Artificial intelligence: The trillion dollar question

The demand shock

Knight Frank forecasts global data centre capacity to expand from 62GW in 2025 to over 110GW by the end of 2028. AI-related volumes, both through self-build development and colocation lease agreements, could expand from c.8GW to c.27GW by 2028, representing a rise from 12.9% of workloads to 24.5% over the forecast period.

Over the course of 2026, it is expected that capital expenditure from US ‘Big Tech’ – Microsoft, Google, Amazon Web Services, and Meta – could surpass $650 billion, up from a combined $376 billion in 2025. This growth is being driven primarily by AWS and Google, which have upped annual capital expenditure forecasts to $200 billion and $185 billion, respectively. Microsoft, on the other hand, has not provided any full-year forecasts, only stating that its quarterly capital expenditure is expected to ‘decrease sequentially’ following a company record period during the quarter ended December 2025, where capital expenditure jumped 54.1% quarter-on-quarter. Despite this, however, Microsoft is still expected to deploy between $110-$130 billion during the year.

The current demand cycle has helped net colocation take-up reach 15.8GW signed in 2025, 37% of which is driven by AI-related demand. A little under 6GW of new AI-related colocation lease agreements were signed during the year, three times the volume signed in 2024.

On the ground, vacancy is hitting record lows. Global wholesale colocation vacancy has dipped to 8.1%, down from 9.1% at the end of 2024, with primary-market vacancy tightening further. Across Europe, the FLAP-D markets hit an overall vacancy of less-than 4%, with London at 3.1% and Frankfurt at <1%, whilst in North America and Asia-Pacific, Ashburn vacancy has dropped to <1% and Singapore hits 2.2%.

Power-first, not location-first

The IEA’s “Energy & AI” analysis emphasises that access to reliable, scalable power is the real bottleneck to “intelligence at scale”, not land or GPUs. As power generation capabilities ramp up across global markets, the widely acknowledged constraint is the transmission grid, which has been subject to decades of underinvestment, creating a structural deficit causing multi-year connection queues, with key areas, such as London, approaching close to a 10-year connection wait time in core submarkets. Grid capacity across core data centre metros is predominantly reserved through to 2030, prompting long-term pre-leasing tendencies and a transition to on-site generation.

To correct this deficit, significant investment is required across most global markets. BloombergNEF estimates requirements in the United States may reach $150 billion by 2030, whilst the UK faces an estimated $107 billion cost to move renewable power generated in the north to core demand regions in the south. China, which operates the worlds largest grid, could require as much as $3.8 trillion by 2050, while for Australia the figure is roughly $300 billion through to 2050. What this means for site selection is that deliverable power has become the prime variable.

A trillion-dollar question

By 2030, global AI-related data centre capacity deployment volumes could hit 60GW, up from 8GW in 2025. This increase of 52GW will come at a development cost of between $676-$780 billion, whilst tenants are likely to spend the same or more to fit out their space with necessary IT infrastructure. The infrastructure necessary to support the AI build-out over the coming 5 years will require an infrastructure investment of $1.4-$1.6 trillion.

The partnership between Oracle and OpenAI, announced in September 2025, quickly became the most high-profile AI infrastructure program. The agreement, a $300 billion, five-year cloud computing contract beginning in 2027 that would make Oracle the primary provider of compute for OpenAI and position it alongside Microsoft, AWS, and Google as a hyperscaler capable of delivering next-gen AI compute infrastructure. Yet, OpenAI’s annualised revenue of $10 billion (as of Q2 2025) was overshadowed by the projected $60 billion yearly cloud bill, potentially forcing it to consider major cost cuts or fresh capital raises. Credit agencies warned the deal could strain Oracle’s balance sheet, with Moody’s highlighting leverage and counterparty risk. By March 2026, the companies cancelled an initial 600MW Abilene expansion, citing funding complications and volatile demand forecasts.

Going forward, OpenAI has $600 billion in projected cloud compute spend by 2030, which was recently cut back from an earlier pledge to spend $1.4 trillion by 2033. Oracle is set to raise $50 billion in debt and equity to fund the OpenAI build-out, with large-scale layoffs expected to help cover costs. Similarly, fellow Stargate backer SoftBank is turning to debt to fund OpenAI’s expansion, seeking $40 billion in loans.

In other news, Anthropic announced a $50 billion investment in US AI infrastructure to build data centres in Texas and New York via Fluidstack. OpenAI completed a $40 billion funding round, the largest ever by a private tech company. OpenAI also completed a $38 billion deal with AWS, providing it with access to AWS’ infrastructure to scale its core AI workloads. Meta & Blue Owl Capital announced a $27 billion joint-venture for the development of its Hyperion AI campus. Nebius announced GPU contracts with Microsoft & Meta for $19.4 billion & $3 billion, respectively, whilst similarly receiving a $2 billion investment from Nvidia to help fund its AI data centre expansion. Similarly, CoreWeave announced a $6 billion AI data centre project in Lancaster, Pennsylvania.

Regarding the cancelled Abilene expansion, and another revealing signal of the instability of the Oracle & OpenAI partnership, came from Nvidia. The company paid a deposit of $150 million to Crusoe to secure the site, ensuring that its hardware is deployed at the site instead of a competitor. Both Meta and Microsoft are considering occupying the unused space.

Sowing the seeds for an overbuild?

With global data centre capacity set to triple by 2030, and AI-based capacity to grow by over ten times, the question remains over whether the current build-out roadmap is steering towards an over-build. At present, there exists an undersupply of data centre capacity, driving record levels of pre-leasing activity from hyperscalers, with multi-year roadmaps, public balance sheets, and a known desire for geographic redundancy underpinning landlord visibility.

Over the next 24-36 months, power availability, not capital, will cap supply, keeping core fundamentals strong across Tier I & Tier II markets and pushing additional development to new power-rich regions.

However, beyond 2028, two potential swing factors will dominate the direction of travel: Inference monetisation, and whether AI can move from promise to profit at scale, and Grid build-out, and whether new generation capacity and transmission networks can catch up. What this argues for is a disciplined, phased build-out with flexible designs that can switch between high-density AI and lower-density cloud, and contract structures that share delivery risk.

What does this mean for interested parties?

For developers and investors, pursue ‘power-first’ land strategies.

Prioritise parcels with realistic pre-2028 energisation paths and proximity to high-voltage transmission and gas networks. Upon which, consider AI-ready shells, with allowances for liquid cooling fitouts and assuming rack densities of 50-100+kW.

For landlords and REITs, lean into deliverable megawatts.

The pricing premium for AI-capable land, with near-term power, is set to remain, with rental growth and multi-phase pre-leasing to remain the norm. Balance tenant concentration risk, stagger expiry schedules, and consider heightened underwriting scrutiny given the sector’s lack of investment-grade credit tenants.

For occupiers, secure power early.

With new-lease rental rates expected to rise by 8-12% annually, take-on phased pre-lease options, particularly in power-scarce regions, rather than betting on a late-cycle market.

Final take

Real estate investors, developers, and operators now sit at the centre of this transformation, where the ability to secure and deliver power, navigate government policy, and build adaptable infrastructure will determine who thrives. The winners will be those who pair disciplined capital deployment with speed, scale, and sustainability.