3 E Network Seeks Global Vendors for Mikkeli, Finland AI Center With 120kW Racks

Hong Kong-based 3 E Network Technology Group Limited (Nasdaq: MASK) said on September 29, 2026, that it has begun a global vendor evaluation and Request for Proposal process for its multi-megawatt AI compute center in Mikkeli, Finland.

The process covers high-density server clusters, liquid cooling infrastructure and core networking equipment.

The company describes itself as a business-to-business information technology solutions provider committed to becoming a next-generation artificial intelligence infrastructure solutions provider.

From Blueprint to Hardware Selection

According to the company, the procurement effort follows the recent release of the Mikkeli Data Center Blueprint, the finalization of multi-megawatt power parameters and the establishment of compute distribution channels.

With those steps complete, the company said, the project moves into the practical phase of physical hardware selection and deployment. The evaluation is organized around four technical requirements set out in the blueprint.

Cooling Requirements for 120 kW Racks

The first requirement concerns thermal management. To address rising Thermal Design Power in Large Language Model training workloads, and in line with previously established thermal management objectives, 3 E Network requires that proposed solutions support rack power densities of 120 kW and above.

The company cited the physical constraints of traditional air-cooling architectures when managing next-generation high-power AI accelerators. As a result, the evaluation will prioritize advanced Direct-to-Chip liquid cooling.

The company described an approach in which micro-channel cold plates are affixed directly to the silicon core, using high specific heat capacity to manage primary component heat.

According to the announcement, the standard aims to mitigate localized thermal hotspots, maintain thermodynamic stability during extended training cycles, extend hardware operational life and optimize the facility's overall Power Usage Effectiveness.

Accelerator Compatibility and Network Design

The second requirement centers on compatibility and network design. The RFP process prioritizes what the company calls "architectural compatibility and stress-test reliability" as key evaluation metrics.

Its engineering teams are benchmarking technical requirements against the spatial, power delivery and data throughput profiles of upcoming flagship AI accelerator architectures, including ecosystems based on Blackwell and Vera Rubin planning. In line with the high-speed cluster objectives in the blueprint, the company has specified standards for low-latency, non-blocking interconnectivity.

Citing the intensive data interaction demands of LLM training, the evaluation will focus on 800G and above Ethernet and InfiniBand-class leaf-spine topology solutions. The stated aim is to optimize internal data flows and ensure that large-scale GPU nodes operate in a highly synchronized environment.

Storage Built Around NVMe Over Fabrics

The third requirement addresses storage. The company stated that in the multimodal model landscape, data transfer efficiency is as vital as underlying computational power.

To overcome Data Input/Output bottlenecks during large-parameter model training and prevent GPU compute idle time, it has designated all-flash NVMe over Fabrics storage arrays and high-performance parallel file systems as standard procurement criteria.

The specification requires the storage architecture to deliver read throughputs at the terabytes-per-second level with microsecond latency. The system must also support efficient, concurrent model checkpointing.

According to the company, this allows rapid preservation of extensive model state data, which facilitates recovery from hardware interruptions, minimizes the loss of training progress and maintains the operational efficiency of compute assets.

Power Delivery Moves Toward 48V DC

The fourth requirement relates to power delivery. To integrate smoothly with the facility's green energy architecture and manage the significant transient power spikes associated with new-generation AI chips, 3 E Network requires rack-level power delivery infrastructure to be compatible with, and capable of evolving toward, a 48V Direct Current busbar architecture.

The company noted that, compared with traditional 12V setups, 48V power delivery lowers line current, which reduces transmission losses and improves end-to-end power conversion efficiency.

The accompanying intelligent Power Distribution Units must provide high conversion efficiency alongside integrated dynamic load balancing and precise energy monitoring.

The company said this design is intended to offer robust reliability for the underlying electrical grid when high-density clusters handle complex inference tasks or initiate large-scale training runs.

Next Steps in the Evaluation

According to the announcement, 3 E Network's management team views the core hardware evaluation as a crucial phase in translating the theoretical blueprint into physical infrastructure.

By outlining specifications that include the 120 kW+ liquid cooling threshold, non-blocking networking, high-throughput I/O, and 48 V power compatibility, the company said it has communicated clear deployment requirements to the hardware supply chain.

The company stated that it is focused on building a robust, industrial-grade technological platform to support the computational demands of large-scale AI models. In the subsequent evaluation period, it will engage in detailed technical discussions with selected vendors to finalize the infrastructure matrix selection.

The company said this is intended to accelerate the capital expenditure rollout and the practical commissioning of the Finnish project.

About the Company

3 E Network Technology Group Limited describes itself as a business-to-business information technology solutions provider. The company trades on Nasdaq under the ticker MASK.