Description
The data centre sector has entered the steepest capacity expansion in its history, and hyperscale self-builds sit at the centre of the story. When we last published this report, the open question was whether the rise of AI would materially alter the hyperscale build-versus- lease calculus. That question has now been answered emphatically. Hyperscale self-build activity has shifted from steady growth into a full supercycle: the hyperscale platforms will bring online more self-build capacity in 2026 alone (~7.2GW) than they did across 2021, 2022 and 2023 combined. What was a ~2GW-per-year market as recently as 2023 has more than tripled its annual build rate, and the forward pipeline indicates this pace will be sustained through to 2030.
The capital behind this is unprecedented. The four largest hyperscale platforms have collectively guided to well over US$700b of capital expenditure for 2026 – AWS at approximately $200b, Microsoft at approximately $190b, Alphabet at $180-190b and Meta at $125-145b – with the majority allocated to data centres, servers and network infrastructure. Individual campuses have crossed thresholds that would have seemed implausible two years ago: Meta’s Hyperion campus in Louisiana is scoped at up to 5GW, Microsoft’s Fairwater sites are being explicitly marketed as the world’s most powerful AI data centres, and OpenAI’s Stargate programme is targeting 10GW+ of dedicated capacity.
Despite the supercycle of self-build activity, it has not come at the expense of colocation leasing. Hyperscale and AI tenants leased more third-party capacity in the third quarter of 2025 alone than in the entirety of 2024, and rental rates continue to climb. AWS itself discloses that its footprint remains roughly half owned and half leased. The central thesis of our prior reports – that hyperscalers will continue to balance self-builds with leasing, and that a rising tide lifts all boats – has been validated more forcefully than we anticipated.
At the same time, the grey area between pure self-builds and traditional colocation has widened into the fastest- growing part of the market. The build-to-suit, powered shell and master-planned data centre park models have been joined by an entire ‘powered land’ asset class and by financing structures – joint ventures, special purpose vehicles, private credit and securitisation – that blur the line between owned and leased infrastructure. The flagship example is the Stargate programme itself: OpenAI is the tenant, Oracle is the primary offtake counterparty, and third-party developers such as Crusoe, Vantage, and STACK Infrastructure actually build and own the campuses. The pre-development process including land, power, entitlement and permitting, has fully become a business of its own, and third-party operators are embedded in the hyperscale value chain in ways that go far beyond traditional colocation.
The self-build landscape has also gained an entirely new class of participants. xAI has become the first genuinely new entrant into hyperscale self-building in over a decade, with its Colossus campus in Memphis, Tennessee. Anthropic has committed $50b to its own dedicated US data centres. OpenAI, having leased and consumed capacity at historic levels, announced its first directly developed campus in Georgia in July 2026. And the neoclouds are pivoting from leasing to owning and building. The definition of ‘hyperscale self-builder’ is expanding in real time.
This report takes a data-driven approach to analysing the hyperscale self-build data centre landscape in this new environment. For 2026, Structure Research estimates that total global hyperscale self-build capacity will reach 29.7GW of operational critical IT load. About 87% of that operational capacity is represented by the world’s four largest hyperscale platforms: AWS, Microsoft Azure, Google Cloud and Meta. Beyond the operational base, we track a further 11.5GW under development and 46.5GW of land bank capacity, for a total identified self-build footprint of 87.1GW. This report is an invaluable resource for any data centre colocation operator, real estate firm, capital group, energy company or hyperscale end user involved in the hyperscale data centre value chain.
Download the executive summary PDF here.
