Density is the new capacity
The datacenter constraint is no longer floor space. It is watts per square metre, and most facilities were designed for a different century.
AI training and inference are pushing rack power from 8 kW toward 40 kW and beyond. That is not an IT problem anymore. It is a civil, mechanical and electrical problem, and the winners are the ones who stopped pretending air could do the job.
The physics always wins
For decades, datacenter design assumed 4 to 8 kW per rack and that air was a reasonable heat transfer medium. At 40 kW per rack, you are moving so much air that the fan power alone becomes a significant fraction of the load. Liquid is roughly a thousand times denser than air as a heat carrier. The engineering case is not subtle. Yet we still meet facility teams treating a liquid cooling proposal as an exotic option to be piloted in a corner of the floor rather than the baseline for anything hosting modern accelerators.
The reluctance is understandable. Liquid cooling touches water chemistry, leak detection, plumbing standards, and a skill set that most facilities teams built their careers avoiding, precisely because air cooling was simpler and forgiving of mistakes. But forgiving of mistakes is not the same as fit for purpose. A facility that cannot deliver 40 kW to a rack is not a smaller version of a modern datacenter; it is a different category of building, and pretending otherwise costs money twice: once in the retrofit, and once in the compute you could not host while you waited.
You can argue with a vendor. You cannot argue with the second law of thermodynamics.
Density changes the business case
- More compute per square metre means the building, not the silicon, becomes the constraint.
- Existing facilities face a mechanical retrofit decision, often more expensive than new build.
- Power availability at the site matters more than floor space; many regions are already capacity constrained.
- Liquid cooling shifts operational skill from CRAC maintenance to plumbing and water chemistry.
- Grid connection timelines are now the long pole in the tent, sometimes longer than construction itself.
The commodity insight is the same as always: you do not need to be a facilities engineer, but you do need to know what you are buying. Cost per kW delivered, cost per useful compute unit, and the real residual value of a building not designed for the next ten years are the numbers that decide whether to renew, retrofit, or relocate. We have watched procurement teams sign colocation renewals on price per rack alone, only to discover eighteen months later that the facility physically cannot support the density their new hardware refresh requires.
What the operators are actually doing
The best run facilities we see treat density as a product specification, not a surprise. They ask every new tenant: what is your kW per rack trajectory over the contract term? A tenant that cannot answer is a tenant that will outgrow the building before the lease ends. The good operators have started pricing that uncertainty into the contract, with power escalation clauses that read almost like insurance products in their own right.
Power procurement has become a boardroom topic in a way it was not five years ago. Sites are being chosen for grid headroom before they are chosen for latency or land price, because a beautiful piece of land next to a fibre exchange is worthless if the local substation cannot deliver another 20 megawatts this decade. We have seen deals fall apart at the last stage of due diligence because nobody checked the grid connection queue until the term sheet was already signed.
The retrofit trap
The hardest conversation is with owners of five to ten year old facilities that were state of the art when built and are now structurally limited. Raising the floor loading, adding chilled water loops, and upgrading electrical distribution is rarely a weekend project; it usually means a phased migration with reduced capacity during the works, which is exactly the period when demand for AI compute is least willing to wait.
- 01Get an independent structural and electrical audit before committing to any retrofit budget.
- 02Model the cost of lost capacity during construction, not just the construction cost itself.
- 03Ask whether a phased liquid retrofit or a clean new build actually costs less over ten years.
- 04Do not let the facilities team make this decision alone; it is a compute strategy decision.
Buy it like a power plant
- 01Specify kW per rack for the full lease term, not just today.
- 02Cooling is not an afterthought; it is the dominant operating cost at scale.
- 03Measure cost per useful compute unit, not cost per square metre.
- 04Treat legacy air cooled floor space as a depreciating asset with a known expiry.
- 05Check the local grid connection queue before you check the fibre map.
More dispatches.
The catamaran is a computer at sea
Multihulls did not win because they are fast. They won because a stable platform lets sensors, models and crew think clearly.
The cat is now a connected device
Pet tech stopped being a gadget niche. It is a health data business with whiskers.
The incubator is a portfolio, not a factory
Ventures do not share a mould. They share a treasury, a network, and a discipline for killing things early.