At a glance
| Booth | T9404 |
| Country | US |
| Website | www.vellex.ai/about-us |
Company profile
Vellex Computing is a Stanford spinout and an analog semiconductor company backed by the U.S. National Science Foundation and Department of Energy. The company is developing AI-training technology around analog circuits and an analog compute architecture, addressing an optimization problem that conventional digital computation approaches through repeated brute-force calculation. Vellex Computing states that its architecture has demonstrated a 17,000x speedup over conventional techniques while operating at milliwatt power levels. Its initial commercial direction is edge AI in energy, industrial IoT, robotics and remote sensing, where it aims to let deployed devices learn continuously, securely and in real time without relying on the cloud. The company reports active commercial pilots and a growing patent portfolio as it develops adaptive, cloud-free industrial intelligence. Vellex Computing also describes its supporters as institutions in science, computing and deep tech. Its stated technology target is practical on-device AI training at the edge, with continuous learning at milliwatt power. The technology targets robotics, satellite systems and industrial IoT among its application areas. In 2026, Vellex Computing reported winning two awards at InnoVEX, including the Plug and Play Taiwan Award at the InnoVEX Pitch Contest, where it was the sole recipient among 15 finalists from 8 countries. The company also reported being one of 11 winners selected from 590 proposals across 56 countries in the 2026 IC Taiwan Grand Challenge, with a $30,000 prize and access to Taiwan's semiconductor ecosystem. At the 2026 Industry Growth Forum, Vellex Computing was named an AI track finalist selected from 273 applicants to pitch before nearly 200 investors with $1.6 billion of capital to deploy.
Exhibits
Vellex Computing presents an on-device AI-training platform designed to train AI at less than 10 mW, which it compares with power below that of a Bluetooth radio, and it also lists 17,000x faster training benchmarked at Stanford and Berkeley Lab. The platform is intended for actual model training on edge hardware rather than only inference or execution of pre-trained models, with the hardware operating at milliwatt power. Vellex Computing says edge devices can use the platform to learn continuously from their own data, adapt to changing conditions in real time, and improve during deployment without cloud retraining, firmware update cycles, or data leaving the device. Its technology description contrasts today's optimizers such as Adam and SGD, which use iterative gradient steps, with an approach that solves the energy cost of the optimization mathematics rather than pursuing higher processor speed. In the company's physics-based approach, analog physics settles toward the solution naturally instead of iterating through millions of calculate-check-repeat steps, reducing the process from watts of power and many steps to a one-step milliwatt regime. Vellex Computing describes traditional AI training as a brute-force optimization loop and maps that optimization onto a physical system that resolves toward its minimum energy state, which corresponds to the optimal solution. The technology page states that model training can converge in microseconds at milliwatt power rather than taking hours at hundreds of watts. The resulting programmable analog IP block is described as delivering optimal model weights at milliwatt power and nanosecond-scale speed while operating alongside standard ARM or RISC-V cores. For integration, Vellex Computing positions its IP as a complement to existing AI development stacks: ARM or RISC-V cores continue to handle functions such as inference, control logic and matrix multiplication, while Vellex handles the iterative optimization at the center of training. Engineers can work in standard machine-learning frameworks including PyTorch, JAX and TensorFlow, while the Code-to-Circuit compiler maps models automatically to Vellex Core, so analog design expertise is not required. The IP Core is delivered as a licensable GDSII hard macro compatible with standard semiconductor manufacturing processes and designed to sit beside ARM and RISC-V cores. For satellite and remote-sensing applications, Vellex Computing describes on-orbit training that separates valuable scenes from noise with higher precision so only relevant information needs to be downlinked. For drones, the company describes mission-by-mission model improvement supporting real-time obstacle avoidance and adaptive search-and-rescue in changing terrain. For industrial robotics, the platform is intended to let robots improve during operation, reducing downtime and mistakes while addressing defects, false alarms and quality issues on factory floors. For wearables and hearables, Vellex Computing describes personalized voice and gesture models that train locally without data leaving the device, while entertainment and gaming examples include private mood-based playlists and game enemies that adapt to the player. The company identifies the existing edge-AI problem as static models that cannot adapt as real-world conditions change, often forcing a loop of collecting data, sending it to the cloud, retraining and pushing an update that can take hours to days, cost thousands per cycle, and fail where bandwidth is limited or data cannot leave the device. Vellex Computing cites a projected $59 billion global edge-AI hardware market by 2030 and positions continuous on-device learning for satellite and remote sensing, industrial robotics, drones and wearable technology where cloud retraining may be too slow, expensive or unavailable.
Capabilities and products
- milliwatt-class on-device AI training
- on-device learning without cloud retraining
- analog-circuit AI training optimization
- Code-to-Circuit model-to-circuit compilation
- licensable GDSII hard macro IP
- PyTorch/JAX/TensorFlow model-development support