The AI data centre is becoming a manufactured product. The question is whether that buys you any electricity.

The most expensive part of an AI data centre was never the chips. The chips are the line everyone watches — the Nvidia order, the allocation fight, the price per accelerator. But by the time a facility is running, the silicon is a minority of the bill and almost none of the delay. The cost and the wait live in the building: the power equipment, the cooling, the concrete, and the eighteen months to four years it takes to assemble all of it on a specific patch of ground. This week Schneider Electric and Foxconn announced a partnership aimed squarely at that part, and it is worth reading not as a press release about AI but as a statement about manufacturing. They want to stop building data centres one at a time and start producing them.

On 15 June the two companies said they would jointly develop modular power and cooling infrastructure for AI data centres, with production of the co-developed systems expected to begin later in 2026. Foxconn brings what Foxconn is — electronics manufacturing at scale, AI server production, system integration. Schneider brings the unglamorous half of a data centre that actually decides whether it runs: power distribution, cooling, and the energy management that ties them together. The plan is to turn that half into standardised, repeatable modules. The interesting question is not whether they can. It is what, in cost-per-kilowatt-hour terms, it actually buys.

What they are actually building

Strip the language down and the deliverable is a set of pre-engineered building blocks: modular power and cooling "skids" — factory-assembled units that arrive as a finished thing rather than a pallet of parts — plus reference architectures, which are standardised blueprints for how a given block of AI compute should be powered and cooled. The pitch is that a data-centre operator stops paying to design each site from scratch and instead orders a known configuration that has already been engineered, tested and costed. Schneider has sold prefabricated and modular data-centre products for years under its EcoStruxure line; what is new is pairing that with Foxconn's manufacturing volume and aiming the whole thing at the specific thermal problem of AI racks.

The logic is the logic of every industry that ever moved from bespoke to manufactured. A custom-built data centre is a construction project, with a construction project's overruns, site-specific engineering and one-off supply chains. A modular one is a product, with a bill of materials, a factory yield and a delivery date. If you are deploying the same kilowatts in twenty places, designing them once and stamping them out is obviously cheaper than designing them twenty times. The companies frame it as cutting the custom engineering required for each new site and deploying faster than traditional construction. That is real, and it is the boring kind of real that tends to actually deploy. But "faster and cheaper to build" is a claim about the building. It is not yet a claim about electricity.

The number that moves: time-to-power

Here is where this partnership does something measurable. The binding constraint on AI build-out, the one I have written about before, is not compute and increasingly not even capital. It is time-to-power — how long from decision to the moment the racks can actually draw the megawatts they need. In large markets that clock is set by things no server vendor controls: interconnection queues measured in years, and grid transformers with lead times that have stretched past the point of comedy. Against that, anything that compresses the schedule has value, because in a shortage the first facility online captures the demand and the second one negotiates with whatever is left.

Modular, factory-built power and cooling compresses one specific stretch of that schedule: the on-site assembly. Instead of building the electrical room and the cooling plant in place, you truck in skids that were assembled and commissioned in a factory in parallel with the site work. Done well, that takes months out of the construction timeline and removes a class of on-site error. For an operator racing a demand window, months are not a convenience; they are the difference between catching the curve and missing it. This is the part of the announcement that is straightforwardly good, and it is good precisely because it is unglamorous.

But compressing the building does not conjure the megawatts. This is the caveat that the word "modular" tends to bury. A factory-built power skid still has to connect to a grid that can deliver power, or to generation built on site, and that connection is the part of the timeline that modularity cannot touch. You can shorten the construction of the thing that uses the power. You cannot shorten the queue for the power itself by manufacturing the cabinet faster. The risk in stories like this one is that "deploy in months" gets heard as "powered in months," and those remain two different sentences with two different bottlenecks. The skid is ready. The interconnection still is not.

Modularity compresses the building, not the megawatts. You can stamp out the power cabinet on an assembly line. The queue for the power going into it does not get any shorter. — On what "deploy faster" actually shortens

The cooling is the part that touches the bill

If the power side of this is mostly about schedule, the cooling side is where it can actually move cost-per-kilowatt-hour — and that is the number that decides whether a data centre is a good business or a marginal one. Cooling is not a footnote in a facility's energy use; it is one of the largest non-compute draws, and in a conventional air-cooled hall it can add a substantial fraction on top of every watt the chips consume. The industry tracks this as PUE, power usage effectiveness: the ratio of total energy drawn to the energy that reaches the computers. A PUE of 1.5 means you spend fifty cents cooling and distributing for every dollar of useful compute. Drag that toward 1.1 and you have cut the overhead by most of itself, on every kilowatt-hour, for the life of the building.

The partnership names closed-loop energy optimisation and the kind of liquid cooling that AI racks now require — the power densities of modern accelerators have simply outrun what moving air can carry. Liquid cooling is more efficient at high density, and a closed loop reduces the water draw that has made some data centres unwelcome in the places that host them. Schneider's earlier work with Nvidia on "AI factory" reference designs in 2025 was pointed at exactly this: standardising the liquid-cooled, high-density block so operators are not reinventing the thermal design each time. The cost-first reading is simple. If the modular design ships a lower, more predictable PUE as a default rather than as a custom optimisation, that is a recurring saving on the operating bill, not a one-off saving on the build. The build is paid once. The PUE is paid every hour for ten or fifteen years. That is where the real money is, and it is the claim worth asking the companies to put a number on.

Where it deploys, and who pays

Standardisation has a second-order effect that the Silicon Valley framing usually misses. A reference architecture is portable. A bespoke data centre needs a bench of specialist engineers to design and commission it; a manufactured one needs a site, a connection and a delivery. That lowers the barrier to building serious compute in places that do not have deep data-centre engineering benches — which is most of the world. If the economics hold, the markets that have been priced out of frontier infrastructure by the cost of bespoke engineering are exactly the ones a product, rather than a project, could reach. Whether that happens depends entirely on where the power is cheap and available, because that, not the cabinet, is what now decides where these facilities land.

And someone pays for all of it. The operator pays for the modules; the operator's customers pay through the price of compute; and where these facilities draw on a shared grid, the public pays through what large new loads do to local power prices and to the queue of everyone else waiting to connect. Industrialising the build does not change that arithmetic — it just makes the build faster and, if the PUE claims hold, the running cost lower, which at the margin is better for everyone on the same grid than the alternative. The honest version of this announcement is therefore narrow and real: two companies are turning the expensive, slow, custom half of a data centre into a manufactured product, which should take time out of the schedule and, if the cooling delivers, cost out of the bill.

What it is not is a solution to the thing actually limiting AI, which remains the supply of power and the years it takes to connect to it. A manufactured data centre is a faster, cheaper box to put compute in. It is still waiting on the same grid as everyone else. The number to watch is not the deployment time the press release leads with; it is the PUE the modules ship with and the time-to-power at the sites they land on. Stamp out the box all you like. The electricity is still the constraint, and the electricity is still the bill.