Key Takeaways
- 3 E Network Technology Group’s Mikkeli facility is designed to native Vera Rubin specifications, while running HGX and MGX clusters until next-generation hardware ships.
- The blueprint decouples the physical plant from the compute layer, reducing the cost of swapping chip generations across a building lifespan of 10-15 years against GPU refresh cycles of one to two years.
- Full-stack liquid cooling and 48V DC power infrastructure address Rubin’s projected rack-level thermal and power density, which makes conventional air cooling and AC distribution impractical. On September 11, 2026, 3 E Network Technology Group disclosed the design of its Mikkeli facility in Finland: infrastructure built to native Vera Rubin specifications, running HGX and MGX clusters now, ready for next-generation hardware when it ships. The central engineering problem is a timing mismatch, data center shells last 10-15 years; the AI chips inside them turn over every one to two.
Built for Rubin’s Demands
The Mikkeli design targets a six-chip, five-rack architecture, per NVIDIA’s own platform description for agentic AI workloads, long-context inference and trillion-parameter Mixture-of-Experts models. (Some analyses count seven, including the Groq LPU NVIDIA added to its lineup following its Groq acquisition.) According to Nvidia‘s own figures, the Rubin platform delivers 3,600 PFLOPS of NVFP4 inference per rack, a five-fold throughput increase over Blackwell at ten-fold lower cost per token, though independent benchmarking has not confirmed those figures. The Rubin GPU is spec’d at 50 PFLOPS FP4 with 22 TB/s HBM4 bandwidth; the Vera CPU runs 88 custom Arm Olympus cores on TSMC’s 3nm process. Those specs push single-rack thermal loads into ranges where conventional air cooling fails.
3 E Network’s answer is full-stack liquid cooling: resilient piping networks, direct-to-chip cooling with Coolant Distribution Unit circulation compatible with blind-mate connectors, and Finland’s ambient climate doing real work on PUE under peak load. The HBM4 memory bandwidth Rubin requires compounds the thermal challenge further, high-bandwidth stacked memory runs hot.
HGX Now, Rubin Later
Building to Vera Rubin spec while shipping HGX and MGX clusters is the core commercial logic. The facility generates revenue today on established hardware while the physical plant sits ready for the next generation. Vera Rubin is reported to have begun shipping to hyperscalers in H2 2026, so the transition window is not distant.
The lifespan mismatch this addresses is real. A building and its mechanical systems are typically written off over 10-15 years; the GPUs inside turn over on one-to-two-year cycles. By decoupling the physical infrastructure from the compute layer, 3 E Network reduces the cost and disruption of swapping generations. The facility does not need to be rebuilt every time Nvidia ships a new architecture.
Networking and Power
Rubin’s 6th-generation NVLink generates substantial intra-rack cabling volume. The Mikkeli blueprint reserves dedicated rack-level routing space for that density, with backbone fiber trays sized and routed for 1.6T non-blocking scale-out network topologies compatible with ConnectX-9. Across a five-rack system at this scale, low-latency interconnect provisioning is not optional.
Power delivery runs a dual-track design. High-voltage smart PDUs handle current MGX and HGX nodes; the facility also carries full forward-compatibility for 48V centralised DC power shelves and rack-level Battery Backup Unit specs. The 48V DC path matters specifically for Rubin’s anticipated power density and transient spikes, which stress conventional AC distribution and can affect grid-level stability at scale. For more on the infrastructure implications of AI’s rising power demands, see our coverage of AI workload power projections through 2030.
Originally published at https://autonainews.com/3-e-network-designs-mikkeli-data-center-around-nvidia-vera-rubin-specs/
Top comments (0)