Somewhere in the last few weeks, big tech’s AI habit became a trillion-dollar one. Add the four largest hyperscalers’ 2026 capital budgets to the $500bn Stargate build-out led by OpenAI, SoftBank and Oracle, and combined AI infrastructure spending has cleared $1tn for the first time. Wall Street had pencilled that figure in for 2027, yet it arrived a year early.
Amazon, Alphabet, Microsoft and Meta are now committing roughly $725bn between them this year alone, up 77% on 2025 and more than triple what they spent in 2024. Bank of America has Amazon at $200bn, Microsoft at $190bn, Alphabet at $185bn and Meta at $135bn, and every one of those figures has been revised upward at least once since January.
The billion-dollar earnings calls
On Meta’s first-quarter call, chief executive Mark Zuckerberg told analysts, “We are increasing our infrastructure capex forecast for this year,” pointing to rising memory-chip costs and additional data centre commitments rather than any change of strategy. Meta’s guidance has since climbed further, to a range of $130bn to $145bn, even as free cash flow fell from $26bn to $1.2bn year on year and the shares dropped sharply in after-hours trading.
That pattern, spending rising while cash generation tightens, is now general across the sector. Analysts at Jefferies wrote in a note to investors that “cap-ex continues to soar as demand outpaces supply and pricing increases,” a reasonably tidy summary of why nobody expects the number to plateau soon.
Alphabet’s cloud backlog has surged past $240bn, Microsoft is sitting on an $80bn queue of Azure orders it cannot yet fulfil, and Amazon’s AWS growth has accelerated rather than slowed. Demand, in other words, is the easy part of this story to justify. Funding it is proving harder, and the sums involved help explain why: Goldman Sachs estimates cumulative hyperscaler capex could reach as much as $7.6tn between 2025 and 2030, while Citigroup has raised its own five-year forecast to $2.8tn.
The mechanics of paying for that are starting to show up in credit ratings rather than earnings decks. Capacity has tracked how Meta’s move to fund its El Paso data centre through off-balance-sheet special purpose vehicles has pushed bond yields north of 7%, a marked repricing from terms it secured only months earlier, while Oracle has taken the opposite path of raising debt directly and has been downgraded to one notch above junk as a result. Morgan Stanley now puts the sector’s off-balance-sheet obligations at close to $1.65tn. None of that is stopping the build-out. It is changing who ultimately carries the risk if the returns take longer to arrive than the spending suggests.
Power, not chips, is the constraint now
Ask anyone building at scale what actually limits them in 2026, and the answer has shifted away from GPU supply and toward electrons. “It’s not just the quantity of power, it’s the density,” one financing executive told a Metro Connect USA panel earlier this year, describing how server rack loads have risen from around 3kW for general-purpose compute to as much as 150kW for AI inferencing, a jump the cooling and power distribution infrastructure built for the last decade was never designed to absorb.
That constraint is reshaping where and how hyperscalers build. JLL expects global data centre capacity to nearly double from 103GW today to around 200GW by 2030, with hyperscalers alone accounting for roughly $1tn of that spend between 2024 and 2026. But the industry consensus that this represents genuine, durable demand rather than speculative excess is not universal.
As one industry executive put it at the same Metro Connect panel, “this is not a bubble, it’s fed by real demand.” Others are less convinced the financing structures underpinning it are as sound as the demand case itself, and the Bank for International Settlements has named circular financing arrangements between hyperscalers, neoclouds and construction lessors as one of three biggest risks to global financial stability in its 2026 Annual Report.
Regionally, the Middle East is one of the clearer beneficiaries of the reallocation. Gartner expects technology spending across MENA to reach $169bn this year, and Ciaran Delaney, chief operating officer at EXA Infrastructure, has said data centre systems specifically would see “the most significant growth over the course of the year” as Gulf states position compute capacity at the centre of sovereign economic strategy.
The construction bottleneck nobody priced in
Capital and power grab the headlines, but the more immediate constraint for anyone with a campus on the ground may be far more mundane: there simply are not enough contractors who can build the thing. Fewer than ten firms worldwide are considered capable of taking a hyperscale project from groundbreaking to live operation, and that shortlist is doing an extraordinary amount of heavy lifting.
In the US, it comes down to AECOM, Turner Construction, DPR Construction, Holder Construction and Skanska USA. In Europe, the picture narrows further, with Skanska again among the handful of firms able to deliver mission-critical facilities at scale.
The backlog numbers make the squeeze tangible. Turner reported a record $44.3bn project backlog at the end of 2025, with an estimated 37% to 40% of that tied directly to data centre work. AECOM has delivered more than 11GW of data centre capacity across 45 countries over 25 years, while DPR Construction has built over 1.5GW of critical load capacity since 1997, and mission-critical infrastructure now accounts for roughly 31% of its revenue.
What stands out is not the individual scale of any one firm, but how often the same small set of names turns up on Microsoft, Google, Amazon and Meta’s project lists simultaneously, across the US, the Middle East, Southeast Asia and Europe, often running near capacity on several continents at once.
That concentration matters because it caps how fast the trillion-dollar spending commitment can actually turn into finished, energised capacity. A hyperscaler can approve a budget in a single boardroom sitting. It cannot conjure a second Turner Construction.
Specialist skills in mechanical, electrical and plumbing installation, high-voltage switchgear commissioning and precision cooling are not fungible with general commercial construction labour, and training pipelines for that expertise take years, not quarters. For operators and investors watching capex guidance rise every earnings season, the contractor bottleneck is arguably a better read on realistic delivery timelines than any capital figure on a slide. Capacity has examined the risk in more detail here, including how exposed the build-out becomes if even one or two of those firms hit capacity limits or pull back.
Local and political resistance compounds the problem rather than sitting apart from it. Oracle’s dispute over a $7bn Wisconsin electricity bill has become a warning shot for data centre finance chiefs well beyond that one project, arriving as roughly $1.2tn in AI infrastructure investment is planned across five major companies for next year. New York’s moratorium on new hyperscale connections, driven by nearly 12GW of interconnection requests sitting in the grid queue, adds a further layer: even where contractors and capital are available, permitting and grid capacity now increasingly decide the pace of delivery.
None of which suggests the trillion-dollar threshold marks a ceiling. It looks more like a staging post. The hyperscalers show no sign of treating this spending as discretionary, and the infrastructure, financing and construction markets that support them are being restructured in real time to keep up. The question worth watching through the rest of 2026 is not whether the spending continues, most indicators say it will, but which link in the chain, power, contractors or capital markets, buckles first under the weight of it.
Related news
Why is contractor concentration becoming a major risk for AI data centre growth?
Meta’s $12bn El Paso bond exposes AI debt repricing
Oracle’s $7bn Wisconsin bill is a warning shot for data centre CFOs
Global data centre capacity set to nearly double by 2030 as AI drives $3tn investment boom
Data centres at a crossroads: Can capital keep up with demand?
New York data centre moratorium: Who will get off unscathed?
