Interview with Mr. Ning CHEN, CEO of Intellifusion Technologies Co., Ltd.
Interview with Mr. Ning CHEN, CEO of Intellifusion Technologies Co., Ltd.
While much of the world is in vacation mode, Geneva—and in particular the International Telecommunication Union (ITU)—is hosting a significant conference entitled AI for Good. Much has been said about artificial intelligence, its rapid development, and how these new technologies will transform our lives.
On this occasion, we had the opportunity to meet with the CEO of one of the pioneering companies in the AI field, who participated in a roundtable discussion dedicated to the ethics of AI and its role in global governance.
The floor is yours, Mr. CHEN.
Q: You represent the companies that are taking us from the more traditional society into the new digital society. How do you see the evolution from here?
From my personal experience, attending the AI for Good Summit organized by ITU reminded me very much of my journey 20 years ago. At that time, I was a telecom engineer, even serving as a delegate for 4G LTE standards for more than four years. Back then, I was deeply involved in standardization work.
But in 2014—eleven years ago—I made a major shift. I founded my own company in Shenzhen, China, called Intellifusion, and moved from telecommunications into the field of artificial intelligence. Our focus has been on designing AI inference chips.
Looking at today, I believe 2025 marks an important turning point: the transition from the training era of AI to the inference era. So what does AI inference mean? Over the past few years, NVIDIA’s general-purpose GPUs have been absolutely essential for training large models — they enabled huge leaps in capability. Training remains compute-heavy and is typically done in large data centers. Inference, by contrast, is about running those trained models in real-world applications: making predictions, recognizing speech or images, translating in real time, driving vehicles, or helping doctors with diagnostics.
Inference brings a very different set of constraints and priorities: latency, power consumption, cost per query, privacy, and deployment footprint. You can’t power every device with a datacenter GPU; many applications need on-device or near-edge inference to meet real-time and privacy requirements. That’s why specialized inference chips — low-power accelerators, efficient architectures, and software stacks that support quantization, pruning, and model compression — are becoming critical. They allow mature models to run widely and affordably at scale.
This shift has practical consequences. Companies can now embed intelligent functions into everyday devices — phones, cameras, medical scanners, factory robots — rather than relying solely on cloud calls. That improves responsiveness, lowers bandwidth and energy use, and helps protect sensitive data. But it also raises new engineering challenges: optimizing models for accuracy vs. size, verifying behavior on constrained hardware, and ensuring robustness in diverse real-world environments.
From a governance and ethics perspective, the inference era changes the landscape as well. Standards and certifications for safety, privacy, energy efficiency, and fairness will matter more because AI will be physically present everywhere. Organizations like the ITU and multi-stakeholder forums have an important role in creating interoperable standards and best practices so that devices from different vendors behave predictably and responsibly.
At Intellifusion, we’re focused on bridging this gap: building inference hardware and toolchains that make deployment straightforward while meeting energy, latency, and security targets. But industry alone can’t solve the broader questions — we need collaboration between governments, academia, civil society, and standards bodies to make sure the inference era benefits everyone.
That, in short, is why I see 2025 as the beginning of a new chapter: training produced the powerful models; inference will determine how—and how widely—those capabilities touch people’s lives.
Q So we’re ready to use AI in our everyday life?
Over the last six months, there has been a noticeable shift in the conversation. Previously, everyone was focused on large language models such as ChatGPT. Now, the focus has moved to agents.
So what do we mean by “agent”? An agent is essentially an application built on top of large language models. It is not just about generating text, but about enabling real interactions, decision-making, and actions in the real world.
This evolution has enormous implications for hardware. AI is set to redefine smart devices: the smartphone, the earphone, even AI-powered glasses. Looking ahead three to five years, it is very likely that most home appliances will be equipped with large models — which means they will be able to communicate with us directly. Imagine your washing machine, your refrigerator, or your car not only executing commands but reasoning, conversing, and adapting in real time.
This is why the concept of the agent is so important: it signals a future where AI moves off the screen and into our daily environment, becoming an integral part of how we interact with the world around us.
Q: How to make that happen?
To make this vision possible, we need to put inference chips directly into smart devices — and at the same time build AI inference data centers distributed across the globe.
This is why I say we are entering an AI inference era. In the past, we lived in the AI training era. To use an analogy: think of electricity. Today, it is unimaginable to live without electricity — it powers our entire civilization. The power plants generate electricity, but what really matters is how it is used, everywhere, by everyone. AI works in a similar way. Training is like the power plant — large, centralized, and resource-intensive. Inference, on the other hand, is like the entire electricity network: ubiquitous, embedded in every device, and transforming daily life.
Just as we only need a limited number of power plants in each country, but an unlimited number of sockets, appliances, and networks to make electricity useful, the same will be true for AI. Inference chips will be everywhere — in homes, in cities, in industries. They will resemble the role of 4G and 5G base stations, or Wi-Fi routers for wireless communication: invisible infrastructure, but absolutely essential.
Over the next five to ten years, AI inference chips, devices, and servers will form a pervasive layer of intelligence. This will allow all of our devices to access and run large language models — tapping into their “tokens” and outputs seamlessly, anytime, anywhere.
That is exactly what we are building at Intellifusion.
Q: How did you get the ideas that led you to develop your company, given your background in telecom? It seems like a very different industry.
Yes, on the surface it seems very different, but in reality I have always been working on the same foundation: parallel computing.
Twenty years ago, as a telecom engineer, I was designing parallel computing chips for wireless communication — baseband chips for cell phone base stations. Even in our smartphones today, parallel computing chips are what make them more and more powerful, because they can handle so many computations at once.
Artificial intelligence, however, requires this on a much larger scale. Large language models have billions of parameters, which means they demand extremely intense computation. And that is what led us to make the transition.
I remember very clearly the moment: around 2012, we came across a paper called AlexNet, written by two PhD students together with Professor Geoffrey Hinton, who later received the Nobel Prize in Physics. They proposed a new kind of deep neural network, and they tested it in the ImageNet competition — an international contest where machines had to recognize images of cats, dogs, plants, and so on.
Up until then, machine accuracy was always below the average human level. But in 2012, AlexNet outperformed human beings for the first time. The machine could identify objects more accurately than people. That breakthrough captured the attention of both academia and industry worldwide.
Looking back, I believe that paper was the ticket to the Fourth Industrial Revolution. It signaled a new era, where AI could truly surpass human capabilities in perception tasks.
That realization is what pushed us to take action. We decided that if deep learning was the future, then we needed to design chipsets specifically optimized for it — more power-efficient and better suited to neural networks like AlexNet. That was the starting point of our company. In a way, we didn’t abandon telecom; we extended our expertise in parallel computation into a new domain. But yes, that was the moment we shifted decisively from telecommunications to artificial intelligence.
Q: In the future, then, you see that more and more things that we do manually will be replaced by robots or AI. So ,what will happen to us?
Honestly, I’m optimistic. I believe AI will create more jobs than it replaces.
If we look back at history, this has always been the case. Think about the past 20 or 30 years: so many jobs exist today that didn’t exist at all back then. For example, the role of a software programmer — today millions of people are working as developers, but 30 years ago this profession barely existed. Every major technological movement has created entirely new categories of work.
The same will happen with AI. In five to ten years, we may see roles like maintenance engineers for robots, prompt engineers for large language models, or product engineers for AI agents. These jobs don’t really exist today, but they will become common as AI integrates into our economy.
So while AI will certainly change the nature of work, I see it as an opportunity for society to evolve — not as something to fear.
Q: Regarding your company, could you tell us in simple terms what you offer — something understandable for non-specialists like me?
At Intellifusion, what we provide are AI inference chips that are low-cost and very power-efficient.
You can think of them a bit like the wireless communication chips inside your cell phone. Those chips made it possible for mobile phones to connect to 3G, 4G, and 5G networks. In the same way, in the AI era, every smart device will need an inference chip.
These chips are what allow a device — whether it’s a phone, a home appliance, or even a pair of AI glasses — to actually run large AI models. Without them, the device can’t really be “AI-powered.” With them, the device can process information, make intelligent decisions, and interact with us in real time.
So in simple terms: our chips are the engines that bring AI into everyday devices.
Q: So if I understand correctly, a company — say in Norway or elsewhere — that wants to build AI-powered solutions should contact you to get the right chips?
Exactly. What we provide are low-cost, power-efficient, yet very high-performance AI inference chips, and they are designed for three main sectors.
First, robotics.
Our chips can serve as the “brains” of robots, enabling them to process information and act intelligently in real time.
Second, cloud computing.
For example, if Norway wants to build AI inference data centers, they would need high-performance chips to power those centers. These centers make it possible for people to use large language models directly on their phones or devices. We design chips that are cost-effective but powerful enough for these large-scale computing needs.
Third, edge computing.
This is what I find most exciting, because it brings AI right to where we live. Edge computing means running AI close to the user — not in a distant cloud but on devices in your home, office, or community. Imagine having your own personal large language model deployed in your home for privacy and security.
Here’s an example: you come home and say, “Show me all the pictures from the first half of this year when I went fishing with my son.” Instantly, your home AI device could scan millions of images, find the right ones, and send them to your phone — or even post them directly on social media.
The same system could also manage your smart home: protecting against intruders, managing alarms, controlling appliances. Your refrigerator, for example, could notice that you’re out of milk or butter, and send that information directly to your home robot — so your household stays one step ahead.
This is the future we are building: AI that is distributed across robots, data centers, and edge devices, seamlessly working together to make life smarter and easier.
Q: How long do you think it will take before these things really happen? And isn’t this mostly something for rich countries?
I would say within five to ten years. By 2030, we will definitely reach a major turning point. In the next five years, AI could fundamentally change the way we live.
Of course, today much of this is happening in advanced economies — countries like the United States or China. Let me give you an example from Shenzhen. A few months ago, the internet influencer IShowSpeed — he has tens of millions of followers worldwide — toured China and live-streamed his experiences in more than eight cities. Shenzhen stood out. Everything he experienced there was about high tech: he drove a water-floating electric car, danced with a humanoid robot, ordered KFC on his phone, and even had it delivered by drone. That kind of everyday integration of AI and advanced technology is already real in some cities.
But the key question is: how do we make this accessible to everyone, not just in rich cities or countries, but also in rural areas, in Africa, in developing regions? That is why we are here at the AI for Good Conference.
To me, there are two dimensions of “AI for Good.” First, how do we build good AI — meaning ethical, safe, and responsible technology? Second, how do we ensure AI for all — so that every human being, regardless of geography or income, can benefit from it?
For this second goal, I believe organizations like the ITU can play a vital role. Just as the ITU once drove international standards for telecommunications, it could now lead an initiative for a global standard on AI inference computing power networks. Such a framework would lower costs and make it easier for developing countries to gain access to AI inference chips and computational power.
If we succeed, then AI will not only transform advanced economies, but truly uplift people everywhere — making everyday life smarter, safer, and more connected for all.
Q: Another thing I was wondering about is the protection of personal data. How do you protect people’s personal lives with AI? Are there any international regulations in the pipeline?
Data privacy is indeed a central topic when we talk about AI for Good. There are already many standards and regulations being developed, including initiatives led by the ITU, to ensure privacy is protected.
Technically, there are also strategies to protect privacy, and in the future, we foresee three main computational scenarios. That’s why we design three different types of AI inference chips with privacy in mind:
1. Terminal inference chips – these are installed directly in devices like your cell phone. The AI large language model runs entirely on the device itself, so there is no need to transfer your personal data to the cloud. I estimate that in the future, 70–80% of tasks could be handled entirely on your phone.
2. Edge computing devices – these are located in your home or office. Your data — images, documents, and other personal information — can be stored and processed locally without ever going to the cloud, giving you full control over your privacy.
3. Cloud-based AI inference – for tasks that require massive computation, data can still be processed in secure cloud environments with strong privacy protections, but the goal is to minimize unnecessary sharing of personal data.
In short, with the combination of regulations and these technical solutions, we can ensure AI benefits users without compromising privacy.
Q: But do we need offices in the future if so much work is done on cell phones? How will people meet in society? Will there be more loneliness?
When I say 70–80% of work will be done on your cell phone, I mean personal tasks, entertainment, and computations that can be handled locally. Your phone itself will take care of that processing power — it doesn’t mean all jobs or all work will disappear.
Offices will still be needed. Some tasks require collaboration, large-scale projects, or more specialized machines. In the office, edge computing devices — like robots, digital agents, or AI-powered workstations — will handle workloads while still protecting privacy. So in the future, there will be three layers of computation: cloud, edge, and terminal, all working together to balance efficiency and privacy.
Q: Second question: we publish a paper magazine. Do you think people will still read printed materials in the future, or will everything be digital?
Honestly, printed media will likely decrease over time. Media consumption is changing rapidly — people might read papers on their AI-powered glasses, for example. What really matters in the future is not the medium but the content: AI-generated content (AIGC), user-generated content, or author-created content. The key challenge will be ensuring that this content is interesting, relevant, and high quality.
Q: Finally, because I know you don’t have much time — our readership is mainly diplomats and people from the international community. What can you offer them specifically, or do you have a special message for them?
We are entering an AI inference era. AI will be everywhere, benefiting everyone — not just rich countries or major cities, but also rural areas around the world. To make this happen, we need global cooperation to build uniform standards.
These standards should not only cover technical aspects — like how we did with 4G or 5G telecommunications — but also ensure that AI is used responsibly for the common good. Regulation is important; we must make sure AI is safe, ethical, and beneficial for all.
This is a moment for international collaboration. Venues like the ITU’s AI for Good Summit are essential: they allow us to communicate, understand each other, and work together to build a positive AI era.
