A device that has been correct for a century

The line-frequency transformer does exactly one thing: it changes voltage by the ratio of its turns. It has no control loop, no firmware, no cooling fan. It is passive, extremely robust, and — measured against what it does — remarkably cheap. Any argument for replacing it has to begin by admitting that. The case is not that the transformer fails. The case is that everything around it has changed.

What changed is the load. A rack that drew five kilowatts now draws sixty. Loads that were uniformly AC are increasingly DC-native. And the constraint on a new facility is less often the energy supply than the floor area available for electrical equipment and the capacity the utility is willing to connect. When the binding constraint moves from cost to volume, the arithmetic changes.

Counting the conversions

Conventional data centre power passes through a chain: medium-voltage AC is stepped down to low-voltage AC by the transformer, converted to DC by the UPS rectifier, stored and re-inverted to AC at the UPS output, distributed as AC to the racks, and converted back to DC inside each server's power supply. That is three AC/DC conversions between medium-voltage AC and the DC the silicon actually consumes; the transformer is a fourth stage, but it only changes voltage.

Each conversion costs two to three percent. Multiplied across the chain, three stages at a nominal ninety-six percent each yield roughly eighty-eight percent, before the transformer, cabling and cooling are counted. At ten megawatts of IT load, every percentage point of system efficiency is worth hundreds of kilowatt-hours per day, continuously, for the life of the facility.

A DC link in the middle of the chain removes two of those three conversions. That is the substance of the argument, and it is independent of any question about the transformer itself.

Frequency is what replaces iron

A transformer's core area scales roughly inversely with the frequency it operates at. That single relationship is the whole reason a solid-state transformer can be small. Raise the link frequency from fifty hertz to twenty kilohertz — a factor of four hundred — and the isolating transformer shrinks from a device occupying a room to one occupying a shelf.

What makes the high frequency possible is the switching device. Silicon carbide MOSFETs tolerate higher voltages, higher temperatures, and higher switching frequencies than silicon IGBTs, with lower switching loss. Without wide-bandgap devices, the topology exists on paper but does not pay for itself; with them, the trade becomes favourable.

It is worth being precise about the sequence of benefits. The transformer gets smaller first, because physics forces it. The control benefits come afterwards, because the conversion is now done by electronics that can be told what to do.

What the architecture does not solve

Replacing several passive devices with one active device concentrates failure modes. A transformer fails loudly and rarely; a power-electronic converter has more ways to fail, and some of them are silent. Redundancy therefore has to be architectural: if the facility requires N plus one, that has to come from parallel units or a designed bypass path, not from a hope that the device is reliable enough.

Protection philosophy changes as well. A power-electronic converter limits its fault current deliberately, which means the protection scheme upstream cannot assume the fault currents it was designed around. This is a design task, not a commissioning detail.

And the certification path differs. Medium-voltage power-electronic conversion equipment is assessed under different standards from conventional switchgear, and that assessment takes time. A project that plans for it late will find the equipment ready before the paperwork.

Where it pays, and where it does not

The useful test is whether the conventional chain is actually binding. If floor area is scarce, if conversion losses are a material operating cost, or if the site needs to control power flow in ways a fixed-ratio device cannot — the architecture is worth evaluating on its numbers.

If none of those are true, the conventional chain is likely the cheaper answer, and a supplier who tells you otherwise is selling architecture rather than solving a problem. There is no virtue in a more modern topology that costs more and solves nothing.

What each one actually does

A grid-following converter treats the grid as a reference. It measures the voltage waveform, locks onto its phase with a phase-locked loop, and injects current in step with it. It behaves, electrically, like a current source that has been told what current to produce. Almost all grid-tied solar inverters and most storage PCS units operate this way, and on a strong grid they work extremely well.

A grid-forming converter does the opposite. It synthesises its own voltage and frequency reference and behaves like a voltage source behind a defined impedance. It does not need to be told what the grid is doing; it defines what the local grid is. This is what allows a converter to operate when there is no grid to follow.

The distinction is often introduced as an off-grid topic. It is not. It matters on any connection where the grid is weak, where a high proportion of generation is inverter-based, or where the site has to ride through a disturbance.

Where grid-following runs into trouble

The phase-locked loop is the weak point, and its weakness is proportional to how weak the grid is. A useful measure is the short-circuit ratio: the ratio of available fault current to the rated current of the converter. When that ratio falls — because the grid is remote, or because inverters have displaced the synchronous machines that used to supply fault current — the phase-locked loop has less to lock onto, and stability margins narrow.

In a system where many inverters are following the same weakening grid, the problem compounds. Units that individually behave correctly can interact through the grid impedance and oscillate together at frequencies that no single unit's control was tuned for. This is a system-level phenomenon, and it is not resolved by improving one converter.

The practical consequence is that as inverter-based generation grows, the grid-following assumption — that there is always a strong reference to follow — becomes less safe to rely on.

Fault current is the part nobody plans for

A synchronous machine delivers several times its rated current into a fault, which is exactly what protection relays are built to detect. A grid-following inverter does not. It is a current-limited source: it will hold its output close to rated current and no more, because exceeding its rating would destroy it. To a protection relay set for conventional fault levels, a fault fed only by inverters can look like a mild overload.

A grid-forming converter can be designed to supply a defined fault current contribution — a multiple of rated current for a defined duration — because it is controlling voltage rather than current. That makes protection coordination possible in a way it is not with current-limited sources alone.

This is why the question is asked at design stage and not later. Protection settings, cable sizing, and the choice of converter control all depend on the answer, and changing the answer after the equipment is ordered means changing the equipment.

Inertia, and what it is a proxy for

Rotating machines store kinetic energy in their mass, which resists sudden frequency change. As synchronous generation is displaced by inverters, that natural inertia falls, and the rate of change of frequency after a disturbance rises. Grid-forming converters can emulate inertia by releasing stored DC-link energy in proportion to the frequency derivative, which slows the fall.

The honest caveat is that emulated inertia is bounded by stored energy. It can slow a frequency excursion; it cannot substitute indefinitely for the kinetic energy a large machine would have supplied. Where the requirement is a defined ride-through duration, that has to be stated as a number and matched to the DC-link capacitance, not assumed as a general property.

What to ask a supplier

Ask what the converter does when the grid disappears, not what it is called. Ask for the fault current contribution in multiples of rated current and for how long. Ask whether the unit can form a reference with no other source present, and under what conditions it hands back to the generator or the grid. Ask what the control does when two units disagree about frequency.

A supplier who answers these in numbers is describing a product. A supplier who answers them in adjectives is describing a brochure.

The datasheet number has a condition attached

A lithium iron phosphate cell advertised at six thousand cycles is not claiming six thousand cycles under all conditions. It is claiming six thousand cycles at a stated depth of discharge, a stated temperature, a stated charge and discharge rate, and a stated end-of-life criterion — typically a residual capacity of eighty percent of beginning of life. Remove any one of those conditions and the number stops being valid.

This is not a trick. It is how cycle life is defined and tested. The difficulty is that the datasheet number travels into proposals and business cases on its own, detached from the conditions it depends on.

Depth of discharge and cycle life are not linearly related

For a typical LFP cell, cycling to a hundred percent depth of discharge might yield something in the region of three to four thousand cycles, while cycling to eighty percent might yield five to six thousand, and to sixty percent considerably more. The relationship is not proportional: each additional percentage point of depth costs more life than the one before it.

The practical consequence is that two systems with identical nameplate capacity can deliver very different total energy throughput over their lives, depending on where the operating window is set. Since throughput is what the business case is actually paying for, the operating window is a financial parameter, not a technical detail.

A system designed to squeeze out maximum daily throughput will cycle deeper and replace sooner. A system designed for longevity cycles shallower and delivers fewer kilowatt-hours per day from the same hardware. Neither is wrong; they are different products being sold under the same capacity label.

Calendar ageing runs in parallel, not in sequence

Cycle ageing is only half of the picture. A cell also ages with time, whether or not it is being cycled, and calendar ageing proceeds faster at higher state of charge and higher temperature. A battery held at full charge in a warm room loses capacity to calendar ageing even if it never cycles.

This matters for residential systems in particular, where the duty cycle is often light. A household that consumes little may cycle the battery only partially each day, yet still lose several percent of capacity a year to calendar ageing alone. The proposal that models purely on cycles will overstate the life of exactly the systems that cycle least.

The honest way to present this is a life estimate bounded by both mechanisms, with the temperature and state-of-charge assumptions stated. A single cycle count is not an estimate; it is a component of one.

Temperature compounds both

Temperature accelerates both cycle ageing and calendar ageing, and the effect is steeper than most people assume. Roughly speaking, each ten degrees of additional cell temperature can halve the calendar life, and it also degrades cycle life. The two effects apply at the same time, so a hot cell does not merely age twice as fast as a cool one in a single respect — it ages faster along two axes at once.

This is why installation location is an electrical parameter. A ventilated utility room and an unheated summer garage differ by far more than ten degrees for much of the year. The same pack, the same duty cycle, and the same capacity can produce very different service lives purely because of where the cabinet was bolted.

A design that does not ask about the installation location is not a design that has accounted for temperature. It has just left it out.

Reserve bands are designed, not leftover

Not all of a battery's capacity should be available for daily cycling. A band at the top is worth reserving so that unexpected solar generation can be absorbed rather than curtailed, and a band at the bottom is worth reserving so that backup capability remains available when it is needed. Both reduce daily throughput and both extend life.

The point is that these bands should be chosen deliberately, as part of the design, rather than left to whatever the inverter's built-in algorithm happens to do. A default state-of-charge window is a decision that was made by someone, for someone; it is worth knowing who and for what.

What to put in a contract

The measurable commitments are operating window, temperature, and end-of-life criterion. A warranty that guarantees a capacity percentage without stating the state-of-charge window and the temperature range it applies to is guaranteeing something the supplier cannot control and the buyer cannot verify.

The better arrangement states all three, and states what happens when the site's measured temperature or duty cycle falls outside the assumed range. That is not a hostile question. It is the question that separates a warranty designed to be claimed from one designed never to be paid.