Martin Snelgrove

Untether AI designs AI inference accelerator chips using a compute-near-memory architecture that places processing elements directly adjacent to memory storage rather than in a separate compute chip, dramatically reducing the data movement bottleneck that limits inference throughput and energy efficiency in conventional von Neumann architectures. The Toronto company develops both the chip architecture and the supporting software compiler toolchain required to deploy AI models on its hardware.\n\nThe company raised approximately $160 million in venture funding from investors including Intel Capital, Export Development Canada, and BDC Capital. Untether AI targets inference-intensive AI applications including edge AI deployment, data centre inference servers, and cloud inference acceleration where power efficiency and throughput per dollar are the primary evaluation metrics. The compute-near-memory approach is architecturally distinct from both GPU-based inference and other AI accelerator designs, with published benchmarks showing significant performance-per-watt advantages for specific inference workload types.\n\nUntether AI competes in the AI inference chip market against NVIDIA TensorRT, Groq, Cerebras, and other AI accelerator companies pursuing different architectural approaches to the inference efficiency problem. Canada position as a global AI research hub, anchored by the Vector Institute in Toronto and Mila in Montreal, provides a strong talent pipeline and research collaboration environment for deep chip design and AI systems work. The company architecture-first approach requires a longer path to production deployment than companies adapting existing GPU infrastructure, but targets a structurally different point on the performance-efficiency curve.