Insights

Designing the future: How AI and silicon engineering converge at the edge

Intelligence is moving out of the data centre and onto the chip. For Southeast Asia, the world’s chip-assembly heartland , that shift is an opening to lead.

By Amar Chhajer · VP-APAC, UST Malaysia

For a decade, the story of AI has been a story of the cloud, ever-larger models trained in ever-larger data centres. That story is not ending, but it is no longer the whole story. The next chapter is being written somewhere much smaller and far more interesting: on the chip itself, inside the device, at the edge of the network. And it is being written by engineers who have stopped treating the AI model and the silicon as separate problems.

This is what convergence means. Not AI running on silicon, that has always been true, but AI and silicon designed together, each shaped by the constraints of the other. It is the difference between renting intelligence from a distant data centre and engineering it directly into the world around us. For Southeast Asia, a region that already manufactures much of the world’s silicon, that difference is not academic. It is a once-in-a-generation strategic opening.

DIVIDER

Why intelligence is moving to the edge

The pull toward the edge is not a fashion. It is the sum of four pressures that centralized compute cannot resolve on its own.

1. Latency: some decisions cannot wait for the cloud

A round trip to a data centre takes time that many AI applications simply do not have. An autonomous machine on a factory floor, a driver-assistance system reading the road, a medical device interpreting a signal, these cannot pause for a network hop. Running inference locally collapses that delay; modern on-device edge inference now operates at sub-50-millisecond latency. When the decision and the data live in the same place, intelligence becomes immediate.

2. Energy: the cloud is hitting a power wall

The scale of AI’s appetite for power is becoming a planning problem for entire economies. McKinsey analysis suggests global data-centre capacity could nearly triple to between 156 and 219 gigawatts by 2030, with roughly 70% of new demand driven by AI workloads. ^[4] Against that backdrop, doing more work on efficient local silicon is not just convenient; it is necessary. Energy-aware edge inference techniques have demonstrated up to 31% lower energy consumption alongside 27% lower latency versus standard approaches.The engineering goal has quietly shifted from raw speed to performance per watt.

3. Cost and bandwidth: stop shipping raw data

Most data generated at the edge is never worth sending to the cloud. Consider a single 4K smart camera running local inference: processing on-device avoids streaming roughly 450GB of raw video per month, turning a continuous bandwidth bill into a trickle of meaningful events. Multiply that across the more than 18.4 billion IoT devices already deployed, a figure forecast to exceed 38 billion by 2030, and the economic logic of processing data where it is born becomes overwhelming.

4. Sovereignty: keeping data where it belongs

In a region as regulatorily diverse as Southeast Asia, where data-localization rules and national digital frameworks vary by market, the edge offers something the cloud struggles to: data that never has to leave the device, the building, or the country. Processing locally is fast becoming a compliance advantage as much as a technical one.

DIVIDER

Convergence is an engineering discipline, not a purchase

The temptation is to treat edge AI as a shopping exercise, buy a faster chip, and deploy a smaller model. That misses the point. The real shift is that the model and the silicon are now co-designed. The constraints of the chip shape the model's architecture; the demands of the model shape the chip's design.

This is why the industry is moving decisively toward purpose-built accelerators. General-purpose processors are versatile but inefficient for AI. Application-specific integrated circuits (ASICs), silicon engineered for one job, are the fastest-growing segment of the edge AI chip market, projected to expand at a 47.2% CAGR as designers chase domain-optimized performance. Further out, neuromorphic architectures that co-locate memory and compute, mimicking the brain’s event-driven efficiency, are forecast to grow at roughly 48.3% CAGR, a signal of how far the co-design philosophy will travel

DIVIDER

Cloud versus edge: a quick orientation

This is not a competition. The mature pattern is a continuum: models trained centrally, distilled and quantised, then deployed to silicon at the edge, with the cloud and the device working as one system.

DIVIDER

Why Southeast Asia is positioned to lead

Here is where the story turns regional. The convergence of AI and silicon is, at root, a hardware story, and hardware is where Southeast Asia already wins.

ASEAN is collectively the world’s largest exporter of integrated circuits and the dominant hub for semiconductor assembly, test and packaging, the back-end stages where chips are finished and made ready for the world.That base is now scaling fast. The ASEAN semiconductor market is forecast to reach USD 247.32 billion by 2034, with Malaysia holding the largest share. The region has attracted approximately USD 60.8 billion in semiconductor foreign direct investment since 2020.

Crucially, the region is no longer content to assemble what others design. Malaysia’s National Semiconductor Strategy is an explicit push up the value chain into design and high-value manufacturing, backed by moves such as a USD 250 million technology-transfer agreement with Arm, while Singapore, which accounts for roughly 10% of global semiconductor output, anchors regional R&D, and Vietnam emerges as a centre for packaging, testing and AI chip design.

DIVIDER

The human discipline behind the silicon

None of this displaces the human-centric principle that runs through everything we build at UST. Earlier, I argued for a human-first approach to AI transformation, the idea that the goal is to augment human judgement, not replace it. Edge AI is where that principle becomes physical.

When intelligence lives in the device a clinician holds, the machine an operator runs, or the vehicle a person drives, the design decisions are not only about silicon. They are about trust, safety, transparency, and accountability, about engineering systems that keep people in command of the decisions that matter. The best edge-AI engineering is not the most autonomous; it is the most accountable.

DIVIDER

What enterprise leaders in the region should do now

For CXOs across Malaysia and Southeast Asia weighing where edge AI fits, four moves separate the leaders from the followers.

1. Map the workloads that belong at the edge. Not everything does. Identify the decisions that are latency-critical, bandwidth-heavy, or privacy-sensitive; those are your first edge candidates, and the ones with the clearest ROI.
2. Design for performance per watt. Treat energy and silicon efficiency as first-class requirements from the start, not as an afterthought. The cheapest inference is the one you never sent to the cloud.
3. Build on the edge-cloud continuum. Architect for models trained centrally and deployed locally, with governance, updates and security spanning both. Avoid silos that trap you in one or the other.
4. Partner where the hardware lives. The convergence of AI and silicon rewards proximity to engineering talent and the semiconductor ecosystem. Southeast Asia offers both; use it.

DIVIDER

Designing the future, deliberately

The phrase “designing the future” is easy to say and hard to earn. But the convergence of AI and silicon at the edge is exactly that: a moment when the choices engineers make about chips and models will determine how intelligent, efficient, and trustworthy the next generation of technology becomes.

Southeast Asia does not have to wait for that future to arrive from somewhere else. It already has the manufacturing depth, investment momentum, and engineering ambition to help design it. The work now is to move with intent, from assembling the world’s chips to designing the intelligence that runs on them.

DIVIDER

Build intelligence where it belongs.

UST helps enterprises across Southeast Asia design AI-ready infrastructure and edge strategies that turn the convergence of AI and silicon into measurable outcomes, with people kept firmly at the centre.

Explore UST in Malaysia & Southeast Asia


DIVIDER

Related reading

From automation to augmentation: The human-centric AI journey in Asia-Pacific

A human-first approach to AI transformation

Amar Chhajer — Boundless Thinkers profile