• Global data centre capital expenditure could reach US$31.6 trillion through 2050, with a plausible upside of nearly $50 trillion if AI adoption accelerates.
• Unlike earlier infrastructure expenditure cycles, data centre investment is expected to keep rising through 2050 because servers, GPUs, and other ICT equipment require refreshes every four to six years.
• Although data centre demand is strong globally, factors such as power availability, data sovereignty requirements, and chip trade flows will determine which regions capture it.

Railways. Electrification. The internet. Each required enormous amounts of capital and defined an era. The AI infrastructure cycle underway dwarfs all three. Perhaps more notably, unlike past build cycles that front-loaded capital expenditures (capex) and tailed off as the network matured, this one resets every four to six years—and shows no signs of ending. 

PwC commissioned Oxford Economics to model data centre capital expenditure across 46 countries/territories and five regions—a footprint that accounts for the lion’s share of global economic output and digital infrastructure spending. In our central scenario projection, data centres attract US$31.6 trillion of capital expenditure globally through 2050, with a plausible upside of nearly $50 trillion, to build the compute capacity required to support the world’s AI ambitions. Although cross-era comparisons are imperfect, those figures are orders of magnitude larger than either the combined amount for UK and US railway construction or the cost of the internet build-out.

In addition to scale, what makes this capex cycle different is that annual spending accelerates over time, rising from roughly $800 billion in 2026 to $1.1 trillion in 2030 to $1.8 trillion in 2050. That’s because the bulk of the spend doesn’t go towards the buildings. Rather, it funds what fills them: servers, storage systems, networking equipment, central processing units (CPUs), and, crucially, the graphics processing units (GPUs) that provide compute power for AI—which age out in a handful of years and will need to be replaced.

The $31.6 trillion estimate is the central scenario within a range of about $22 trillion to $50 trillion; the outcomes depend largely on the pace and scale of AI adoption. But across the scenarios we constructed, the direction is the same: capex rises significantly over the course of the forecast period. The trajectory holds even if geopolitical tensions constrain the supply of advanced chips or shift demand towards locally serviced workloads (two scenarios we’ll explore below). What changes underneath is the geography of the spend.  

Although an enormous amount of funding is required to meet mounting compute needs from a widening set of users, capital should be available in sufficient quantities. Power availability, however, is another matter. AI demand is running up against the limits of grids built decades ago, and whether our forecast becomes reality depends on how fast, and how reliably, electricity can power data centres.

What follows is an analysis of where this capital is most likely to go, what could redirect it, and what the resulting map means for the investors, operators, and governments with stakes in the outcome. 

A structurally different cycle

The forecast for capex requirements is shaped like a cone. Near-term demand is more predictable given its visibility in cloud migration, hyperscaler expansion, and the first wave of AI-native businesses. Farther out, the cone widens, with the spread between optimistic and pessimistic AI adoption outlooks equivalent to roughly the size of US GDP. Faster AI adoption pushes capex to about $50 trillion through 2050; a slower adoption path removes roughly $10 trillion from the cycle and leaves stadium-sized data centres around the world sitting half empty. 

Across the plausible range of outcomes, annual capex remains elevated and continues to rise. The reason lies in the capital stack of a data centre. The building shell and core site infrastructure—for example, utility connections and fibre routes—may have long, useful lives, but the technology inside doesn’t. What’s still underpriced by much of the market is that every $1 of construction capex effectively commits the market to roughly $12 of capex for future information and communications technology (ICT) equipment—servers, storage systems, networking equipment, GPUs, and other hardware that must be installed, refreshed, and upgraded. Across the market as a whole, ICT equipment rises from 70% of total capex in 2026 to 93% by 2050. 

AI is what makes that equipment cycle so capital-intensive. Traditional cloud infrastructure was built largely around central processing units (CPUs). AI workloads, by contrast, rely on GPUs and accelerators that are more expensive, are more energy-intensive, and follow a faster innovation curve. GPUs and servers typically turn over every four to six years, meaning a single data centre may require three to five rounds of ICT investment over a 20-year asset life. Higher rack densities can also force power and cooling upgrades, but the largest and fastest-recurring share of capex is the compute equipment itself. Every prior infrastructure wave—railways, electrification, the internet—front-loaded construction capex and tailed off as the network matured. This wave inverts the pattern. The data centre is essentially a chip-replacement subscription with a building wrapped around it. Build the box once, refill it every four to six years. That pattern has implications not just for how much capital is needed, but for who is servicing the demand it meets.

From one market to many

The cloud era served primarily one customer profile: hyperscalers. The AI era adds at least five more—neoclouds, model developers, inference platforms, enterprises, and governments. Their requirements pull in different, and sometimes competing, directions. Each buyer brings a different demand and risk profile; each workload brings a different set of technical and geographic requirements. Where those two forces line up will help determine where the next wave of investment flows.

Different buyers, different demand and risk profiles

For most of the past decade, demand could be understood largely through cloud growth, with hyperscalers at the centre of gravity. AI has made the market more complex by ushering in a new set of buyers. Each has a different demand profile and risk mix. Hyperscalers underwrite scale but depend on speed. Neoclouds and model developers create dense GPU demand but face exposure to chip access and AI monetisation. Enterprises generate durable demand but have tighter integration and governance requirements. Governments anchor domestic capacity but at the pace of procurement and policy. All are sensitive to the same physical bottleneck: whether sufficient, reliable power can reach the right place at the right time. They share a commercial consideration, too. Unlike past infrastructure waves, in which spending fell once the network was built, this one requires capex to keep rising for decades. And that means the revenue AI generates also has to keep rising. If adoption stalls or pricing weakens, the later years of the build-out become harder to fund, and each buyer in this list will be exposed to that risk.

Different workloads, different geographic requirements

The workload matters just as much as the buyer. Traditional cloud capacity tends to follow population, GDP, enterprise density, and existing digital services markets. AI inference is more location-sensitive. As models move into critical enterprise workflows, latency, data access, privacy, security, and sovereignty requirements can pull capacity towards local data centres. AI training sits closer to the other end of the spectrum. It is compute-heavy but less tied to end-user proximity, so it can gravitate towards markets in any region that has cheap and reliable power, access to advanced chips, large-scale sites, technical talent, established AI ecosystems, and the ability to execute at speed. PwC estimates that about 30% of workloads today carry localised requirements, and that share is rapidly growing.

The result is a more uneven map of demand.

The uneven map of capex

What separates one region from another is not just how much demand they generate, but also how much demand they can actually capture. That depends on a set of factors different from the ones that shaped the cloud era.

Five forces will determine where capex flows. Power sits at the top of the list because affordable, reliable, and increasingly low-carbon electricity at scale is the hardest requirement for many markets to meet—and delivering it quickly is harder still. Transmission capacity, substation availability, and multiyear transformer lead times are becoming the bottlenecks that determine which projects break ground and when. Operators are increasingly bringing their own generation to sites, which helps individual projects but doesn’t remove the need for grid build-out at market scale. 

The next three forces—latency and connectivity, security and trusted-region hosting, and GPU access and ecosystem depth—largely sort markets by the type of workload they can credibly serve, whether that’s consumer inference, sovereign and regulated data, or frontier model training. 

Finally, policy certainty and community consent aggregate the other forces, turning underlying advantages into capacity that actually gets built—or blocking projects that the market would otherwise expect to be completed. 

Sustainability isn’t modelled as a separate variable in this analysis, but it runs through the forces above—sometimes as an advantage, sometimes as a constraint. Markets with renewable-heavy grids, cooler climates, and credible decarbonisation pathways carry a structural comparative advantage that the model picks up through their power positions. Markets with binding grid, land, or planning constraints carry disadvantages. 

The result of these forces is a picture in which capex shares diverge materially from a region’s economic weight, and the gap itself provides the most telling indicator of who is pulling ahead and who is falling behind. 

The Americas: The region accounts for $16.5 trillion in cumulative capex through 2050, with the US alone responsible for a whopping $15.1 trillion—around 48% of the global total and materially ahead of the region’s GDP share. The lead in AI infrastructure is wider than in any major industrial category since postwar manufacturing. That’s because the US remains central to the advanced-chip ecosystem and is home to the largest AI model developers, hyperscalers, and AI-native businesses. Talent, capital, and new ventures continue to cluster around that base, and facilitative state-level policy compounds the country’s lead. The Americas have the largest absolute uplift if AI accelerates, with cumulative capex through 2050 rising to $27.1 trillion, and the largest absolute shortfall if it doesn’t, reflecting the GPU-intensive composition of the US pipeline.

Behind the US, Chile and Canada emerge as the region’s clearest sustainability-anchored plays: Chile has competitively priced renewable energy underpinned by abundant solar resources, and Canada has grid stability and a renewable energy base that support northward expansion of US hyperscaler infrastructure.

Asia-Pacific: The region accounts for $8.2 trillion in cumulative capex, meaningfully below its share of global GDP. China and India are the largest sources of incremental demand, supported by large populations, rapidly expanding digital economies, and substantial headroom for AI to embed in business and consumer activity. China remains one of the few markets with the scale and strategic focus to support large-scale training workloads. The region has the widest proportional swing of any in the forecast, rising 69% under faster adoption and falling 34% under slower adoption. Outcomes within the region are uneven. Markets with deeper domestic demand and more diversified workloads, including Japan and Australia, prove more resilient; those most dependent on internationally mobile AI workloads are more exposed.

Europe: Europe is punching below its economic weight. The region’s $5.6 trillion in cumulative capex represents a share lower than its proportion of global GDP. The reasons are familiar and largely self-inflicted: power constraints, planning friction, and fragmented regulation across countries in the region. Amsterdam’s 2025 ban on new data centres, which cited land and grid limits, is illustrative rather than exceptional. The same factors mean Europe sees the smallest proportional uplift on acceleration, at 23%, because the central scenario is already supply-constrained. Stringent data-sovereignty rules across the EU help anchor a domestic floor, however, and parts of Europe—particularly markets close to advanced manufacturing equipment supply chains—retain a stronger position on chip access than the headline number suggests.

The Nordics are emerging as credible alternatives to more constrained Western European hubs, supported by renewable-heavy grids, climates that reduce cooling loads, and electricity prices that are 40–50% below those in other parts of Europe. Ireland’s Large Energy-User Action Plan signals the direction of travel elsewhere in the region: data centre siting is increasingly conditional on renewable integration and grid investment commitments.

The Middle East: Although the region’s $1.1 trillion in capex is small in dollar terms, it’s a share roughly in line with the region’s contribution to GDP. The Middle East’s share is also the fastest-growing on a CAGR basis, due to both its lower existing installed base of data centres and its ability to compress building timelines by aligning energy, capital, planning, and developer pipelines through a single coordinated front door. The build-out is GPU-heavy and orientated towards attracting internationally mobile workloads in addition to serving regional demand, an approach that delivers the largest CAGR in the central case but concentrates the region’s exposure to anything that disrupts global chip supply.

Africa: The region’s $255 billion in cumulative capex sits slightly below its share of global GDP, but the growth profile tells a more constructive story. Africa is the only region in this forecast whose central scenario isn’t a bet on AI. The region is largely buying foundational digital infrastructure that pays off regardless of which AI scenario plays out. That makes it, paradoxically, one of the lowest-risk capex stories on the map. South Africa anchors the region with the most established data centre base. Kenya, Nigeria, and Ghana are among the most promising emerging markets. Kenya has carved out a particularly distinctive position through a power grid that is approximately 95% renewable, becoming one of the most sustainably powered data centre markets in the world. 

In every region, the same underlying rule holds: power is the binding constraint, and policy is what determines how quickly it can be brought to bear. This raises the question of what happens if the geopolitical environment in which everything is being built becomes meaningfully more fragmented than the central case assumes.

Two geopolitical ‘what-ifs’

The central scenario projection assumes a reasonably open trading system in which chips move across borders, workloads are serviced where conditions are most favourable, and capacity is built according to comparative advantage. The demand cone described earlier depends on these assumptions and captures uncertainty about how much capacity will be built. Geopolitics introduces a different question—what happens if those assumptions don’t hold?

Two scenarios test the question from different angles. The first constrains the supply of advanced chips. The second constrains where critical workloads can be serviced. A more fragmented world would likely see elements of both. However, the scenarios redistribute capex along distinct fault lines, and each rewards and penalises a different set of regions.

What if chip flows become constrained?

In the first scenario we considered, the export controls already in place between the US and China escalate to peak trade war intensity and remain there. Advanced GPUs become harder to procure across a much wider set of markets, and retaliatory restrictions on critical raw materials propagate friction along the entire semiconductor supply chain. CPU supply, which is more diversified, is largely unaffected. The build-out slows and narrows, and although all regions feel the impact, country markets with secure chip access suffer least.

The headline impact is severe in the near term and partially recovers over time. Annual capex falls to roughly half the central scenario by 2030, before recovering as supply chains adapt. By 2050, annual capex exceeds the central case by 8% as the market catches up. But the cumulative shortfall over 2026–50 is around $6 trillion, reducing the global total from $31.6 trillion to $25.5 trillion.

The shock is felt first in markets with AI-heavy near-term pipelines, because GPU-intensive projects are the most exposed to constrained chip access. The Middle East is the most proportionally exposed region of all. Its cumulative capex falls 29%, and the impact is concentrated in Saudi Arabia, Qatar, and the UAE, where the pipeline depends most heavily on attracting internationally mobile AI workloads. China absorbs one of the greatest national shocks for the same reason, dragging the Asia-Pacific region’s cumulative capex from $8.2 trillion to $6.4 trillion. The Americas record the largest dollar reduction, falling from $16.5 trillion to $13.8 trillion as the US pipeline’s GPU weighting takes a near-term hit.

Recovery, when it comes, is uneven. Markets with depth in the chip supply chain bounce back fastest. The US recovers strongly enough that the Americas emerge with the smallest proportional loss of any major region, reflecting the depth of the US chip ecosystem and the policy capacity to support domestic semiconductor production. Within Asia-Pacific, the Taiwan region, Japan, and Singapore prove resilient on the strength of their supply chain roles. Europe’s cumulative capex falls from $5.6 trillion to $4.3 trillion, with Germany emerging as one of the more resilient national markets given its position in the precision equipment and materials that underpin advanced semiconductor fabrication. Africa isn’t a direct target of export controls, but owing to its limited semiconductor depth and weaker access to advanced ICT equipment, it still absorbs a meaningful share of the second-order disruption; its cumulative capex falls from $255 billion to $193 billion.

What if sovereignty becomes the organising principle?

The second scenario we considered tests a different fault line. In this scenario, demand for cloud, data, and AI capability remains broadly unchanged, but governments and regulated industries become less willing to rely on foreign infrastructure for workloads considered essential to national resilience, such as those for public services, financial systems, healthcare, and sovereign AI. Non-critical workloads continue to move freely. 

This time, total capex barely moves. Cumulative global capex through 2050 falls from $31.6 trillion to $29.5 trillion, a reduction of just 6.7%. The story is one of redistribution rather than reduction. Established global hubs lose part of the international servicing premium, whereas emerging markets that have significant domestic demand but limited current capacity build more than they otherwise would have.

The gains accrue to regions with deep domestic demand and underbuilt capacity. Asia-Pacific is the largest beneficiary in absolute terms; its cumulative capex rises 7% above the central scenario by 2050. India, Vietnam, Indonesia, the Philippines, and Thailand all record material uplifts, reflecting large domestic demand bases that have so far been serviced disproportionately from regional hubs. By contrast, Hong Kong SAR, Singapore, Japan, Australia, and South Korea record lower capex than in the central case, reflecting the loss of part of the regional and international workload they are expected to capture under the central scenario.

Africa records the largest proportional uplift of any region, as its cumulative capex rises roughly 12%, from about $255 billion to $284 billion, supported by onshoring of cloud storage in the near term and a developing AI market over the longer term. It is one of the few scenarios in which Africa is a net winner.

The losses fall on regions whose central scenario depends on servicing demand from elsewhere. The Americas record the largest absolute reduction, with cumulative capex falling from $16.5 trillion to $13.7 trillion as US-serviced cloud, data, and AI workloads are repatriated. The US alone loses around $2.9 trillion, though within the region, Brazil, Chile, and Canada gain on the back of stronger domestic servicing of critical workloads. 

The Middle East sees a more modest reduction, cumulative capex falling from $1.1 trillion to $1.0 trillion. Saudi Arabia, the UAE, and Qatar lose part of their workload-attracting premium as the region’s most internationally focused markets, while smaller Gulf markets such as Kuwait and Oman gain from domestic onshoring. The region as a whole is partially insulated because much of its near-term build-out is already orientated towards serving regional demand rather than attracting internationally mobile workloads. 

Europe tracks broadly with the central scenario, as stringent EU data sovereignty rules already mean that much of the region’s demand is serviced domestically and the marginal redistribution is therefore small. The regional total, however, masks meaningful internal redistribution: the United Kingdom records the largest European uplift on the scale of its public and financial sectors, and Türkiye and Poland gain from their greater need to host sovereign workloads domestically. Meanwhile, established hubs such as Ireland, the Netherlands, and Germany lose part of the cross-border share they would otherwise capture.

Positioning for the cycle

The analysis points to a cycle that is large, structural, and recurring, but unevenly distributed and vulnerable to geopolitical redistribution. For the institutions that are building capacity, committing capital, or shaping the conditions, the question is what to do with that picture. The recommendations below apply to both the demand cone and the two geopolitical ‘what-if’ scenarios.

For investors

• Underwrite data centres as a hybrid asset. The three layers—property, utilities, and semiconductor exposure—have different durations and different risk profiles. Obsolescence and residual-value risk on the semiconductor layer is the most frequently underpriced of the three.

• Look through the structure to where risks actually sit. Capital, development, leasing, and counterparty risk flow through the system differently. And the counterparty mix itself is now plural, spanning hyperscalers, neoclouds, sovereign vehicles, and regulated industries—each has a different durability across scenarios. Treating the asset as a single exposure misses these dimensions.

• Diversify away from concentration in established hubs. The largest hubs capture the most upside if AI acceleration plays out, but are also the most exposed to slower growth, chip-access constraints, and sovereignty-driven redistribution. 

For operators

• Decide which competition you are in. The two geopolitical scenarios in this analysis describe two distinct games. One is a global-hub competition for chip access, hyperscaler scale, and training workloads. The other is domestic and regional competition for sovereignty-anchored inference and foundational digital infrastructure. Success factors, off-takers, and design implications differ materially between them, and a strategy built for one will struggle in the other.

• Compete on speed to power. Power is the binding constraint under every scenario in this analysis, including the ones where total demand is lower. The operators that can secure megawatts faster than competitors will capture a disproportionate share of capacity regardless of which path the market takes.

• Design for optionality. Consider phased grid connections, modular capacity, and hardware-flexible cooling. Optionality is itself a hedge against scenario uncertainty: it preserves value if acceleration plays out and limits exposure to stranded capacity if it doesn’t.

• Sign long-dated leases that transfer obsolescence risk. The off-takers willing to sign lengthy leases—hyperscalers, sovereign vehicles, regulated industries—are also the counterparties whose demand is most resilient under slower growth and sovereignty-driven scenarios. Underwriting them well is the single most effective hedge available at the asset level.

For policymakers

• Treat data centres as national strategic infrastructure. The most successful jurisdictions coordinate energy, capital, planning, and developer pipelines through a single front door rather than leaving them to parallel private actors. The Middle East’s policy coordination is the working illustration. Other markets may compete through tax incentives, accelerated approvals, subsidised land, or targeted support for digital infrastructure. What matters is the combined policy package. Jurisdictions that make data centre development faster, more predictable, and more investable are more likely to pull ahead.

• Build the capability to pre-procure long-lead power and grid components. Transmission equipment, substations, and transformers now carry multiyear lead times that determine which projects break ground and when. Pre-procurement is one of the few timeline tools available that can compress that bottleneck—but it should be calibrated to demand visibility, since pre-procurement without a credible pipeline produces stranded grid investment of its own.

• Plan for a low-carbon build-out from the start. The markets that will hold long-run advantage are those where renewable and firm low-carbon power capacity scales alongside data centre siting, rather than lagging behind it. The Nordic markets, Chile, and Canada show what that can look like—competitively priced, clean power that becomes a comparative advantage rather than a constraint.

The $31.6 trillion question isn’t whether the capital exists. It does. Nor is the question whether the demand is real. It is. The question is which regions, operators, and institutions are positioned to capture it and which aren’t. Investors who price the chip replacement cycle rather than the building the chips sit in, operators that pick a lane between global hyperscale and sovereign regional, and governments that treat electricity as the strategic asset it has become—they are the ones that will win. The hubs coasting on yesterday’s advantages, the regions whose regulators are still drafting frameworks for the cloud era, and the capital allocators underwriting data centres as though they were warehouses will discover that the AI build-out isn’t a rising tide. The map of global compute in 2050 is one of redistribution based on such factors. It will look very little like the map of 2026, and the difference between the two will be measured in trillions of dollars.