NTT has announced the world's first low-power AI inference chip that can process 4K video in real time for self-driving cars and surveillance drones.
Surveillance systems and self-driving cars are literally figuratively hitting the wall. This is especially true when it comes to real-time decision-making in power and bandwidth constrained environments. Today's autonomous AI solutions rely heavily on cloud connections with energy-intensive GPUs, and often struggle to process ultra-high resolution videos at the edge. Japan's telecommunications giant at the Upgrade 2025 conference
The chip allows real-time 4K video inference at just under 20 watts to provide effective performance without the energy burden typical of GPU-based solutions.
“The new AI inference chip allows autonomous systems to be seen further, react faster and operate independently. This is a critical step towards the more critical deployment of safer, sustainable AI in the real world.” “We are currently testing drone chips for surveillance applications, but in the future, we can also integrate them into autonomous vehicles where real-time inference with low power consumption is important.
By running advanced object detection AI models like Yolov3 directly on-device, the chip eliminates the need for video compression or cloud-based processing. This boosts both efficiency and operational independence. For example, drones could allow people and infrastructure to be identified from the 150-meter Japanese legal flight cap. This is five times the current solution allows.
“High-level reasoning at the edge is a disruptive force and a true game-changer for a potentially diverse industrial and commercial use case,” Carlos Moreno Alvarez, senior cloud architect for the cloud infrastructure platform ESCALA24X7, told me. “The impact is on enabling data that is traditionally processed centrally at the edge, relying on near-real-time analytics and generating more actionable insights for sectors that manage vast amounts of data.”
A smarter, more sustainable AI competition
The meaning of AI chips goes far beyond technology demonstrations, and could change the way AI is deployed for on-field use cases, from sensible surveillance to real-time inspection of remote bridges and pipelines. In disaster zones where there is a shortage of seconds and power supply, NTT chip-equipped drones can help identify survivors, monitor environmental threats, assess structural damage, and assess structural damage without eliminating high-bandwidth connectivity or valuable battery life.
“Advances in edge technology, especially in model distillation, allow small models to run comparable models with larger models, and even consumer-grade devices, almost anywhere.” “The ability of AI models to operate offline makes them particularly valuable in scenarios where rapid local decisions are important. Edge AI models are still not reliable enough to fully automate sophisticated tasks, but they can help you respond quickly to new situations.”
NTT's chips leverage in-house AI inference engines to dramatically improve computational efficiency. Its main differentiators are proprietary machine learning technologies such as interframe correlation and dynamic pre-bit control, enhancing computer vision and image processing while maintaining high levels of accuracy.
Garbage further emphasizes, “AI infrastructure is growing so large and energy-intensively, making it unsustainable. With the new AI inference chip, NTT research will help fill that gap.
However, NTT is not alone in the race to deploy AI chips for surveillance and autonomous vehicle applications. Xpeng, a leading Chinese EV manufacturer, has developed an autonomous driving chip.
The future of AI, photonics and networking
NTT plans to commercialize chips within fiscal 2025 through its subsidiary NTT Innovative Devices Corporation. But this chip is just one of the broader visions. It is also being investigated as part of NTT's innovative Optical and Wireless Networking (IOWN) initiative. It is a next-generation data infrastructure that uses photonics to increase speed and energy efficiency.
“Traditional GPUs are often idle at around 50% of the time, and processors and memory are bottlenecked by transmission delays,” Gomi said. “Photonics can bridge that inefficiency. With analog-style optical circuits, you can ultimately replace GPUs in some scenarios and relieve pressure on the data center.”
He also emphasized that the rise of powerful edge AI does not eliminate the need for human surveillance. “Before AI or AGI replaces skilled human-level decision-making, it must be trained and tested with extraordinary care,” he said. “Technology can only go so far, and human judgment is still essential, especially when safety is on the line.”
As autonomous systems evolve, we believe that conversations must expand beyond their capabilities to include responsibility. Who is accountable when AI makes a mistake? How do you build trust in a system that operates beyond immediate control? NTT's approach to these challenges includes interdisciplinary collaboration, from encryption and secure data sharing to secure data sharing with NTT data, to exploring photonics as a long-term solution to today's data center power and latency bottlenecks.
With the AI chip, NTT not only allows smarter surveillance and safer driving, but the company shows that intelligence at the edge is both strong and principled.
