AI capex set to hit $31.6tn by 2050 as chips overtake concrete - Capacity

Standfirst: Global spending on AI infrastructure is on course to reach $31.6tn by 2050, according to new baseline projections from PwC’s Global Data Centre Outlook, published this week. The figure marks the first long-range capex forecast of its kind, running through to mid-century rather than the usual five-year horizon, and it lands at a moment when the industry is already grappling with the practical limits of power, permitting and public consent. Annual data centre capital expenditure is expected to rise from around $800bn in 2026 to $1.8tn by 2050.

The United States is forecast to take almost half of that spend, at $15.1tn, with Asia Pacific accounting for a further $8.2tn led by China and India, while sovereign AI strategies pull increasing volumes of capital into Europe and the Middle East. What sets this cycle apart from previous infrastructure booms, PwC argues, is that it does not taper off once the buildings go up. ICT equipment, chips, servers and networking gear, is expected to grow from 70% of total investment today to 93% by 2050, as operators replace hardware every few years while the shells around it last for decades. “AI infrastructure is becoming one of the defining capital allocation challenges of the next generation,” said Clara Cutajar, PwC Australia’s global infrastructure leader. “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.” A mismatch the industry is already pricing in PwC’s equipment-versus-construction split echoes a problem Capacity has heard articulated directly by financiers on the ground. Speaking at Datacloud USA in Austin this week, Jonathan Mauck, senior managing director at Digital Bridge Holdings, described data centres as 20-year industrial infrastructure housing GPUs with a useful life of five to seven years. “You’re effectively a 20-year creditor,” he said, pointing to the risk facing anyone who commits capital to a facility on the assumption a tenant will keep paying rent over two decades against hardware that needs replacing several times over. That mismatch is precisely what Nvidia’s recently announced financing platforms with Apollo, BlackRock, Blackstone, Brookfield, Goldman Sachs and KKR are designed to solve, turning GPU capacity into a financeable asset class in its own right, separate from the buildings that house it.

PwC’s forecast suggests that trend has further to run: as the equipment share of spend keeps climbing, so does the pressure on financing structures built around chip refresh cycles rather than building lifespans. Power, and increasingly fibre, decide where the money goes PwC names power as the chief factor shaping where AI infrastructure investment flows globally, ahead of connectivity, security, policy certainty, community consent and GPU access. “The AI buildout is not a rising tide that will naturally lift all boats,” Cutajar said. “Capturing this investment requires active positioning. Investors should recognise data centres as hybrid assets with a complicated risk profile.” That framing broadly tracks what Capacity has heard from the industry itself.

Andy Lipman of Morgan Lewis told delegates at Metro Connect Fall this week that power is now the binding constraint on US data centre growth, with the average facility more than tripling in size since 2020 and electricity consumption from data centres set to climb sharply by 2030. There is a wrinkle worth flagging, though. A separate keynote panel at the same event found that fibre availability, not power, is now viewed as the single biggest constraint on new builds, a reversal moderator Andrej Danis of AlixPartners described as a shift from just a few years ago.

It is not a direct contradiction of PwC’s finding so much as a difference in scope: PwC is modelling where global capital lands over 25 years, while the Metro Connect panel was polling site-selection priorities for projects being planned right now. Executives weighing near-term development decisions may find permitting timelines for fibre routes just as decisive as grid connection dates. Trade policy could redraw the map, not shrink it PwC also tested two scenarios against its baseline forecast.

In the first, tighter export controls disrupt chip supply chains badly enough that annual investment halves relative to the central forecast by 2030, before gradually recovering. Even with that recovery, cumulative investment through 2050 lands around $25.5tn, roughly $6tn below the baseline. The second scenario is less about shrinking the total and more about where it goes.

Greater emphasis on digital sovereignty barely dents overall spending, PwC found, but shifts capital toward countries with strong domestic demand and underdeveloped data centre capacity, as governments and regulated industries prioritise local infrastructure over imported capacity. That tallies with what Capacity reported from the G7 summit in Évian earlier this year, where sovereign AI infrastructure emerged as one of the defining phrases of 2026, with the US, EU and Canada each pursuing distinct approaches to chip access, cloud infrastructure and procurement. PwC’s modelling suggests that fragmentation is less likely to shrink the overall pool of AI capital than to redirect an increasing share of it away from the US and toward markets asserting greater control over their own compute.

What it means for operators and investors For data centre executives, the practical takeaway is less about the headline $31.6tn than about what it implies for underwriting. If equipment costs come to dominate lifetime capex, financing structures, lease terms and depreciation assumptions built around the building rather than the chips inside it will increasingly look out of step with the asset’s real economics. Power access remains the clearest differentiator between markets that capture this investment and those that do not, but fibre availability, permitting speed and community consent are closing the gap fast enough that none of them can be treated as secondary.

RELATED STORIES Nvidia now agrees to rent back unused GPU capacity from neocloud operators if customer demand falls short CoreWeave’s debt hits $35bn: What it means for the neocloud refinancing wall What is circular financing in AI infrastructure, and should telecoms and data centre operators be worried? What is a neocloud, and why does it matter for digital infrastructure? Metro Connect USA 2027 08 February 2027 Metro Connect USA is the largest executive-level digital infrastructure event in the U.S. The only one of its kind, this 25-year-strong gathering is where decision makers come together to make deals happen.