Building the world's AI infrastructure will require $31.6 trillion in investment between now and 2050, according to modelling commissioned by PwC from Oxford Economics across 46 countries. The projection, published this week, puts annual capital expenditure on the trajectory to nearly double from roughly $800 billion this year to $1.8 trillion by mid-century.
The figure lands amid a heated debate about whether AI data center spending resembles a productive buildout or a speculative bubble. PwC's numbers do not settle that argument, but they quantify its scale with unusual precision — and they point to a structural shift in what a data center actually is as a financial asset. For more context on this story, see our ongoing AI trends.
The US Dominates the Spending Map
The United States accounts for $15.1 trillion of the projected total, or 48 percent of global spending over the 24-year horizon. Asia Pacific follows with $8.2 trillion, led by China and India, while Europe and the Middle East make up the remainder.
That concentration means nearly half of the world's AI infrastructure investment flows into a single country — a fact that will not be lost on European policymakers, who have watched sovereign AI strategies struggle to translate ambition into delivered capacity.
PwC describes Europe as a region where sovereign AI strategies are driving growing investment, though that is a considerably more modest claim than the US share implies. The continent's €30 billion gigafactory programme, its largest coordinated response so far, has been running into delays.
Equipment, Not Buildings, Is Where the Money Goes
Perhaps the more consequential number in the report is not the headline total. Equipment currently represents about 70 percent of data center capital expenditure, and PwC projects that share will rise to 93 percent by 2050.
That shift changes what a data center looks like as an asset. A building can depreciate over decades and be financed with long-dated, low-interest debt. A rack of AI accelerators can become obsolete within a few years and must be funded through cash flow or debt priced against much shorter horizons.
The distinction also matters for how public and private spending should be compared. Sovereign AI budgets are typically intended to create capacity for research, defense and public administration rather than commercial cloud services, which means they cannot simply be stacked on top of hyperscaler capex when tallying national investment races.
PwC describes data centers as "hybrid assets" — an acknowledgment that they no longer fit neatly into traditional property or equipment categories. Infrastructure funds, which typically buy long-lived assets with predictable cash flows, may find that a facility needing substantial re-equipment every five years does not fit their mandate as cleanly as an office tower or a toll road.
"AI infrastructure is becoming one of the defining capital allocation challenges of the next generation," said Clara Cutajar, PwC Australia's global infrastructure leader, in the report.
The Bottleneck May Not Be Money at All
The projection arrives alongside a separate reality check: the constraint on AI buildout is increasingly physical rather than financial. Transformer lead times are measured in years. Grid connection queues in Texas and Denmark stretch far into the future. Cooling equipment and planning approvals all move more slowly than capital.
As The Next Web's analysis of the report put it, the industry has plenty of money — it is waiting for the physical infrastructure to catch up. None of those bottlenecks can be solved by increasing a capex forecast.
The PwC figures are broadly consistent with other estimates. McKinsey has separately projected nearly $7 trillion in data center investment by 2030, a number that lines up with PwC's trajectory given the different time horizons involved.
Read Long-Horizon Projections With Care
A 24-year model for capital expenditure on a technology that has only existed commercially for a few years rests on assumptions about demand, chip prices and the continuation of the current investment cycle — none of which can be known with much confidence.
There is also a disclosure worth noting: PwC, like other consultancies producing infrastructure research, advises companies and investors on the very transactions and projects covered by that research. It is not a disinterested observer.
Still, two numbers from the report are worth holding onto. The first is $800 billion — roughly what the world is already spending on AI infrastructure per year, a figure that grounds the projection in observable reality rather than speculation. The second is 48 percent, the share of the next quarter-century's spending headed for one country.
The first number describes an industry that has already been built. The second describes the geopolitical shape of everything still to come — and it is likely to be the harder figure for policymakers outside the United States to digest.
If equipment reaches 93 percent of data center capital expenditure by 2050, as PwC projects, the financing models behind the AI boom will have to change with it. Buildings can be financed over 30 years. Chips cannot.
---
Stay Ahead of AIGet the latest AI news, analysis, and breakthroughs — all in one place.
Read more AI news →
