Building the world’s AI infrastructure will cost $31.6 trillion between now and 2050, according to modeling commissioned by PwC from Oxford Economics across 46 countries. Annual capital expenditure is projected to rise from $800 billion this year to $1.8 trillion by 2050.
The United States accounts for $15.1 trillion of that spending, or 48%; next in line is Asia Pacific with $8.2 trillion, led by China and India, while Europe and the Middle East make up the remainder.
The more interesting figure is not the total. Equipment currently represents about 70% of data centre capital expenditure, rising to 93% by 2050. That changes what a data centre actually looks like as an asset.
A building can depreciate over decades, while a rack of AI accelerators can become obsolete within a few years. A business whose costs are 93% equipment starts to look much less like a property business, whatever its balance sheet says.
“AI infrastructure is becoming one of the defining capital allocation challenges of the next generation,” said Clara Cutajar, PwC Australia’s global infrastructure leader. The firm describes data centres as hybrid assets, which is a polite way of acknowledging that they do not fit neatly into traditional categories.
The shift also changes who can afford to own them. Buildings can be financed over 30 years at relatively low rates, while equipment that needs to be replaced every few years has to be funded through cash flow or debt priced against much shorter time horizons.
Europe’s position in the headline numbers is the part worth considering. PwC describes the continent as a region where sovereign AI strategies are driving growing investment, which is a much smaller claim than saying Europe will account for 48% of global spending.
That matches what has been happening on the ground. Europe’s €30 billion gigafactory programme is its largest coordinated response, but it has been running into delays.
Sovereign AI spending is also different from hyperscaler capex. Public money is intended to create capacity for research and public administration rather than commercial cloud services, and the two cannot necessarily be treated as interchangeable when comparing investment totals.
Projections covering this kind of timeframe deserve some skepticism. A 24-year model for capital expenditure in a technology that has only existed commercially for a few years requires assumptions about demand, chip prices, and the continuation of the current investment cycle, none of which can be known with much confidence.
PwC is also not a disinterested observer in the conventional sense. Like other consultancies producing infrastructure research, it advises companies and investors on the transactions and projects covered by that research.
The Practical Challenges
What the modeling does provide is a useful sense of scale for something that is already visible. McKinsey has separately projected nearly $7 trillion in data centre investment by 2030, and the figures are broadly consistent given the different time horizons.
The 93% figure may prove more durable than the $31.6 trillion total. Chip generations are getting shorter rather than longer, and operators are finding that some of the most expensive components of an AI data centre are also the ones that become outdated fastest.
The main constraint may not be capital at all. Transformers, grid connections, cooling equipment, and planning approvals all move more slowly than money, and none of them can be solved simply by increasing a capex forecast.
Grid connection queues in Texas and Denmark, transformer lead times measured in years, and a European gigafactory programme running late are different versions of the same problem. The industry has plenty of money; it is waiting for the physical infrastructure to catch up.
Cet article a été rédigé avec l'assistance de l'IA.
News Factory APP - actualités agentiques pour booster votre SEO et AEO.