If you’ve been following AI infrastructure at all over the past year, you know that most of the conversation has been centered around training, which, of course, refers to the massive systems used to build these models in the first place.
That’s where all the talk of bigger clusters, more GPUs, and more power has come from.
That story is still definitely playing out, but it’s starting to feel like only one of the layers of the massive onion that is AI.
Data center buildouts continue as AI demand accelerates
Reuters reported this week that I Squared Capital is acquiring 10 data centers from Cogent Communications, with plans to invest another $1 billion to expand capacity for AI workloads.
So, rather than concentrating everything in one massive campus, which would be the natural assumption, this leans more toward spreading capacity across different locations.
Basically, this brings infrastructure closer to where AI is actually used, so it can run faster and more reliably day-to-day.
As the report states, the investment aims to support demand for AI infrastructure, particularly as workloads shift toward inference rather than just training.
The rise of inference infrastructure
Training models is only one part of a much bigger and annoyingly complicated equation. Once those models are deployed, they have to respond in real time across apps, customer interactions, and internal systems.
Enter… inference.
Several industry reports point to a growing need for infrastructure that sits closer to end users. Data Center Dynamics and others are starting to describe a different kind of facility built for inference, with smaller, more regional setups that can keep up with steady, real-time demand.
S&P Global and JLL have also flagged this trend, noting that edge and regional data centers are becoming increasingly important as AI moves into production environments.
The requirements are quite a bit different from those for training clusters, with more focus on responsiveness and proximity than on raw compute.
A more distributed buildout
The shift toward inference changes how infrastructure gets deployed. Instead of relying solely on large, centralized regions, organizations are starting to spread workloads across multiple locations.
This has pretty practical implications, including that latency becomes harder to ignore, data residency and compliance requirements start to influence where systems run, and networking and integration work grow more complex as environments span regions.
The Cogent acquisition lines up with that direction. Instead of centralizing everything, it spreads capacity across a bunch of existing sites, which makes sense when you think about how these systems are actually used.
For companies working through AI deployments, the focus is definitely shifting from building models to how they run in practice, which is where things get a little spicier.
Where the compute sits, how it connects, and how far it has to travel start to matter a lot more.
It’s still early, but this phase (or onion layer, if you will) is already shaping up a little differently than the first.
Catch up on other important things MSPs should know about AI infrastructure demand in this episode of Channel Insider: Partner POV, featuring SotaTek’s US CEO, MK Tong.