PwC Puts $31.6 Trillion Price Tag on Global AI Data Center Build Through 2050

The number that keeps circulating this month is $31.6 trillion, and it deserves more than a quick read.

On September 2, PwC released what it calls the inaugural Global Data Centre Outlook, a long-range capital expenditure model covering 46 countries and territories through 2050. The modeling was commissioned from Oxford Economics. According to the PwC press release, annual data center spending is forecast to rise from roughly $800 billion this year to $1.8 trillion by 2050, with the U.S. capturing nearly half the cumulative total at $15.1 trillion.

Those top-line figures are striking enough, but the structural argument underneath them is what makes the report worth reading carefully. Most infrastructure booms front-load spending and then taper. Railways, electrification, fiber rollout all followed that pattern. PwC argues this one won't, because the thing driving the spending isn't construction. It's hardware refresh cycles.

As Data Center Frontier noted in its coverage of the report, the economics are shifting toward recurring GPU and server replacement every four to six years. That's the architectural fact hiding inside the headline number. If PwC's model is right, you're not building toward a finish line. You're committing to a perpetual capex cycle that resets on a schedule tied to chip generation timelines.

The report also maps three scenarios, and the spread between them is wide enough to matter. The central case lands at $31.6 trillion. An accelerated adoption scenario, per reporting by Quartz, pushes that figure toward $50 trillion. A disrupted trade scenario, in which chip supply chains face tighter export controls, cuts cumulative investment by close to a fifth, settling at roughly $25.5 trillion before supply chains adapt.

That trade scenario is the one I'd watch. The report's central case, as PwC acknowledged in the release, assumes a relatively open trading system where semiconductors move freely across borders. That assumption is doing a lot of work. Right now it's a debatable premise.

On geography: the Asia-Pacific region is projected at $8.2 trillion cumulative, led by China and India, with Europe at $5.6 trillion. According to the Taipei Times' coverage of the report, sovereign AI strategies are accelerating investment in Europe and the Middle East, which means the regional split is as much a policy question as a demand question.

The binding constraint that PwC names, in every region, is power. Not chips, not land, not permits, though all three matter. Power. Affordable, reliable, low-carbon electricity at scale is, as the report puts it, the hardest requirement for most markets to meet. That's a conclusion that has real implications for grid policy and for the siting decisions data center operators are making right now.

PwC's Clara Cutajar, the firm's global infrastructure leader, described AI infrastructure in the report's press release as "one of the defining capital allocation challenges of the next generation," cutting across technology, energy, real estate, supply chains, and regulation.

That framing is accurate and worth taking seriously. The question for anyone in infrastructure is whether they're treating power procurement as a first-order engineering problem or still as a procurement afterthought. The data, and this report, says it's the former.

Sources cited:
- PwC Global Data Centre Outlook (press release) (https://www.pwc.com/gx/en/news-room/press-releases/2026/global-investment-in-ai-infrastructure.html)
- Data Center Frontier (https://www.datacenterfrontier.com/machine-learning/article/55403120/pwc-maps-316-trillion-ai-data-center-buildout-through-2050)
- Quartz (https://qz.com/pwc-data-center-ai-investment-forecast-2050-090226)
- Taipei Times (https://www.taipeitimes.com/News/biz/archives/2026/09/04/2003863640)
- TheStreet (https://www.thestreet.com/technology/pwc-ai-data-center-capex-forecast-power-constraint)