EdgeCortix is a Japan-headquartered AI semiconductor company that develops energy-efficient hardware and software for running advanced AI models at the edge, particularly in physical-world environments such as robotics, industrial automation, aerospace, telecom and space.
What was the vision behind the company when it was founded in 2019?
Fundamentally, EdgeCortix is focused on bringing near-cloud-level or data-center-level AI performance into the real world, or what we now call the physical world. These are systems constrained by size, weight, power and, most importantly, the need for fast processing while remaining energy and cost-efficient.
Rather than focusing purely on data-center-grade AI, where those constraints may not be as real, we focus on the real world, which is extremely challenging. The question is whether we can combine software and hardware without sacrificing performance, so AI can be deployed across devices.
From the beginning, we saw that many companies were either focused on cloud-centric processing or purely on hardware. Very few were taking a software-centric approach to building AI focused hardware. That is the fundamental reason we started the company: to bring silicon to market that makes AI processing extremely efficient from both a power, compute, and cost perspective, with software as a core focus from day one.
Would it be fair to say that many AI deployments fail to become profitable because of reliance on data centers, which are energy-intensive and costly?
I do not think that would be an entirely fair statement. If you look at the 2019 to 2022 timeline, much of the AI world was dominated by training, because AI itself was still in an early phase. Training requires a lot of compute, because it involves feeding large amounts of data into models in order to build them. Inference, which is what we focus on, is much more solution-centric. If you are ready to do inference, you are ready to deploy a solution.
That shifts the conversation from building the model to deploying the model, and deployment means real-world constraints. Historically, the data center dominated the conversation because AI was focused on training. Today, as models have matured and become more efficient from an algorithmic standpoint, companies like EdgeCortix are building silicon that is also efficient. We can bring those pieces together and focus on inference and deployment, which are much more real-world ready. Inference now happens both in the data center and at the edge. The edge now becomes more prominent because, in the end, it is a physics problem: there will always be far more devices at the edge than data centers in the world.
What results can a company, drone or camera achieve using your chips compared with the status quo?
Historically, much of the hardware in the market was designed around traditional convolutional models, which worked well for vision processing. If the model remained a convolutional model, the hardware performed very well. But the moment the model changed from convolutional neural networks (CNNs) to generative AI or transformer-based models, which are now mainstream across large language models and vision-language models, hardware fine-tuned only for vision-centric processing no longer worked as well in a multimodal world.
That is the challenge we are solving. Rather than designing hardware fixed around one type of AI architecture, we build hardware with the understanding that the AI model landscape will continue to change. Our current product, SAKURA-II, enables customers across the edge AI landscape to run traditional convolutional models, such as object detection and scene understanding, as well as multimodal language models. A robotic system, for example, might run a model that understands images or video, another model like Whisper for speech-to-text and text-to-speech, and a RAG system that allows the robot to access a knowledge base. These scenarios require audio, video and text models to work together under limited power or battery constraints without sacrificing performance. That is what our hardware and software combination is solving.
What is the commercial significance of your partnership with NASA?
One of the things we believe is that if you design solutions for the hardest constraints, the easier constraints fall out naturally. There is no more difficult environment than space. It is the final frontier, or the ultimate edge.
Working with NASA, we discovered that our chip, designed as a commercial off-the-shelf product, could withstand 10 times higher radiation compared with similar commercially available products in its category.
That means our chip can operate in harsh environments without requiring special packaging, which can be extremely cost-prohibitive for space applications. The same solution and cost profile can support use cases from low Earth orbit and geosynchronous orbit to future lunar missions. This includes future landers that may require AI guidance during landing, or rovers operating on the lunar surface under extreme conditions. It has also opened opportunities in areas such as orbital data centers, which are still in their early stage, but require compute platforms that are highly energy-efficient. Traditional GPUs require a lot of power to deliver AI performance, and that is where our hardware and software combination can excel.
Beyond space, are you already in conversations with businesses, and where are you seeing the most interest by industry and geography?
Absolutely. Space is the final frontier for us, but our near-term opportunities are in environments where we are already generating the most revenue. Industrial and factory automation is a major area. That includes robotic systems or camera-based systems used for anomaly detection and predictive maintenance across manufacturing facilities and factory floors. Japan has historically been a leader in manufacturing, so we have seen strong resonance with multiple customers in that segment.
Robotics and physical AI are also becoming more prominent. While we are not yet seeing a lot of appetite for humanoid robotics, we have seen design wins in social robotics applications, where robots may have a wheeled structure and a human-like upper torso with cameras and sensors. These systems can be used in hospitals, factories or logistics facilities, and they need localized, efficient multimodal AI. Telecom infrastructure is another area of early growth, particularly at the intersection of AI, 5G and post-5G communications. We are also seeing activity in aerospace, defense and mission-critical environments, including work connected to the U.S. Defense Innovation Unit, Japan’s Ministry of Economy, Trade and Industry (METI), Japan’s New Energy and Industrial Technology Development Organization (NEDO) under METI, and the Japan Ministry of Defense.
How do you see EdgeCortix’s role in the market two or three years from now?
Today, the market tends to think about inference in a binary way. On one side, there is the data center, where the amount of energy and power required for AI processing is becoming enormous. On the other side is what people colloquially understand as the Internet of Things or the edge, meaning very low-power, sensor-grade devices that deliver lower performance. We believe the largest part of the market sits in between, in what we call the “thick edge.”
The thick edge requires hardware capable of multiple trillions of operations per second, allowing users to run 10, 20 or even up to 100 billion parameter AI models under the constraints of the physical world. These could be robotic systems, aerospace systems, satellites, factory environments, smart infrastructure, or servers and racks deployed outside traditional data centers. Our approach is not just about hardware, although the hardware must be extremely energy-efficient. It is really about software. This is a fragmented market with many different types of systems and applications. Customers do not want to replace everything they already have, so our SAKURA-II chip is designed as a co-processor that augments existing systems. That makes adoption easier, and it is where our software becomes fundamentally important.
Are there industries of specific interest for you to target?
Robotics is clearly one where we are very bullish. I have also worked in robotics myself, so I hope robotics finally catches up. As humans and robots working together become more mainstream, I believe our software and hardware combination will be a critical enabler.
Another area is the high-performance edge market, which we can address with our next-generation silicon. Micro data centers are one example. Not every data center needs to be a power-hungry system equivalent to a nuclear power plant. There is a space for micro data centers, whether for commercial applications or mission-critical applications, and we believe this is another area where we can disrupt the market with our technology.
