The whole arc converges on the AI factory

Someone asked me recently what changed when AI arrived. Honestly, less than you would think. The models are new. The economics are new. The physics is the one I have been designing around since a refinery pad in 2009. The AI factory is where the whole arc converges: the thermal work from the mine farms, the fabric and storage from the enterprise years, the governance from the regulated migrations, all pointed at one question I have been answering my whole career. Where does the compute physically live, and what has to be true for it to survive there.

Storage is where AI clusters actually stall

My current work is as the enterprise storage subject-matter expert for AI and HPC platforms at a Tier 1 Australian telco, and senior solution designer across on-prem, hybrid cloud and edge. Storage gets treated as an afterthought, and then a very expensive GPU supercluster sits idle waiting for data. A training run is only as fast as the slowest thing feeding it, and past a certain scale that is almost always the storage and the fabric, not the accelerators.

The trap is designing storage as if AI were one workload. It is at least two, and they pull in opposite directions. Training wants sustained sequential bandwidth, enough aggregate throughput to keep every accelerator saturated, and it wants it to survive checkpointing, where the whole cluster tries to flush state at once and turns a bandwidth problem into a burst problem. Inference wants the opposite: low-latency random reads, small objects, and a tight tail, because the ninety-ninth percentile is what the user actually feels. Design for the average and you starve both.

So the design authority I hold is over a data path, not a box. NVMe flash close to the accelerators for the hot set, a parallel high-throughput file system across it so many GPUs can read the same corpus without contending, a capacity tier behind that, and object storage for the lake, with data placement and lifecycle policies deciding what lives where rather than leaving it to accident. The fabric is the other half of the same design. InfiniBand or RoCEv2 for the east-west collective traffic, sized so all-reduce across the cluster does not choke, and NVMe-oF and GPUDirect-style paths so data moves storage-to-GPU without bouncing through host memory and burning a copy. NVIDIA reference architectures, the DGX and HGX and GB200-class building blocks, give you the node. Making a supercluster out of nodes is a storage and fabric problem, and it is the one people under-budget.

I stay deliberately vendor-neutral on the silicon. NVIDIA is the centre of gravity, but I track the AMD stack too, ROCm and the Instinct line through MI300 and MI400, because an AI-factory architect who can only design for one vendor is a procurement risk, not an asset. The job is to keep the option open, not to pledge allegiance.

Data-centre design governance is the old craft, renamed

The other half of the role is governance over the physical envelope: space, power, cooling, fabric and redundancy for AI-class racks. This is the refinery discipline and the mine-farm discipline, wearing a new label. When a rack goes from fifteen kilowatts to well over a hundred, air cooling stops being a choice. I have run the air-to-liquid transition before, on mining rigs a decade ago, and now I run it on liquid-cooled AI compute where the stakes are a national platform rather than a hobby loss. Direct-to-chip cold plates for the accelerators, the coolant distribution and heat rejection to match, and the honest question of whether the facility can even feed and cool what the design wants before a single GPU is ordered. The first question about a dense GPU cluster is still the cooling plant.

Value has to land on the business

Architecture that cannot show its worth is just expense, and the part of this job that most people in deep infrastructure avoid is the part where you defend the spend. I led an architectural review of more than a thousand applications, deciding what to modernise, rehost or rebuild on economics rather than dogma, and returned north of twenty million dollars a year in savings. I have carried multi-year, multi-million-dollar investment cases into steering committees and had to make an AI-factory build legible to people who will never touch a fabric switch: the total cost of ownership, the FinOps model that keeps it from becoming an open tab, the value realisation that justifies the next tranche. Then the same week I sit with the low-level design and the reference blueprint and make sure the thing we sold is the thing we can actually build. That translation, commercial intent into technical roadmap and back again, is the work. I built an automated pipeline for cloud account provisioning so new environments came up governed by default, and an object-based backup pattern into AWS and Azure, and I have started bringing AI agents into the internal practice itself, because the fastest way to trust this technology is to run it on your own work first.

Where the compute lives is the whole game

The pattern underneath all of it is placement. Centralised training in the AI factory, where the power and cooling and fabric justify the density. Distributed inference at the edge, close to where the data is made and the latency budget is tight. Sovereign AI where the data is never allowed to leave the ground it sits on, which is the same data-sovereignty problem I first met on an industrial site fifteen years ago. MLOps to keep the whole loop honest.

I do not design this from whitepapers. I run a DGX and a cluster of Mac Studios in my own lab, loading and quantising local models, building evaluation datasets, and watching where they fail, so that when I stand in front of a steering committee and say a design will hold, I have already broken a smaller version of it myself. The NVIDIA certifications put a formal edge on instincts I built the long way. The hardware changed across every chapter. The scale changed. The thermodynamics, the fabric, the storage physics never did. Twenty-five years of deciding where compute physically lives, all pointed now at the AI factory.