Editorial Article
The Data Centre UPS Is Losing Both of Its Jobs, and It Is Being Paid Less for the Ones Left
There is a single line diagram behind every data hall, and on it the uninterruptible power supply is the only device the load cannot route around. Switchgear can be bypassed. Cooling can be economised, and increasingly it can be moved to a dry cooler in a car park. Generators can be paralleled into a bus. The UPS sits in the path, and if it fails, the hall stops. That is not a technical detail. It is the reason the product has been bought on specification rather than price for thirty years, sold with a service contract, and replaced reluctantly.
It is also the reason the next eight years are going to be awkward for the people who make it.
Two rates, and the gap between them
The global data centre UPS market is worth USD 7.40 billion in 2025 and reaches USD 15.30 billion in 2033. That is a compound annual growth rate of 9.5 per cent, and USD 7.90 billion of new annual revenue.
Now the same market measured in what it physically handles. Kilovolt-amperes shipped rise from 17.5 GVA to 42.3 GVA over the same eight years — 11.7 per cent a year.
Those two numbers describe one market and they are not the same number. The 220 basis point gap is realised price per kVA falling from USD 424 to USD 361. A supplier in this market can grow its output by eleven per cent and its revenue by nine. That is not a rounding difference. It is the difference between a business that compounds and a business that runs to stand still, and it is the first thing anybody planning capacity in this industry has to internalise.
Why the price falls is not mysterious, and none of it is cyclical.
The first force is density. A single modern rack draws 132 kW in NVIDIA’s GB200 NVL72 reference design. The same total load now sits in a fraction of the floor space and needs a fraction of the enclosures, because one 2 MW frame replaces twenty 100 kW boxes. Fewer breakers, fewer terminations, fewer engineering hours, fewer metres of copper. The megawatt is unchanged. The bill of materials wrapped around it is not.
The second is the collapse of the middle of the market. The 20 kVA three-phase box sold to every enterprise server room carried the highest price per kVA in the industry, because it was specified by someone who did not buy them often and installed by someone who charged by the day. That segment is being eaten from both ends — downward by battery backup units inside the server power shelf, sideways by colocation, where the tenant buys a resilient feed instead of a device. The high-margin middle is going, and the blended realisation goes with it.
The third is Chinese supply. Huawei is a tier-one vendor everywhere it is permitted to sell, and behind it sit Kehua, Kstar and East Group, all of whom price a double-conversion frame well below the Western three. In any procurement not constrained by politics, that is the reference price. Which means the same product sells at two very different prices depending on where the buyer is standing — a market with a politically segmented price structure is not one market, and anyone planning capacity on a blended global price is planning on a number nobody actually transacts at.
What the box is bought for
To see where this goes, it helps to be precise about what a UPS does. IEC 62040-3 classifies static systems on three axes. The first is input dependency, and its top grade is voltage and frequency independent — VFI, double conversion. The UPS rectifies incoming AC to DC, holds it on a battery bus, and inverts it back into a clean sine wave. The load never sees the grid at all. It is protected against all seven classes of power problem, including frequency variation and waveform distortion.
That is an excellent engineering answer to a problem that has largely stopped existing.
A modern server power supply rectifies to DC within the first few centimetres of the chassis. It takes a wide input range. It does not care about phase. It has its own power-factor correction stage. The harmonics and the waveform distortion that VFI exists to remove arrive at a component that was going to remove them anyway. What a data centre still buys from a double-conversion UPS is not conditioning. It is ride-through — the thirty seconds to two minutes that bridges the gap between a utility event and a generator reaching speed.
And ride-through is moving.
In August 2026 NVIDIA published an 800-volt direct current architecture, and it is worth reading as a schedule rather than as a specification. Three stages. A hybrid 800 VDC power rack in the second half of 2026, which slots into existing AC infrastructure and delivers DC to compute racks within the row — purely additive, the building untouched, the facility UPS keeps its job. A row power centre in 2027, an overhead 800 VDC busway supporting up to 2 MW per row, which starts to displace the distribution downstream of the UPS. And a DC power block: a facility-scale unit that converts grid power directly to 800 VDC in a single step, enabling direct medium-voltage conversion at scale.
Read the third one twice. A facility that takes medium-voltage AC and produces 800 VDC in one stage has no place in its single-line diagram for an AC double-conversion UPS. The rectifier has moved into the facility transformer. The inverter has no job, because the load wants DC.
The honest qualifications matter. This is an explicitly decade-long transition. The installed base of AC halls is enormous and will be maintained for twenty years. Every hybrid product in the roadmap exists to preserve infrastructure already paid for. And no published DC architecture yet replaces the one thing a facility UPS actually sells, which is a documented, testable availability guarantee. But direction counts in capital allocation even when timing does not, and the direction is now published by the single most influential buyer-designer in the industry rather than sketched on a vendor’s slide. A supplier building ten-year capacity on the assumption that the AC double-conversion frame is the permanent shape of data centre power is building for a market its largest customer has already announced a plan to leave.
The rule that does not apply
Here is a fact that ought to be better known. A UPS converts power continuously, every hour of every day, for the life of a data hall, and loses three to four per cent of everything that passes through it. Across the global fleet that is one of the largest single pools of avoidable electricity consumption in the industrial economy. And Europe’s product efficiency regulation for this family of equipment does not apply to it.
The Commission Regulation on ecodesign requirements for external power supplies and related products, adopted on 13 October 2025 and repealing Regulation (EU) 2019/1782, says so at Article 1(2)(a): the regulation does not apply to “uninterruptible power supplies, meaning devices that automatically provide backup power from storage when the electrical power from the mains power source drops to an unacceptable voltage level”.
The exclusion is defensible. A UPS is not a passive adapter. Its efficiency depends on the load profile, the redundancy configuration, whether eco mode is enabled and a duty cycle nothing like the constant load of a phone charger. Regulating it properly would need a bespoke methodology. The consequence, though, is that no such methodology exists in binding form, and this market has never had an efficiency floor imposed on it.
Europe regulates the building instead. Directive (EU) 2023/1791, Article 26, and Delegated Regulation (EU) 2024/1364 capture power usage effectiveness — total facility power divided by IT power. The UPS loss is inside that ratio. So the operator is measured on a number that contains the UPS’s inefficiency, while no instrument, label or threshold tells it what a good UPS looks like. The product is a loss term in somebody else’s efficiency metric.
What that is worth: across 100 MW of IT load, five points of efficiency is 5 MW of continuous loss. That is 43.8 GWh a year, which at USD 100/MWh is roughly USD 4.4 million a year, every year, for the life of the hall. It is a purely commercial decision, and no regulator requires anybody to take it. The hyperscalers take it anyway, because they can measure it. Most of the rest of the market does not, and that gap is where the incumbents earn their margin.
The annuity that is quietly being dismantled
The most underweighted risk in this market is not architectural. It is a chemistry change, and it is converting recurring revenue into capex.
On Schneider Electric’s own published comparison, valve-regulated lead-acid runs a three to six year service life against 10 to 15 years for lithium. Maintenance is budgeted at roughly 8 per cent of capex a year for VRLA against 1.5 per cent for lithium. Steady-state trickle losses are 0.2 per cent against 0.1 per cent. Over a ten-year window that is two battery replacements for lead-acid and none for lithium.
Read that as a supplier and it is uncomfortable. The lithium UPS is a better product, a larger initial sale, and a smaller lifetime relationship. Whoever sells the cells captures the replacement revenue that used to accrue to whoever sold the frame. And the recurring revenue shrinks as a share of the total at exactly the moment the field engineering network — thirty countries, service vans, spare modules — needs more funding, not less.
The industry’s answer will be to substitute monitoring, remote diagnostics and availability guarantees for the lost battery annuity. Whether a software subscription can carry the margin of a battery replacement is the open question in this business, and nobody has demonstrated it at scale.
The tell is in the acquisitions
Where the incumbents have spent their M&A budgets says more about their beliefs than any investor deck. Eaton’s thermal position was assembled by buying Boyd. Schneider’s was built by taking a controlling stake in Motivair. Both are thermal. Neither is power.
That is a revealed preference, and it is worth being blunt about what it reveals. The companies with the best information about the data hall have bought the half of it that still requires fluid dynamics, engineering complexity and a service engineer, and have not bought the half where the conversion stage is being designed out of the single-line diagram. It is not a prediction that the UPS disappears. It is a statement about which half of the hall the people who can see the order book think is worth owning in 2035.
What survives, and what to watch
Ride-through survives, because physics requires stored energy between the grid and the load. What changes is where it sits. Some of it moves down into the rack, as battery backup units inside the power shelf holding the load for the milliseconds a downstream transfer takes. Some moves up into the facility, as on-site battery storage that performs the UPS function and simultaneously earns money on the grid through arbitrage and demand response — a far easier business case to approve than a battery that sits at float for a decade waiting for an outage.
What also survives, and is badly underrated, is the transfer function itself. Somebody has to detect the event and move the load, reliably, every time, for twenty years, and be able to prove it afterwards to an insurer. That is a controls and switchgear problem rather than a power-conversion problem, and it is where the incumbents’ engineering depth is genuinely hard to copy. The frame is commoditising. The assurance is not.
The market will be bigger in 2033 than it is today, by nine per cent a year. But it will be worth less per kilovolt-ampere, the product will be doing less of what it was bought for in 2020, and the recurring revenue that funded the service network will be smaller. The suppliers who understand that are already repositioning around integration, prefabrication and assurance. Eaton’s data centre orders in the fourth quarter of 2025 were up approximately 200 per cent year on year and its Electrical Americas backlog up 31 per cent; Schneider’s Systems line, a third of group revenue, grew 19 per cent organically on prefabricated data centre offers. Those are not frame sales. They are the sound of a supply base quietly changing what it sells.
The full sizing, construction, value chain, technology, competitive, regulatory, investment and risk analysis is in the Global Data Center Uninterruptible Power Supply (UPS) Market Report, Size, Share and Forecast 2026–2033.
Frequently Asked Questions
Why does the data centre UPS market grow at 9.5 per cent when data centre power demand is growing at 15 per cent?
Because volume and value are separating. Pheonix models kVA shipped growing at 11.7 per cent a year, from 17.5 GVA to 42.3 GVA, while realised price per kVA falls about 2 per cent a year, from USD 424 to USD 361. Eleven point seven per cent less two per cent is 9.5 per cent. Three forces push price down: rising power density, which means one 2 MW frame replaces twenty 100 kW boxes; the collapse of the small and medium three-phase segment, which carried the highest price per kVA; and Chinese supply setting a reference price in every market it can reach.
Is 800 VDC actually a threat to the AC uninterruptible power supply?
In direction, yes; in timing, not immediately. NVIDIA's published architecture stages a hybrid 800 VDC power rack in the second half of 2026 that leaves building AC distribution untouched, a row power centre in 2027 using an overhead busway up to 2 MW per row, and then a facility-scale DC power block that converts grid power directly to 800 VDC in a single step. The third stage removes the AC double-conversion UPS from the facility's single-line diagram, because the rectifier moves into the facility transformer and the inverter has no job when the load wants DC. The transition is explicitly a decade long and the installed base of AC halls will be maintained long after, so the realistic outcome is a split market rather than the disappearance of the product.
Why is the UPS market not covered by EU energy efficiency rules?
Because the Commission Regulation on ecodesign requirements for external power supplies and related products, adopted 13 October 2025 and repealing Regulation 2019/1782, expressly excludes uninterruptible power supplies at Article 1(2)(a). The exclusion is technically defensible — UPS efficiency depends on load profile, redundancy configuration and duty cycle in ways an external power supply's does not — but it means no binding efficiency methodology exists for the product. Europe regulates the data centre instead, through Directive (EU) 2023/1791 Article 26 and Delegated Regulation (EU) 2024/1364 on reporting, both of which capture power usage effectiveness. The UPS loss sits inside that ratio, so the operator is measured on a number containing the product's inefficiency while nothing tells it what a good UPS looks like.
How much of the UPS market is replacement rather than new build?
By Pheonix's construction, roughly 42 per cent of the kVA shipped in 2025 replaces existing capacity rather than adding to it. The installed UPS base is modelled at about 87 GVA in 2025; at a twelve-year average service life, replacement accounts for 8.3 per cent of that base each year while new capacity adds about 11.7 per cent. That replacement floor is why the market is unlikely to contract in absolute terms even during a hyperscale capital spending pause — the replacement cycle is driven by the age of the installed fleet rather than by anyone's capital plan.
