Krishna RangasayeeFounder & CEO

I have 30 plus years in semiconductors, software, and AI. I spent 18 years at Xilinx and learned the physical AI space. About six and a half years ago I started SiMa.ai and the journey has been fantastic.

I see 2015 to 2025 as cloud and consumer AI. I see 2025 to 2035 as the decade of physical AI. AI will show up in robotics, automotive, medical, aerospace and defense, and smart vision. We built a purpose-built company to scale AI in the physical world.

We are full stack. We make our own chips and the AI software on top. We support multimodal input and output across audio, video, and text. We run CNNs, transformers, and LLMs on one chip.

Two things. We run a wide range of models on a single chip under 10 watts, which is critical at the edge. And our software is easy to use. From a Hugging Face repo you can pick a supported model and run it with a button push.

Security and privacy, latency, and total cost of ownership. Some markets cannot tolerate cloud privacy risk. Many applications cannot accept round-trip latency. Cloud can be expensive. Processing near the data helps you manage and monetize it.

The same ingredients matter everywhere, but in aerospace and defense they are critical. You need ruggedization, safety, privacy, security, and accuracy. The cost of mistakes is high.

We moved from rule-based compute to statistical ML. Now we add reasoning like humans. You can converse in audio and video. Context and memory improve accuracy and personalization. The human-machine interface changes.

We grew up with SWaP-C. Our strategy is to upgrade existing equipment. We stay inside tight power and thermal envelopes so you do not need new cooling or safety changes.

We are AI-focused compute, not a general CPU or GPU. We do not run every workload. For pre-processing, post-processing, and AI compute we aim to be world class. We chose the 3 to 10 watt segment to be best in class there.

We are engaged in drones, robotics, mobility on land and sea, and satellites. We bring multimodal sensing, redundancy, and accuracy. Our power efficiency lets customers pack more intelligence without breaking thermal limits.

We detect the delta on the satellite and transmit only the difference. That reduces battery use, compute, and latency. It shifts from data movement to information management.

Build power-efficient silicon that is upgradeable. The same chip runs CNNs, transformers, and LLMs. Today we run an 8B Llama model around seven watts. With RAG you can talk to the machine and get steps from embedded docs.

Real machines need audio, video, and text together. Think predictive maintenance by sound, computer vision for assistance, and text guidance. Extend that to tanks, submarines, aircraft, and more.

Pick a lane and stay with it. We focus on edge AI at 3 to 10 watts, easy software, and multimodality. We avoid distractions that dilute the vision.

AI will live in everything. Soldiers will have personalized compute in what they wear and see. Machine-to-machine information sharing will improve. Cycle times will shrink.

You need product longevity of 12 to 16 years, ruggedization, and qualification. You must bridge fast AI changes with slower platform cycles. You must treat privacy, security, and cybersecurity as first-order problems.

Accept trade-offs. Stay purpose-built. Put ease of use first. Respect power and thermals. Help customers upgrade what they have. Assume change is constant and design for it.