Field report E-19 / J8 · sheet /strategy

Corporate StrategyFiled from the territory

Ceva and Mobilint put 10-TOPS edge AI in shipping hardware

Mobilint’s shipping MLD-R1 accelerator uses a 10-TOPS REGULUS processor, giving Ceva’s NeuPro family its first production deployment; volume, pricing and benchmarks remain undisclosed.

Grid
E-19 / J8
Surveyed
November 20, 2026
Reading
4 min read
Correspondent
Sophie Lindqvist
Ceva and Mobilint Bring Physical AI Closer to Everyday Devices
Fig. E-19 / J8Ceva and Mobilint Bring Physical AI Closer to Everyday Devices — blickpixel / pixabay

Field notes

  1. REGULUS delivers 10 TOPS of AI performance below 3 watts.
  2. REGULUS supports up to 8 gigabytes of memory.
  3. REGULUS is the first production deployment of Ceva’s NeuPro NPU family.
  4. The MLD-R1 AI USB accelerator was previewed at COMPUTEX 2026 and is shipping.
  5. REGULUS supports PyTorch, ONNX, Hugging Face, TensorFlow, TFLite and Keras.

By Sophie Lindqvist

Mobilint is shipping the MLD-R1 AI USB accelerator, built around a REGULUS processor that delivers 10 TOPS (trillion operations per second) of AI performance below three watts.

The chip marks the first production deployment of Ceva’s NeuPro NPU family and puts Ceva’s intellectual property into production silicon. Mobilint has disclosed no shipment volumes, customer names, pricing or comparative benchmarks. REGULUS therefore demonstrates feasibility more clearly than market scale.

What does this change for device makers?

It gives manufacturers a route to add local AI (inference performed on the device) to existing systems without designing every processing block from scratch. Mobilint’s MLD-R1 AI USB accelerator is one route; it was first previewed at COMPUTEX 2026 and is now shipping.

REGULUS serves as a standalone edge AI system-on-chip (SoC), meaning a processor designed for workloads near a device’s sensors rather than in a remote data center. It supports up to eight gigabytes of memory and pairs that capacity with sub-three-watt operation, specifications aimed at products where energy use, space and immediate response constrain design choices.

Cloud computing remains valuable for demanding workloads, the announcement says. Yet robots, drones, cameras and industrial equipment often must interpret changing conditions close to their sensors. Local processing can improve responsiveness while keeping computation within practical power and space limits.

Why the architecture matters

Physical AI—software that senses and acts in the physical world—rarely depends on one calculation. A machine may need to:

  • prepare incoming sensor data;
  • recognize an object;
  • interpret its surroundings; and
  • trigger an appropriate response.

Mobilint assigns those stages to different processing engines. SensPro, an AI digital signal processor (DSP), provides programmable sensing and signal processing, while NeuPro-Nano, a neural processing unit (NPU), handles neural processing for always-on inference.

Mobilint’s proprietary AI processing architecture adds a third layer. The combination gives developers a flexible platform for workloads that involve understanding sensor inputs and making decisions. It also lets product designers balance performance and efficiency as application requirements change.

Mobilint CEO Dongjoo Shin described REGULUS as combining “specialized capabilities for sophisticated AI under tight power constraints.” Ceva’s contribution adds sensing, signal processing and inference resources that can serve a broad range of customer needs across device categories.

Yaron Galitzky, Ceva’s Executive Vice President and General Manager of its AI Division, said efficient inference and flexible sensing provide “an important enabler for devices that must operate locally and respond in real time.” The rationale is straightforward: local capability must fit a product’s power and space limits, not merely demonstrate laboratory performance.

What the platform can reach

REGULUS supports six named AI development frameworks: PyTorch, ONNX, Hugging Face, TensorFlow, TFLite and Keras. The stated compatibility gives developers familiar starting points for connecting existing AI work with edge applications. The announcement does not specify supported versions, model sizes or framework-specific performance.

The accelerator provides a route for businesses exploring computer vision, video analytics and generative AI (models that create content). Potential end uses span smart factories, robotics, machine vision, security, medical devices and digital signage. These are target sectors, not disclosed deployments. The announcement names no customer operating those systems in production and provides no cost, accuracy or latency comparison with competing hardware.

For Mobilint, the integration strengthens its ability to serve customers developing next-generation intelligent edge products, meaning devices that process information locally. Mobilint contributes expertise in energy-efficient AI semiconductors and integrated SoC design. Ceva supplies established silicon and software intellectual property for intelligent-edge applications.

What remains unproven

The production milestone gives Ceva a concrete demonstration that its AI technology has reached commercial silicon. It also gives Mobilint a shipping product rather than only an architectural proposition. That is the firmest part of the announcement.

The broader promise remains conditional. The announcement projects more capable devices, more productive workplaces, more responsive systems and wider access to intelligent technology across applications and environments. It offers no evidence yet that REGULUS-based hardware will deliver those outcomes at scale.

The 10-TOPS figure measures compute capacity, not application performance. Without named workloads, latency measurements, accuracy thresholds, sustained-power tests or independent reviews, buyers cannot judge whether REGULUS will outperform alternatives or meet a particular product specification. The announcement also gives no distribution terms or unit volumes.

The practical conclusion is narrower than the promotional framing: Ceva and Mobilint have combined three processing approaches in a production processor and a shipping accelerator. Whether that foundation produces commercially meaningful volume remains open.

via semiwiki.com (Source)

Share this article:

More from Sophie Lindqvist

Sophie Lindqvist

Show full bio

Correspondent covering money, funding and market shifts around corporate strategy and chief-executive leadership at CEO Frontier.

515 articles

Filed under

Adjacent sheets

« Previous articleNext article »

End of sheetE-19 / J8 · November 20, 2026