Taiwanese startup Skymizer has introduced the HTX301, a PCIe accelerator card that supports AI models with up to 700 billion parameters using older 28nm semiconductor technology. The company claims this design significantly reduces power consumption compared to hyperscale GPU clusters typically used for large language models.
The HTX301 features 384 GB of memory on a single card, enabling it to run massive models like GPT-3 (175 billion parameters) or larger. Skymizer's approach leverages mature 28nm process nodes, which are cheaper and more energy-efficient than cutting-edge 3nm or 5nm chips, though performance trade-offs may exist.
According to Skymizer, the accelerator is designed for inference tasks, not training, and targets edge deployment or smaller data centers. The company has not disclosed pricing or availability dates as of May 2026. Industry analysts note that while 28nm chips are less powerful, their low cost and power efficiency could democratize AI access for smaller enterprises.
Skymizer's announcement comes amid global chip shortages and rising energy costs, positioning the HTX301 as a potential alternative for organizations seeking to run AI models without massive infrastructure investments. However, independent benchmarks are not yet available to verify performance claims.
