123NET: Inside the Data Centre With 4MW GPU Power for AI

Data centre operators across North America are competing to add capacity for GPU-intensive workloads as enterprises deploy AI systems at scale. 

Now 123NET, a Michigan-based colocation and network services provider, is completing an expansion of its Southfield DC1 facility that adds 4MW of power capacity designed for organisations running graphics processing units for AI inference – the process of using trained machine learning models to make predictions or generate outputs.

The project makes DC1 the only facility in Michigan to offer high-density GPU infrastructure with both liquid and air cooling systems alongside on-site peering. 

The facility now supports organisations running GPU clusters for real-time AI applications that depend on both processing power and low-latency connections.

What is peering and Detroit Internet Exchange’s role in it?

The Southfield data centre houses the Detroit Internet Exchange, a switching platform that enables direct peering between network operators. 

Peering allows networks to exchange traffic directly rather than routing through third-party carriers, reducing costs and latency. The exchange operates on a fee-free model.

“Having DET-iX inside the data centre gives our clients instant access to a powerful peering ecosystem and a more direct path to end users,” says Dan Irvin, Founder and CEO of 123NET.

“It's the only place in Michigan where massive GPU power meets free peering, significantly cutting network costs compared to traditional colocation.”

DET-iX serves 85 members including Apple, Microsoft and Google. 

The exchange has recorded traffic peaks above 2.5 terabits per second and offers port capacities of up to 400 gigabits per second. 

Housing the exchange within the data centre allows organisations to collocate equipment and connect to the peering fabric in the same building – which matters for AI deployments that process data from distributed sources and serve requests from users across different networks.

Targeting GPU deployment requirements

The facility provides space for organisations to deploy GPU clusters that require power densities higher than traditional server equipment. 

GPUs consume more electricity per rack than central processing units and AI workloads push power requirements beyond what older data centres support. 

The DC1 expansion includes electrical and cooling systems built for these higher densities.

As a result, 123NET designed the expanded space to support both AI inference and ML training.

This means that the facility offers organisations the option to deploy either workload type based on their requirements.

“We're seeing explosive demand for GPU clusters that require both substantial power and ultra-low latency,” says Dan.

“This expansion, paired with our connectivity, creates an ideal environment for real-time AI workloads. We've built the density and network performance that serious GPU deployments require.”

Why Michigan?

The Southfield facility serves organisations in the Midwest region of the US. 

Michigan has attracted data centre development due to access to power and fibre connectivity, though it remains a smaller market compared to Northern Virginia and Northern California. 

DET-iX, one of the larger fee-free peering platforms globally, provides network operators in the region with a local interconnection point that reduces the need to route traffic through exchanges in other states.

The facility expansion includes both liquid and air cooling options, giving organisations flexibility in thermal management. 

“This expansion, paired with our connectivity, creates an ideal environment for real-time AI workloads,” says Dan.