Global Data Center Uninterruptible Power Supply (UPS) Market Report, Size, Share and Forecast 2026–2033: Eleven Per Cent More kVA and Nine Per Cent More Money
What this report measures
A data centre uninterruptible power supply is a serial device. Power enters the building, passes through it, and only then reaches anything that computes. Nothing else in the hall has that property. Switchgear can be bypassed, cooling can be economised, generators can be paralleled into a bus — but the UPS sits in the path, and if it fails, the hall stops. That is why, for thirty years, it has been bought on specification rather than price, sold with a service contract, and replaced reluctantly.
This report sizes the global market for that device and forecasts it to 2033. Its argument is not that the market is small or shrinking. It is that the two jobs the device is bought for are both being taken away from it — one by a change in power architecture, one by a change in battery chemistry — while the volume it handles continues to grow. The money and the kVA therefore separate, and by 2033 they have separated by a fifth.
Sized from the electricity up
There is no audited anchor for this market. Not one participant discloses UPS revenue as a reportable segment: Schneider Electric reports Energy Management, Eaton reports Electrical Americas, Vertiv reports by geography and by product-and-service line but never UPS alone. Any figure for this market, including the ones produced by large research houses, is a construction. What differs between constructions is whether the assumptions are shown. Ours are below, and every step is reproducible.
| Step | Input | Value | Source or disclosure |
|---|---|---|---|
| 1 | Data centre electricity consumption, 2024 | 415 TWh | IEA, Energy and AI, base case; about 1.5% of global electricity, grown 12% a year over five years |
| 2 | Implied average draw, 2024 | 47.4 GW | 415 TWh / 8,760 hours |
| 3 | Consumption, 2025 | 477 TWh, 54.5 GW | IEA base case growth of about 15% a year to 2030 |
| 4 | Blended power usage effectiveness | 1.40 | Assumption. Hyperscale runs near 1.15–1.20; enterprise and colocation 1.5–1.8. Range tested 1.25–1.55 |
| 5 | Average IT load, 2025 | 38.9 GW | 54.5 / 1.40 |
| 6 | Average utilisation of installed IT capacity | 62% | Assumption. Range tested 55–75% |
| 7 | Installed IT capacity, 2025 | 62.8 GW | 38.9 / 0.62 |
| 8 | UPS sizing factor | 1.25× at 0.9 power factor | Assumption. Reflects N+1 redundancy typical of Tier III design |
| 9 | Installed UPS base, 2025 | 87.2 GVA | 62.8 × 1.25 / 0.9 |
| 10 | Average service life | 12 years | Assumption. Replacement is 1/12 of the installed base each year |
| 11 | kVA shipped, 2025 | 17.5 GVA | Growth in the base (11.7%) plus replacement (8.3%), applied to 87.2 GVA |
| 12 | Realised price per kVA | USD 424 | Assumption: module, battery, switchgear, controls, excluding installation labour |
| 13 | Market value, 2025 | USD 7.40 billion | 17.5 GVA × USD 424 |
| 14 | Data centre electricity, 2030 | 945 TWh | IEA base case, about 3% of global electricity |
| 15 | Extrapolation to 2033 | 1,157 TWh | Assumption: 7% a year after 2030, roughly half the 2024–30 rate. The IEA base case stops at 2030 |
| 16 | Installed UPS base, 2033 | 211 GVA | Steps 4–8 applied to 1,157 TWh |
| 17 | kVA shipped, 2033 | 42.3 GVA | Growth in the base plus replacement |
| 18 | Realised price per kVA, 2033 | USD 361 | Assumption: minus 2% a year on efficiency, power density and Chinese competition |
| 19 | Market value, 2033 | USD 15.30 billion | 42.3 GVA × USD 361 |
Result. USD 7.40 billion in 2025, USD 15.30 billion in 2033, a 9.5 per cent CAGR and USD 7.90 billion of new annual revenue. kVA shipped grows at 11.7 per cent a year over the same window.
Sensitivity. The 2033 figure is most sensitive to price per kVA and to the post-2030 electricity path. Holding price flat at USD 424 gives USD 17.9 billion. Letting price fall 4 per cent a year gives USD 12.9 billion. The band is therefore USD 12.9 billion to USD 17.9 billion around a USD 15.30 billion midpoint. The growth rate is far better anchored than the level, because it follows from the electricity series rather than from the price parameter.
Why the kVA outruns the money
Two rates describe this market and they are not the same rate. Installed capacity, and therefore the kVA a supplier can ship, follows the electricity curve at 11.7 per cent a year. Revenue follows it at 9.5 per cent, because the price of handling a kVA is falling. Three things push it down, and they are structural rather than cyclical.
The first is power density. A single modern rack draws 132 kW in NVIDIA's GB200 NVL72 reference design. The same total load now occupies a fraction of the floor space and a fraction of the UPS units it once needed, because one 2 MW frame replaces twenty 100 kW boxes. Fewer enclosures, fewer breakers, fewer terminations, fewer engineering hours per megawatt. The megawatt is the same; the bill of materials around it is not.
The second is the collapse of the small and medium three-phase segment. The product that used to be sold to every enterprise server room is being replaced by rack-mounted battery backup inside the power shelf, and by colocation, where the tenant buys a resilient feed rather than a box. That segment carried a materially higher price per kVA than hyperscale multi-megawatt systems, and its disappearance drags the blended realisation down.
The third is Chinese supply. Huawei is a top-tier supplier everywhere it is permitted to sell, and behind it sit Kehua, Kstar and East Group, all of which price a double-conversion frame well below the Western three. In markets where procurement is not politically constrained, that is the reference price.
Where the demand comes from, and how concentrated it is
The demand side of this market is now four balance sheets. Microsoft's capital expenditure forecast for the year stands at USD 170 billion, reduced from USD 175 billion on third-party accelerator delays that push some capacity into the first half of 2027. Alphabet held guidance at USD 195 to 205 billion. Meta guided USD 135 to 145 billion. On the reported figures, the four largest buyers are on track to spend between USD 803 billion and USD 828 billion in 2026.
Not all of that is a data hall, and not all of a data hall is a UPS. But the UPS is ordered against the hall, which means two things follow. First, the demand signal is unusually legible: capital expenditure guidance published in October is a UPS order book in the following year. Second, it is unusually fragile: a delay of two quarters in accelerator supply becomes a delay of two quarters in power equipment, and the equipment vendor carries the inventory. Microsoft's own explanation of its capex trim is precisely that mechanism, and Eaton's data centre orders rising approximately 200 per cent year on year in the fourth quarter of 2025 shows what the same mechanism looks like when it runs the other way.
Segmentation
By topology. Standby and line-interactive units serve edge and enterprise loads; double conversion, the VFI class in IEC 62040-3, serves essentially all of the data centre market that matters for value. The value is overwhelmingly in the third category, and the first two are where volume growth in unit terms comes from.
By capacity. The economically significant divide is not small against large but modular against monolithic. Prefabricated and modular systems, assembled and tested off site and delivered as a skid, are taking share from stick-built rooms because they compress the construction schedule — and schedule is the binding constraint on a data centre programme.
By end use. Hyperscale and wholesale colocation, retail colocation, enterprise, and the edge. Hyperscale dominates value; enterprise dominates unit count and is shrinking in value.
By geography. We do not publish a regional percentage split for this market, because no primary series supports one at UPS level. The single observed geographic data point is Vertiv's disclosure that 62 per cent of its USD 10.23 billion of 2025 net sales was transacted in the Americas — a genuine signal for where the spend is, but a reflection of one company's footprint rather than a market split. The directional facts are not in doubt: the United States hosts the largest concentration of installed data centre capacity, China is the second and the largest single growth market outside it, and Europe's share of new capacity is falling even as its regulatory apparatus for data centres becomes the world's most developed.
What this market is not
It is not the data centre power market. Switchgear, busway, rack power distribution and generators sit outside this perimeter. On this site the adjacent thermal plant — USD 18.7 billion for data centre cooling in 2025, with liquid cooling separately sized at USD 4.30 billion — is roughly two and a half times the size of the UPS market, and the ratio is the right way to read the two together: a UPS is one component of a power train, while the cooling figure covers the whole thermal chain from chiller to cold plate. A supplier that reads the UPS number as the size of the power opportunity will understate its own addressable market by a factor of three or more.
It is also not a growth market in the sense the demand figures suggest. The electricity consumed by data centres doubles by 2030. The dollars earned by the companies that keep those data centres energised do not.
Table of Contents
Competitive Landscape
Structure
The market is moderately consolidated at the top and long-tailed beneath. Three Western vendors — Schneider Electric through APC, Vertiv through Liebert, and Eaton — plus Huawei hold the majority of global value, with ABB, Siemens, Mitsubishi Electric, Delta Electronics and Legrand filling out the tier below and a large Chinese and regional tail underneath them. Concentration is highest in hyperscale, where a handful of framework agreements decide multi-year volume, and lowest at the edge, where the buyer is a distributor and the product is a box.
Competitive intensity is high, and it is high in an unusual place: not in the rectifier, where the top four are technically level, but in three adjacent competitions — service attach, prefabrication capacity, and supply security. A buyer choosing a UPS in 2026 is choosing a delivery date and a network of field engineers at least as much as an efficiency figure.
The three Western incumbents, read from their own disclosures
Vertiv is the only pure play, and therefore the closest thing to a read on this segment. Net sales for 2025 were USD 10,229.9 million, with 62 per cent transacted in the Americas. Its portfolio is broader than UPS — thermal, busbar, switchgear, racks, services — but all of it sells into the same purchase order as the UPS, and the shape of the company’s growth is the shape of the data hall’s bill of materials. That 62 per cent Americas figure is the most useful geographic data point available in this market, and it is a company footprint rather than a market split.
Eaton reported record 2025 sales of USD 27.4 billion, up 10 per cent, with segment margins of 24.5 per cent. The relevant detail is in the fourth quarter: Electrical Americas sales of USD 3.5 billion, up 21 per cent, with data centre orders up approximately 200 per cent and data centre revenue up approximately 40 per cent year on year. Electrical sector backlog was up 29 per cent, and Electrical Americas backlog up 31 per cent. Backlog growth of that order means the constraint on this business in 2026 is manufacturing capacity, not demand.
Schneider Electric crossed EUR 40 billion for the first time in 2025 at EUR 40,152 million, up 8.9 per cent organic, with Energy Management up 10 per cent organic and led by data centre. Systems — 34 per cent of group revenue — grew 19 per cent organic in the fourth quarter, and the company specifically attributes the strength to prefabricated data centre offers. That is the integration layer turning into the growth engine, and it is a strategic choice: Schneider is competing on how the hall is assembled rather than on the efficiency of the frame.
Huawei is the price setter outside the markets it is excluded from, and its exclusion from the United States and much of Europe is itself a structural feature of the market. The practical consequence is two price levels for the same class of product, and a Western vendor base protected from the deflation Chinese supply would otherwise have imposed.
The tell is in the acquisitions
Where these companies have spent their M&A budgets says more about their view of the future than anything in their investor decks. Eaton’s thermal position was assembled by acquiring Boyd. Schneider’s was built by taking a controlling stake in Motivair. Neither bought a UPS business, because neither needed one — but both were willing to pay strategic multiples for thermal assets.
Read that as a statement of relative belief. The incumbents are consolidating the part of the data hall where engineering complexity, fluid dynamics and service attach still defend a margin, and they are not consolidating the part where a double-conversion frame is being designed out of the facility’s single-line diagram. That is not a prediction that the UPS disappears. It is a revealed preference about which half of the hall the people with the most information think is worth owning in 2035.
What competition looks like through 2033
Service attach on the installed base. This is the profit pool, and it is under threat from the chemistry change: a lithium battery does not need replacing twice a decade, so the annuity that funds a field engineering network shrinks with every frame that ships with lithium instead of lead-acid. Vendors will respond by bundling monitoring, remote diagnostics and availability guarantees into subscription contracts. Whether that replaces the lost battery revenue at the same margin is the single most important open question in the industry.
Prefabrication capacity. Factory floor space, not engineering, decides who can serve hyperscale in 2027. This favours the large diversified vendors and disadvantages specialists.
DC-native product. Whoever sells the first credible facility-scale DC power block with the same availability guarantees as an AC UPS takes the topology transition rather than being taken by it. The incumbents have the switchgear and controls capability to do it. Whether they have the willingness to cannibalise their own frame business is a different question, and it is the one that decides which of them is still growing in 2033.
Value Chain
The chain, and where the margin sits in it
Six positions, and value concentrates at the two ends rather than in the middle where the hardware is made.
Semiconductor. The insulated-gate bipolar transistor and, increasingly, the silicon-carbide module are the cost and performance drivers. A double-conversion frame is a rectifier, a DC bus, an inverter and a controller; the switching devices set the efficiency ceiling and the price. This layer is not specialised to UPS at all — the same parts go into drives, solar inverters and electric vehicles — so the UPS industry is a price taker in its own most important input, and periodically a supply-constrained one.
Magnetics and assembly. Transformers, inductors, busbars and boards. Copper is the volatility here, and a multi-megawatt frame is a substantial copper purchase. Assembly is increasingly contract-manufactured, and increasingly in Southeast Asia and Eastern Europe rather than in the vendor’s home market.
Module manufacture. Where the three Western incumbents and Huawei actually compete. A scale business with a real learning curve, and the layer at which price per kVA is set.
System integration. Adding switchgear, breakers, maintenance bypass, battery, monitoring and controls into something that can be commissioned. This is where the gross margin is, because it is engineering rather than manufacturing — and it is the layer prefabrication is reorganising. Schneider’s Systems line grew 19 per cent organically in the fourth quarter of 2025 with prefabricated data centre offers specifically called out, on a base that is 34 per cent of group revenue. That is the integration layer becoming a construction business.
EPC, commissioning and service. Installation, testing under load, and then a maintenance contract for the following decade or two. Service is the highest-margin revenue in the chain and the most defensible, because it is attached to the installed base and requires local presence.
The operator. Hyperscale, colocation, enterprise. In 2026 the four largest operators set the terms.
The battery is a separate chain, and it is changing hands
The energy storage inside a UPS is not made by the UPS vendor in most cases. Valve-regulated lead-acid cells come from a small set of manufacturers, and lithium iron phosphate cells come from the same supply chain that serves vehicles and grid storage. That shift moves the bill of materials toward the cell supplier and away from the UPS vendor, and it moves the vendor’s revenue from an annuity to a one-off.
The economics are stark and they are disclosed. On Schneider’s own published comparison, VRLA runs a three to six year service life against 10 to 15 for lithium; maintenance is budgeted at roughly 8 per cent of capex 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. 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.
Three concentrations worth naming
Switching devices. Silicon-carbide capacity is tight and allocated. A vendor without a supply agreement is a vendor without a delivery date, and in a market where the customer is building to a schedule, the delivery date is the product.
Commissioning labour. The binding constraint on a data centre programme in 2026 is not equipment, it is the electricians and commissioning engineers required to energise it. Prefabrication exists to convert site labour into factory labour. Its adoption is a labour-market story before it is a technology story.
Single-vendor exposure to China. Huawei’s presence is large outside the United States and parts of Europe, and its absence from those markets means the same product is sold at two very different prices depending on where the buyer stands. A market with a politically segmented price structure is not a single market, and any participant planning capacity on a blended global price is planning on a number nobody transacts at.
Investment Activity
The demand-side number that matters
Investment in this market is not really investment in UPS. It is investment in data centre capacity, of which the UPS is a small and late-ordered fraction. So the number to watch is capital expenditure at the four companies that build the halls.
For 2026, Microsoft has guided capital expenditure of USD 170 billion, reduced from USD 175 billion after third-party accelerator delays pushed some capacity into the first half of 2027. Alphabet held guidance at USD 195 to 205 billion. Meta guided USD 135 to 145 billion. On reported figures the four largest buyers together are on track to spend between USD 803 billion and USD 828 billion. There is no prior period in the history of industrial capital spending that resembles this, and it is the entire demand-side case for this market.
The lag structure matters more than the total. A data centre programme orders its UPS roughly twelve to eighteen months before energisation, and energisation follows the IT equipment. When Microsoft trims capex because chips are late, the UPS order moves too — but the vendor has already committed the factory slot. That asymmetry, between a demand signal published quarterly and a supply chain that plans annually, is where inventory builds and where price concessions are given.
What the supply side is spending money on
Vertiv’s 2025 net sales of USD 10,229.9 million, with 62 per cent in the Americas, describe a company that has grown faster than the electrical equipment market for several years and has converted that into capacity. Eaton’s record USD 27.4 billion with Electrical sector backlog up 29 per cent and Electrical Americas backlog up 31 per cent describes a company whose problem in 2026 is factory throughput.
Schneider’s EUR 40,152 million, with Systems at 34 per cent of revenue and growing 19 per cent organic in the fourth quarter, describes the strategic bet: not more frames, but more prefabricated power trains arriving complete.
The acquisitions point the same way. Eaton’s thermal position came from Boyd; Schneider’s from Motivair. Both are thermal. Neither is power. The companies with the best information about the data hall have spent their acquisition currency on the half of it that still requires fluid dynamics and a service engineer, and have not spent it on the half where the conversion stage is being removed from the single-line diagram.
Where the capital is not going
Nobody has consolidated the Chinese tail, and that is the investment gap. Kehua, Kstar and East Group compete on price in the markets Huawei is permitted to serve, and a Western vendor cannot buy them, cannot easily compete with them on cost, and cannot ignore them in Asia, Africa, the Middle East and Latin America. The result is a market where the strategic response to the low-cost producer is political protection rather than commercial competition, which is not a durable position.
What we would want to see and cannot
Not one participant in this market discloses UPS revenue. Schneider reports Energy Management; Eaton reports Electrical Americas; Vertiv reports by geography and by product-and-service category, never by product line. There is therefore no audited series anywhere against which any published estimate of this market — ours or anyone else’s — can be checked.
That is a genuine limitation and it should be stated as one. The consequences are three. Every figure for this market is a model, and models differ by choice of parameter rather than by access to better data. The market cannot be valued on a revenue multiple with confidence, because the denominator is unknown. And a supplier cannot benchmark its own share, which weakens the discipline competition normally imposes.
The one observable that would change this is segment disclosure. If any of the three Western incumbents began reporting data centre power separately — as several have begun to do for thermal — the market would acquire an anchor overnight. Until then the reliable investment signals are order growth, backlog and factory capacity announcements, which is why the fourth quarter of 2025 is worth reading closely: Eaton’s data centre orders up approximately 200 per cent year on year and revenue up approximately 40 per cent is the clearest public statement anybody has made about the demand trajectory of this product.
Valuation logic, briefly
Because the segment is not disclosed, the market prices these businesses on blended growth and on the thermal story that is disclosed. Vertiv, as a pure play, is the closest thing to a segment multiple and therefore the reference asset for anyone wanting exposure. The risk embedded in that multiple is the one this report keeps returning to: the thermal business is acquiring the margin, and the power business is the one facing architectural substitution. A buyer of the pure play is buying both and cannot separate them.
Technology & Innovation
What a UPS actually does, and which half still matters
IEC 62040-3 classifies a static UPS on three axes, written as a code: input dependency, output waveform in normal mode and in stored-energy mode, and dynamic output performance. Input dependency has three grades — voltage and frequency dependent, voltage independent, and voltage and frequency independent. VFI, double conversion, is the top grade. It protects against all seven classes of power problem, including frequency variation and waveform distortion, because the load never sees the input at all: AC is rectified to DC, held on a battery bus, and inverted back to a clean AC waveform.
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 accepts a wide input range, it does not care about phase, and it has its own power-factor correction stage. The harmonics and waveform distortion that VFI exists to remove arrive at a device that was going to remove them anyway. What the data centre is still buying from a VFI UPS is not conditioning. It is ride-through: the thirty seconds to two minutes of autonomy that bridges the gap between a utility event and a generator reaching speed.
Which means the product’s defensible function has narrowed to energy storage and transfer, and everything else in the frame is being re-evaluated against its cost. That is the correct frame for reading every development below.
Efficiency, and the trade nobody wants to price
Double conversion costs efficiency because the power is converted twice. A modern transformerless frame reaches roughly 96 to 97 per cent in double conversion and around 99 per cent in a high-efficiency or eco mode that lets the load run on conditioned utility power and transfers to the inverter only on an event. The gap is real money at scale: across 100 MW of IT load, five points of efficiency is 5 MW of continuous loss, which is 43.8 GWh a year, which at USD 100/MWh is approximately USD 4.4 million a year, every year, for the life of the hall.
The hesitation is transfer time. Eco mode introduces a few milliseconds of exposure, and the buyer’s willingness to accept that depends on what the load is. For a batch training cluster it is a straightforward trade. For a payment switch it is not. The interesting thing is that the decision is made entirely privately: no regulator imposes an efficiency floor on this product, and no label discloses which mode a frame was specified in. The consequence is that the same 100 MW hall can carry USD 4.4 million a year of avoidable electricity cost and nothing in the procurement record will say so.
Modularity and prefabrication
Modular frames — a chassis populated with power modules that can be added as load grows — have become the default in new build. The technical merit is real: run the modules at their efficient point rather than part-loaded, add capacity only when needed, and replace a module rather than a frame when one fails. The commercial merit is better: it converts a capital decision into a series of smaller ones, which is easier to approve and easier to finance.
The larger shift is prefabrication of the whole power train, and it is a schedule innovation rather than an electrical one. A skid that arrives assembled and tested removes weeks from a programme where weeks are the scarcest input. It also moves value from the site to the factory, which favours vendors with manufacturing capacity and disadvantages electrical contractors.
The development that actually threatens the market
In August 2026 NVIDIA published a power architecture that takes conversion stages out of the path. The proposal is 800 volts direct current, distributed from the facility toward the rack, on the argument that fewer conversion stages between the grid and the accelerator means more of the available power reaches the compute.
Three products carry it forward, and the third is the one to read carefully.
- A power rack, MGX-compatible, arriving in the second half of 2026. It slots into existing AC infrastructure and delivers 800 VDC to compute racks within the row. This is additive: the building’s electrical system is untouched, so the facility UPS keeps its job.
- A row power centre in 2027 — a centralised station for a full rack row, using an overhead 800 VDC busway, supporting up to 2 MW per row. This begins to displace the row-level AC distribution downstream of the UPS.
- 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. This is the one that matters. 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 disappeared, because the load wants DC.
It would be wrong to read this as the end of the market. The transition is explicitly staged over a decade, the installed base is enormous, and every existing hall keeps its AC architecture. But it establishes the direction of travel, and it establishes it inside the roadmap of the single most influential buyer-designer in the industry rather than in a vendor’s slide. A supplier building capacity on the assumption that the AC double-conversion frame is the permanent shape of data centre power is building for a market the largest customer has already published a plan to leave.
What survives
Ride-through survives, because physics requires stored energy somewhere between the grid and the load. What changes is where it sits. Some moves down into the rack: battery backup units inside the power shelf, holding the load for the few hundred milliseconds a downstream transfer takes. Some moves up into the facility: grid-scale battery storage on site, which can serve the UPS function and simultaneously provide utility services, arbitrage and demand response. A battery that earns money when the grid is fine is a much easier business case to approve than a battery that sits at float for a decade waiting for an outage.
What also survives, and is underrated, is the transfer function itself. Somebody has to detect the event and move the load, reliably, every time, for twenty years. That is a controls and switchgear problem rather than a power-conversion problem, and it is where the incumbents’ engineering depth is genuinely difficult to replicate. The frame is commoditising. The assurance is not.
Innovation intensity, honestly assessed
Topologically, this is a mature product. The rectifier-inverter-battery architecture has not changed in essence in four decades, and patent activity reflects that: steady and incremental, concentrated in control algorithms, paralleling and battery management rather than in new circuit topologies. The high innovation intensity in this market sits in integration — prefabrication, thermal coupling, DC distribution and controls — not in the box. That is an uncomfortable place for a hardware vendor, because integration advantages are easier to copy than a semiconductor process and are worth less each year they are held.
Market Risk
1. Architectural substitution, and why it is the top risk
The 800 VDC roadmap published by NVIDIA is the most consequential document in this market’s recent history — not because it changes anything in 2026, but because it establishes that the largest buyer-designer in the industry has a plan to remove the AC conversion stage from the facility. The staged path is a plan to eliminate, at the top of the range, the product this report is about.
The mitigating facts are substantial. The transition is explicitly a decade-long programme. The installed base of AC halls is enormous and will be maintained for twenty years. Every hybrid product in the roadmap is designed to preserve existing infrastructure. And the facility UPS performs a function — ride-through with a documented, testable availability guarantee — that no published DC architecture yet replaces with equivalent assurance.
But direction matters in capital allocation even when timing is uncertain. A vendor making a ten-year investment decision on AC double-conversion capacity is making it against a roadmap the most important customer in the market has published. The risk is not that the market disappears. It is that growth accrues to whoever sells the DC-native replacement, and that the incumbent frame business is left with the installed base and the replacement cycle.
2. Price realisation, the most sensitive parameter in the forecast
The forecast assumes realised price per kVA declines 2 per cent a year, from USD 424 in 2025 to USD 361 in 2033. That assumption alone moves the 2033 market between USD 12.9 billion and USD 17.9 billion depending on whether it settles at minus 4 per cent or zero. It is the least evidenced input in the model and should be treated as such.
Four forces push it down: power density reducing bill-of-materials content per megawatt; the disappearance of the small and medium three-phase segment where price per kVA was highest; Chinese supply setting a reference price in every market it can reach; and hyperscale procurement, where four buyers with published target returns negotiate multi-year framework agreements. One force pushes it up: the increasing engineering content of prefabricated power trains, which is genuinely more valuable per kVA than a bare frame.
On balance the decline wins, but a supplier planning capacity should stress the downside case rather than the base case, because the downside case is where the volume assumptions stay true and the revenue still misses.
3. Customer concentration
Four companies will spend between USD 803 billion and USD 828 billion on capital expenditure in 2026, and they are the reason this market is growing. That is a demand base with no diversification and with a published, quarterly, publicly negotiated capital plan that can be revised downward on a single earnings call — as Microsoft’s was, by USD 5 billion, over third-party chip delays.
A historical note that should temper the fear: the installed base is now so large that replacement demand alone is roughly 42 per cent of kVA shipped in 2025 by this model’s arithmetic. That is a floor. Even a sharp hyperscale pause would leave the replacement cycle, and the replacement cycle is driven by the fleet’s age rather than by anyone’s capital plan. It is the reason this market is unlikely to contract in absolute terms, even in a bad year.
4. The service annuity is being converted to capex
This is the slowest-moving risk and the one most likely to be underweighted. A UPS sold with lead-acid batteries returns a replacement sale every three to six years and a maintenance contract priced at roughly 8 per cent of capex a year. The same frame sold with lithium returns a larger initial sale, a maintenance contract at roughly 1.5 per cent, and no replacement for a decade.
Multiply that across the installed base and the industry’s recurring revenue shrinks as a share of total revenue, at exactly the moment it needs to grow to fund a field engineering network that has to be present in thirty countries. Vendors will attempt 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, and the honest answer today is that nobody has demonstrated it at scale.
5. Supply chain and geopolitical exposure
The exposure is high and concentrated in three places. Silicon-carbide and IGBT modules are allocated in tight markets and the UPS industry is a small buyer relative to automotive and solar. Lithium iron phosphate cells come from a supply chain dominated by Chinese manufacturers, which makes a European-assembled UPS a Chinese-content product for trade purposes. And Huawei’s presence means the global market is already segmented into two price zones, so a vendor’s cost position depends on where it is permitted to sell rather than on how well it manufactures.
The asymmetric risk is not a shortage. It is a tariff or an origin rule that changes the landed cost of a frame by 15 per cent in a market where realised price is falling 2 per cent a year and the contract is fixed. A supplier with thin margins and a fixed-price framework agreement absorbs that entirely.
6. The counter-risk: this market may be more protected than the thesis implies
It would be dishonest to present substitution as certain. The UPS holds a position very few industrial products hold: it is the component whose failure is most visible, most expensive, and most likely to end a career. Operators are conservative about it for good reason. A 100 MW hall losing power for two minutes can destroy equipment, void warranties, breach service level agreements and, in some jurisdictions, trigger reporting obligations.
That conservatism is a genuine moat, and it is why a new architecture has to be adopted by an entire supply chain — chip designer, OEM, colocation provider, insurer and the authority having jurisdiction — before it displaces an incumbent design. The realistic scenario is not that the AC UPS disappears but that the market splits: DC-native for new hyperscale build from the late 2020s, AC retained for the installed base, for enterprise, for regulated facilities and for anything where an insurer has an opinion. In that scenario the market’s value peaks later and lower than the base case, and the competitive set changes shape without the product disappearing.
Risk summary
Overall market risk is moderate: demand is strong and structurally underpinned by a replacement cycle, while the product faces a genuine architectural substitution over a decade rather than a contraction. Substitution risk is high. Geopolitical and supply-chain exposure is high, because the input chain is concentrated and the output market is politically segmented. Regulatory complexity is moderate — the compliance burden is real but stable, and product-efficiency regulation is conspicuous by its absence rather than by its severity.
Regulatory Landscape
The instrument that should govern this market, and does not
An uninterruptible power supply converts power continuously, every hour of every day, for the life of a data hall, and in doing so loses three to four per cent of everything that passes through it. At the scale of the global data centre fleet that is one of the largest single pools of avoidable electricity consumption in the industrial economy. It is worth stating plainly, then, that the European Union’s product efficiency regulation for this family of equipment does not apply to it.
The Commission Regulation laying down ecodesign requirements for external power supplies, wireless chargers, battery chargers and USB Type-C cables — adopted on 13 October 2025 and repealing Regulation (EU) 2019/1782 — sets out its scope at Article 1. Article 1(2)(a) excludes, in terms, “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 definition is precise, it describes exactly the product this report is about, and it removes it from the only ecodesign instrument in the power-electronics family revised this cycle.
The exclusion is defensible on technical grounds. A UPS is not a passive adapter; its efficiency depends on the load profile, on the redundancy configuration, on whether eco mode is enabled, and on a duty cycle nothing like the near-constant load of an external power supply. Regulating it properly requires a bespoke methodology. The consequence, though, is that no such methodology exists in binding form, and the market has therefore never had an efficiency floor imposed on it.
What does regulate the data centre, and why it does not reach the UPS
Europe has regulated the data centre instead. Directive (EU) 2023/1791, Article 26, requires Member States to ensure that data centres above a threshold of rated energy input use their waste heat or are connected to a heating or cooling system. Commission Delegated Regulation (EU) 2024/1364 established the reporting framework. Both instruments regulate the facility.
The link to the UPS is indirect and, from a product point of view, perverse. The reporting framework captures 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, but has no instrument, label or threshold telling it what a good UPS looks like. The product is treated as a loss term in somebody else’s efficiency metric rather than as a regulated product. A vendor selling a 96 per cent frame against a competitor’s 99 per cent eco-mode frame is arguing in a room where nobody is required to keep score.
Safety and availability codes: where the compliance burden actually sits
UL 1778 is the United States safety standard for uninterruptible power systems — construction, marking and test. It is the gate to the North American market and it is not optional.
IEC 62040 is the international family: Part 1 for general and safety requirements, Part 2 for electromagnetic compatibility, Part 3 for performance and test methods. It is Part 3 that produces the classification code buyers specify against, and the reason a datasheet can say VFI-SS-111 and mean something specific.
NEC Article 708, critical operations power systems, is the provision that matters most in the United States and is the least understood outside it. It applies where a facility is designated as requiring continuous operation for public safety or national security, and imposes requirements beyond ordinary standby power: a formal risk assessment, physical separation and fire protection for the critical distribution, and documentation of the system’s capability. Where Article 708 applies, the UPS is not a product choice — it is a code obligation, and the purchase becomes a wiring and documentation project as much as an equipment one.
NFPA 110 governs the generator the UPS bridges to, and its acceptance testing regime usually determines how much ride-through is actually needed. Designers who specify UPS autonomy without reading the generator’s tested transfer time are specifying a number, not a capability.
The regulatory risk, ranked
Low in the forecast window: a European efficiency floor. The exclusion in the 2025 regulation is a scope decision and the technical argument for it is genuine, so a bespoke UPS measure is a long project rather than an imminent one. The realistic path is indirect — through the facility reporting framework tightening PUE expectations, which makes UPS efficiency matter commercially before it matters legally.
Medium: tariff and origin rules. The supply chain runs through China for cells, magnets and increasingly for whole frames. Measures that change the landed cost of a Chinese-assembled UPS, or of the lithium cells inside a European-assembled one, move price per kVA directly — and because realised price is the most sensitive parameter in the forecast, this is the regulatory lever that moves the number most.
Medium: procurement exclusion and sovereignty. Huawei’s exclusion from US federal and much of European critical infrastructure has already segmented the global market into two price zones. Further segmentation would make the consolidated global figure progressively less meaningful as a planning number.
High over time, and under-priced: battery safety codes. Lithium-ion energy storage inside a building attracts fire code attention that lead-acid never did, and the requirements are still being written in most jurisdictions. A substantial lithium installation in a data hall is governed by provisions that differ by city, which raises design cost and slows approvals. The direction is toward more constraint, and this is the one regulatory area where the product’s own chemistry creates the exposure.
The honest summary
This is a mature, safety-regulated, efficiency-unregulated product. Its compliance burden is high but stable; its regulatory risk is concentrated in trade policy and battery safety rather than in product performance rules; and the absence of an efficiency floor is a commercial rather than a legal fact, which means it will be resolved by buyers with enough scale to measure the loss rather than by a regulator. The hyperscalers already do. The rest of the market mostly does not, and that gap is where the incumbent vendors earn their margin.
