Data centres are infrastructure-grade real estate riding an AI demand wave, but whether an owner compounds or collapses comes down to one thing: the balance sheet behind the building.
Key takeaways
- Market size: ~$383.8bn in 2025, heading to ~$425.3bn in 2026; capacity forecast to nearly double to ~200 GW by 2030.
- The demand engine: hyperscaler capital spending. Eight of them guided to ~$371bn in 2025, forecast to exceed $600bn in 2026.
- The main vehicles: listed REITs Equinix (EQIX) and Digital Realty (DLR) trade globally; Asia adds Keppel DC REIT (SGX: AJBU) and Digital Core REIT (SGX: DCRU); GDS (NASDAQ: GDS, HKEX: 9698) covers China; London offers Cordiant Digital Infrastructure (LSE: CORD); and the Global X Data Center REITs & Digital Infrastructure UCITS ETF (LSE: VPN / PNG) gives one-ticker global exposure in £, € or $.
- The core constraint: electricity. US data-centre power demand is forecast to climb from ~50 GW in 2024 to ~76 GW by 2026.
- The cautionary tale: Cyxtera filed for bankruptcy in June 2023 after rate hikes doubled its interest bill, because it leased rather than owned its buildings.
The 60-Second Version
Every time someone runs a search, streams a film, backs up a phone, or asks a chatbot a question, the work happens inside a building. That building is a data centre, a warehouse-scale structure packed with computers, wrapped in power and cooling, and connected to the world by fibre. For most of the last two decades this was infrastructure nobody thought about, the way nobody thinks about the sewer. Then artificial intelligence arrived, the machines got hungrier, and the buildings that house them turned into one of the most contested pieces of real estate on the planet.
The scale of the money is what makes this worth attention. The global data centre market sat at roughly $383.8bn in 2025 and is tracking toward $425.3bn in 2026. Capacity is forecast to nearly double to around 200 gigawatts by 2030. And the four largest American hyperscalers, Amazon, Microsoft, Google and Meta, spent more than $200bn of capital in 2024 alone, a 62 per cent jump on the year before, most of it on the buildings and machines that make up this asset class. When the richest companies in the world redirect their balance sheets at a single category of physical asset, it is worth understanding what they are buying.
For an outside investor the appeal is that this is real estate with a technology tailwind and utility-like contracts, wrapped in vehicles you can buy on a public exchange. Against that appeal, it is capital-hungry, power-constrained, and one operator has already gone bankrupt doing it wrong. The buildout mechanics, the vehicles that give you exposure wherever you invest from, the unit economics, and the risks that decide whether an owner survives all follow below.
Data centres are physical real estate that behaves like infrastructure: long leases, credit-worthy tenants, high barriers to entry, now riding a structural demand wave from cloud computing and artificial intelligence. You get exposure mainly through listed real estate investment trusts (REITs) and infrastructure funds. The sector’s defining strengths are supply scarcity and contracted cash flow; its defining risks are the cost of capital and the availability of electrical power. An operator that owns its assets and keeps its debt disciplined tends to compound. One that leases its assets and carries floating-rate debt can end up in Chapter 11.
I. What It Is
A data centre is a purpose-built facility whose entire job is to keep computers powered, cooled, connected and secure, and it is one of those assets that hides in plain sight. You have used a hundred of them today without seeing one. They range from a single room in an office block to campuses the size of a small town, and the industry sorts them by who uses them and how.
At the top sits the hyperscale facility, a very large data centre, typically owned and operated by one of the cloud giants for its own workloads. Then come colocation centres, buildings where an operator rents out space, power and cooling to many tenants who bring their own servers (“colo” is the trade shorthand). Below those sit enterprise data centres, run by a single company for its own use, and increasingly edge facilities, small centres placed close to users to cut the delay, or latency, of data travelling long distances.
For an investor, what matters is not the size but the ownership model, because it determines who captures the economics. When you buy shares in a data-centre REIT, you are usually buying the landlord, not the tenant. The REIT owns the building, the power connection, the cooling plant and the security. It signs long leases with tenants, often the same hyperscalers spending those hundreds of billions, and collects rent. A REIT is a real estate investment trust, a listed company that owns income-producing property and passes most of its profit to shareholders, and it is the primary way ordinary investors touch this asset class.
What you own, then, is not a bet on which artificial intelligence company wins, but a claim on the physical floor those companies must rent regardless of which of them wins. It is a picks-and-shovels position. The buildings get paid whether the gold is real or not, as long as the miners keep showing up, and Section XIII deals with what happens if they stop.
II. Market History and Growth
The data centre is nearly as old as networked computing, but the asset class is young. For decades these were cost centres, buildings companies grudgingly built because they had to, depreciating boxes full of depreciating boxes. The shift came when computing moved to the cloud in the 2010s and a handful of operators realised there was a real estate business hiding inside the technology business. Own the building, standardise it, lease it to the giants, and you had an infrastructure asset with a technology growth rate.
The scale today is what separates this cycle from everything before it. The global market reached approximately $383.8bn in 2025 and is projected to hit $425.3bn in 2026. That figure comes with a health warning. Analyst houses disagree sharply on the single revenue number because they draw the boundary of “the market” in different places. A more useful measure, and the one the property advisers favour, is capacity measured in gigawatts of power, because power is what genuinely constrains the sector.
On that measure the trajectory is stark. Property adviser JLL forecasts the sector will nearly double to around 200 GW of capacity by 2030, adding roughly 100 GW of new capacity between 2026 and 2030 at a compound annual growth rate of about 14 per cent. To translate: a gigawatt is a billion watts, roughly the output of a large power station. The industry is proposing to add a hundred power stations’ worth of demand in five years, purely to run computers.
The capital required to build that is where the eye-watering numbers live. JLL frames it as a roughly $3 trillion investment “supercycle” by 2030, of which the 100 GW of new capacity alone implies around $1.2 trillion in real-estate asset value creation. JLL titled that piece “not a bubble”. When the property industry feels the need to say so, the bubble question is clearly live, and Section XIII takes it up directly.
III. Demand Drivers
Growth forecasts are cheap. What makes this one credible is that you can see the demand in the leasing data, and the leasing data is about as tight as commercial property gets.
Start with vacancy. In North America’s primary markets, vacancy fell to a record-low 1.9 per cent at the end of 2024, down from 3.3 per cent in the first half of 2023. For context, a healthy office market runs vacancy in the low teens. Sub-2 per cent means almost nothing sits empty. And it is not because nobody is building. It is because everything gets spoken for first. Of the 3,871.8 MW under construction in the first half of 2024, around 80 per cent was already pre-leased; in Dallas/Fort Worth the figure hit 94.5 per cent. Tenants are signing for buildings that do not yet exist.
The construction pipeline confirms the same thing from the supply side. There was a record 6,350 MW under construction in US primary markets at the end of 2024, more than double the 3,077.8 MW a year earlier. Doubling the pipeline in a year, while vacancy falls to a record low, is the signature of demand outrunning supply.
Where is the demand coming from? Follow the capital. The four largest US hyperscalers spent more than $200bn in capital in 2024, a 62 per cent year-on-year jump, most of it on data-centre and artificial-intelligence infrastructure. Widen the lens and eight hyperscalers guided to roughly $371bn in 2025 capex, up 44 per cent, with the figure forecast to exceed $600bn in 2026, around $450bn of it tied to artificial intelligence. These are the tenants. When your tenant list is the most cash-generative businesses in history and they are collectively raising spending toward $600bn a year, the landlord’s rent roll looks secure.
The deepest driver, though, is power, and it deserves its own frame because it is both the demand engine and the binding constraint. US data-centre power demand is forecast to rise from around 50 GW in 2024 to roughly 76 GW by 2026, on Deloitte’s numbers. Push the horizon out and the artificial-intelligence slice alone could grow more than thirty-fold to 123 GW by 2035, up from just 4 GW in 2024. A thirty-fold rise in a decade is not a normal infrastructure growth rate. It is why one operator, Microsoft, is now the world’s largest corporate clean-power buyer, with roughly 40 GW contracted as of late 2025, a technology company buying power on the scale of a national grid.
IV. The Players
An asset class is easier to understand through the people and firms that dominate it. Data centres split into the listed landlords, the private capital, and the tenants who set the pace.
The listed landlords. Two REITs sit at the front of the public market, and both trade globally rather than only in America. Equinix (NASDAQ: EQIX) is the largest listed data-centre REIT by revenue, posting FY2024 revenue of $8.748bn, up 7 per cent, with adjusted EBITDA of $4.097bn. Its business is interconnection, being the place where networks physically meet, which gives it stickier tenants and pricing power. Its CEO and President is Adaire Fox-Martin. Digital Realty (NYSE: DLR) is the other giant, more focused on large-scale wholesale space; it reported FY2025 funds from operations of $7.10 per share, up from $6.27 in FY2024. “Funds from operations” (FFO) is the REIT world’s preferred profit measure: net income with property depreciation added back, because buildings on the books lose value on paper faster than good ones do in reality. Both names carry international portfolios spanning Europe, Asia and the Americas, and both are held inside the global ETFs covered in Section VI, so a non-US investor gets them without buying an American single stock directly.
The Asian and European operators. Outside America the sector has its own listed champions. In Singapore, Keppel DC REIT (SGX: AJBU) was the first pure-play data-centre REIT in Asia and held a SGD 6.3 billion portfolio of 25 properties across 10 countries in Asia and Europe as at 31 March 2026, while Digital Core REIT (SGX: DCRU), sponsored by Digital Realty, ran about US$1.8bn of assets across the US, Canada, Germany and Japan at the end of 2025. In China, GDS Holdings (NASDAQ: GDS, HKEX: 9698) is the largest independent operator, with 2025 revenue up 10.8 per cent to RMB 11.43 billion and its first-ever annual profit. In London, Cordiant Digital Infrastructure (LSE: CORD) owns data centres, fibre and towers across Europe and North America and entered the FTSE 250 after net asset value reached £1.12 billion (about US$1.52 billion). The point is that the listed universe is genuinely global, not an American club you have to buy into from outside.
The private capital. The biggest single force in data centres is not a REIT at all. It is a private equity firm. Blackstone, through its funds, owns QTS, the largest independent operator, bought in 2021. Blackstone’s President and COO Jonathan Gray has become one of the sector’s most closely watched voices, because his firm has so much at stake. His line on build discipline, quoted in Section XVII, captures how the survivors in this business think.
The tenants. The pace-setters are the hyperscalers, Amazon (AWS), Microsoft (Azure), Google (Cloud) and Meta, whose combined 2024 capex topped $200bn. They both rent from the landlords above and build their own. That dual role matters: they are the REITs’ best customers and their largest potential competitors at once, a tension Section XIII returns to.
Here is Equinix’s chief executive, on the company’s third-quarter 2025 earnings call, on pricing and strategy:
“We’re certainly not seeing any dilution in our pricing. Very firm… We are advancing our build bolder strategic move where our intent is to double capacity by 2029.”
Adaire Fox-Martin, CEO & President, Equinix, Q3 2025 earnings call, 29 October 2025
Firm pricing and a plan to double capacity within four years is roughly the whole bull case, stated by the person responsible for delivering it.
V. Geography
Data centres are global, but they cluster, because they need three things that do not exist everywhere at once: cheap reliable power, fibre connectivity, and land near enough to users to keep latency low. Where those three overlap, capacity concentrates.
North America is the deepest market and the one the data tracks most closely. US primary markets, Northern Virginia (the largest single cluster on earth), Dallas/Fort Worth, Chicago, Phoenix and Silicon Valley, carry that record 6,350 MW construction pipeline and 1.9 per cent vacancy. Northern Virginia alone routes a large share of global internet traffic. The American market’s advantage is scale, capital depth and, for now, power availability, though the climb from 50 GW to 76 GW by 2026 is straining local grids.
Europe clusters around the “FLAP-D” markets, Frankfurt, London, Amsterdam, Paris and Dublin, where connectivity and financial infrastructure concentrate demand. Europe’s constraint is power and planning: several jurisdictions have restricted new grid connections, which paradoxically strengthens the value of existing sites. Scarcity, again, favours the incumbent owner, and it is why a London-listed operator like Cordiant can build value from European sites that are hard to replicate.
Asia-Pacific is the fastest-growing region by several measures, with major hubs in Singapore, Tokyo, Sydney, Mumbai and the Chinese metros. It is also where the sector’s own listed vehicles sit closest to the growth: Keppel DC REIT and Digital Core REIT on the Singapore Exchange, GDS in China, and Australia’s NextDC among the ASX names. Singapore’s moratorium on new capacity, since partly relaxed, is the textbook case of a government deciding data centres were consuming too much of a small nation’s power and water. The region’s growth is what pushes the global forecast toward that 200 GW by 2030.
Emerging markets, the Gulf, Latin America and Africa, are the frontier, drawn by cheap energy (the Gulf) or fast-growing user bases (India, Nigeria, Brazil). These carry higher political and currency risk but also the steepest growth curves.
The investment lesson from geography is that the constraint travels. Wherever power and permitting tighten, existing sites gain value and new supply slows, which is good for owners and bad for tenants. An investor in listed vehicles gets this diversification automatically; the big REITs and the global ETFs already span these regions.
VI. How to Actually Invest
Most investors reach this asset class through one of a handful of practical routes, which trade off access, cost and control against each other. The routes below work whether you invest from London, Frankfurt, Singapore or New York; where a route is genuinely restricted to one market, it says so.
| Vehicle | Ticker / listing | What you own | Fee / cost | Minimum |
|---|---|---|---|---|
| Data-centre REIT (interconnection focus) | EQIX (NASDAQ, global portfolio) | Shares in the largest listed operator | Brokerage commission only | One share |
| Data-centre REIT (wholesale focus) | DLR (NYSE, global portfolio) | Shares in a large-scale landlord | Brokerage commission only | One share |
| Asia-listed pure-play REIT | AJBU / DCRU (SGX) | Keppel DC REIT / Digital Core REIT | Brokerage commission only | One share |
| China operator | GDS (NASDAQ / HKEX: 9698) | Shares in the largest independent China operator | Brokerage commission only | One share |
| UK-listed digital-infra trust | CORD (LSE, FTSE 250) | Data centres, fibre and towers, Europe + N. America | Trust management fee | One share |
| Global sector UCITS ETF | VPN / PNG (LSE, also Xetra/SIX/Milan) | A global basket (Equinix, Digital Realty, American Tower, NextDC…) | 0.50% TER | One share, in £/€/$ |
| US sector ETF | SRVR / DTCR (US) | A US-listed basket of digital-infra names | 0.49% / 0.50% | One share, no fund minimum |
| Broad REIT ETF | VNQ (US) | Diversified property, small data-centre slice | Low expense ratio | One share |
| Listed infrastructure closed-end fund | UTF (US) | Infrastructure incl. data-centre exposure | Fund management fee | One share, no fund minimum |
| Private / institutional | via Blackstone (owns QTS) | Direct stake in operating assets | Private-fund fees | Institutional / accredited only |
Route one, individual REITs. Buying Equinix or Digital Realty shares gives the purest, most concentrated exposure. You own the landlord directly, you collect the dividend, and you carry single-company risk. If you would rather own the growth where it is fastest, the Asian pure-plays are the direct route: Keppel DC REIT (SGX: AJBU) and Digital Core REIT (SGX: DCRU) are both listed in Singapore, and GDS (NASDAQ: GDS, HKEX: 9698) gives China exposure on either exchange. This is the highest-conviction, highest-variance way in, wherever you buy it.
Route two, the sector ETF. For most investors the cleanest one-ticker option is a global fund. The Global X UCITS ETF lists on the London Stock Exchange (VPN in dollars, PNG in sterling) as well as Xetra, the SIX Swiss Exchange and Borsa Italiana, at a 0.50 per cent total expense ratio, and holds a global basket led by Equinix, Digital Realty, American Tower and Australia’s NextDC. American investors have the equivalent US-listed funds: the Pacer Data & Infrastructure Real Estate ETF (SRVR) at a 0.49 per cent expense ratio and the Global X Data Center & Digital Infrastructure ETF (DTCR) at 0.50 per cent. All of them blend data centres with communications towers, which is “digital infrastructure” as a theme rather than pure data centres. That is a feature for diversification and a bug for purity.
Route three, broad real estate. A diversified vehicle like the US-listed Vanguard Real Estate ETF (VNQ), or a global REIT tracker in your own market, gives you data centres as one slice of a property portfolio; for reference, data-centre REITs were about 6.9 per cent of iShares’ ESG-aware real estate ETF as of 31 January 2025. This is the lowest-conviction, lowest-concentration route: the tailwind is present but diluted.
Route four, infrastructure and private. In London, a listed digital-infrastructure trust such as Cordiant Digital Infrastructure (LSE: CORD) trades like a share and owns data centres, fibre and towers directly, targeting total returns of at least 9 per cent a year; it trades at a premium or discount to net asset value, which is worth checking before you buy. American investors have the Cohen & Steers Infrastructure Fund (UTF), a listed closed-end fund with data-centre exposure that also trades like a share. The genuinely private route, direct operating stakes, runs through firms like Blackstone, which acquired QTS for roughly $10bn at a 21 per cent premium, closing in September 2021. That route is open only to institutional and accredited investors through private funds, worth knowing exists, out of reach for most.
For the overwhelming majority of investors, anywhere in the world, this asset class is bought on a public exchange, through a REIT or an ETF, for the price of a single share. The private tier is where the largest returns have been made, and it is also the tier you cannot access without size.
VII. Unit Economics
The economics are easiest to follow built from the ground up, using the industry’s own figures rather than round abstractions.
Start with construction cost. A standard build runs about $10 to 12m per megawatt of capacity; an artificial-intelligence-optimised facility costs more than $20m per megawatt because of the density of the chips and the cooling they demand. Megawatts, again, are the unit that matters, because power, not floor space, is what a data centre sells.
Now scale it. Take a 30 MW facility at roughly $10m per megawatt. That is $300m of capital expenditure to build. Energy consultancy Thunder Said Energy‘s worked model says that facility must generate around $100m a year in revenue to earn about a 10 per cent internal rate of return, the annualised return on the capital, the number that decides whether the project clears its cost of money. Of that roughly $100m in annual cost, around 40 per cent is maintenance and 15 to 25 per cent is electricity.
That cost split is worth reading closely, because it describes the whole business. Electricity is a large, variable, and largely uncontrollable input, up to a quarter of the cost base and rising with power prices. Maintenance is the biggest line. And the capital is spent up front, all $300m, before a single dollar of rent arrives. The asset has enormous fixed capital, a long payback, and thin operating flexibility. It rewards patient owners with cheap capital and punishes leveraged owners when money gets expensive, which is the story of Section X’s cautionary case.
For the artificial-intelligence facilities driving the current cycle, the total cost of ownership runs around $8.5m per megawatt per year for a 1 GW data centre. Multiply that across a gigawatt and the annual cost of ownership runs into the billions. A facility at that scale is closer to a power station that happens to compute than to an ordinary building.
The maths governs everything that follows: in this business the cost of capital is not a background variable, it is the thing that decides whether a project works. A project that clears a 10 per cent return at 4 per cent borrowing costs can go underwater at 8 per cent borrowing costs without a single tenant leaving, and that mechanism drives the next two sections.
VIII. Macro Sensitivity
Every asset behaves differently across the economic cycle. Data centres are defensive on the demand side and dangerously sensitive on the financing side, and that split shows up clearly when you sort the behaviour by interest-rate and growth regime.
| Regime | What happens to data centres | Why |
|---|---|---|
| Falling rates, growth | Best case. Cheap capital funds the buildout; contracted cash flows compound; REIT valuations re-rate up. | The 10% IRR maths works easily when borrowing is cheap. |
| Rising rates, growth | Mixed. Demand stays strong (1.9% vacancy) but financing costs bite; leveraged, floating-rate operators are exposed. | Rate-sensitivity is the sector’s defining macro risk. |
| Rising rates, recession | Worst case for the weak. Refinancing walls plus a demand pause can be fatal, see Cyxtera. Strong owners with fixed debt survive. | Interest expense can double in a year as debt reprices. |
| Falling rates, recession | Resilient. ~80% pre-leasing insulates cash flow from a short-cycle demand shock while cheaper money eases refinancing. | Contracted revenue is booked years ahead. |
The same pattern runs through all four rows. On the demand side, data centres are genuinely defensive: record-low 1.9 per cent vacancy and around 80 per cent pre-leasing mean cash flows are contracted years ahead, insulating owner-operators from short-cycle demand shocks. Very few real assets can say their revenue is largely booked before the building opens.
On the financing side, though, the sector is as rate-sensitive as any leveraged real estate. Interest-rate exposure is its defining macro risk, illustrated most cleanly by Cyxtera, whose annualised interest expense doubled from $35.9m to $75.7m in a single year as rates rose. Nothing about its buildings changed. The cost of its money changed, and that was enough to sink it.
What decides outcomes, then, is less good economy versus bad economy than cheap money versus dear money, and whether the operator’s balance sheet is ready for the switch. An investor who ignores the debt structure of what they own is ignoring the only variable that has actually bankrupted an operator in this cycle.
IX. Tax
Tax treatment is where the REIT structure earns its keep, and where jurisdiction-specific rules make blanket statements dangerous. This section stays jurisdiction-neutral and general. It is not tax advice. Confirm every point against your own jurisdiction and a qualified adviser.
The foundational feature is the pass-through. A qualifying REIT generally avoids entity-level income tax by distributing the bulk of its taxable income to shareholders, who are then taxed on those distributions. The tax point shifts from the vehicle to the investor. That is the whole appeal of the REIT wrapper: the income is taxed once, in the investor’s hands, rather than twice (once at the company, once at the shareholder). The exact minimum-distribution threshold that a REIT must pay out to qualify is set locally and varies by jurisdiction. The US requires 90 per cent, the UK and Singapore run their own REIT regimes with their own thresholds, so treat it as a country-specific rule.
The practical consequence for a holder is less flattering. REIT distributions are typically taxed as ordinary income to the holder rather than at the lower rates that apply to qualified dividends, and the split of any distribution between ordinary income, return of capital, and capital gains varies by fund and by year. In plain terms: the tax efficiency lives at the company level, not necessarily at your level, and you may pay your ordinary income rate on the cash you receive. The fund’s annual tax statement is what tells you the actual breakdown. Do not assume; read it.
Two more general points that hold across most jurisdictions. First, holding these vehicles inside a tax-advantaged account (an ISA or SIPP in the UK, an IRA or 401(k) in the US, or the equivalent your country offers) can neutralise much of the ordinary-income drag, which is often the single biggest lever an individual has. Second, cross-border holdings can attract withholding tax on distributions, and treaty relief varies. If you are buying an American REIT from outside America, a Singapore REIT from Europe, or any listing outside your home market, the withholding question is not optional homework, and the rate you actually keep depends on the treaty between the two countries.
None of this argues against the asset class. It argues for knowing, before you buy, which pocket the tax comes out of, because for an income-focused holding the return that counts is the one left after tax.
X. Case Studies
Case one, the private compounder (Blackstone / QTS). Blackstone bought QTS for roughly $10bn in 2021. By late 2024 the platform had expanded more than eightfold to a roughly $70bn portfolio with a development pipeline of about $100bn; Blackstone refinanced ten QTS centres with a $3.5bn commercial mortgage-backed securities loan and two more with a $1.5bn loan. The lesson is not “buy QTS”, because you cannot. The lesson is what patient private capital, with a long horizon and disciplined debt, did with the asset: took it off the public market at the start of the demand wave and rode the whole thing. It is the clearest evidence that the biggest returns in this asset class were made in the tier most investors cannot reach.
Case two, the listed compounder (Equinix). The public-market equivalent shows the same trajectory at a pace a retail investor could actually own. From FY2024 revenue of $8.748bn, Equinix carried into 2026 with Q1 2026 revenue of $2.444bn, up 10 per cent, a record 51 per cent adjusted-EBITDA margin, and adjusted funds from operations of $1.065bn, up 12 per cent, guiding to 9 to 10 per cent revenue growth and 8 to 10 per cent AFFO-per-share growth for 2026. Not explosive. Steady, contracted, margin-expanding growth in a business you can buy for the price of one share. That is what the listed version of this asset class delivers when it works.
Case three, the cautionary tale (Cyxtera). And here is what it looks like when it fails. Cyxtera filed for Chapter 11 bankruptcy in June 2023. The mechanism was brutal and simple: rate hikes more than doubled its annualised interest expense from $35.9m in Q1 2022 to $75.7m in Q1 2023, against roughly $1.02bn of loan maturities converging on 2024. But the fatal flaw was structural: Cyxtera leased rather than owned most of its real estate. It carried the debt of an owner and the assets of a tenant, the worst of both. When rates rose it had no property to borrow cheaply against and no owned buildings to anchor its value. The assets were eventually sold to Brookfield for $775m, closing on 12 January 2024.
Put the three side by side and the pattern is clear. The two winners owned their assets and managed their debt on their own terms. The loser leased its assets and let its debt reprice against it. They faced the same asset class and the same demand wave and ended up on opposite sides, decided almost entirely by capital structure. It is the lesson that matters most across everything else here.
XI. The Core Constraint
The constraint that most determines who wins and how fast the sector can grow is not demand, capital, or land. It is electricity.
The demand for computing is effectively unlimited at current prices; the capital is flowing at $600bn-plus a year; land can be found. What cannot be conjured on demand is a grid connection large enough to power a facility that draws as much as a small city. US data-centre power demand rising from 50 GW to 76 GW by 2026, and the artificial-intelligence slice heading toward 123 GW by 2035 from 4 GW in 2024, collides directly with grids that take years to upgrade and communities that increasingly object to the load.
This is why Microsoft became the largest corporate clean-power buyer in the world, with ~40 GW contracted. Securing power ahead of need is now a competitive weapon. It is why Singapore paused new builds and several European jurisdictions restricted grid connections. And it is the reason the sector’s scarcity, which is so good for existing owners, is likely to persist: you cannot flood the market with new supply when the supply is gated by transformers and transmission lines that take half a decade to build.
For an investor, the constraint cuts both ways. Power scarcity protects incumbents, since the operator that already has a powered, permitted site owns something genuinely difficult to replicate, and its rents reflect that. Power scarcity also caps the sector’s growth and imports a new dependency, because energy prices and grid policy now sit inside the data-centre thesis. Buying this asset class means also taking a position, quietly, on the price and availability of electricity. Data centres are not a pure technology play once you look at the cost stack, where electricity is up to a quarter of the bill.
XII. Inside the Asset
It helps to know what you actually own, physically, when you buy a slice of this asset class. Strip away the finance and a data centre is a machine for turning electricity into computation without letting anything overheat or go dark.
Power comes in from the grid at high voltage, steps down through transformers, and splits into redundant paths so that no single failure can cut it. Backup generators and banks of batteries stand ready to carry the load in the seconds and hours a grid outage lasts, because for a hyperscale tenant, downtime is measured in lost millions. The industry grades reliability in “tiers”, from a basic single-path facility up to fully fault-tolerant designs with no single point of failure. Higher tiers command higher rents.
Then there is heat. Every watt that goes into a chip comes out as heat, and the artificial-intelligence chips driving this cycle run hot enough that traditional air cooling is reaching its limit, which is why the newest facilities are moving to liquid cooling, running coolant directly to the processors. This is why an artificial-intelligence-optimised build costs more than $20m per megawatt against $10 to 12m for a standard one: the density of the computing and the cooling it demands. The physical difference between a 2015 data centre and a 2026 one is mostly about how it moves heat.
The “white space”, the raised-floor rooms full of server racks, is what the tenant rents, measured in megawatts of power draw rather than square feet. Around it sits the fibre, the network connections that make a data centre useful, because a building full of computers with no connectivity is a very expensive space heater. The most valuable sites are the ones where the most networks physically meet, which is precisely the interconnection advantage that lets Equinix keep its pricing “very firm”.
Knowing the physical asset changes how you read the finance. When a REIT reports capital spending, that is transformers, generators, cooling plant and fibre, all long-lived and hard to replicate. When it reports maintenance as roughly 40 per cent of annual cost, that is the price of keeping all of it running without a second of unplanned darkness. That hard-to-replicate infrastructure is what the tenants are paying to rent.
XIII. The Central Dilemma
The tension a serious investor has to sit with, and the one most capable of unravelling the thesis, is that the data-centre landlords’ best customers are also their most capable competitors.
The hyperscalers spending $600bn a year both rent from the REITs and build their own facilities. Every megawatt Amazon or Google builds for itself is a megawatt it does not rent from Equinix or Digital Realty. So far demand has been overwhelming enough that both models grow together, with more than enough to go round, which is why vacancy is at 1.9 per cent. The structural question does not go away, though: what happens to landlord pricing power the day the giants decide to own rather than rent?
Layered on top is the demand-durability question, the “bubble” word JLL felt compelled to address when it called the $3 trillion supercycle “not a bubble”. The entire step-change in demand rests on the assumption that artificial-intelligence workloads keep growing at the rate the 123 GW by 2035 forecast implies. If a more efficient generation of models needs far less computing per unit of output, the demand curve bends, and even Blackstone, the most committed private player, has publicly admitted it is watching efficiency concerns “very closely”.
Near-term and long-term risk sit at different distances here. The pre-leasing data means near-term cash flows are contracted and safe regardless of how the debate resolves. The long-term thesis that justifies today’s valuations depends on the demand wave being structural rather than a spike. A buyer of this asset class is taking a view, whether they admit it or not, that computing demand compounds for a decade. That view may well be right, but it is not a certainty, and the vehicles that survive being wrong will be the ones that, like the winners in Section X, own their assets and control their debt, so that a demand pause is survivable rather than fatal.
XIV. The Next Frontier
Several shifts in where this asset class is heading are already visible in the data, and each changes what you would be buying.
Power becomes part of the asset. The electricity constraint is pushing operators to stop treating power as an external input and start owning it: on-site generation, direct deals with power producers, and the kind of 40 GW clean-power book Microsoft has built. The data centre of 2030 may come bundled with its own power source. That blurs the line between a data-centre investment and an energy-infrastructure investment, and it is why the “digital infrastructure” ETFs blending towers, power and data centres may end up describing the sector better than “data centres” alone.
The build gets denser and more specialised. Artificial-intelligence facilities at more than $20m per megawatt with liquid cooling are a different asset from the general-purpose colocation building. As the split widens, investors will increasingly be choosing between two sub-classes: the high-density artificial-intelligence facility with concentrated tenant risk, and the diversified interconnection hub with many smaller tenants. They will not carry the same risk profile.
Geography spreads to the power. As mature markets tighten on grid connections, new capacity follows cheap energy, to the Gulf, to the Nordics, to wherever power is abundant. The map of the sector redraws itself around electricity rather than around users. For an investor in global vehicles, that is diversification arriving automatically; for a single-market bet, it is a reason the “primary markets” of today may not be the growth markets of 2035.
Running through all three, “data centres” is quietly becoming “digital-and-power infrastructure”. An investor who sees that early is buying the right thing under a label that has not caught up yet.
XV. Lessons From History
This is not the first infrastructure buildout financed on a wave of belief, and the earlier ones are worth reading against it.
The closest parallel is the fibre-optic boom of the late 1990s. Telecom companies borrowed enormous sums to lay cable in the certainty that internet traffic would grow forever. Traffic did grow forever, so the demand thesis was correct. But the buildout ran ahead of it, financed with too much debt, and when the timing gap opened the over-leveraged builders went bankrupt even as the demand they had bet on arrived as predicted. The cable stayed in the ground and made fortunes for whoever bought it out of the wreckage rather than whoever built it. Cyxtera’s Chapter 11 and $775m sale to Brookfield is the same story on a smaller scale: the right asset and the right demand, undone by the wrong balance sheet, with the patient buyer collecting.
The second lesson is that capital structure decides survival, not the quality of the thesis. Being right about demand is not enough. Cyxtera was right about demand and still failed, because its interest expense doubled in a year and it had leased rather than owned. Blackstone and Equinix were right about the same demand and thrived, because they owned and financed conservatively. The demand wave is available to everyone; surviving long enough to collect it is not.
The third lesson is that the biggest gains often go to the buyer, not the builder. Brookfield bought Cyxtera’s assets out of bankruptcy. Blackstone bought QTS at the start of the wave and grew it eightfold. In infrastructure, the discipline to wait for the right entry price frequently beats the enthusiasm to build first. It sits awkwardly in a sector where the narrative pushes hard to get in now, which is why it is easy to forget and worth keeping in view.
XVI. The Case For It
The bull case for data centres as an asset class comes down to four things.
First, the demand is real and visible, not projected. You do not have to believe a forecast to see 1.9 per cent vacancy and 80 per cent pre-leasing. The cash flows are contracted years ahead. This is a rare case where the demand shows up in signed leases before the building exists.
Second, the tenants are the strongest in the world. When your rent roll is Amazon, Microsoft, Google and Meta collectively spending toward $600bn a year, tenant default risk, the thing that sinks most commercial property, is about as low as real estate offers.
Third, the barriers to entry are genuine and rising. The electricity constraint, the $10 to 20m-per-megawatt build cost, and the years-long permitting mean existing powered sites are hard to replicate. Scarcity is structural, not cyclical, and scarcity is what protects an owner’s rents.
Fourth, it is accessible from anywhere. Unlike most infrastructure, you do not need institutional size. A single share of a REIT or ETF buys a slice of the same asset class Blackstone is deploying billions into, and you can buy it in London, Singapore, Frankfurt or New York in your own currency. That combination, infrastructure economics at retail access, is unusual.
Underneath all four sits the compounding demonstrated by Equinix’s steady 9 to 10 per cent growth guidance and Blackstone’s eightfold platform growth. This is an infrastructure asset with a genuine growth rate rather than a speculative punt, and the demand behind that growth is visible in signed leases today rather than only in forecasts.
XVII. The Risks
Four risks carry real weight, and they are not evenly matched. The first has already bankrupted a company in this cycle; the rest are live but slower-moving.
First, and above all others, the cost of capital. The unit economics turn on borrowing costs, and Cyxtera proves the risk is not theoretical: interest expense doubled from $35.9m to $75.7m in a year and bankrupted a business whose buildings were full. Any operator with too much floating-rate debt and a refinancing wall carries the same failure mode. This is the risk that has actually killed a company in this cycle.
Second, the power constraint could bite the other way. The same electricity scarcity that protects incumbents also caps growth and raises costs. Rising power prices flow straight into the 15 to 25 per cent of costs that electricity represents, and grid or planning restrictions can strand a half-built project.
Third, the demand thesis could soften. If artificial-intelligence workloads grow more efficient and need less computing than the 123 GW by 2035 forecast assumes, today’s valuations look stretched. This is the concern Blackstone itself flagged.
On that discipline, here is the sector’s most-watched private investor:
“I’d start with our data center business, which is the largest in the world. We have $80 billion of leased data centers… We do not build data centers speculatively anywhere in the world.”
Jonathan Gray, President & COO, Blackstone, on a 2025 earnings call
“We do not build speculatively” is the whole risk-management approach compressed into a sentence: build against a signed lease rather than a hope. The operators who ignore that rule are the ones who become the next cautionary case.
Fourth, tenant concentration and the build-versus-rent tension. The hyperscalers who are the best tenants can become competitors, and a facility leased to one giant carries concentrated risk if that giant changes strategy. Diversified interconnection hubs mitigate this; single-tenant artificial-intelligence facilities do not.
XVIII. The Alternative Fortune Verdict
Data centres are a genuine infrastructure asset class riding a real, visible demand wave, and at the same time a leveraged real-estate bet whose fate turns on the cost of money. The demand wave is the part everyone sells; the cost of money is the part that actually decides who survives it.
The demand is as well-evidenced as anything in property: 1.9 per cent vacancy, 80 per cent pre-leasing, the strongest tenants on earth spending toward $600bn a year. The vehicles are accessible from almost any market, from a single REIT share to a 0.50 per cent global UCITS ETF you can buy in sterling. And the winners, Equinix compounding at 9 to 10 per cent and Blackstone’s QTS growing eightfold, are real and repeatable. Against that, the cost of capital can bankrupt a full building, the power constraint caps growth and imports energy risk, and the demand thesis rests on artificial-intelligence appetite compounding for a decade. It is neither a sure thing nor a bubble, but an infrastructure asset with a strong tailwind and a sharp, specific downside.
Where the edge actually is. It is not in guessing which artificial-intelligence company wins, since the landlord gets paid either way. It is in reading capital structure. The pattern proven three times over in Section X is that owned-and-conservatively-financed assets compound while leased-and-floating-rate assets collapse. So the edge is fairly unglamorous: prefer operators that own their buildings and finance them on fixed, long-dated terms, and be sceptical of anyone building speculatively against a demand curve rather than a signed lease. Jonathan Gray’s “we do not build speculatively” works as the filter to apply to any operator you consider.
Questions to ask, by vehicle:
- Buying an individual REIT? Does it own or lease its buildings? What share of its debt is fixed-rate, and when does it mature? What is its power position, does it have secured connections or is it exposed to grid constraints? How concentrated is its tenant base?
- Buying a sector ETF? What exactly is inside it, pure data centres or a “digital infrastructure” blend of towers and power? What is the expense ratio, and does the 0.50 per cent buy you the concentration you actually want?
- Buying a broad REIT fund? How small is the data-centre slice really, under 7 per cent in some funds, and is that enough exposure to matter to your thesis?
- Buying an infrastructure or closed-end fund? Does it trade at a premium or discount to its net asset value, and what is the fee for the active management you are paying for?
- Every vehicle: how will the distributions be taxed in your jurisdiction, what withholding applies if the listing is abroad, and does holding it in a tax-advantaged account change the after-tax return enough to matter?
The asset class is real and the demand is visible, and the mistake that undoes an operator is almost always the same one: the wrong balance sheet against the right thesis. Ask the questions above before you buy and you are looking at digital infrastructure with your eyes open rather than through the hype.
This is an educational overview of an asset class, not investment or tax advice, and not a recommendation of any named security. Figures are sourced as of the research date; verify current data and your own jurisdiction’s rules before acting.
For more on where data centres sit within property as an asset class, see the Alternative Fortune real estate category: alternativefortune.com/investments/real-estate.
