3 E Network has unveiled a novel approach to data center design that decouples long-lived physical infrastructure from rapidly evolving GPU hardware, allowing facilities to adapt to next-generation chips without costly renovations. The company's Mikkeli, Finland facility is engineered to support NVIDIA's upcoming Vera Rubin architecture, which will demand tens to over 100 kilowatts per computing rack, a dramatic increase from current systems.

The core challenge driving this redesign is straightforward but rarely addressed directly: data centers typically last 10 to 15 years, while AI chip technology cycles every 1 to 2 years. This mismatch forces operators into an uncomfortable choice: either invest heavily in facility upgrades every few years or watch their infrastructure become obsolete. By designing the Mikkeli facility's civil, mechanical, electrical, and environmental systems as a flexible foundation separate from compute hardware, 3 E Network aims to preserve long-term capital investments and reduce the disruption of future upgrades.

What Makes This Data Center Different from Traditional Designs?

The Mikkeli blueprint incorporates several engineering innovations specifically tailored to handle extreme power and thermal density. The facility features direct-to-chip liquid cooling with Coolant Distribution Units, a system that circulates coolant directly to GPU chips rather than cooling the surrounding air. This approach is essential because future Vera Rubin GPU Superchips paired with HBM4 high-bandwidth memory will generate heat at levels that traditional air cooling cannot manage.

The facility also reserves extensive routing capacity for 6th-generation NVLink and 1.6 terabit-per-second lossless network topologies, meaning data can move between GPUs without losing packets or requiring retransmission. This is critical for training large AI models, where communication bottlenecks between thousands of GPUs can waste enormous amounts of computing power and energy.

Structurally, the data center anticipates computing racks weighing nearly two tons when filled with liquid-cooled hardware. The facility's flooring and support systems are reinforced during the civil engineering phase to handle these loads safely and reliably.

How Does the Facility Handle Power Delivery for Future GPUs?

Power architecture represents one of the most forward-looking aspects of the design. The facility adopts a dual-track approach that supports both current high-voltage smart Power Distribution Units (PDUs) for existing NVIDIA HGX and MGX nodes, while simultaneously preparing for a shift toward 48-volt direct current (DC) power shelves with rack-level Battery Backup Units. This flexibility allows the facility to transition to more efficient power delivery systems as the industry evolves, without requiring a complete infrastructure overhaul.

The hybrid power architecture is designed to handle high power density and transient power spikes, which occur when thousands of GPUs suddenly increase their computational load. By incorporating battery backup at the rack level, the facility can smooth these spikes and enhance grid-level stability, reducing strain on Finland's electrical infrastructure.

Steps to Future-Proof AI Data Center Infrastructure

• Separate Infrastructure from Hardware: Design facility systems (cooling, power, networking) independently from compute hardware, allowing chip upgrades without facility modifications or extended downtime.
• Implement Full-Stack Liquid Cooling: Deploy direct-to-chip liquid cooling with redundant piping networks and blind-mate connectors to handle thermal loads of tens to over 100 kilowatts per rack.
• Reserve Excess Network Capacity: Build physical routing space and fiber cable trays for 1.6 terabit-per-second scale-out networks and next-generation Ethernet architectures before they become standard.
• Reinforce Structural Systems Early: Design flooring and support systems during the civil engineering phase to accommodate near two-ton liquid-cooled computing racks and future weight increases.
• Plan Hybrid Power Delivery: Support both current high-voltage PDUs and future 48-volt DC power systems with rack-level battery backup to handle power spikes and grid stability.
• Integrate Environmental Monitoring: Deploy out-of-band monitoring and micro-leak detection systems to identify cooling issues before they cause hardware damage or facility downtime.

What Workloads Will the Mikkeli Facility Support?

3 E Network plans a dual-track commercialization strategy that divides the facility into two operational zones, each optimized for different AI workloads. The HGX Core Zone will focus on large-model pre-training, the computationally intensive process of training foundational AI models on massive datasets. This zone features high-density deployment of HGX baseboards with centralized power and liquid cooling, designed to handle trillion-parameter training tasks with lossless interconnects that minimize data communication bottlenecks.

The MGX Elastic Zone addresses commercial inference, vertical fine-tuning, and agile deployment. This zone supports flexible, heterogeneous combinations of CPUs, GPUs, and Data Processing Units (DPUs), allowing enterprises to run Retrieval-Augmented Generation (RAG) systems, industry-specific model fine-tuning, and edge computing tasks with a more cost-effective structure. As AI moves from research into production, demand for these inference and customization workloads is growing rapidly.

"Building modern AI computing infrastructure is a dual test of capital planning and engineering foresight. For the Mikkeli project, we are designing the facility's physical infrastructure to accommodate the anticipated power, cooling and interconnect requirements of the Vera Rubin architecture and support future hardware upgrades," stated Dr. Tingjun Yang, Chief Technology Officer of 3 E Network.

Dr. Tingjun Yang, Chief Technology Officer at 3 E Network

Why Does Location Matter for This Data Center?

Finland's geographic and climatic advantages play a significant role in the facility's efficiency. The country's cool climate and abundant hydroelectric power enable natural cooling resources that reduce the energy required for cooling systems, a major operational expense in data centers. By leveraging these advantages, 3 E Network aims to integrate high energy efficiency with computing power, balancing the facility's massive power demands with sustainable operations.

The Mikkeli facility represents a shift in how the industry thinks about data center design. Rather than treating infrastructure as a static foundation that must be replaced when chips evolve, 3 E Network has engineered a system that anticipates change and builds flexibility into every layer. As GPU power consumption continues to climb and AI models grow larger, this approach to decoupling infrastructure from hardware may become a template for the next generation of AI computing facilities worldwide.