On 14 October 2026, BrainChip Holdings Ltd. announced the availability of Sequora™ a processor intellectual-property (IP) core expressly built for edge devices that run on a battery. The offering expands the company’s established Akida™ platform, which already excels at ultra-low-power event-based inference, by adding the capability to host generative artificial intelligence workloads such as large language models (LLMs). The key promise is that a single chip can perform both conventional sensor classification and on-device language reasoning without streaming data to the cloud, preserving privacy and reducing latency.
Sequora targets a spectrum of markets that have long struggled with the memory-bandwidth and energy demands of generative AI. From earbuds that could converse with users to aerospace sensor nodes that not only detect radar signatures but also explain them, the core promises to keep the entire processing pipeline inside the device’s modest power envelope. By eliminating the need for external compute, manufacturers can design slimmer, lighter products while still delivering sophisticated interactive experiences.
How Sequora overcomes memory and power bottlenecks
The breakthrough stems from a state-space architecture that fixes memory usage regardless of how long a conversation or data stream persists. In practice, this means the core can scale from models with a few million parameters up to roughly three billion parameters, all while consuming the same amount of SRAM. Such predictability is critical for sectors like automotive or medical devices, where certification requires deterministic execution paths. Moreover, the design sidesteps the per-word cost explosion typical of traditional transformer models, because the internal cache does not grow with context length.
In addition to the architectural efficiency, Sequora integrates a native static-graph compiler and a lightweight C++ runtime library that together shave off unnecessary compute cycles. The SDK ships with a curated model zoo of pre-trained generative workloads, allowing developers to prototype voice-to-voice assistants, on-device translation, or AI agents that react to sensor inputs without ever leaving the chip. The result is a solution that meets the strict size-weight-power (SWaP) constraints demanded by wearables, drones, and defense-grade ASICs.
Licensing, ecosystem and market outlook
Sequora is delivered as a fully synthesizable RTL package, complete with verification IP, test-bench files, synthesis scripts and timing constraints. Customers can embed the core into their own system-on-chip (SoC) designs and fabricate it on any foundry and process node that suits their volume and cost targets. This process-portable approach accelerates time-to-market for chip makers who wish to differentiate their products with on-device generative AI while avoiding a costly design-in effort.
Early adopters already span a diverse set of verticals. ForwardEdge ASIC’s CEO highlighted the relevance for “mission-critical edge ASICs” in defense, while edgeAI Inc. noted that utility meters and mobility platforms will soon require both predictive inference and natural-language interaction. With the additional privacy benefit of keeping data on-device, regulators and end-users alike gain confidence in deployments ranging from medical wearables to industrial IoT gateways. As the AI race pushes more workloads to the edge, analysts expect the demand for low-power, versatile IP such as Sequora to grow sharply.
