model customisation via fine-tuning, supervised fine-tuning (SFT), continued pre-training, reinforcement learning, and model distillation.

However, one of the most valuable messages from the session came before any discussion of these techniques:

• Define the success criteria first.

• Build an evaluation set that represents the task.

• Establish a baseline.

• Then start changing the system.

The session also emphasised the ongoing importance of human evaluation, especially when automated assessments are ambiguous.

This may seem like a no-brainer. However, organisations often reverse this sequence:

“We have access to this model. What can we do with it?”

Versus

“We have this business problem. How will we know whether AI has improved it?”

This difference becomes increasingly important as enterprises customise models with proprietary data.

NVIDIA’s Nemotron resources now include model weights, open datasets, training recipes, evaluation resources and deployment guidance. NVIDIA says its datasets cover more than 10 trillion tokens across pre-training, post-training, personas, safety, reinforcement learning and retrieval use cases.

The broader implication is substantial. Access to a model is now only one aspect of AI advantage.

Evaluation, data quality, customisation, deployment, and operational learning are increasingly determining outcomes beyond the model itself.

The infrastructure discussion is increasingly focused on economics.

NVIDIA’s emphasis on inference was another valuable aspect. When AI was primarily about training models, infrastructure conversations naturally focused on compute.

Agentic AI is changing the economic landscape. An agent may call multiple models, retrieve information, invoke tools, run code, test outputs and repeat those steps many times before completing a task.

Consequently, the number of inference calls can increase significantly.

Latency, tokens, storage, networking, memory, power consumption, and cost per inference are now business considerations as well as technical concerns.

NVIDIA now describes this infrastructure as an AI factory, integrating energy, chips, infrastructure, models, and applications to produce intelligence at scale.

For enterprise leaders, this means the AI conversation must address economics as well as capability.

It is not simply: Can the model do this?

It is also: Can we do it reliably, securely and economically thousands or millions of times?

Singapore’s role in the AI ecosystem is also evolving.

A distinctly Singaporean perspective was evident throughout the event. The components below are becoming more interconnected.

Government, Enterprise, Infrastructure, Research, Local models, Public-sector deployment, Global technology companies, Talent, and Start-ups

NVIDIA described AI Day Singapore as showcasing advances across Southeast Asia, including efforts to move governments and organisations from AI pilots to production while developing capabilities that reflect local languages, cultures, and economic priorities.

This aligns with broader observations about Singapore’s innovation ecosystem. The advantage may not reside in any single programme, company, or institution.

The advantage lies in the connections among them.

Research connects to infrastructure. Infrastructure connects to models. Models connect to enterprises and government.

Enterprises provide real problems and environments in which solutions can be tested.

Policy and governance create conditions for responsible adoption.

Talent, capital and regional connectivity create pathways for scaling beyond Singapore.

AI further increases the importance of these connections.

The AI conversation is moving beyond the question of “Which model?”

My main takeaway from the event wasn’t a specific product announcement; instead, it was a shift in the nature of the conversation.

The first phase of generative AI encouraged organisations to experiment with models. The next phase increasingly involves systems engineering and organisational redesign.

Models still matter enormously. Infrastructure, proprietary data, orchestration, inference economics, evaluation, security, governance, human oversight, change management, and workflow design are equally critical.

Most importantly, clearly defining the desired outcome is essential.

For organisations, the key imperative is shifting from:

“How do we use more AI?” To increasingly:

“Where can AI create meaningful value, and what complete system needs to exist around it to make that value reliable, scalable and trusted?”

This is a far more meaningful and valuable goal.

Thanks to NVIDIA for a meaningful day of learning, discussion and a clearer view of where enterprise AI is heading.