How ASUS is Rethinking AI Infrastructure Efficiency for the Next Generation of AI?

As AI evolves beyond simply responding to human prompts and begins to autonomously plan, reason, and execute tasks, can today’s AI infrastructure keep pace with the rapidly growing demands of next-generation workloads?

The evolution from prompt-driven AI to agentic AI, with its ability to make autonomous decisions and execute multi-step tasks, is fundamentally changing the role of artificial intelligence — from responding to requests to proactively completing tasks. As AI capabilities continue to advance, AI is expanding into robotics, automotive systems, manufacturing lines, and industrial environments. Emerging workloads such as digital twins, physical AI, and industrial autonomy are accelerating this transformation, driving AI systems toward continuous perception, reasoning, action, learning, and optimization.

For ASUS, this transformation is changing more than just model scale. It is changing the fundamental nature of AI workloads.

As autonomous AI systems require longer periods of continuous computation, higher-speed data exchange, and increasingly dense compute resources, power delivery, thermal management, connectivity, and large-scale deployment capabilities are all facing unprecedented pressure.

This leads to a fundamental principle: as AI evolves at unprecedented speed, infrastructure must evolve ahead of it.

The next generation of AI breakthroughs will not depend solely on more powerful compute silicon. They will increasingly depend on whether the underlying infrastructure can overcome existing physical and engineering constraints — and transform rapidly increasing compute demand into stable, scalable, and usable AI output.

The ASUS Vision: Token Factory Economics

If infrastructure must evolve first, should the way we measure AI data center efficiency evolve as well?

ASUS believes the answer is yes.

Historically, the industry has measured AI infrastructure capability through metrics such as GPU count, peak compute performance, or overall system scale. But as agentic AI and autonomous AI factories continue to expand, the real value of an AI infrastructure investment will depend on more than how much compute capacity is deployed.

The critical question is:How much usable, value-generating token output can be produced from the power, capital, and time invested?

Based on this principle, ASUS introduces the concept of Token Factory Economics (TFE) — a new way to evaluate the overall economics of AI infrastructure by focusing on the conversion of infrastructure resources into productive AI output.

This approach addresses one of the most practical challenges in high-density AI computing: hardware-level performance does not necessarily translate into sustained, usable compute capacity.

An enterprise may invest 100% of its capital in high-performance computing infrastructure. But if thermal, power delivery, or system-level constraints limit the infrastructure to sustaining only 70% of its potential performance, the impact extends beyond compute efficiency. A portion of the capital investment remains underutilized.

ASUS therefore shifts the objective from simply delivering higher specifications to continuously optimizing usable performance.

By optimizing power, thermal management, connectivity, and automation as an integrated system, ASUS aims to convert every watt of power, every unit of capital, and every stage of deployment time into more valuable Token output — ultimately improving the productivity and investment efficiency of the AI factory.

Engineering Beyond the Limits: Four Pillars of ASUS AI Infrastructure Innovation

Making TFE a reality requires more than improving the performance of individual components. It requires rethinking the data center infrastructure from the ground up.

As power density and system complexity continue to rise, ASUS is advancing four critical areas — power, thermal, connectivity, and automation — to push beyond existing physical and engineering constraints and convert more infrastructure resources into usable AI compute capacity.

Power — Turning Every Watt into More AI Compute

As AI GPU and rack-level power densities continue to increase, conventional power architectures face growing limitations in efficiency, space utilization, thermal performance, and transient load response. ASUS is therefore re-engineering the power architecture across multiple levels.

At the facility level, ASUS is developing 600kW-to-1MW 800V HVDC sidecars with N+1/N+N redundancy and battery backup units (BBUs), targeting 90% power conversion efficiency by 2026. At the same time, ASUS is actively advancing solid-state transformer (SST) technology as a path toward replacing conventional large-scale transformers, with the goal of achieving 96% overall conversion efficiency by 2028.

Closer to the compute silicon, ASUS integrates 48V/12V High-Speed Current Carrying (HSCC) architectures, coupled inductors, vertical voltage regulation (Vertical VR), and gallium nitride (GaN) power conversion.

These technologies shorten power delivery paths, mitigate voltage droop and resistive losses, and improve power delivery efficiency at high current densities. The objective is to extract more usable compute performance from existing grid capacity — maximizing the value generated by every watt of power.

Thermal — 3kW Microchannel Liquid Cooling and Fanless Design for Sustained Compute Performance

As power consumption per chip continues to rise, thermal management must evolve from a component-level challenge into a system- and data center-level capability.

At the chip level, ASUS leverages microchannel cold plate technology, bringing micron-scale cooling channels closer to the primary heat sources. This reduces thermal resistance by more than 35% and enables single-chip thermal dissipation of up to 3kW, pushing cooling performance toward the physical limits of high-density compute.

At the server level, the flagship ASUS rack-scale solutions can achieve 100% liquid cooling with a completely fanless design, eliminating up to 90% of fan-related energy consumption and mechanical vibration while supporting compute densities exceeding 200kW per rack.

At the data center level, ASUS applies high-temperature liquid cooling to support thermal management at the 5MW scale, reducing reliance on energy-intensive chillers. Combined with optimized facility-level infrastructure, this approach can help drive PUE below 1.1, balancing extreme compute density with energy efficiency and long-term sustainability.

Connectivity 3.0 and Algorithmic Expertise — Pushing Signal Integrity Beyond Conventional Design Limits

As large-scale AI systems enter an era of increasingly high-speed interconnects, Connectivity 3.0 has become another critical capability.

ASUS transforms years of engineering expertise into a Predictive Virtual Lab, enabling engineers to simulate and predict signal transmission stability, signal integrity, and design margins across different AI platform topologies before physical systems enter mass production.

This predictive approach improves the probability of achieving first-pass production success, reducing the risks associated with complex high-speed system design.

At the level of extreme signal integrity, ASUS advances both physical architecture and algorithmic optimization to mitigate risks arising from impedance mismatch, crosstalk, and manufacturing variations.

Proprietary ASUS mathematical and optimization algorithms were recognized with a Best Paper Award at the 2025 IEEE/ACM International Conference on Computer-Aided Design (ICCAD). The algorithms achieved a 6.9x speed improvement over conventional commercial tools, demonstrating the ability of ASUS to go beyond established design standards and apply algorithmic and engineering expertise to push toward the physical limits of high-speed computing.

Automation — Predictable Large-Scale Delivery Through L10/L11 Automated Manufacturing

Building an AI factory is not simply about speed. At large scale, predictability of delivery becomes equally critical.

ASUS has progressively implemented L10 and L11 automated production lines, significantly reducing assembly, integration, and validation time. These capabilities have enabled 50% to 100% improvements in overall production capacity while achieving yields above 98%.

Whether customers require hundreds of servers or thousands of racks, ASUS leverages standardized, modular, and highly automated manufacturing processes to deliver AI infrastructure with greater scalability and predictable lead times and quality.

Partner with ASUS to Build the Future of AI — From Compute Deployment to High-Efficiency, Sustainable, and Reliable AI Factories

As AI evolves from individual model applications to large-scale agentic AI and autonomous AI factories, the next stage of competitive advantage will no longer be determined by hardware performance alone.

The critical factor will be whether a complete infrastructure stack can continuously transform compute resources into productive, value-generating AI capacity.

With decades of expertise spanning server design, system engineering, thermal management, power delivery, high-speed connectivity, and automated manufacturing, ASUS is expanding its role from a leading hardware manufacturer to an AI infrastructure integration expert — helping customers transform compute, energy, and capital investments into sustainable AI production capacity.

ASUS believes the next generation of AI will be enabled by infrastructure.

As agentic AI scales, the industry faces increasingly complex challenges across compute capacity, energy consumption, and large-scale deployment. Now is the time to build the infrastructure ahead of that demand.

ASUS is ready to work with customers and industry partners to prepare the infrastructure before the next wave of AI arrives — building a new generation of AI Factories that combine high efficiency, sustainability, reliability, and scalability, and ensuring that every investment in compute can be transformed into a foundation for AI innovation and business growth.