Every era-defining build-out gets measured against the ones before it. Railways. Electrification. The internet. PwC's new Global Data Centre Outlook puts a number on the AI cycle and it dwarfs all three: $31.6 trillion in data centre capital expenditure through 2050. The catch that makes it a paradox rather than a prophecy is inside the number. Most of that money buys machines that stop being competitive within four to six years. So the build-out never ends. It only resets.
Annual spend rises from ~$800B in 2026 to $1.1T in 2030 and $1.8T by 2050; by mid-century 93% of the total goes to ICT equipment, not buildings.
Power availability, chip trade flows and data sovereignty decide which regions capture the build-out, not capital availability.
The report, published on 2 September by PwC with modelling from Oxford Economics across 46 countries and five regions, is the first long-range capex forecast for the sector. Its central scenario assumes relatively open chip trade. Its two stress scenarios move trillions on policy choices alone. That makes the document less a prediction than a map of where the money is exposed.
AI infrastructure build-out
Cumulative capex under PwC's central scenario, rising from ~$800B a year in 2026 to $1.8T in 2050. · PwC Global Data Centre Outlook, 2026
The capex cycle that resets every four to six years
Railways, electrification and the internet front-loaded their capital and then tapered as the network matured. The AI infrastructure cycle does the opposite. Annual spending accelerates, from roughly $800 billion in 2026 to $1.1 trillion in 2030 and $1.8 trillion by 2050, because the bulk of the money does not build the building. It fills it.
The firm estimates that every $1 of construction capex commits the market to roughly $12 of future ICT equipment spend. Servers, storage, networking, the GPUs. Each ages out in a handful of years. Over a 20-year data centre asset life, the facility can absorb three to five rounds of ICT investment. The economic life of the building and the financial life of the hardware inside it have effectively decoupled.
Capex composition shift
Every dollar of shell-and-site construction commits ~$12 of future ICT spending; equipment rises from 70% of capex today to 93% by 2050. · PwC, 2026
For an investor the composition thesis matters more than the headline. If chip refresh dominates long-term spend, the companies that control the supply chain and the hardware refresh capture the majority of this capital flow, not the companies pouring concrete. The AI infrastructure story at this scale is a semiconductor and supply chain story wearing a real estate costume.
Why the refresh cycle never ends
Power decides where the money lands
The US captures almost half of the central-scenario total, $15.1 trillion or 48%. Asia Pacific follows at $8.2 trillion, Europe at $5.6 trillion, the Middle East at $1.1 trillion and Africa at $255 billion. China and India drive the largest share of incremental demand. But the report is explicit that capital is not the constraint. Affordable, reliable, increasingly low-carbon electricity at scale is the hardest requirement most markets cannot meet.
The report names five factors that direct where investment flows, with power first: transmission capacity, substation availability and multiyear transformer lead times increasingly decide whether projects break ground at all. On-site or behind-the-meter generation can help individual projects but does not remove the need for grid expansion where entire markets try to add gigawatts of load. Renewable-heavy grids and cooler climates, the Nordics being the clean example, become structural advantages for data centre siting.
AI infrastructure is becoming one of the defining capital allocation challenges of the next generation. It cuts across technology, energy, real estate, supply chains, regulation and financing. This changes how infrastructure investors need to think about capital requirements, risk and returns.— Clara Cutajar, Global Infrastructure Leader, PwC Australia
Two policy scenarios that move trillions
The headline number assumes chips move freely across borders. The report models what happens if they do not. If the export controls already in place escalate to peak trade-war intensity and stay there, advanced GPUs become harder to procure across far more markets and retaliatory restrictions on critical raw materials propagate friction through the entire semiconductor chain. Cumulative capex falls to $25.5 trillion by 2050, a shortfall of roughly $6 trillion. Annual spend drops to about half the central case by 2030 before recovering.
The second scenario is quieter and more structural. If governments and regulated industries refuse to host essential workloads on foreign infrastructure, cumulative capex barely moves, from $31.6 trillion to $29.5 trillion. The story is redistribution rather than reduction. Asia Pacific gains about 7% above the central case because it combines deep domestic demand with underbuilt capacity. Established global hubs lose part of their international servicing premium.
The difference between the two maps of 2050 is measured in trillions of dollars, and both are policy outcomes rather than technology outcomes.
What happens to the capex cycle a decade from now?
Probability: 75% — the ICT share of capex already rises from 70% today toward 93% by 2050 under PwC's model, and refresh cycles of four to six years are shorter than any prior infrastructure era.
✅ Arguments for
Confirmation criteria: hyperscaler 2027 guidance holds at or above 2026 levels while GPU and accelerator refresh cadence stays at four to six years.
❌ Arguments against
Disconfirmation criteria: interconnection queue lengths stop growing, or announced capacity keeps sliding to later in-service dates.
Development scenarios
🟢 Optimistic scenario (35%)
Implications: GPU and accelerator vendors, power equipment makers and grid developers capture an outsized share of a multi-decade capital wave.
🟡 Base-case scenario (50%)
Implications: returns concentrate in the recurring hardware refresh layer and in companies that can secure power ahead of competitors.
🔴 Pessimistic scenario (15%)
Implications: regions with secure chip access and domestic capacity gain relative advantage; projects without firm power commitments face the deepest delays.
Hyperscaler 2027 capex guidance versus 2026, since a step-down would be the first sign the refresh cycle is compressing.
Interconnection queue lengths and transformer lead times, the fastest public proxy for real deployable capacity.
Whether sovereign funds and national champions begin citing the $31.6T figure in public capital commitments, which would show the projection is steering real allocation.
Data centre power deals shifting from grid-dependent to on-site generation, the leading edge of the power constraint.
As we wrote in August, the value in the AI infrastructure stack sits in the layers closest to the silicon. The outlook quantifies that intuition at trillions of dollars. This build-out resets every four to six years, which makes it a permanent procurement machine for the compute industry rather than a construction cycle. The winners are whoever controls the hardware refresh and whoever secures the power first.