Power, Not GPUs, Is the Constraint on Your Next Cluster
Most cluster plans are sized on GPU count and then discover the room cannot power them. Working the other way round produces a design that actually runs.
Insights
Technical guides, reference architectures, product announcements and industry analysis from the team that builds these systems.
Featured
Latest
Deskside or Cluster: Where Model Work Should Actually Happen
Most organisations put development on the cluster because that is where the GPUs are, then wonder why training queues never clear.
PCIe or SXM: What the Choice Actually Costs
Eight GPUs in slots and eight GPUs on a baseboard are not two price points of the same thing. One is serviceable, the other is faster between cards, and the workload decides which matters.
Running AI Where There Is No Fibre
A site without terrestrial connectivity can still run inference locally. What it cannot do is go unmanaged, and that is the problem worth solving first.
Sustainability Reporting for Accelerated Computing
Measuring and reporting energy use credibly, and why PUE alone understates the picture for AI workloads.
Scheduling a Shared Cluster Without Starving Anyone
Queue design, fair-share policy and why moving development work off the cluster is usually the highest-return change.
The Economics of Serving Large Models
Batch size, memory bandwidth and interconnect all determine cost per token. Modelling the tradeoffs.
The Total Cost of AI Infrastructure
Capital, power, cooling, floor space and staff, modelled across a five-year deployment rather than a purchase price.
Sizing PDUs for GPU Racks Without Tripping Breakers
Transient peaks run well above steady-state draw. How much headroom to leave and why nameplate ratings mislead.
When Air Cooling Stops Being an Option
The threshold is around 40kW a rack, and the reason is airflow volume rather than efficiency. What changes at that point and what it costs.
Cooling Strategy: Air, Liquid-to-Air and Liquid-to-Liquid
A direct comparison at 20kW, 40kW and 80kW per rack, including operational cost and failure modes.
Choosing a Fabric Topology for Your Node Count
When a single tier suffices, when a fat tree is required, and the cost of over-designing for growth that does not arrive.
Planning an AI infrastructure build?
Tell us the workload and the power envelope you have to work within. We will come back with a configuration, a lead time and a written quote.