TAIPEI (Taiwan News) — Skymizer on Friday introduced the HTX301, an artificial intelligence accelerator card designed to move large-model computing from cloud data centers to on-premise environments.
The company said the system reflects a broader industry move toward running AI models within private infrastructure rather than relying on external cloud providers such as OpenAI. In an on-premise setup, organizations operate and manage AI systems internally, giving them greater control over data and security, according to Iguazio.
The push toward local deployment has been reinforced by growing concerns over data security and governance, as employees increasingly use public AI tools such as ChatGPT to process sensitive materials. Using these tools for strategy documents, technical reports, or internal incident summaries may expose confidential information outside company-controlled systems, according to Everpure.
Skymizer said regulated and latency-sensitive sectors, including healthcare, banking, government, and manufacturing, are accelerating efforts to run large AI models locally. However, adoption remains constrained by cost, power consumption, and system complexity.
The HTX301 is designed to address these constraints. It focuses on memory bandwidth and capacity during the inference phase of large language models, rather than repurposing training-oriented silicon.
The inference phase is the stage where a trained AI model is used. After training on large datasets, inference is when the model applies what it has learned to generate outputs or decisions, according to Google Cloud.
The HTX301 is a single-card architecture that scales from one to eight inference chips per card. Each card also carries up to 512 GB of DRAM memory and is designed to run trillion-parameter-scale AI models without requiring GPU rack-level infrastructure.
Skymizer said the architecture is intended for flexible deployment. A single card can power edge AI nodes or compact systems, while multiple cards can be combined into on-premise racks or small data center–scale clusters, allowing enterprises to scale incrementally without redesigning infrastructure.
Skymizer Chief Marketing Officer William Wei (魏國章) said the HTX301 represents Taiwan’s pathway “from silicon to system scale” in the global enterprise AI market. The company is also working with the government-backed Institute for Information Industry to promote adoption of on-premise AI systems.
The HTX301 is being demonstrated alongside agentic AI, smart home, and smart vehicle applications at Skymizer’s booth at Computex Taipei, which runs Tuesday through Friday.
