Why the $6.7 Trillion Data Center Buildout Still Won’t Be Enough

The global data center industry has committed more capital than at any point in its history. Announced investment totals for the decade through 2030 have crossed $6.7 trillion, per McKinsey analysis of spending plans across hyperscalers, colocation providers, and national sovereign AI programs.

By Q1 2026, the four largest hyperscalers confirmed a combined $650 to $725 billion in AI infrastructure capital expenditure for the year alone. That total is more than double the figure cited in early projections. The capital commitment has never been larger.

The AI infrastructure supply gap is widening.

Vacancy rates across primary North American data center markets have fallen to 1%, the lowest on record per JLL’s year-end 2025 analysis. Across EMEA’s five established hubs (Frankfurt, London, Amsterdam, Paris, Dublin), vacancy has dropped to 6.3%, down from 16.9% in 2021. Hardware lead times for GPU clusters currently run 36 to 52 weeks from order to rack. Some data-center configurations are now approaching a full year.

Meanwhile, new construction timelines for hyperscale facilities average 3 to 5 years from groundbreaking to first power-on. Of the 35 gigawatts currently under construction globally, 92% is already pre-committed before a single rack is installed. By the time the bulk of the $6.7 trillion reaches operational status, AI compute demand will have grown well past its current level.

Capital alone does not fix the supply problem.

What 1% Vacancy Means

A functioning commercial real estate market operates at vacancy rates between 8% and 15%. At 10%, there is enough excess capacity for tenants to find space quickly, negotiate terms, and plan infrastructure without 18-month lead commitments.

The global data center industry has committed $6.7 trillion in capital. North American data center vacancy stands at 1%. Of the 35 GW under construction globally, 92% is already pre-committed. Goldman Sachs estimates a 9.3 GW supply shortfall in 2026 alone. The AI infrastructure supply gap is not closing on the timeline announced investment implies. For enterprise AI teams making decisions today, understanding why matters more than the headline number.

At 1%, there is no discretionary supply. Every enterprise team looking for GPU capacity is competing for space that was contracted 18 to 24 months ago. The capacity available today was reserved before most enterprise AI programs existed.

Goldman Sachs estimates the global supply shortfall at 9.3 GW in 2026 alone, widening to 10 GW by 2028. That figure assumes announced construction projects deliver on time. The downstream consequence of this scarcity is already reshaping how enterprises access capacity.

Pre-leasing

Pre-leasing, once reserved for the largest hyperscale operators, has become a baseline requirement. In Europe’s core FLAP-D markets (Frankfurt, London, Amsterdam, Paris, Dublin), 83% of the construction pipeline is already pre-let. Consequently, finding contiguous space of 10 MW or more without a multi-year forward commitment is effectively impossible.

Hardware lead times for GPU clusters currently run 36 to 52 weeks from order to rack. Some data-center configurations are now approaching a full year.

JLL’s year-end 2025 analysis tracks absorption and availability across primary and secondary markets. The pattern is consistent: primary markets have effectively no available supply. Secondary markets absorb capacity faster than new supply comes online. The vacancy floor is set by construction timelines, not by demand.

Global capital commitment

By Q1 2026, the four largest hyperscalers confirmed a combined $650 to $725 billion in AI infrastructure capital expenditure for the year alone. That total is more than double the figure cited in early projections. The capital commitment has never been larger.

Meanwhile, new construction timelines for hyperscale facilities average 3 to 5 years from groundbreaking to first power-on. Of the 35 gigawatts of data center capacity currently under construction globally, 92% is already pre-committed before a single rack is installed. By the time the bulk of the $6.7 trillion reaches operational status, AI compute demand will have grown well past its current level.

This is what sold out looks like at infrastructure scale. For AI teams running production workloads in 2026, the next available reservation in most primary markets is for capacity that will not be provisioned until 2027 or later.

The Build Timeline vs. the Demand Curve

The $650 to $725 billion in hyperscaler AI capital expenditure confirmed for 2026 is frequently interpreted as evidence that the supply problem is being solved. It is worth examining what that money is actually buying.

More than 60% of combined hyperscaler infrastructure capex in 2026 is going into power infrastructure, cooling, and data center construction, not compute hardware. Microsoft confirmed $190 billion for the year. Meta raised its full-year guide to $125 to $145 billion. Amazon and Alphabet reported first-quarter capex of $44.2 billion and $35.7 billion respectively, putting their full-year totals on track to exceed $175 billion and $140 billion. The binding constraint has shifted from chips to electricity, and capital allocation has followed.

A greenfield hyperscale data center requires 3 to 5 years from site acquisition to operational status. The critical path moves through four stages, each carrying independent delay risk.

Four Stages of Building Hyperscale Data Centers

Site selection and permitting: 6 to 18 months, depending on jurisdiction. Markets with high existing data center concentration, Northern Virginia being the clearest example, now face longer permit timelines as local governments manage grid load and zoning concerns.

Grid interconnection: 4 to 7 years in most US markets. The queue for new power connections to major utilities has lengthened significantly, driven by data center demand layered on top of EV charging and industrial electrification. A facility that broke ground in 2025 may not receive its full power allocation until 2029 or beyond.

Equipment procurement: 36 to 52 weeks for GPU clusters from the point of confirmed power availability, with some configurations now approaching a full year. Cooling infrastructure, power distribution units, and physical security systems add further lead time.

Integration and commissioning: 3 to 6 months.

Best case from investment decision to operational GPU capacity: 36 months. For large facilities in constrained markets, 48 to 60 months is the realistic range.

According to the IEA Electricity 2026 report, AI data center power demand is growing at 30% annually in the base case. At that rate, demand compounded over four years produces roughly a 2.9x increase. The supply that $6.7 trillion is building will enter a market that has moved well past its original design point.

The AI compute market context behind these demand projections is covered in detail in AI Compute Market 2026: What the Numbers Actually Show.

Power Is the Actual Constraint, Not Capital

The $6.7 trillion figure represents committed capital across construction, hardware, cooling, and networking. Capital, however, does not move faster than the electrical grid.

The IEA Electricity 2026 report projects global AI data center power demand reaching 1,000 TWh by year-end, equivalent to Japan’s entire electricity consumption. The US power grid, outside a small number of high-voltage transmission additions, was designed for industrial loads that change slowly. A hyperscale GPU cluster drawing 50 to 100 megawatts requires grid connections, substation upgrades, and often dedicated transmission lines. Each of those is a project measured in years, with permitting timelines that parallel the data center construction schedule itself.

Grid interconnection queues in the United States currently run 4 to 7 years in most regions. Hyperscalers with the resources to sign long-term power purchase agreements and fund transmission upgrades can partially accelerate this. Organisations without those resources cannot. In practice, the power constraint is a barrier to entry that entrenches the hyperscaler position in AI infrastructure.

Bypassing Public Grid Entirely

Faced with interconnection queues stretching to a decade in some markets, hyperscalers have begun bypassing the public grid entirely. Microsoft’s 20-year power purchase agreement with Constellation Energy covers 835 megawatts from the restarted Three Mile Island Unit 1 reactor. The project is backed by a $1.6 billion restart investment, including a $1 billion DOE loan to Constellation Energy. It is the most prominent example of how far up the energy supply chain major infrastructure operators are now forced to move.

Lawrence Berkeley National Laboratory estimates that US data center electricity consumption will reach 6.7% to 12% of national electricity use by 2028. The scale of the power gap is further visible in BloombergNEF’s project tracker, which shows that early-stage data center projects with unconfirmed grid connections have more than doubled since 2024. Those projects represent tens of gigawatts of announced capacity with no confirmed path to power. They will not deliver compute on their announced timelines.

Northern Virginia, the largest data center market in the world, has paused new facility approvals in multiple jurisdictions due to grid saturation. Ireland’s moratorium on new data center grid connections was lifted in December 2025. Notably, any new facility must now install on-site generation or battery systems capable of meeting full electricity demand before connecting to the grid. The full mechanics are covered in The Power Problem: Data Center Energy Constraints Reshaping AI Infrastructure.

Geographic Concentration vs. Global Demand

More than 60% of US data center capacity sits in five metropolitan markets: Northern Virginia, Silicon Valley, Dallas, Phoenix, and Chicago. The announced buildout is directed primarily at those same markets. Land is understood, permit processes are established, utility relationships exist, and skilled workforces are concentrated there.

As a result, Data Centre Review reports that FLAP-D capacity pricing is set to rise 12% in 2026 as the gap between available supply and committed demand continues to widen.

AI demand, however, is not concentrated in those five markets. European enterprises face data residency requirements under GDPR and the EU AI Act that prevent routing workloads through US infrastructure. Healthcare and financial services firms in APAC operate under jurisdiction-specific sovereignty rules. Latin American AI adoption is growing faster than any available local capacity can absorb.

The Middle East is the only significant geography where sovereign-backed buildout is adding capacity ahead of existing demand. Nine metros now have approximately 1 GW of live capacity, with 2.2 GW under construction and 12 GW planned. Riyadh alone has 1.4 GW under construction. That pipeline is government-funded and not directly accessible to most enterprise AI teams.

For enterprises with regulatory or latency requirements in specific geographies, the announced buildout provides no near-term relief. The supply that matters is the supply within jurisdiction. That supply is not where the capital is going.

The result

Start from the current state: 1% vacancy across primary North American markets, hardware lead times of 36 to 52 weeks climbing toward a year, the IEA projecting 30% annual growth in AI data center power demand, and 92% of the global construction pipeline already pre-committed before it comes online.

The optimistic case: $6.7 trillion in committed investment delivers on schedule. Power constraints resolve through grid upgrades and on-site generation. Hardware supply normalises. All announced facilities reach operational status by 2030. In this scenario, total global data center capacity grows by 2x to 2.5x over the current base.

AI data center power demand, at 30% annual growth per the IEA base case, grows by approximately 2.9x over the same four-year period. In the IEA accelerated scenario, the multiplier is higher still.

Supply doubles. Demand nearly triples. Goldman Sachs puts the shortfall at 9.3 GW this year and rising. Vacancy falls further. Lead times extend.

The $6.7 trillion closes a meaningful portion of the gap while demand moves beyond it. The investment is real and the capacity will eventually come online. Sufficient supply, however, is measured in half-decades, not years. For enterprise AI teams making infrastructure decisions in 2026, the buildout is not a solution to the question of where to run workloads today.

What This Means for Enterprise AI Teams

Enterprise teams waiting for hyperscaler capacity to free up, or for new colocation space to open, are planning around a timeline that will not arrive before their AI programs require compute. With 92% of the global construction pipeline already pre-committed, waiting for new supply to arrive is a plan that ends at the back of a queue that does not move.

Axe Compute operates across 200+ locations worldwide on existing GPU infrastructure that is live. It requires no new grid connections. Hardware is already provisioned. Capacity reserves in approximately 48 hours. The 400,000+ GPUs available on the network are not part of a 2028 completion schedule.

In summary, the AI infrastructure supply gap will not be resolved by announced investment. It will be managed, quarter by quarter and workload by workload, by teams that plan their compute the same way they plan headcount: from what is available now, not from what is promised for 2030.

For a framework on how enterprise AI teams are structuring infrastructure decisions around the supply constraint, see Enterprise GPU Strategy in 2026. View live capacity at dashboard.axecompute.com.

---

Sources

1. McKinsey Global Institute — The Cost of Compute: A $6.7 Trillion Race to Scale Data Centers, 2025
2. JLL — North American and EMEA Data Centre Market Overview, Q4 2025
3. Data Centre Review — EMEA Data Centre Vacancy Hits Record Low as AI Demand Outpaces Supply, April 2026
4. Goldman Sachs Research — Data Center Supply Shortfall and Power Demand 2026-2028
5. IEA — Electricity 2026: AI Data Centre Power Demand Projections
6. Tom’s Hardware — Google, Microsoft, Meta, and Amazon Capex to Hit $725 Billion in 2026, Q1 Earnings Analysis
7. The Next Web — Q1 2026 Big Tech Earnings: $650 Billion in AI Capex and Compute Constraints
8. CBRE — North American Data Center Trends Report H2 2024
9. Lawrence Berkeley National Laboratory — United States Data Center Energy Usage Report 2024
10. Data Centre Review — FLAP-D Data Centre Capacity Pricing Set to Rise 12% in 2026, May 2026
11. BloombergNEF — AI and the Power Grid: Where the Rubber Meets the Road
12. S&P Global — 2026 Trends in Data Center Services and Infrastructure
13. Uptime Institute — Global Data Center Survey 2025