Loading
PMO Hive acts as the independent control tower for hyperscale delivery. We integrate cost, schedule, and risk governance across every tier of a programme — ensuring clarity, compliance, and control from day one.
The next three years: the technology about to reshape data centre delivery
A few weeks ago SpaceX filed an application with the US regulator for up to one million satellites, each one designed to carry AI compute. Not communications. Compute. The plan projects 100 gigawatts of processing capacity in orbit, and Elon Musk has said the cheapest way to generate AI compute will be in space within two to three years.
It is easy to file that under science fiction and move on. The problem is that most of the technology reshaping data centres over the next three years sounds like science fiction right up until someone signs a purchase order. Solid-state transformers, factory-built nuclear reactors, racks drawing more power than a street of houses. All of it is moving from whitepaper to procurement faster than the industry can absorb.
This post looks at what is actually coming, how near it is, and why every one of these shifts lands squarely on the people responsible for delivering the buildings.
Start with the rack, because everything else follows from it.
A legacy server rack drew 5 to 10 kW. Today's AI racks run at 50 to 70 kW. That already forced the industry to abandon air cooling. What is coming next makes those numbers look modest. At Data Centre World London this year, vendors showed pod systems rated to 250 kW per rack. NVIDIA's Rubin Ultra "Kyber" rack, a 600 kW test unit, is slated for around summer 2027.
That is roughly a hundredfold increase in power density inside a decade. No power or cooling architecture designed for today's loads survives contact with it.
"If you design for the density you need today, you have already designed a building that is obsolete before it opens." — /pmo
The practical consequence is that anyone breaking ground now has to build for loads that do not yet exist, using standards that are not yet finalised. That is a governance problem long before it is an engineering one.
Here is a shift most people outside the industry will never notice, but it changes the electrical design of every new facility.
Data centres have always run on alternating current, the same AC that comes out of the grid. Every conversion step from AC to DC inside the building wastes energy. At hyperscale, those losses are enormous. So the industry is moving to 800V direct current distribution, where power is converted once and stays as DC all the way to the chip.
One vendor at Data Centre World demonstrated that total power loss in a DC system runs around 10 percent, against 25 percent in a conventional AC setup. That is not a marginal gain. At 100 MW of load it is the difference between needing 110 MW or 125 MW from your supply.
The enabling technology is the solid-state transformer, a device that replaces the room-sized copper-and-iron units that have barely changed since the 1880s. Startups in this space (DG Matrix, Amperesand, Heron Power) have raised serious money, and incumbents like Eaton, ABB and Hitachi Energy are all prototyping. Expect the first SST-native, all-DC greenfield campuses to reach anchor-customer deployment in the 2026 to 2027 window.
For a delivery team, this is a new equipment class, a new set of interfaces, and a new commissioning discipline arriving all at once.
Liquid cooling is no longer the future. It is the baseline. Air simply cannot remove the heat a 56 kW rack produces, let alone a 250 kW one.
The near-term progression runs through three stages, and most operators will touch all of them inside three years:
Each stage demands specialist commissioning that most data centre contractors have never performed. A cooling system that is 95 percent right does not run at 95 percent. It does not run. And in a water-constrained region, regulators are increasingly mandating dry or hybrid cooling and recycled water, adding another constraint to an already tight design envelope.
The single biggest constraint on data centre growth is not chips or cooling. It is power. Grid connection timelines now stretch to the back end of the decade in saturated markets, and that has pushed operators toward an answer that would have seemed absurd five years ago: build your own reactor.
This is no longer speculative. Through 2025 and into 2026 the agreements have become real:
Small modular reactors run at around 300 MW or less; microreactors at 20 MW or less, small enough to sit on site and bypass the grid entirely. The core appeal is that by making nuclear smaller and factory-built, it becomes far less prone to the cost and schedule overruns that have plagued conventional nuclear for decades.
That last point deserves a pause. The entire promise of SMRs rests on delivery discipline: repeatable, modular, series-manufactured units that behave predictably. The technology only pays off if the programme management around it is world class. This is energy megaproject territory arriving inside the data centre fence line.
Which brings us back to those satellites.
The case for orbital data centres is genuinely compelling on paper. In the right orbit the sun delivers around 1,361 watts per square metre, continuously, with no night and no clouds. There is no grid to connect to, no community objection, no water for cooling. Google's Project Suncatcher plans to launch two prototype satellites carrying its TPUs in early 2027. Starcloud has already run an NVIDIA H100 in orbit and trained a model there.
The obstacles are equally real. Cooling in a vacuum is hard, because heat cannot be carried away by air and must be radiated through enormous panels. Launch cost is the whole economic argument: space compute only competes with the ground if launch drops toward 200 dollars per kilogram, a target that depends on SpaceX's Starship scaling in a way it has not yet demonstrated.
The honest read is that orbit will not replace terrestrial data centres this decade. The likely outcome is specialisation, with orbital systems taking on energy-heavy, latency-tolerant workloads while the ground keeps everything real-time. But the direction of travel tells you how hard the terrestrial constraints have become. When serious companies are looking at space to escape grid queues and planning objections, the delivery problem on Earth has clearly reached a breaking point.
Step back and a pattern emerges. Every one of these shifts (higher density, DC power, exotic cooling, on-site nuclear, even orbit) is a response to the same underlying problem. Demand is outrunning the physical ability to deliver.
None of these technologies solves the delivery challenge. They intensify it. Each one introduces a new equipment class with its own lead time, a new interface between contractors who have never worked together, and a new commissioning discipline that few teams have performed at scale. A facility being designed today has to accommodate power densities, electrical architectures, and power sources that did not exist when the last generation of standards was written.
This is precisely the environment where independent programme governance earns its keep. Not by choosing the technology, but by making sure the cost, schedule, and risk across every new work front stay under control while the ground keeps shifting. The tools change. The discipline does not.
PMO Hive's programme management approach is built on decades of delivering large, capital-intensive projects with complex, parallel work fronts, the same conditions now defining hyperscale. From cost governance and schedule risk analysis to commissioning readiness and EPC oversight, the Hive platform is designed to keep control when the technology underneath the programme is changing faster than the playbook.
The frontier keeps moving. The need to deliver it on time and on budget does not.
Related Posts
