Hepzibah: Delivering Energy-Efficient, Memory-Centric AI for the Edge and Beyond
At COMPUTEX 2026, the Canadian startup challenges the capital-intensive "silicon-first" status quo with an energy-efficient, IP-licensing model built for ubiquitous AI.
As the global technology industry converges in Taipei for COMPUTEX 2026, Hepzibah, a Canadian IC design startup, is planning to showcase a paradigm shift in AI hardware architecture. Under the strategic leadership of co-founders Raymond Chik and Martin Snelgrove, the company is pivoting away from the traditional, capital-intensive “silicon-first” model that often plagues hardware startups.
This shift is backed by a proven silicon lineage, as the leadership team brings deep expertise from their prior foundational work at Untether AI to this new venture. Instead of chasing a standard test chip, Hepzibah is championing a flexible, application-specific IP licensing and “white glove” service strategy designed to meet the explosive demand for localized AI intelligence. Following its successful debut at SEMICON 2025, Hepzibah arrives at COMPUTEX 2026 with a clear mission: providing the most energy-efficient underlying engines for a future where AI compute is ubiquitous, appearing in everything from drones and phones to massive industrial sensors.
A Memory-Centric Architecture for a Ubiquitous AI Future
Hepzibah’s core technology is built upon a highly modular, memory-centric tile-based architecture specifically engineered to scale across a vast spectrum of performance requirements. This design allows for a seamless transition from compact, power-constrained edge devices to expansive, high-throughput data center racks.
The brilliance of this modularity lies in its granular customization; customers can choose to integrate a single tile for simple edge tasks or scale up to thousands of tiles on a single high-performance chip. Beyond single-chip scaling, the architecture supports further expansion through an array of chiplets tied together by a custom Network-on-Chip (NoC) architecture.
This hardware efficiency is driven by a series of internal innovations, including custom low-power SRAM blocks and a specialized transpose unit. This unit is particularly critical as it allows the hardware to perform the mathematical transpositions required for AI training and inference with extreme efficiency, effectively eliminating the energy-heavy process of moving data back and forth across the chip. Furthermore, Hepzibah is emphasizing a path of technological independence, as the architecture can be implemented across a range of process technologies, including mature nodes that offer an attractive balance of performance and cost.
Target Markets: Vision, Radar, and Edge AI
Rather than chasing every segment of the market simultaneously, Hepzibah is focusing on workload-driven validation to ensure its architecture perfectly maps to high-demand, high-impact applications in vision, sensing, and radar.
The company views AI compute as something that will eventually be in every chip, from cars and drones to phones and laptops. Their focus is on providing the best energy-efficient platform for running any computational or data-intensive workload.
The company has already gained significant early traction in radar signal processing and detection systems, a field where precision is paramount for the detection of drones and other incoming objects. Hepzibah’s architecture is uniquely suited for this because radar processing is mathematically akin to high-speed image recognition.
Internal modeling indicates that for radar and other data-intensive signal processing workloads, Hepzibah’s architecture has the potential to significantly outperform conventional GPU-based approaches while retaining greater programmability than fixed-function ASIC solutions.
Beyond defense and sensing, Hepzibah is targeting the burgeoning field of on-prem security and anti-hallucination systems, processing massive volumes of local data like legal documents with high throughput and low latency.
The Partnership-Driven Path to Silicon
One of the most ambitious prongs of Hepzibah’s strategy at COMPUTEX involves the consumer electronics and PC ecosystem. The company is actively seeking dialogues with brand-name PC and laptop makers who may be interested in following the trend of industry giants like Apple and Amazon by developing their own proprietary silicon.
Hepzibah offers a unique partnership-driven path to silicon through a “white glove” service, acting as a general contractor that provides the front-end design and IP licensing while managing partnerships with backend specialists like Alchip to handle the physical manufacturing flow. This model lowers the barrier to entry for “box makers” to become chip designers, allowing them to differentiate their products with custom AI features tailored to their specific hardware brands.
Hepzibah is currently operating in a pre-silicon phase, a deliberate choice that prioritizes capital efficiency and customer traction over premature tape-outs. The company is moving rapidly from architectural design to full workload validation on a hardware emulation platform that is expected to be available for customer demonstrations and workload validation by Q4 2026.
This allows potential customers to execute parts of their actual application code on Hepzibah’s compute cores to verify performance before committing to silicon production.
Book a meeting with Hepzibah at Computex or online before the exhibition!