Citi Outlook on Taipei International Computer Show: AI Infrastructure Heads Towards the 'Composable Factory' Era

Citi Research pointed out in its latest report that the 2026 Taipei International Computer Show will witness a significant shift in the AI infrastructure narrative. The future core focus will move comprehensively from the power of a single GPU to the overall optimization of CPU scheduling, high-speed networking, efficient cooling, and high-voltage power supply systems.

This means that AI infrastructure will no longer be just a simple GPU cluster, but will evolve into a complete "AI factory" model. Under this trend, the assessment criteria for the semiconductor supply chain will also be completely rewritten.

The core competitiveness of enterprises will no longer depend on the exposure to individual components. Companies that can provide a complete set of AI system delivery capabilities, including cabinet integration, liquid cooling, and high-voltage power systems, will gain more market dividends.

From Server to Cluster System Design

Citi analysis believes that NVIDIA is promoting the evolution of AI infrastructure architecture from traditional server level to cabinet level and even cluster level design. Under this new model, the traditional server deployment method is being disrupted, and the entire cabinet is evolving into a new system boundary.

The next-generation Rubin architecture to be showcased will further extend this design logic. For example, the brand new Vera Rubin Ultra NVL576 architecture is a representative product of cluster-level system design.

It integrates eight independent MGX cabinets into a massive NVLink compute domain with 576 GPUs through copper cables and direct optical connections. This indicates that the traditional "cabinet" is becoming an independent module in a larger cluster computing domain.

Vera CPU Positioning the Scheduling Layer of Intelligent Agents

The report points out that NVIDIA's Vera CPU is not intended to directly compete with traditional chip manufacturers in the general-purpose server market. Its core mission is to precisely control the overall business orchestration and scheduling layer of the AI factory.

In the era of agent-based AI, this CPU will act as the traffic controller for workflow control, tool invocation, memory indexing, and retrieval management. As inference loads become longer and continuous, memory bandwidth and capacity on the CPU side also become crucial.

Supply chain information shows that the production preparation volume of Vera CPU in 2026 is expected to reach about 2 million units. Mainstream ODMs such as Foxconn and Quanta are expected to showcase related cabinets and complete systems at the computer show.

Power Consumption Soars, Forcing a Paradigm Shift in Electrical and Thermal Management

With the surge in compute density, the overall power consumption of the Rubin NVL72 cabinet is expected to soar to more than 180 to 220 kilowatts. Compared to the Hopper era, which was only 40 to 60 kilowatts, this figure has increased severalfold.

Such extreme power consumption has made full-cabinet direct liquid cooling technology an undisputed industry standard. At the same time, the traditional 54V bus is approaching its physical limit at such high power, making it difficult to meet the thermal management needs of copper cables and connectors.

This is forcing data centers to accelerate their transition to 800V direct current power architectures and more efficient power racks. Faced with dual challenges in electromechanics and thermodynamics, the dividends of the Taiwanese supply chain are rapidly spreading in all directions towards integrated system solutions such as cooling modules, high-voltage power supplies, and buses.