Your datacenter is a warehouse with opinions

Three variables, endlessly repackaged

A datacenter converts electricity into computation and heat. Everything else is packaging. The variables that decide whether yours is good are boring and countable: how many watts you can deliver per rack, how efficiently you remove the resulting heat, how quickly a dead component is replaced, and how much of your installed capacity is actually doing work.

Industry average power usage effectiveness has hovered around 1.5 to 1.6 for years, while purpose-built modern facilities run near 1.1. That spread is not a technology gap. It is a management gap, airflow discipline, containment, set-point courage, and someone whose job is to care.

We have walked facilities where the cold aisle was set to 18 degrees because that was the number written down in 2011, with nobody able to say why. Moving that set point to the 24 to 27 degree range recommended by ASHRAE for years, with proper containment, cut cooling load by close to a fifth in one site without touching a single server. That is a decision, not an investment, and it sat unmade for over a decade because nobody was accountable for the power bill line by line.

The zombie audit

Every consolidation project we have joined starts the same way: an inventory nobody trusts. Run a two-week audit of CPU, network and disk activity per host and a familiar shape appears. A fifth to a third of the estate is comatose, kept alive because a decommission requires a decision, and decisions require an owner.

• Tag every host with a named owner and a business service. No owner, no power.
• Measure utilisation for 14 days, including a month-end and a batch window.
• Anything under 5% average CPU with no inbound sessions goes on the candidate list.
• Power it off for 30 days before you remove it. If nobody screams, decommission.

This is not glamorous work. It is also the highest-return work available in most infrastructure estates, because it removes cost in three places at once: hardware refresh, licensing, and the power and cooling underneath.

The resistance to this audit is rarely technical, it is political. Every zombie server has a name attached to it somewhere in an old change ticket, and the person who deployed it in 2017 does not want to be the one who admits it was never decommissioned properly. Running the audit as a blameless exercise, with amnesty for anyone who flags their own orphaned box, moves this from a three-month standoff to a three-week project in our experience.

Density changed the maths

For twenty years a rack meant 4 to 8 kW and air cooling was fine. GPU-heavy workloads pushed racks toward 40 kW and beyond, and at that point air stops being a serious answer. Liquid cooling is not a futurist talking point anymore; it is a facilities decision with a spreadsheet attached.

The commodity argument still holds. You do not need to own the building to get the benefits, you need to know your kW per rack, your cost per kW, and your cost per useful unit of work. Colocation, cloud and on-premise are then a comparison instead of a religion.

The mistake we see most often is retrofitting a legacy hall for high density rack by rack, chasing the workload instead of the plan. Direct-to-chip liquid cooling needs a coolant distribution unit, piping, and a facilities team trained to service it, none of which are cheap to bolt on piecemeal. If you know GPU density is coming in the next 24 months, it is materially cheaper to zone a section of the hall for it now than to retrofit a rack at a time later.

Build, colocate or rent: the actual comparison

Most of the build-versus-buy debate in datacenters is conducted on vibes. Owners cite control; colocation providers cite speed to deploy; cloud vendors cite elasticity. The number that actually matters is cost per useful compute-hour delivered at your real utilisation rate, not at nameplate capacity, and almost nobody calculates it that way.

• Owned facility: lowest marginal cost at high, steady utilisation; worst option if your load is spiky or your growth is uncertain.
• Colocation: pay for space and power, own the equipment; sensible middle ground for most enterprise estates above a few racks.
• Public cloud: highest per-unit cost at steady state, cheapest option for anything spiky, seasonal, or genuinely uncertain.

A hybrid answer is usually correct and usually unpopular internally, because it means running three operating models instead of one and explaining that to an audit committee. We think that discomfort is a bad reason to pick the wrong model. Pick per workload, not per company.

Buy it like a warehouse