ASUS Unveils Hybrid AI Architecture Slashing Inference Costs by 70%

In a notable advancement for AI infrastructure, ASUS has unveiled a hybrid AI architecture that could slash inference costs by up to 70%. This approach incorporates Phison's aiDAPTIV technology, allowing businesses to deploy generative AI applications efficiently across various devices, including laptops, desktops, and mini PCs.

Addressing Rising Inference Costs

With the growing adoption of large language models (LLMs) and AI agent-based solutions, businesses face increasing financial pressures from inference-related token costs. Traditional cloud-based models can lead to unpredictable expenses and operational hurdles, hindering the wider deployment of AI technologies. ASUS's hybrid architecture aims to address these challenges by enabling some AI processing to occur locally on devices, while reserving cloud resources for more complex tasks. This dual strategy boosts both performance and cost-effectiveness, making it an attractive option for companies looking to expand their AI capabilities.

Bryan Chang, General Manager of Commercial PC BU at ASUS, highlighted the need to balance performance with cost. "Balancing performance and cost has become a critical challenge as enterprises scale their AI adoption," he said. The new architecture seeks to transfer more processing responsibilities to local devices, decreasing reliance on cloud resources and improving efficiency in AI applications.

Enhancing Local Processing Capabilities

Central to this hybrid model is Phison's aiDAPTIV memory extension technology, which empowers devices with limited resources to support mid- to large-scale language models locally. This innovation eliminates traditional memory constraints, allowing commercial PCs to handle AI workloads that previously required high-end infrastructure. The architecture also features a gateway-based routing mechanism that intelligently allocates tasks based on complexity, prioritizing local execution to further optimize efficiency.

Benchmark results from PinchBench highlight the effectiveness of this hybrid inference method, indicating it can reduce costs for mid- to large-scale models—such as those with 26 billion and 35 billion parameters—by up to 70% without compromising performance. K.S. Pua, CEO of Phison, remarked, "aiDAPTIV enables local execution of larger AI models by overcoming traditional memory limitations, while hybrid inference significantly reduces overall compute costs. Our collaboration with ASUS demonstrates how this technology can be effectively deployed across commercial platforms, delivering a scalable and cost-efficient solution for enterprise AI adoption."

Broadening Use Cases Across Enterprises

ASUS's hybrid architecture broadens its capabilities across a wide range of enterprise applications, including multilingual translation, business email drafting, meeting note summarization, and automated customer service. By managing these tasks locally, companies can reduce their dependence on cloud token consumption, enhancing data privacy and response efficiency. This comprehensive approach not only improves operational flexibility but also supports the transition from pilot projects to large-scale AI implementations.

Looking ahead, ASUS is dedicated to further enhancing its AI PC and commercial computing platforms. The company plans to integrate hardware, software, and ecosystem partnerships to provide AI infrastructure that balances performance, cost, and adaptability. As businesses continue their digital transformation, ASUS's hybrid AI architecture may play a crucial role in facilitating scalable AI adoption.