Our team at Sunway iLabs has been closely following the evolving robotics landscape. Signals are becoming clearer: robots are steadily moving beyond research labs and controlled demonstrations into real operational environments. Even industry leaders like NVIDIA CEO Jensen Huang has suggested that robots could be only a few years away from widespread use.

As more robotics companies enter the field and promise similar capabilities, how do we determine which robot actually makes sense for a given use case? The situation is not unlike the early days of smartphones, when dozens of devices appeared with seemingly similar features. The real question then becomes:

What should we really look for when evaluating a robot?

How to Choose a Robot

Let’s take a look at some of the key criteria to consider when surveying the robotics landscape: the factors that help determine whether a robot is truly ready for real-world deployment, or just an impressive demo. To make sense of this, it helps to evaluate robots across a few key questions: how well they understand their environment, how they move, whether they can operate reliably over time, and whether they are already being used in real-world settings.

Embodied Intelligence

The first test of any robot is intelligence: does it understand its surroundings, respond appropriately, and adapt in variable conditions in the environment it is meant for?

The emergence of AI has marked a shift in robotics. Embodied AI combines intelligence with physical structure—the robot’s body, sensors and surroundings—so that it can perceive, decide, and act in the real world.

Rather than solely relying on fixed data, these systems are also trained extensively in simulated environments. This allows robots to learn faster and at scale, while continued learning from real-world deployment improves how well they adapt to changing conditions and handle tasks that require both precision and coordination.

Platforms such as AgiBot’s Real-World Reinforcement Learning (RW-RL) system combine large-scale pre-collected data with on-site training. It allows their robots to learn tasks directly on the factory floor within minutes and adapt to changing production conditions without major reconfiguration, marking a step toward intelligence that improves through real-world use rather than solely relying on fixed instructions.

This shift is not limited to a single company. Other leading full-stack robotics players are also moving in the same direction, overcoming one of robotics’ biggest constraints: the difficulty of collecting enough data to train systems effectively.

By combining simulation and real-world learning, they reduce reliance on slow and costly data collection while improving how quickly robots can be deployed and adapted in practice.

Progress in robotics is increasingly defined not just by hardware performance, but by how well systems can operate in semi-structured or unstructured environments, where situations can be unpredictable, tasks vary, and decisions must be made in real-time.

In other words, look for a robot meant for messy reality, one that can handle unpredictable human behavior, adapt to dynamic settings, recover safely from errors, and improve continuously through experience.

Mobility

Once a robot can decide what to do, the next question is whether it can physically do it.

Mobility is more than whether a robot can walk. It includes balance, payload capacity, movement speed, joint coordination, and how well the machine handles obstacles such as slopes, stairs, tight corners, or uneven ground.

Recent demonstrations, such as Unitree’s G1 humanoid performing complex martial arts movements offer a useful way to interpret these capabilities. While such demos may appear theatrical, they reveal underlying attributes that matter in real-world use, including dynamic balance, precise coordination across multiple joints and the ability to maintain stability while executing fast, continuous movements.

These capabilities are reflected in the robot’s design. The G1 has 43 degrees of freedom, allowing for fine-grained control across its body, while its high joint torque enables it to generate and withstand force during movement.

Its dexterous hands also matter because mobility is not only about locomotion, but about whether the robot has the flexibility and control to manipulate objects and carry out specific tasks. Hand design, force control and degrees of freedom help indicate how precisely it can grasp, hold, and handle objects once in position.

Details like these offer useful clues about whether the robot is meant to fit into specific environments. When evaluating mobility, ask whether the robot’s movement capabilities actually match the physical conditions of the job.

Operational Readiness

A capable robot still needs to operate as part of a real-world system.

Operational readiness is about whether the robot can run consistently within the constraints of your workflow. This includes practical factors such as battery life, charging time, uptime, maintenance requirements, availability of spare parts and support and how the robot handles failures or interruptions.

A robot that performs well in controlled settings but cannot be supported or maintained efficiently may not yet be ready for real deployment. For instance, a robot that operates for two hours but requires four hours to recharge may struggle to sustain continuous operations without additional units or human intervention.

Though often overlooked, this factor ensures whether a robot is practical or just impressive. UBTECH’s Walker S2, for example, is designed to swap its own battery pack and continue operating with minimal downtime.

When evaluating operational readiness, consider whether the robot can run consistently, be maintained efficiently, and operate as part of a larger system.

Real-World Adoption

Even the most advanced robot means little if it hasn’t proven itself outside the lab. It’s important to look at whether the robots being evaluated are already deployed in real environments, who their customers are, and whether those use cases resemble your own needs.

Service robots are already operating in everyday settings. Companies such as Pudu Robotics have deployed robots across restaurants, retail stores, hotels and healthcare facilities worldwide. These robots assist diners in places from Canada and Hong Kong to Malaysia, while others operate in shopping malls across Poland, Thailand and Jordan.

UBTECH has also announced plans to scale production from hundreds to several thousand units over the next few years with automotive manufacturers such as BYD, Geely Auto, FAW-Volkswagen, and Dongfeng Liuzhou Motor beginning to integrate its robots into industrial operations for round-the-clock use.

These figures indicate how quickly robots are moving from experimental prototypes toward commercial deployment. Look for signs of real-world traction: how many units have been shipped, whether production is scaling up, and whether the robots are being used by real organisations.

Where This Leaves Us

Robots are useful because they can take on routine tasks with consistency and minimal supervision. With companies ramping up production and preparing for wider deployment, robots are beginning to show up across a growing number of settings, from service environments to industrial operations: in different forms, for different kinds of work.

So if you are considering a robot, don’t forget to ask a few fundamental questions:

1. Can it move reliably in your environment?

2. Can it understand and respond to what’s happening around it?

3. Can it operate consistently within your workflow over time?

4. Is it already being used outside the lab?

The best robot is not the most futuristic one; it is the one whose capabilities match the job you need done.

Acknowledgements: Thank you to the Sunway iLabs team for their invaluable contribution and insights in preparing this article.

References

FutureX Insights. (2026). Robots Among Us. Futurexinsights.news. https://www.futurexinsights.news/p/robots-among-us?r=6i8ih6&utm_campaign=post&utm_medium=web

Jijo Malayil. (2025, December 30). UBTECH robot dances, kicks box as 1,000 Walker S2 milestone marked. Interesting Engineering. https://interestingengineering.com/ai-robotics/china-ubtech-builds-1000-humanoid-robots

South China Morning Post. (2025, November 19). South China Morning Post. https://www.scmp.com/tech/article/3333427/ubtechs-2026-humanoid-robot-output-grow-10-fold-costs-plunge-scale-economics

‌

‌

But look at what just happened during the humanoid robot showcase at China’s state broadcaster, CCTV’s Spring Festival Gala. Chinese firm Unitree Robotics stunned audiences with fluid movements, agility, and impressive martial arts performance on stage alongside young human artists during a live national broadcast.

A stark contrast to the stiff dance routines we saw just a year ago in 2025, this performance signals how quickly humanoid capabilities are evolving.

Now, we shift the lens. As we move through 2026, the industry is no longer just fighting over how robots move; it’s fighting over how they think.

The Brain and the Body

At a basic level, many modern robots can be understood through two parts: a brain and a body.

The brain is the thinking part. It’s what lets a robot see the world, make decisions, and plan its next move. It’s made up of the AI models, software, and computing systems that turns data from sensors into action.

The body is the doing part. It’s the motors, sensors, batteries, connectors and mechanical components that enable movement, balance and interaction with the real world.

This distinction shows that not all robotics companies build the same thing. Some focus entirely on the brain, while others build full-stack systems, tightly integrating the brain with the body to deliver complete robots.

Selling the Universe Mind - The “Android” Play:

If the last decade was about mechanical engineering, 2026 is about Foundational Intelligence. A growing faction of the industry argues that the physical hardware is a solved problem; the real bottleneck is the “ghost” in the machine — the intelligence.

Brain Providers are the Google of this new era. They don’t want to build the metal, they want to build the Operating System that runs on everyone else’s hardware. That’s where some companies are placing their bets.

Companies like Skild AI and Physical Intelligence are building foundational models for the physical world. Their goal? A universal “robot brain” that can be dropped into a forklift, a humanoid, or a robotic arm and just work.

Others, such as PrismaX, work on the data and learning layer that helps robots improve through real-world use. Even firms like HappyRobot AI show how powerful software systems can automate complex work without being tied to a specific machine.

It’s similar to how NVIDIA powered the AI boom. Companies like OpenAI, Anthropic, and Meta, run their large language models (LLMs) on NVIDIA chips. Different applications, same underlying layer. Brain providers in robotics are aiming to become that shared intelligence layer, powering many different machines.

The Full-Stack Titans: The “iPhone” of Robotics

While the brain camp bets on a universal OS, the full-stack builders are playing a different game entirely. They aren’t just writing the software, they are forging the steel. This is the Apple model of robotics: a closed-loop system where the “Mind” and the “Machine” are designed as a single, inseparable unit.

Some companies choose to design and build both brain and body together, tightly integrating AI with hardware from the start. These full-stack builders control the entire system, from the motors and mechanics to the software that drives them.

Companies like Figure AI, Apptronik, UBTECH, Unitree, LimX Dynamics, and Fourier Intelligence fall into this category. They design robots where intelligence and physical design evolve side by side. The primary advantage of this approach is total optimisation:

• Lightning-Fast Reflexes: In the physical world, a split-second delay is the difference between a robot catching a falling object or causing an expensive accident. By building both the sensors and the software, these companies achieve a level of precision that universal brains can’t match.

• Physical Intuition: These builders argue that true intelligence must be “embodied.” This means the AI isn’t just a generic program, it intimately understands the unique physics and “muscles” of its own custom-built body.

When the same company builds both the body and the brain, it can optimise performance, balance, and behaviour as one cohesive system rather than stitching mismatched components together. If brain providers aim to create reusable intelligence platforms, full-stack builders aim to deliver complete, turnkey machines ready to operate in the real world.

The Missing Middle: The Data Pipeline

In 2026, the real bottleneck isn’t the “Brain” or the “Body”, it’s the data pipeline connecting them. Without Robot Operations (RobOps), even the most advanced machine is just an expensive paperweight.

Data is the new oil, and the industry is drilling for it in two ways:

• Synthetic Data: To bypass the slow, costly process of real-world training, companies use hyper-realistic simulations. Here, robots practice tasks millions of times in seconds before touching a factory floor.

• Data Sovereignty: As robots enter homes and hospitals, privacy is paramount. Firms like PrismaX are leading the charge in ensuring that the sensitive data robots collect remains secure and private.

While Silicon Valley perfects the “Mind,” China is building the infrastructure to feed it. By leveraging its massive manufacturing footprint and deploying “cyber-labourers”, human operators in VR rigs who act out tasks for AI to learn, China has become the world’s largest sandbox for RobOps.

Robotics: Ready or Not?

Big players in the industry suggest that robotics is approaching a breakthrough, though more cautious voices point to the constraints.

At CES 2026, NVIDIA’s Jensen Huang said the “ChatGPT moment for robotics is nearly here,” signalling that advances in AI could soon unlock a step change in robotics capability.

Yet Morgan Stanley’s 2026 report offers a grounded reminder, that robots today remain costly. Wider adoption will depend on hardware becoming cheaper, systems becoming more reliable, and AI models improving enough to operate consistently beyond controlled settings.

Where This Leaves Us

The robotics industry is currently in its “dial-up” phase; it is functional but waiting for its broadband moment. Whether the winner is a universal “Android-style” Brain or a specialised “Apple-style” Full-Stack machine, the prize is the same: the automation of physical labour.

Acknowledgements: Thank you to the Sunway iLabs team for their invaluable contribution and insights in preparing this article.

References

FutureX Insights. (2026). Robots Among Us. Futurexinsights.news. https://www.futurexinsights.news/p/robots-among-us?r=6i8ih6&utm_campaign=post&utm_medium=web

Morgan Stanley. (2026, January 26). Embodied AI and the Rise of Humanoid Robots. Morgan Stanley. https://www.morganstanley.com/

The Star Online. (2026). China’s humanoid robot makers pivot from “body” to “brain” as commercial race heats up. The Star. https://www.thestar.com.my/aseanplus/aseanplus-news/2026/02/07/chinas-humanoid-robot-makers-pivot-from-body-to-brain-as-commercial-race-heats-up

‌

‌

Sam Altman on Switching Sides of the Table

One of the most interesting parts of this recent a16z podcast conversation with Sam Altman wasn’t actually about Sora, AGI timelines, or even OpenAI’s next big product move.

It was about something far more human, and for many of us in the startup world, far more relatable: the transition from being an investor to becoming an operator.

As someone who has spent time on both sides of the table, I found myself reflecting on the same question Sam’s journey raises:

Am I better suited to investing in companies, or building them?

And perhaps more importantly:

What does it really take to make that switch?

From Investing to Operating: Different skills, different realities

Sam Altman is one of the most high-profile examples of someone who has successfully moved from investor to entrepreneur and CEO. But what this conversation highlights is how fundamentally different those roles can be.

Being an investor requires a particular set of muscles. Operating, however, is an entirely different experience.

Investors can step back. Operators cannot.

There is something relentless about being responsible not just for ideas, but for outcomes. The feedback loop becomes immediate, personal, and constant.

Sam also admits he was not particularly good at being a CEO at first, and that the role taught him humility. That honesty matters, because operating is not an extension of investing, it is a different discipline entirely, one you grow into through responsibility and repetition.

Why OpenAI Wasn’t a Typical Product Company Transition

Another fascinating point in the conversation is Sam’s view that moving into a research lab environment felt like a more natural evolution than running a pure product company.

That stood out to me.

OpenAI is not a traditional startup in the sense of optimising funnels or iterating on SaaS metrics. It sits somewhere between a frontier research institution, an infrastructure builder, and a consumer technology company.

Research labs reward a different kind of long-term thinking:

• Breakthroughs rather than incremental iteration

• Mission-driven talent density

• A longer horizon for impact

• Discovery as much as delivery

In that sense, OpenAI’s trajectory is less about building “a product” and more about building a platform for the future.

Sam’s transition makes sense in that context. The move from investor to operator may be difficult, but moving into a research-led organisation where the mission is as central as execution may feel more natural than jumping into a traditional product CEO role.

It reflects how entrepreneurship itself is evolving. Some of the most important companies today are not just selling software. They are building new technological frontiers.

A Personal Reflection: Enjoying Both Sides of the Table

This part of the podcast resonated with me because I’ve wrestled with the same question.

I’ve been an entrepreneur. I’ve been an investor. And I still ask myself what suits my skillset better: building companies or backing them.

The truth is, I genuinely enjoy both.

I find the act of investing intellectually exciting: identifying emerging technologies, meeting ambitious founders, and supporting companies at the edge of innovation.

But I also know that investing becomes far more grounded when you’ve experienced what it actually takes to run a company.

Operating teaches you things no pitch deck can:

• How hard hiring really is

• How messy execution becomes

• How difficult scaling culture can be

• How lonely decision-making feels at the top

That experience stays with you. And I believe it makes you a better investor, because you don’t just evaluate ideas, you understand the realities behind them.

The Best of Both Worlds: Investor, Intrapreneur… and Working with the Next Generation

What I love about my work today is that I get to combine multiple worlds that are often kept separate.

On one hand, I have the privilege of investing in next-generation technology companies shaping the future of AI, innovation, and entrepreneurship.

On the other hand, I also act as an intrapreneur: working within Sunway’s ecosystem to build and scale solutions that connect corporates, startups, academia, and government.

But what makes the work especially meaningful is that it doesn’t stop there.

I also get to spend time with students and researchers at Sunway University, engaging with the next ideas being developed in the lab, often at the earliest stage, before they are even companies or products.

There is something uniquely energising about being close to that frontier: watching curiosity turn into experimentation, and experimentation into innovation.

It is a reminder that entrepreneurship is not only about venture and execution, but also about learning, discovery, and the long-term cultivation of talent and ideas.

In many ways, this blend of investing, building, and working alongside emerging researchers feels like the most complete seat at the table.

It reinforces that the investor vs entrepreneur question doesn’t always need a binary answer.

Sometimes the most meaningful careers happen in the overlap, where you can support builders, while also building yourself, and helping shape what comes next.

Closing Thought

Sam Altman’s journey from investor to operator is a powerful example of how career paths in tech and entrepreneurship are rarely linear.

It also raises an important question for anyone in this space:

Are you energised more by identifying potential, or by executing against it?

And perhaps the best answer is that both perspectives, when combined, create a rare kind of leverage.

What about you?

Have you ever wrestled with whether you’re better suited to building companies or backing them? And do you think the best investors are those who have operated before?

REFERENCE:

In Star Wars robots weren’t special…they were useful.

Long before real-world robots showed up, sci-fi imagined a world where robots handled everyday work.

The arrival of robot vacuum cleaners in the early 2000s marked the dawn of service robotics. Designed to operate autonomously, they quickly gained traction in busy households, with early models like the Roomba selling over a million units within two years.

This opened up new possibilities for automation: service robots handling routine tasks around the clock without needing constant supervision.

Broadly speaking, this is what service robots are designed for, to work in everyday environments, helping organisations deal with labour shortages and long, repetitive tasks that humans would rather not do themselves, without getting tired.

They aren’t robots confined to factories or dealing with manufacturing products, but instead they are there to support daily tasks such as in hotels, hospitals, malls, airports and offices.

Beyond the Gimmick

For a long time, robots moved through the world by following fixed paths. They knew where to go, but not much about what they were looking at.

That has started to change.

Today’s robots often appear in demos, videos, and public showcases — from making appearances at shopping malls like Sunway Pyramid and i-City to dancing on shows like America’s Got Talent — capturing attention and showing what robots can do in motion.

With recent advances in AI, robots are increasingly able to make sense of spaces rather than just move through them. By combining vision, language, action (VLA) and large language model (LLM), today’s robots can interpret their surroundings and adjust how they move or work in real time.

You may already be familiar with some humanoid robots like Agibot, and robot dogs such as Unitree’s Go2 and Boston Dynamics’ Spot.

Still, robots aren’t just built to impress people, they’re there to do real work.

What These Robots Are Really Used For

Service robots have been taking on more practical roles from robot vacuums like iRobot’s Roomba cleaning floors at home to cat-eared waiter robots like Pudu Robotics’ BellaBot delivering your plate of roti canai at mamak restaurants.

So who else is already putting robots to work?

The Trouble with Keeping Things Clean

Cleaning is more complicated than it looks.

Primech AI’s Hytron robot is built for this exact challenge. It is designed to operate in tight, wet environments like restrooms, where automation is particularly difficult.

It can scrub floors, clean toilet bowls, and sanitise surfaces while navigating large facilities. This makes Hytron suitable for places where hygiene standards are high and cleaning needs to be consistent.

Hytron robots are already deployed in some of Singapore’s prominent public hospitals and shopping malls, and large facilities in parts of Asia and Europe, where cleanliness is non-negotiable.

Some other cleaning robots like Pudu Robotics’ CC1 and SoftBank Robotics’ Whiz are also designed for routine cleaning in commercial spaces.

The Trouble with Getting Things Delivered

In a hospital, things are constantly on the move.

Medications travel from pharmacies to wards, and meals, linens and lab supplies move between floors all day long.

Aethon’s robots such as the T3 and Zena RX can transport medical supplies securely, along scheduled routes across hospital floors, operating continuously throughout the day. By doing so, these robots can reduce the time nurses and staff spend moving items and free them to focus more on patient care. Relay and Ottonomy’s delivery robots also explore how supplies can be delivered with care across hospitals, hotels and other large facilities.

Aethon’s robots are already in use across major hospitals in the U.S and Singapore, including the largest fleet of over 90 robots deployed at Children’s Healthcare of Atlanta.

The Trouble with Parking

Parking is one of those problems that most people just put up with.

In dense cities, space can be limited, ramps are tight, and a lot of room is wasted just to accommodate human driving and door-opening.

Shenzhen Shanyi Technologies tackles this problem by removing the human driver from the parking process entirely. Their robots lift and move cars within dense parking facilities, slotting them into spaces that would usually be difficult, sometimes even impossible, to navigate manually. With its automated parking robot, cars can be parked faster and closer together.

Shenzhen Shanyi Technologies automated parking robots have moved beyond pilot trials into production, with systems adopted in commercial facilities in China, including large parking complexes, transport hubs and industrial sites.

Other automated parking systems like Parkie also focus on increasing space efficiency and reducing the hassle of parking.

Where This Leaves Us

By now you can probably picture the kinds of work robots are already taking on. Service robots aren’t meant to be shiny centerpieces, they’re built to do the work. There are many more service robots already operating and still being developed.

As technology matures and challenges around safety and deployment are worked through, we may find ourselves closer to the world of sci-fi…much like Star Wars, where robots live among us as helpers woven into everyday life.

Coming Up Next

Up next, we’ll zoom in to look at the robot value chain and how AI is reshaping robotics, from the brain that makes decisions to the body that carries them out.

Acknowledgements: Thank you to the Sunway iLabs team for their invaluable contribution and insights in preparing this article.

References:

International Federation of Robotics (IFR). (2025). Service Robots See Global Growth Boom. IFR International Federation of Robotics. https://ifr.org/ifr-press-releases/news/service-robots-see-global-growth-boom

Primech AI. (2025, June 12). The Real-World Use Cases of Robotics in Commercial Cleaning - Primech AI. https://primech.ai/the-real-world-use-cases-of-robotics-in-commercial-cleaning/

Robots Authority. (2024). The History of the Roomba: The First Commercially Successful Robot Vacuum. Your Guide AI and Robotics. https://robotsauthority.com/the-history-of-the-roomba-the-first-commercially-successful-robot-vacuum/

Shanyi Technologies（深圳閃移技術有限責任公司）৹ 最先端レーザー、ロボット、計測技術のカンタム・ウシカタ. (2024). Kantum.co.jp. https://www.kantum.co.jp/manufacturer/Shanyi

‌ST Engineering. (2025). Stengg.com. https://www.stengg.com/en/innovation/innovation-stories/transforming-hospital-operations-with-autonomous-mobile-robots/

‌

‌

‌

At first glance, AI feels like instant magic. Type a prompt, get an answer.

But when you look closer, things are not as simple as they seem. Behind every response sits a massive amount of computation, and it all starts with the brain.

In AI, the brain, also known as hardware accelerators, are specialised chips designed to handle the heavy mathematical operations behind AI training and performing inference. These chips process enormous amounts of data in parallel.

Graphics Processing Units (GPUs) dominate today’s AI ecosystem, while custom chips like Tensor Processing Units (TPUs) and Language Processing Units (LPUs) are built for more specialised AI tasks.

From School to the Workforce

A useful way to think of AI training is to compare it to how humans grow up.

As children, we spend years learning. We absorb knowledge, but only later does that learning begin to show its value, when we enter the workforce.

AI pretty much works the same way.

Before it can be useful, an AI model has to be trained. This training phase can take months and consumes enormous computing power (Epoch AI, 2025). Costs pile up, but there’s little direct payoff at this stage. Like education, training is expensive…but necessary.

Once an AI model has gone through training, the real question is this:

How does it deliver answers in real time—quickly and reliably?

Because AI chips don’t just train models, they’re also responsible for inference—the process of generating answers from prompts. This depends on what’s happening inside the AI’s brain at the moment the request comes in. So let’s break down how GPUs retrieve and process data during inference.

GPU, TPU and LPU: Where is Data Stored?

Picture AI running inference like chefs busy preparing cheesecakes for new orders. The bakery kitchen represents the GPU, while high-bandwidth memory (HBM) is the large storeroom next door where ingredients are stored and retrieved.

• The storeroom = HBM, where ingredients (data) are kept

• The counters, shelves, and clear walkways = bandwidth, allowing ingredients to move quickly and in bulk

• The bakery kitchen = the GPU, where all the baking (inference) happens

• The chefs and baking stations = the cores that do the work

The chefs first fetch the ingredients they need from the storeroom, carry them across the kitchen, and bring them to their stations, where the actual baking happens.

They can prepare many cheesecakes at the same time, as long as ingredients keep flowing smoothly from the storeroom to the chefs. This setup isn’t unique to GPUs. Other hardware accelerators, such as TPUs, also rely heavily on fetching data from HBM.

But here’s the catch.

When too many chefs are baking at once, they often need the same ingredients simultaneously. Even with wide walkways and counters, chefs start waiting on one another. As orders pile up, movement slows, bottlenecks form, and the kitchen becomes congested.

In short, GPUs rely on a large, shared storeroom that many chefs must access at once. This is why GPUs can experience heavy memory traffic and higher inference latency when generating responses.

But relying on a shared storeroom (HBM) isn’t the only way to prepare an order.

LPU: Ingredients at the Baking Station

Now imagine a different kitchen.

Instead of chefs constantly walking to a shared storeroom, each baking station has its own small shelf of ingredients. This shelf represents SRAM, much smaller than the main storeroom (HBM), but placed right where baking happens.

Because the ingredients are already within arm’s reach, there’s no waiting, no queueing, and no kitchen congestion. Chefs can start baking immediately, resulting in faster and more predictable response times.

This approach has delivered around 2× faster real-time inference in independent tests, which helps explain why LPUs are attracting attention as a complement to GPUs (CNBC, 2025).

The trade-off is capacity.

Each station’s shelf can only hold a limited number of ingredients, so it can’t replace a large shared storeroom. Fitting much more of this fast storage directly into the kitchen remains a major technical challenge (Asia Business Daily, 2026).

Also, LPUs are designed mainly for inference. Training large AI models still depend on GPUs, TPUs and other training-focused hardware.

The Brain Now Needs a Traffic Controller?

As we know by now, AI runs on massive computing power, AI chips aren’t getting any cheaper, supply is tight, and no existing architecture fits every job. As more specialised chips enter the mix, new questions emerge:

How do you decide which chips to use (US or China), how do you get the most out of the chips you already have and what if you want to use chips from both US and China?

Something has to manage the flow of work between them. This is where software steps in.

These traffic controllers decide which jobs run on which chips and when. The goal is to keep expensive AI chips busy doing useful work rather than sitting idle or waiting in line. In practice, this means spreading workloads more evenly and making sure each chip is used for what it does best.

This is where platforms like Infinigence, Exostellar, Run:ai and Flex:AI come in:

Infinigence and Exostellar try to solve a broader problem. Both focus on optimising and coordinating across different chip types from different vendors within the same platform.

Infinigence supports both U.S. and China AI chips, while Exostellar focuses on U.S-based GPU ecosystems.

Meanwhile, Run:ai and Flex:AI, focus on optimising AI chips within a specific hardware ecosystem. Run:ai works best with NVIDIA chips, ensuring that all GPUs are used as efficiently as possible, whereas Flex:AI, developed by Huawei, does a similar job but for Huawei’s Ascend chips and other China-based AI chips.

Where This Leaves Us

AI has moved beyond its early days. The focus is shifting from training models to making inference faster, cheaper and more reliable. Much like the internet, progress was slow and access was expensive, but over time, costs fell and usage scaled. In the next phase of AI, efficiency won’t be a nice-to-have, it will be a competitive advantage.

Coming Up Next

Now that we’ve unpacked the brain behind AI, we turn to where the brain is deployed. We’ll explore more on physical AI — where AI systems are embedded into machines that see, move and act in the real world.

Acknowledgements

Thank you to the Sunway iLabs team for their invaluable contribution and insights in preparing this article.

References

Asia Business Daily (2026). Welcome To Zscaler Directory Authentication. Asiae.co.kr. https://cm.asiae.co.kr/en/article/2025123016015748006

Emberson, L. (2025). Frontier training runs will likely stop getting longer by around 2027. Epoch AI. https://epoch.ai/data-insights/longest-training-run

‌Faber, D. (2025). Nvidia buying AI chip startup Groq’s assets for about $20 billion in its largest deal on record. CNBC. https://www.cnbc.com/2025/12/24/nvidia-buying-ai-chip-startup-groq-for-about-20-billion-biggest-deal.html

FutureX Insights. (2026). Smartphones are Getting Pricier Because of AI?. https://www.futurexinsights.news/p/smartphones-are-getting-pricier-because

‌Laurent, A. (2025). Nvidia’s $20B Groq Deal: Strategy, LPU Tech & Antitrust. IntuitionLabs. https://intuitionlabs.ai/articles/nvidia-groq-ai-inference-deal

‌

For corporates, this allows for efficiency, reducing turnaround time, and making better use of legal expertise across the organisation.

What Is Legal Tech?

Legal tech refers to software and AI-based systems designed to help legal professionals work more efficiently by handling routine tasks such as research, drafting, contract review, due diligence, compliance checks, and post-execution management, while humans remain in control of all legal decisions.

How Legal Tech Adds Professional Value in a Corporate Setting

For corporates, the value of legal tech lies in its ability to support legal teams handle large and complex workloads with consistency. AI-enabled tools can review contracts and legal materials at scale, applying the same standards set by legal teams across documents, which can be useful for companies dealing with recurring agreements, regulations, or operations.

Legal tech saves time by taking on repetitive, low-value tasks that requires days or even weeks, such as initial contract reviews or information extraction. This frees internal legal teams and external legal advisors to focus more on important work such as analysis, risk assessment, negotiation, and strategic advice that directly support business decision-making.

Just as importantly, some legal tech tools are designed to spot potential risks areas by flagging unusual clauses, key obligations, or deviations from standard terms. When used appropriately, this can support quality control and help reduce the likelihood of issues being overlooked due to workload volume.

Key Legal Tech Tools and Their Roles

Choosing the Right Legal Tech

For corporates, adopting legal tech works best when it starts with everyday realities rather than technology itself. A helpful first step: Identify where legal processes most often slow the business down, whether that is waiting for contract reviews, multiple drafting iterations, or time spent clarifying legal positions.

Matching Tools in Practice

Different legal AI tools are designed for various types of legal work. Solutions such as Harvey, CoCounsel, or Lexis+ AI can help legal teams and external legal advisors work more efficiently on research and drafting.

At the same time, platforms like Luminance or Libra are better suited for handling high volumes of contracts and compliance-related reviews, such as tenancy agreement, MOUs, Non-Disclosure Agreement (NDA), as well as contract comparison, particularly when benchmarking documents against a consistent standard set by the legal team.

Most importantly, successful adoption respects how legal teams work. Tools that fit naturally into familiar environments, such as Microsoft Word or existing document systems, are far more likely to be used effectively.

A Real-World Perspective

That shift is already visible in practice. A Fortune article published in May 2025 highlights how AI-powered legal tools such as Harvey are taking on routine, time-insensitive tasks, allowing junior lawyers to engage earlier in more analytical and problem-solving work.

More broadly, this reflects how AI can change the balance of work within teams, creating more exposure to practical, real-world experience, while routine tasks are handled more efficiently in the background.

Major Industry Signals: AI Investment and Strategic Acquisitions

Harvey’s $8 Billion Valuation Backed by Andreessen Horowitz

Andreessen Horowitz (a16z) a venture capital firm behind Airbnb, Facebook, Figma, and GitHub, has taken major step into legal tech by leading a $150 million funding round into Harvey, valuing the company at around $8 billion (Bloomberg, 2025). This investment suggests growing market validation that AI for legal workflows is becoming core infrastructure for modern legal practice.

Wolters Kluwer Acquires Libra, Strengthening Integrated AI for Legal

Another significant development came from Wolters Kluwer’s — a European-based professional information and software company — acquisition of Libra Technology GmbH for up to approximately €90 million (RTTNews, 2025). The move integrates Libra’s AI legal assistant capabilities with Wolters Kluwer’s extensive legal content and tools, aiming to deliver a unified platform for legal research, drafting, and analysis built to support lawyer workflows.

Together, these developments show that legal tech is not a trend, but an evolving ecosystem shaped by market demand and professional standards.

Legal Tech Is an Enabler, Not a Replacement

Legal tech does not diminish the role of lawyers, it amplifies professional expertise. These tools help legal professionals focus on what humans do best; judgment, interpretation, strategy, and client care.

The latest investments and acquisitions, from Harvey’s billion-dollar valuation boost to Wolters Kluwer’s integration of Libra, reflect confidence that legal tech will partner with legal professionals to elevate practice standards, not undercut them.

Legal tech is not a shortcut. It is a strategic ally that strengthens how legal work is done, how insights are generated, and how clients are served.

Disclaimer: This article is for general information only and does not constitute legal advice.

References

Clark, K. (2025). Andreessen Horowitz Invests in Legal AI Startup Harvey at an $8 Billion Valuation. Bloomberg. https://www.bloomberg.com/news/articles/2025-10-29/andreessen-horowitz-invests-in-legal-ai-startup-harvey-at-an-8-billion-valuation

Mollman, S. (2025). An AI tool has become so useful to junior lawyers at PwC that “there’d be a riot” if the firm took it away, says partner. Fortune. https://fortune.com/2025/05/09/ai-legal-tool-harvey-junior-lawyers-careers/

‌RTTNews. (2025). Wolters Kluwer To Acquire Libra Technology For Up To EUR 90 Mln. Nasdaq.com. https://www.nasdaq.com/articles/wolters-kluwer-acquire-libra-technology-eur-90-mln

‌

Recently, word has spread that the AI boom is driving a severe shortage of memory chips, the same components used in both AI data centres and everyday devices like smartphones.

SK Hynix and Samsung, two of the world’s leading suppliers of memory chips, have already sold out their high bandwidth memory (HBM) supply through 2026, with industry expectations that the memory shortfall could stretch into late 2027 (Reuters, 2025). As supply tightens while demand remains red-hot, prices rise. Real inflation pressure.

If artificial intelligence (AI) demand is strong enough to push up prices in everyday tech such as phones and laptops, it tells us something important, that our ChatGPT subscription isn’t getting cheaper anytime soon.

What’s really keeping AI expensive?

To make sense of it, the Sunway iLabs team stepped back and compared notes, and one thing became clear. This isn’t a single-issue problem. AI needs three main things to work at scale, and all three are under strain. To paint the picture, AI is intelligence with dependencies: it needs a body to carry it, a brain to think and fuel to run it.

Infrastructure: The body that holds everything together

Infrastructure is the body of AI, the physical foundation everything else plugs into. This includes data centres, and those data centres depend on a whole chain beneath them, such as grid connections that deliver electricity, transformers that manage it, and power plants that generate it in the first place. AI chips can’t run without buildings to house them.

One reason AI isn’t getting cheaper is the sheer cost of building this body. Big tech companies are spending hundreds of billions of dollars building AI-ready data centres, so much so that data centre construction is set to overtake office buildings by 2026 (The Wall Street Journal, 2025c). These aren’t ordinary buildings, they require specialised cooling, power distribution, and built-in backups, all of which drive costs higher.

And then there’s time. Even with unlimited funding, data centres and grid connections can’t be built overnight. Physical infrastructure expands slowly, in years rather than months. That heavy, long-build reality puts a natural limit on how fast AI can scale, and helps explain why costs stay high, even as the AI technology itself continues to improve.

Hardware accelerators: The brain doing the thinking

If infrastructure is the body, then hardware accelerators are the brain. AI semiconductors like GPUs and TPUs do the heavy mathematical work behind training and running AI models.

The challenge is that this brain is hard to scale. Most of the world’s AI chips, such as GPUs and TPUs are manufactured by a single company, TSMC. As demand has surged, supply hasn’t been able to keep up, making shortages and pricing pressure hard to avoid.

A major bottleneck isn't just making the chips themselves, but packaging them. AI chips require complex packaging processes that depend on specialised equipment from many vendors and expanding this capacity takes time (Epoch AI, 2024). And because NVIDIA's GPU chips have become the industry standard, many companies are effectively competing for the same limited supply.

Alternatives like Google's TPUs are either tightly controlled, while others such as Huawei's Ascend chips are still catching up (The Wall Street Journal, 2025a). So even as AI software advances quickly, the brain it runs on doesn't scale at the same speed, which helps explain why costs are unlikely to fall meaningfully anytime soon.

Power: The fuel that keeps the lights on

AI doesn’t just think, it consumes. Training AI models takes enormous amounts of electricity, and once those models are deployed, every question asked, image generated, or chatbot reply — known as “inference” — adds to the load. Together, training and usage keep the power meter running almost constantly.

This is where costs start to balloon. Some countries are better positioned than others. In this case, China has a relative edge, with far greater power-generation capacity than the US and cheaper electricity in many regions (The Wall Street Journal, 2025b). The US, by contrast, has strong demand but slower grid expansion, creating congestion and rising costs (Epoch AI, 2024).

At the same time, emerging, fast-growing markets like Malaysia are accelerating data centre expansion precisely because electricity costs are significantly lower than in the US (ACCEPT, 2025).

But here’s the catch: cheaper power doesn’t solve everything. AI remains extremely energy-hungry. Unless AI systems become far more energy efficient, electricity costs alone will keep AI expensive, even in energy-rich countries.

Where this leaves us

When you line up the clues, things start to make sense. AI isn’t expensive just because innovation has stalled. It’s expensive because multiple parts of the system are under strain at the same time.

The most visible constraints are physical: a body that takes years to build, a brain that’s hard to scale, and fuel that keeps burning. Together, they help explain why AI costs remain high. But the physical constraints aren’t the whole story, software, operations and geopolitical conditions matter too.

We’ve seen this pattern before. In the early days of the internet, access was limited, hardware was pricey, and bandwidth felt like a luxury. Over time, as networks expanded and technology matured, costs came down and the internet became accessible everywhere.

AI may follow the same path. But for now, this isn’t a one-bottleneck problem. It’s a system problem.

Coming Up Next

In the upcoming piece, we’ll zoom in on hardware accelerators, the brains behind AI. We’ll take a a closer look at GPUs, TPUs and emerging alternatives like LPUs, explore how these chips differ, where each excels and what their evolution could mean for the future scalability of AI.

Acknowledgements

Thank you to the Sunway iLabs team for their invaluable contribution and insights in preparing this article.

References

• Huang, R., & Qin, S. (2025a). Huawei Unveils AI Chip Pipeline as Tech Rivalry Heats Up. The Wall Street Journal. https://www.wsj.com/tech/huawei-unveils-ai-chip-pipeline-as-tech-rivalry-heats-up-07600dc8?mod=article_inline

• Huang, R. & Spegele, B. (2025b) China’s AI Power Play: Cheap Electricity From World’s Biggest Grid. The Wall Street Journal. https://www.wsj.com/tech/china-ai-electricity-data-centers-d2a86935?st=6h2Mq9&reflink=desktopwe…

• Jin, H., Potkin, F., Lee, W.-Y., Bridge, A., & Cherney, M. A. (2025). The AI frenzy is driving a memory chip supply crisis. Reuters. https://www.reuters.com/world/china/ai-frenzy-is-driving-new-global-supply-chain-crisis-2025-12-03/

• Maurer, M. (2025c). AI Construction Costs Can Be an Accounting “Black Box.” The Wall Street Journal. https://www.wsj.com/articles/ai-construction-costs-can-be-an-accounting-black-box-3c197b09

• Nathania Azalia. (2025). The Rise of Data Centres, Artificial Intelligence, and ASEAN’s Decarbonisation Goal - ASEAN Climate Change and Energy Project (ACCEPT). ASEAN Climate Change and Energy Project (ACCEPT). https://accept.aseanenergy.org/the-rise-of-data-centres-artificial-intelligence-and-aseans-decarbonisation-goal

• ‌Sevilla, J. (2024). Can AI Scaling Continue Through 2030? Epoch AI. https://epoch.ai/blog/can-ai-scaling-continue-through-2030

If you told a toy executive 15 years ago that a Chinese collectible brand would be worth more than Hasbro, Mattel and Sanrio combined, they’d probably laugh you out. Yet as of mid-2025, Pop Mart’s valuation stood at roughly US$44.43 billion, surpassing the old-school giants of the global toy industry.

The secret isn’t plastic. It’s feelings.

Gen Z and the Rise of Emotional Value

And now, increasingly, Gen Z consumers are shopping for emotional return on investment (including some GenX and Y, ahem...). Blind box toys, where buyers don’t know which character they’ll get, perfectly tap into this mindset. The thrill of uncertainty, the dopamine hit of surprise, and the comfort of collecting, all play into a deeper desire, emotional regulation.

Characters like Labubu, Molly, Skullpanda, and Crybaby aren’t just toys. They’re mood avatars. One Skullpanda for a melancholic day. A Crybaby is clipped onto a bag when life feels overwhelming. A Labubu guarding your phone like a mischievous emotional support gremlin. These figurines travel with their owners as accessories, tiny reflections of inner states.

Then the blind box concept hasn’t stayed in toys. It’s quietly resurfaced in entertainment—especially music. Buying albums with randomised celebrity photo cards, forcing fans to purchase multiple copies to get the idol they want, feels curiously like the 1990s all over again: same psychology, new packaging.

Buying something with “little practical value” suddenly makes perfect sense when it contributes heavily to emotional well-being.

But emotional economies are fragile. By December 2025, Pop Mart’s stock had slid nearly 44% in just four months, wiping out over US$25 billion in market value. One primary concern: overexposure. Analysts warned that aggressive production, especially of Labubu, could destroy the scarcity that made the brand desirable in the first place.

When everyone has the same emotional anchor, it stops feeling special.

What’s the next emotional investment after blind boxes?

Enter CES 2026, the Consumer Electronics Show — the world’s biggest stage for consumer technology. This year, one signal became impossible to ignore: physical IP is converging with AI companionship. It’s not so much about the clunky, humanoid robots anymore but shift toward something far cuter — a plushie that interacts with you and… talks to you (yes, slightly Chunky vibes).

• Ropet by Mengyou Intelligence first appeared at CES 2025. Pre-sold in JD.com China, and covered by Forbes, BBC, and CNBC, it is emotionally responsive. Then, more in CES 2026, the floodgates opened.

• Leimeng Technology’s Bibo looks like an electronic cousin of Labubu. It recognises faces, tone, and movement, reads emotional states, and proactively interacts using emotional AI.

• Lovipeer by LOVEAXI goes deeper, combining micro-expression recognition, NLP, adaptive learning, and bilingual conversation. It can comfort you when you’re sad, and tutor your child in English when you’re not.

• Honghuo Moya, wrapped in plush fur, focuses on sleep and “active embodied empathy,” offering emotional support alongside practical assistant functions.

• AiMOON by Lynxaura Intelligence (灵犀智能), which looks similar to skullpanda, is powered by zodiac culture and emotional memory systems. It tracks moods, remembers preferences, and shifts personality accordingly: therapist, hype partner, or professional complainer.

Even legacy giants are paying attention. Panasonic’s NICOBO responds to touch, recognises faces and voices, and exists purely to make homes feel less lonely.

Where This Leaves Us

Surprise boxes were never really about toys. They were about outsourcing emotion, letting objects carry feelings we didn’t know how to process ourselves. AI companions take this idea one step further. Instead of silent figurines, we now have responsive, learning, emotionally aware entities.

In the future, every household will likely own one or two companion robots. Not as gadgets, but as emotional infrastructure, perhaps, digital extensions of the human.

But one thing is clear, the next wave of consumer tech won’t just solve problems. It will feel something while doing it, and maybe think with us. Cute or creepy? We’ll find out soon, won’t we.

Across Southeast Asia, the startup ecosystem is entering a tougher funding environment. Recent high-profile startup failures have reinforced a more cautious approach to funding, bringing recurring red flags into sharper focus, including poor governance, inflated revenues, and unsustainable burn rates.

This current retreat is a much-needed correction that is also reshaping what investors look for next. Investors are increasingly backing start-ups with strong intellectual properties and differentiated artificial intelligence (AI) capabilities compared to those focused solely on operational execution, says Karen Lau, chief operations officer of Sunway Innovation Labs (Sunway iLabs), the innovation lab and corporate venture capital (CVC) arm of Sunway Group.

The Edge — December 29, 2025
Read More...

Sunway iLabs and and JETRO concluded the J-StarX Malaysia Deep Tech Programme (MYTech) 2025, accelerating four Japanese startups’ expansion into Malaysia and the Asean region. Marking its seventh year of collaboration, the accelerator programme leveraged on Sunway’s living laboratory for market access and JETRO’s extensive network, reinforcing Malaysia as a launchpad for Japanese innovators seeking regional growth opportunities.

The 2025 cohort featured Japan’s diverse applications of AI and deep tech including Buddy Cloud’s mobile-enabled veterinary diagnostics, FiberCraze’s sustainable nanofibre textiles, BioPhenolics’ bio-based phenolic compounds, and I.W.G’s interoperable healthcare data platform, DoCloud — each reflecting how innovation can create meaningful, sustainble impact across industries.

Digital News ASIA — November 27, 2025
Read More...

FutureX Ventures Fest redefines the traditional conference:

curating an exclusive AI / deep tech beachside gathering of 200 selected leaders from startups, investors, government, and corporates to spark meaningful conversations shaping the future of AI / deep tech while strengthening Malaysia’s position as a regional innovation hub.

Future of AI / Deep Tech Driving Real-World Impact

This year, FutureX is anchored around three key themes:

1. Reimagining Healthcare with AI

2. AI Frontier & Emerging Breakthroughs

3. AI as a Catalyst for Ecosystem Growth

With perspectives from leading players like a16z Perennial, one of the most successful tech investors in the world; OpenAI, the company behind the world-renowned frontier model, ChatGPT; and thought leaders from global investment and innovation ecosystems, the conversations go beyond hype into what it truly takes to build, fund, and scale the next wave of technology.

Good Vibes To Expect

Set against a beachside backdrop, FutureX Ventures Fest blends intellectual exchange with experiential moments, from the Sands of Surprise and Networking Soiree to Fire Show and DJ Soundwave Session, creating a dynamic setting where meaningful connections happen both on and off stage.

Join us at FutureX Venture Fest 2026 and be part of the conversations shaping what’s next: let’s make waves together.

🌊 REGISTER NOW🌊

Wondering what FutureX is all about? Catch a glimpse at last year’s highlights:
Watch video...

The FutureX Connect event brought together cross-border startups and visionary leaders from Malaysia, Japan, South Korea, China and beyond to dive deep into the world of AI and deep tech 🌐

At the 22 October 2025 event, leading minds in AI and Deep Tech unite with one mission: to foster cross-border innovation that pushes boundaries and transforms industries.

Throughout the event, industry leaders and experts got the chance to move beyond showcasing innovation and into real connection. Founders, corporates, and ecosystem builders came together to question, collaborate, and co-create — turning global perspectives and local ambition into meaningful momentum.

Watch Video...

That momentum carried into broader industry conversations. At the panel “From Borders to Breakthroughs: AI’s Role in Global Industry Transformation,” ecosystem stakeholders and industry experts shared how early AI adoption is already seeing measurable impact. As Malaysia advances its NIMP 2030 vision, the discussion highlighted the need for stronger public-private collaboration, smarter factories, and a workforce ready to adapt and innovate.

Read More...

🌊 Making Waves in AI / Deep Tech 🌊

FutureX Ventures Fest, a high-energy beachside convergence of AI and deep tech minds brings together visionary founders, forward-thinking investors, and ecosystem builders to exchange ideas on trends, opportunities, and what it takes to build successful early-stage deep tech startups.

It is a space where people are showing up for:
💧Connections that matter
💧Insights that spark breakthroughs
💧To walk the AI & Deep Tech talk while having the time of your life

At the 2025 edition of FutureX Ventures Fest, held on 15 April 2025, the energy, conversations and connections spoke for themselves. Somewhere along the way, it earned the nickname “the Coachella of investors and startups” and we’re not arguing.

Missed it? Drop us a message to catch the next one.
Watch video...

Remember when ChatGPT rolled out its GPT-4o image generator earlier this year and suddenly half the internet became ghiblified? Social media feeds were flooded with AI-generated images that looked straight out of a Studio Ghibli film, from celebrities and politicians to movie scenes, memes, and even people’s pets.

If AI can imitate creative works so convincingly that it feels almost… too good, where do we draw the line between inspiration and imitation — and who owns it?

Read More...

Where We Left Off

Previously, we covered how the law treats AI training as fair use. What’s still off-limits, though, is imitation. If generated content too closely resembles existing copyrighted material, it can still amount to infringement. In other words, training an AI model and using its output are two very different things. And for organisations using generative AI in real-world settings, that difference matters.

Precedents in the Making

While legal frameworks around AI and copyright are still evolving, some companies are choosing a more proactive route: working directly with rightsholders.

One of the most high-profile examples is the partnership between the Walt Disney Company and OpenAI, which enables OpenAI’s video platform, Sora, to generate videos featuring more than 200 characters from Disney, Pixar, Marvel, and Star Wars.

In the music industry, KLAY Vision Inc. has taken a similar approach, having signed licensing deals with major record labels including Universal Music Group, Sony Music Entertainment, and Warner Music Group, enabling AI training on authorised music rather than relying on unverified data sources.

These emerging patterns point to leading AI firms deliberately adopting responsible approaches to creative works. That, in turn, shifts the focus back to the users and companies using AI today:

How should organisations approach using AI tools responsibly in our own work?

When using generative AI, the risk often isn’t intentional copying, but somewhat accidental resemblance. Because AI outputs are based on learned patterns, what it produces can sometimes end up substantially similar to existing copyrighted works.

That distinction matters. Intentionally prompting an AI to replicate specific copyrighted materials is a conscious choice, not a technical accident, and responsibility for that choice sits with the user.

Responsible GenAI Use 101

There’s no single checklist that eliminates IP risk, but a few practical measures can make a meaningful difference and allow teams to continue using generative AI productively.

1. Originality over convenience

AI should support human creativity, not replace it. When using these tools, organisations should prioritise meaningful human input, such as creative direction, editing and judgement, to ensure outputs don’t drift too close to existing works.

This aligns with the guidance issued by the U.S Copyright Office in early 2025, which clarified that for any part of a work to be copyrightable, there must be a meaningful human authorial contribution. If a machine does most of the heavy-lifting without specific human guidance, the output remains in the public domain.

One practical approach to this is at the prompt level. Highly specific prompts that reference recognisable styles or aim to replicate a particular aesthetic increase the risk of outputs resembling protected works. Avoid asking AI to create content “in the style of” identifiable creations, such as:

• “Write like [famous author]”

• “Draw like [artist]”

• “Make a logo like [brand]”

The goal is to create something new, not a lightly modified version of someone else’s work. Thoughtful prompting, combined with human refinement, helps keep AI-generated content on the right side of originality.

2. Rules before tools

Organisations should lay out their AI-use policies — which AI tools may be used, for what purposes, and which data can or cannot be used. One approach to this is to group AI use into three risk-based levels:

• Low risks:

  Everyday tasks that do not involve sensitive or proprietary information, such as brainstorming ideas, summarising non-confidential text, or drafting internal emails, can generally be considered acceptable. Organisations can mandate a simple disclosure statement for any AI-assisted work to maintain transparency.

• Medium-risk:

  Content intended for external use, such as marketing drafts, images, or code prototypes, should go through basic review, including plagiarism scans, reverse image searches and license or open-source code checks. These steps help ensure the output is appropriate and doesn’t closely resemble existing copyrighted materials.

• High-risk:

  Confidential information, trade secrets or client data should never be entered into public AI platforms without safeguards. These materials require greater control. They should only be handled in secure environments, such as internal LLMs hosted on-premises or private enterprise AI instances where data remains within the organisation’s network.

Organisations with the highest security requirements can also adopt confidential computing. This cloud technology keeps data encrypted even while the AI is actively processing it, helping protect sensitive information from both external leaks and unauthorised internal access. These guidelines keep decision-making simple, help teams work confidently with AI, and reduce the risk of accidental misuse (WIPO, 2024).

3. Know the fine print

Not all AI tools treat ownership and liability the same way. Reviewing the terms and conditions of AI platforms helps organisations make informed decisions about which tools are appropriate for their use.

Some AI platforms, like Midjourney, for instance, disclose that they may reuse user-generated content, including for purposes unrelated to AI training, and these terms can change quickly as the sector evolves.

In practice, organisations should know:

1. Who owns the output

2. How the content may be reused

3. What responsibilities does the AI platform cover or not

That said, because platform terms can change quickly, organisations may also consider conducting regular reviews of their AI portfolio, such as quarterly, to stay aware of any updates to data usage, ownership rights, or liability terms that could affect how AI tools are used internally or externally.

Moving Forward

As AI tools become part of everyday workflows, it’s worth noting that while AI training may be allowed, copyright rules still matter. AI can speed up creation, but it doesn’t remove the need for review, judgment, or responsibility.

For organisations, this means being mindful of how data is handled, understanding what rights platforms retain over generated content and ensuring outputs are reviewed before external use.

And for all of us, the principle is the same: use AI to support creativity, not shortcut it. Prioritising originality over imitation remains key. AI is a powerful tool, but how responsibly it’s used by companies and individuals alike is what ultimately shapes its impact.

Disclaimer: This article is for general information only and does not constitute legal advice.

Acknowledgements: Thank you to the Sunway iLabs team for their invaluable contribution and insights in preparing this article.

References:

Corcoran, N. (2025). Major labels sign licensing deals with AI music company Klay. Pitchfork.

Generative AI. (2025). Authors Alliance. Guan, L. (2025, June 17). Securing Gen AI With Confidential Computing. Accenture.com; Accenture.

U.S. Copyright Office Clarifies AI Authorship Rules in Part II of Its Copyright & AI Report. (2025).

World Intellectual Property Organization. (2024). Generative AI Navigating Intellectual Property IP and Frontier Technologies.

What happens when you give the same magical prompt to Veo 3.1, Hailuo 2.3 and Qwen 3-Max? A whole lot of sparkle, surprises, and wildly different interpretations. Here’s our fun little GenAI showdown.

Read More...

Sunway iLabs partners Alibaba to strengthen regional AI startup ecosystem

Sunway iLabs and the Alibaba Entrepreneurs Fund (AEF) – NextGen Fund has entered a strategic partnership to deepen collaboration between Malaysia, Hong Kong, and the Greater Bay Area (GBA). The partnership opens a gateway for Sunway iLabs to tap into cutting-edge AI developments from Hong Kong, while positioning Sunway’s integrated ecosystem as a living lab where AI startups can explore real-world applications and regional pathways for adoption.

Read More...

Sunway iLabs and the Alibaba Entrepreneurs Fund (AEF) – NextGen Fund has entered a strategic partnership to deepen collaboration between Malaysia, Hong Kong, and the Greater Bay Area (GBA). The partnership opens a gateway for Sunway iLabs to tap into cutting-edge AI developments from Hong Kong, while positioning Sunway’s integrated ecosystem as a living lab where AI startups can explore real-world applications and regional pathways for adoption.

Digital News Asia — December 10, 2025
Read More...

Sunway iLabs and China International Capital Corporation (CICC) announced partnership at FutureX Connect in Kuala Lumpur to advance AI technology and strengthen innovation ties between China and Southeast Asia. The partnership bridges China’s mature AI ecosystem with Malaysia’s growing digital economy.

Digital News Asia — October 24, 2025
Read More...

Despite only 26% of manufacturers having embraced AI, early adopters are already seeing clear productivity and efficiency gains. According to FMM’s Helena Eian, companies using AI across areas like customer service, quality control, and predictive maintenance report up to 60% higher productivity and faster decision-making.

Backed by Malaysia’s New Industrial Master Plan 2030, the nation aims to transform 3,000 smart factories, fostering innovation and high-skilled jobs. Experts from Sunway Group and industry leaders stress that collaboration and regulatory reform are key to unlocking AI’s full potential in Malaysia’s manufacturing sector.

The Edge Malaysia — October 23, 2025
Read More...

Sunway iLabs has partnered with Cradle Fund to launch the Startup Deep Tech Ventures Lab (DTVL) to build and scale high-impact artificial intelligence (AI) and deep tech startups. DTVL is a transformative six-month co-creation programme, aims to catalyse breakthrough solutions to some of the region's most critical challenges, setting the stage for a new era of technological advancement and entrepreneurial excellence.

New Straits Times — August 4, 2025
Read More...