By Adam King, VP & General Manager of Personal Devices Business Unit at MediaTek. Follow Adam's newsletter to learn more about the latest Innovation Talk.
The AI Inflection Point
2022 was the year that generative AI went mainstream, with the introduction of ChatGPT. This year is even more dramatic, as agentic AI has quickly become part of your everyday workflow, driving orders-of-magnitude increases in token demand. The entire industry is running at the redline to add more data center capacity and everything in the supply chain is constrained, especially memory for consumer products, and for CPUs, which are enjoying a renaissance as agentic AI requires them for orchestration.
The Questions Shaping the Next Wave of AI
The most urgent question for my business is how supply chain costs and shortages are affecting it. Beyond that, the most important questions are: How will agentic AI change the consumer device landscape? How will it affect existing devices? What new kinds of devices will arise? How will workloads be balanced across the cloud and the edge? And what technologies need to be developed or evolved?
Will the Smartphone Remain the Center of AI?
The starting point is the most successful consumer device ever: the smartphone. Will phones evolve to become agentic, and remain your primary, can't-live-without device? Or will they ultimately be disrupted by wearable products that are built closer to your senses, such as smart glasses? Or will they coexist, with glasses serving as an accessory or extension of your phone?
Actually, smart glasses and augmented reality glasses were conceived before the new AI era. Is AI a key enabler for turning them into daily-use, must-have products? I am sure the answer is yes, but the how and when is what is interesting.
When Will Agentic AI Move to the Edge?
Pre-deep learning AI has existed in consumer products for years, supporting applications such as sensor fusion, speech recognition, image enhancement, power management, and connectivity optimization. However, with agentic AI driving a 100-fold increase in token usage, alongside continued advances in smaller models and more capable hardware, there is both the motivation and the opportunity to bring LLM-based AI to the edge. Everyone I speak with agrees it is no longer a question of if, but when and how.
Why Physical AI Is Different
Physical AI presents a clearer path. Autonomous vehicles and robots process vast amounts of real-world sensor data, particularly visual information, and require millisecond-level latency that makes local processing essential. They also have the size, power, and cost budgets to support significantly more computing capability than smartphones or AI glasses.
Building the Next Generation of AI Hardware
What new technologies will be required? Will consumer devices need higher-performance CPUs? How will the balance between CPUs, GPUs, NPUs, and memory evolve? Will existing system architectures be sufficient, or will they need to be fundamentally redesigned? These are the questions being actively debated within our business and across the wider industry as AI continues to reshape computing.
