Computex 2026: the hardware pivot and the new reality of AI infrastructure

The Next Evolution of Data Center Infrastructure: Rack-Scale AI Platforms

As agentic AI pushes traditional server limits, manufacturers are standardizing on fully integrated rack-scale architectures to maximize density and power efficiency.

The rise of agentic AI is forcing enterprise data centers to evolve. To maximize compute performance within strict power and physical footprints, manufacturers are standardizing on fully integrated rack-scale architectures. Supermicro is leading this modular transition with its Data Center Building Block Solutions, enabling hyperscalers to scale infrastructure seamlessly from single racks to entire facilities.

At Computex, Supermicro unveiled the AMD Helios platform, a 72-GPU double-width rack-scale system optimized for frontier model training and high-throughput inference. Powered by AMD Instinct MI455X GPUs and 6th Gen AMD EPYC CPUs, the system integrates AMD Pensando networking and the open AMD ROCm software stack to deliver exceptional compute density.

Power density and thermal management remain the primary engineering constraints. Supermicro is deploying rack-scale systems powered by 136-core Arm AGI CPUs to minimize overhead and sustain unthrottled performance. For extreme density, its liquid-cooled multi-node solution for OCP ORV3 environments supports up to 20,672 cores per rack, with individual nodes carrying up to 6TB of DDR5-8800 MT/s memory.

Concurrently, Giga Computing introduced its NVIDIA Vera Rubin NVL72 platform, achieving ten times more throughput per watt than previous systems. To accelerate installation, their Accelerated Deployment Unit integrates compute nodes with dedicated power and cooling modules for rapid, plug-and-play setup.

The Operator’s Lens: Bleeding Edge AI Physical Data Center Expansions

Audit your current data center lease terms to ensure facility power and cooling capacities can support upcoming rack-scale investments without triggering penalty clauses.
Pivot upcoming hardware refresh cycles toward direct-to-chip liquid cooling systems to handle the extreme thermal density of 72-GPU configurations.
Evaluate vendor roadmap stability and liquid cooling readiness in your facility capacity planning to mitigate future hardware lifecycle risks.

Wall Street Outlook on AI Infrastructure Winners

Investors are expanding their focus beyond chip designers to reward the server manufacturers and specialized clouds powering a multi-trillion-dollar infrastructure boom.

Compute is direct revenue. This was the core message from NVIDIA CEO Jensen Huang’s Computex keynote, a perspective that sparked a notable sector-wide rally. Wall Street is expanding its focus beyond silicon designers to reward the physical infrastructure providers enabling this multi-trillion-dollar transition.

The scale of this infrastructure expansion is unprecedented. Microsoft, Google, Amazon, and Meta are on track to spend a combined $700 billion this year on physical infrastructure, with projections climbing past $1 trillion by 2027. Goldman Sachs estimates indicate significant capital expenditures, with projected spending reaching $765 billion in 2026 and potentially $1.6 trillion by 2031.

Physical server manufacturers are experiencing corresponding growth. Dell Technologies is projecting $50 billion in AI revenue for the current fiscal year, representing a 103% year-over-year increase supported by $34.1 billion in new AI server orders booked in their recent fourth quarter. ABI Research estimates that Dell currently controls roughly one-fifth of the global AI server market.

This surge in spending is also benefiting specialized infrastructure providers. Following Huang’s keynote, shares of CoreWeave rose approximately 14%, Nebius gained 14.5%, and Oracle increased 9.9%, reflecting investor confidence in the broader AI infrastructure ecosystem.

However, rapid scaling has exposed severe supply chain chokepoints. High-Bandwidth Memory, which requires 12-layer stacked modules, is currently limited to three major global suppliers, with production capacity sold out years in advance.

Thermal and power limits are also emerging as critical constraints. Modern AI racks can draw up to 50 times the power of legacy systems. To manage this heat, the industry is transitioning to direct-to-chip liquid cooling, while hyperscalers are securing long-term energy agreements to guarantee continuous, gigawatt-scale electricity for future AI workloads.

The Operator’s Lens: Physical Data Center Compute Expansions

Allocate a dedicated contingency budget for High-Bandwidth Memory premiums to protect critical AI initiatives from unpredictable supply chain pricing surges.
Redesign physical facility blueprints to accommodate the massive weight requirements and specialized cooling loops of next-generation hyperscaler server racks.
Evaluate long-term energy availability and utility expansion plans when selecting future AI infrastructure locations.

Advancing Industrial and Rugged Edge AI Solutions

Hardware innovators are deploying physically resilient, autonomous computing platforms capable of orchestrating complex manufacturing tasks without persistent cloud connectivity.

While hyperscalers concentrate AI compute into increasingly dense facilities, industrial organizations face the opposite challenge: bringing intelligence closer to physical processes where latency, connectivity, safety, and operational continuity make cloud dependence impractical.

Enterprise AI is migrating from centralized cloud models to physical, localized environments that require real-time perception, reasoning, and autonomous action. This transition demands resilient edge computing frameworks capable of operating in demanding industrial settings without persistent cloud connectivity.

Advantech USA is addressing this need with its WISE-Edge Developer Architecture. By integrating NVIDIA’s three-computer framework—including Omniverse, Isaac Sim, and NemoClaw—Advantech has built an AI Factory Brain that enables autonomous software agents to manage visual inspection and factory operations. To simplify scaling across diverse environments, the company introduced seven standardized AI Foundry modules to transition intelligent systems from pilot phases to mass deployment.

Physical hardware resilience is a key constraint on the factory floor. IEI Integration Corp. showcased a specialized portfolio focused on high workloads, precise motion control, and cyber resilience. Their GAIA-5040A server runs vision models directly at the edge, while the TANK-XM811 platform integrates Mitsubishi Electric motion control with ROS2-based robotics. For hazardous or harsh conditions, IEI delivers DNV-certified maritime systems and IP69K-rated stainless steel servers designed to withstand rigorous washdowns.

Operational safety is similarly receiving a hardware upgrade. IEI consolidates mobile collaborative robot management and AI-driven intrusion detection on its TANK-XM813 system, powered by Intel OpenVINO, to identify personnel and execute safety shutdowns automatically. Concurrently, a partnership between Blaize and Winmate Inc. integrates low-latency programmable Graph Streaming Processor AI acceleration into rugged mobile platforms, ensuring autonomous operation in remote or disconnected locations.

The Operator’s Lens: The latest Industrial OT Implementation

Model your edge computing financial returns based on the reduction of unexpected manufacturing downtime rather than just the raw cost of hardware acquisition.
Standardize your factory floor deployments on fanless, modular systems designed to maintain local AI inference during network disruptions and harsh environmental conditions.
Upskill your operational technology workforce to manage ROS2-based robotic control frameworks before deploying autonomous inspection agents at scale.

The Physical Constraints of the AI Era

Computex 2026 highlighted a reality that many organizations are only beginning to recognize: AI scaling is no longer primarily a software challenge. Power availability, cooling capacity, hardware supply chains, and edge deployment resilience are becoming the new limiting factors.

Whether securing gigawatts of energy to support centralized AI infrastructure or deploying ruggedized edge systems on factory floors, enterprise intelligence increasingly depends on physical infrastructure resilience.

The organizations securing their cooling pipelines, energy contracts, and hardware supply chains today will be best positioned to scale the frontier models of tomorrow.