Artificial intelligence is rapidly moving beyond screens and cloud services into the physical world. The next wave of AI devices will not simply capture images or process commands—they will continuously interpret their surroundings and respond in real time.
Mosaic SoC believes this shift requires a fundamentally different computing architecture. Rather than treating perception as one application among many, the company is building a dedicated platform optimized for always-on spatial intelligence, combining custom silicon, software toolchains, and AI deployment frameworks.
In this interview, Mosaic SoC Co-founder Alfio Di Mauro discusses why perception has become the critical missing layer in wearable AI, how companion AI glasses could emerge as the first mass-market spatial computing device, and where Korea's semiconductor and hardware ecosystem fits into the future of Physical AI.
The following is a Q&A with Alfio Di Mauro, Co-founder of Mosaic SoC.
You’ve said the next generation of devices will “understand” the world rather than just capture it. What does that shift look like in real-world user experience?
What changes is that a wearable device stops being a passive “camera" and becomes an active interpreter of the user physical space. Instead of opening separate apps to translate a sign in a foreign language, check a train departing time, navigate a station, or search for context, the device can understand the scene in real time and return the one useful action: take the south exit, buy your train ticket at this machine, your train leaves in three minutes. The experience becomes contextual, heads-up, and far less menu-driven. That is what I mean by devices that understand the world instead of just recording it.
Most of today’s systems rely on GPUs or application processors. Why is a dedicated perception chip the right architectural answer now?
Because perception has become an always-on workload, and always-on workloads punish any architecture that wastes power moving data around. GPUs and application processors are excellent when you have a larger thermal and battery budget, but glasses and wearables do not. A dedicated perception chip lets us sit close to the sensor, turn raw images into semantically rich information locally, and do it in a sub-100 mW class power envelope and very small footprint.
Your multi-core architecture emphasizes performance per watt. What specific limitations of existing ARM-based designs were you trying to overcome?
The limitation we were attacking was not just raw compute; it was system efficiency. In this class of workloads, the energy penalty often comes from memory traffic, bus bottlenecks, and the separation between compute and usable low-latency memory. Our unique multi-core architecture with a unified memory fabric, tightly coupled memory, and configurable NPU was designed to remove that data-movement tax.
Mosaic SoC positions itself as reducing complexity for ODMs. In practical terms, what disappears from the development stack when your chip is adopted?
In practical terms, a lot of engineering cost disappears. The ODM no longer has to stitch together an image processing path, map it to a vision processor, and implement this on top of fragmented AI deployment flow just to ship one perception feature. We integrate the compute, on-chip memory, runtime, compiler flow, and accelerated computer-vision and AI libraries, so teams can focus on the product experience instead of rebuilding the stack underneath it. Just as important, ODMs can deploy proprietary code without disclosing it to us, because our stack will be open.
Among AR glasses, smartphones, and wearables, where do you see the first meaningful adoption at scale — and why?
If the question is where we expect spatial intelligence to first become a real product category rather than just a feature, I would say companion AI glasses that look like regular passive glasses. Smartphones definitely have volume, but glasses have the sharpest and most tangible new value proposition: hands-free navigation, translation, indoor space mapping, and always-on contextual assistance. The market is already moving in that direction: EssilorLuxottica said AI glasses sold more than 7 million units in 2025, and Samsung and Google have announced intelligent eyewear for fall 2026; by contrast, more immersive display glasses are still heavier, more expensive, and battery-constrained. Snap’s newly launched Specs are priced at $2k and rated for about four hours. So I see meaningful scale starting with lightweight companion glasses, then broadening into other wearables.
You are building not just chips, but also compilers and deployment toolchains. How critical is the software layer in determining whether Mosaic SoC can become a standard platform?
It is decisive. Great silicon becomes a standard because developers can take real models and applications to deployment quickly, predictably, and repeatedly. That is why we are building the compiler flow, runtime, and APIs alongside the chip, using open pieces like Open-Source AI runtimes and compiler infrastructures where it makes sense and adding Mosaic’s own runtime and libraries where differentiation matters. If that layer works, we become a platform. If it does not, we remain a custom component.
Large incumbents are also moving into edge AI. Where do you believe their approaches fundamentally fall short?
I would frame it less as them being wrong and more as them optimizing for a different center of gravity. Most incumbents come from smartphone, GPU, or general XR architectures, so they tend to treat perception as one workload among many. We start from the opposite premise: to Enable Physical AI, continuous perception is the pre-requisite, so the memory hierarchy, power budget, and software stack have to be designed around always-on processing first. Even the newest smart-glasses platforms from incumbents emphasize smaller size, lower power, and more on-device AI, which validates the direction; where we think the gap remains is in building the full perception platform end to end, not just a component.
Optional angle:
Is spatial intelligence ultimately a hardware problem, a software problem, or a system integration problem? Has the industry been solving the wrong layer?
I think spatial intelligence is fundamentally a system-integration problem, but the integration only works if hardware and software are co-designed around perception. The industry has spent a lot of time solving display, graphics, and cloud-assistant layers. Our view is that the missing layer is the local perception substrate: continuous sensing, local scene understanding, and low-latency action within a wearable power envelope. Once that layer exists, much better applications and much better form factors become possible.
As always-on perception becomes viable, what new constraints emerge — power, privacy, or something else?
Power is still the first constraint, because power turns directly into heat, battery size, and wearability. But privacy and trust move from secondary concerns to product requirements very quickly. The acceptable version of always-on perception is one that stays on-device, with low latency, no transfer of raw data to the cloud for processing, and very clear user control. That is why the shift toward on-glass and on-device AI is the only acceptable option.
If potential collaboration with Korean companies were explored, how would you assess their core strengths — and what structural challenges might they need to address to compete effectively in this space?
Korean companies would bring real advantages: deep strength in hardware engineering, consumer-device industrialization, optics and camera modules, and semiconductor scale. You can see that in Samsung’s Android XR eyewear work, Samsung Electro-Mechanics’ camera-module business, and LG Innotek’s breadth across camera modules, 3D sensing, and LiDAR. The structural challenge is that this category also rewards fast fabless iteration, tight hardware-software co-design, and early bets on system-semiconductor platforms, which is precisely why Korea has been expanding targeted support for fabless startups and the broader semiconductor ecosystem. So I would say Korea has many of the right ingredients; the open question is how quickly those capabilities can be aligned around end-to-end platforms rather than individual components.
