The $31 Trillion AI Data Center Boom: Inside PwC’s 2050 Forecast

PwC published its Global Data Centre Outlook on 2 September 2026, projecting cumulative global data centre capital expenditure of $31.6 trillion through 2050.

For scale: US annual GDP is around $30 trillion. PwC’s own framing is that the buildout dwarfs railways, electrification, and the internet — each of which required enormous capital and defined an era.

The headline number is doing the rounds. It is also the least interesting finding in the report.

What the Report Actually Says

The core figures, from PwC’s release and the accompanying coverage:

The forecasts were commissioned from Oxford Economics and cover 46 countries and territories across five regions.

Note the range before the point estimate. The central scenario sits inside a plausible band running from roughly $22 trillion to nearly $50 trillion — a spread of more than two to one. Most coverage is reporting $31.6 trillion as a finding. It is the midpoint of a very wide distribution, and the report is explicit about that.

The More Important Number

The structurally significant finding is not the total. It is what the money is being spent on.

Recurring chip upgrades, not construction, drive the majority of long-term capital investment. Servers, GPUs, and other ICT equipment require refreshing every four to six years, which means this is not a buildout that completes. As PwC puts it, the cycle resets every four to six years with no sign of ending.

Reporting on the underlying figures puts equipment at around 93% of costs. That single proportion changes what a data centre is.

A building depreciates over decades. A rack of AI accelerators can be obsolete within a few years. A business whose costs are overwhelmingly equipment is not a property business, whatever its balance sheet says — and PwC’s own description of data centres as “hybrid assets” is a polite acknowledgement that they do not fit existing categories.

Why That Classification Matters

This is where the finding becomes commercially consequential rather than merely interesting.

Buildings can be financed over thirty years at relatively low rates. Equipment that must be replaced every few years has to be funded from cash flow or debt priced against much shorter horizons.

That difference determines who can afford to own these assets. Infrastructure funds and REITs are structured around long-lived, slowly depreciating assets with predictable yields. An asset base that turns over every half-decade demands a different capital structure entirely, and a different tolerance for reinvestment risk.

If the 93% figure holds, it will outlast the $31.6 trillion headline. Chip generations are getting shorter rather than longer, and the most expensive components of an AI data centre are frequently the ones that obsolesce fastest.

The Scenarios

PwC modelled two alternatives to the central case, both worth understanding because they point in different directions.

Tighter export controls. Chip supply chains face disruption, annual investment falls to roughly half the central forecast by 2030, then gradually recovers as supply chains adapt. Cumulative investment through 2050 lands near $25.5 trillion — around $6 trillion, or close to a fifth, below the central case.

Digital sovereignty. Greater emphasis on trusted domestic infrastructure does not shrink the total. It redistributes it, pushing capital toward markets with substantial local demand and limited existing capacity.

The distinction is useful. Trade disruption reduces the pool; sovereignty reshuffles it. Anyone modelling exposure to this cycle should treat those as separate risks rather than a single geopolitical variable.

Power Is the Binding Constraint

PwC identifies power availability as the foremost factor shaping where investment flows, alongside data sovereignty requirements and semiconductor trade.

That deserves emphasis, because it points at the report’s main unstated tension: the constraint may not be capital at all.

Transformers, grid connections, cooling equipment, and planning approvals all move considerably more slowly than money does, and none of them can be resolved by raising a capex forecast. A projection of $1.8 trillion in annual spending by 2050 assumes the physical world can absorb it.

There is early evidence it may not, at least smoothly. Reporting on the first quarter of this year cited at least 75 projects stalled by local opposition, representing roughly $130 billion in investment. That figure comes from a single source and is worth treating as indicative, but the direction is consistent with what utilities and planning authorities have been signalling for two years.

The Comparison PwC Invites

PwC frames the buildout against railways, electrification, and the internet — each an era-defining capital cycle. The comparison is deliberate and mostly fair on scale, but it breaks in one important place.

Those earlier cycles were front-loaded. You laid the track, strung the grid, or trenched the fibre, and the asset then produced returns for decades with maintenance rather than replacement. Capital intensity fell sharply once the network existed. That is why railway and utility assets became the archetype for long-duration infrastructure investment in the first place.

The AI cycle inverts that. The physical shell behaves like traditional infrastructure — long-lived, slow to permit, expensive to build. The contents do not. On PwC’s own reading, the contents are where most of the money goes, and they turn over roughly every half-decade.

So the honest version of the analogy is that this resembles railways in scale, and something closer to a manufacturing capital cycle in structure. That distinction is not pedantic. It determines whether the spending curve eventually flattens, as every previous infrastructure cycle did, or whether it behaves as PwC projects and keeps climbing through 2050.

The report takes the second view. It is a defensible position and it is also the single assumption on which the entire headline figure rests.

Read the Forecast With the Author in Mind

Worth stating plainly and without cynicism: PwC, like other consultancies producing infrastructure research, advises companies and investors on the transactions and projects its research covers.

That does not make the modelling wrong. Oxford Economics is a credible modelling partner, and the figures are broadly consistent with independent work — McKinsey has separately projected nearly $7 trillion in data centre investment by 2030, which reconciles reasonably with PwC’s trajectory given the different horizons.

But a twenty-four-year forecast is a scenario, not a measurement. Treat it as a well-constructed sense of scale for something already visible, rather than as a number to plan a balance sheet around.

What It Means Depending on Who You Are

Investors. The asset-classification question is the live issue. Equipment-heavy, fast-depreciating assets do not behave like the infrastructure most infrastructure capital is structured to hold. Underwriting assumptions built on thirty-year depreciation will not survive contact with a four-to-six-year refresh cycle.

Operators. The refresh cadence is the business model, not an overhead. Sites are competing for grid capacity years before they compete for tenants, which makes power procurement a strategic function rather than a facilities one.

Enterprises buying capacity. Regional concentration matters for cost and for sovereignty compliance. Europe’s $5.6 trillion share against the US $15.1 trillion tells you where capacity, and therefore pricing leverage, will sit.

Policymakers. The sovereignty scenario says explicitly that domestic-infrastructure priorities redirect capital rather than destroying it. That is an unusually direct argument that industrial policy in this sector can work, and it will be quoted accordingly.

Final Thoughts

The $31.6 trillion figure will be repeated for the next year, usually without its range and usually without its assumptions.

The findings likelier to matter are narrower. Data centres are equipment businesses wearing property clothing. The spending never completes because the hardware obsolesces. And the ceiling on all of it is more probably a transformer queue than a shortage of willing capital.