The Scale of the Build-Out
Start with a number that doesn’t get smaller the more you examine it.
More than 23 gigawatts of data center IT capacity was under active construction globally as of September 2025, per BloombergNEF’s analysis released March 24, 2026. The Americas region alone accounts for 17 of those gigawatts across 311 locations. Twenty-three gigawatts is the generating equivalent of roughly 20 large nuclear reactors. All of it is being built to run AI workloads.
The capital behind the construction is what makes the scale historically unusual. BloombergNEF estimates the combined capital expenditure of the 14 largest publicly owned data center operators could reach approximately $750 billion in 2026. Futurum Research puts a related figure at $690 billion, measuring a different scope of operators. Both are analyst projections, not audited spending. The variance between the two estimates reflects the methodological ambiguity in measuring “AI infrastructure” investment when hyperscalers don’t break out AI-specific capex in their public filings. What both figures agree on: the order of magnitude. This is hundreds-of-billions territory, not tens of billions.
Meta’s announcement this week is the most visible single-company anchor for that macro figure. Bloomberg confirms Meta is increasing its El Paso, Texas data center investment from $1.5 billion to more than $10 billion. CNBC reports the facility targets 1 gigawatt of capacity by 2028. Meta’s 2026 capex guidance spans $115 billion to $135 billion, a range, not a fixed commitment, but a range that signals a company treating AI infrastructure as a core operating expenditure at the scale of a major utility’s annual capital program.
Meta is not alone in that posture. Its announcement this week is one data point within a broader hyperscaler commitment pattern. The BloombergNEF figure covers 14 operators, not one. The infrastructure sprint is systemic.
The Energy Equation
Infrastructure at this scale requires power that the current grid wasn’t designed to supply.
The IEA projected in April 2025 that global AI data center energy consumption would grow from 485 TWh in 2024 to 945 TWh by 2030, a 95% increase in six years. Climate Home News places that trajectory at approximately 3% of global electricity demand by 2030. These are IEA projections, based on data that is now a year old and in a market moving faster than most projection models can track. The directional conclusion is more reliable than the specific figures: AI data center energy demand is growing faster than grid infrastructure can expand.
Texas is the most visible pressure point. The ERCOT grid, which manages the Texas electricity market, has been navigating data center power requests at a scale that strains near-term capacity planning. Specific ERCOT figures from this cycle haven’t been independently confirmed and are excluded from this analysis. The principle holds across public reporting: the power requirement for the AI infrastructure build-out is now large enough to be a binding constraint on construction timelines, not a peripheral concern.
The energy constraint creates a structural tension that most coverage of AI infrastructure misses. Companies are committing capital to facilities that require power agreements, grid connections, and permitting timelines that run 2 to 5 years. The construction timeline for a 1 GW facility like Meta’s El Paso site is shorter than the timeline to secure the power that running it requires. Hyperscalers are beginning to invest in dedicated power generation, solar and nuclear agreements, specifically because utility grid capacity can’t keep pace with their build schedules.
This is a regulatory story waiting to happen. The EU is already discussing data center energy regulation in the context of industrial policy and grid stability. The United States has no equivalent federal framework, but state-level pressure, particularly in Texas, Virginia, and Georgia, is increasing as data center power requests become visible in rate-setting proceedings. Watch this space.
The Workforce Paradox
Capital is accelerating into AI infrastructure. Human roles in adjacent support functions are contracting.
Meta’s layoffs this week are the sharpest illustration. Approximately 700 jobs cut across Reality Labs, Facebook, global operations, recruiting, and sales, fewer than 1,000 total, per NBC News, in the same week as a $10 billion data center commitment. Meta hasn’t publicly attributed the layoffs directly to AI investment. Multiple sources frame the events together contextually. The Seattle Times headline put it directly: cuts happening “amid record AI spending.”
The causal relationship between hyperscaler infrastructure investment and workforce reduction isn’t confirmed at the company level. But the NBER working paper published this week, drawing on a survey of 750 CFOs conducted with Duke University and the Federal Reserve Banks, adds a macro dimension that the Meta event alone couldn’t provide. According to reporting on the paper, it projects AI-driven job cuts in 2026 at approximately 502,000, roughly nine times the approximately 55,000 in 2025. Many of those projected cuts are characterized by analysts reviewing the paper as preemptive, driven by CFO expectations about AI’s future capabilities, not just current replacement.
That distinction matters more than the headline number. Companies aren’t waiting for AI to demonstrably replace workers before restructuring. They’re acting on the expectation. The infrastructure investment and the workforce reduction may be expressions of the same strategic logic: building the AI-native operating model before the full capability set exists, rather than adapting after it arrives.
Who Wins and Who Bears the Cost
The infrastructure sprint is producing clear beneficiaries and less visible costs.
Hyperscalers benefit from first-mover advantages in compute capacity, companies with more processing power, closer to more users, will have structural advantages in model serving costs and latency that compound over time. Data center operators, construction firms, and power equipment manufacturers benefit from the capital deployment. Utilities, in markets with regulatory frameworks that allow cost recovery, will recoup infrastructure investments through rate-setting.
Workers in support functions, operations, recruiting, sales, administrative roles, are the category most exposed to the restructuring that accompanies infrastructure expansion. These are not the roles being eliminated by AI capability in a direct sense; they’re the roles being reduced as organizations resize their human workforce in anticipation of AI taking a larger share of operational throughput. The 16% employment decline among entry-level workers aged 22 to 25 in AI-exposed roles, reportedly found in the NBER paper, is the data point that deserves the most scrutiny if confirmed, entry-level erosion affects workforce pipelines in ways that only become visible years later.
Regulators are the actors furthest behind. Energy regulators didn’t design their frameworks for data center power requests at this scale. Labor regulators are operating with definitions of AI-driven displacement that don’t capture preemptive restructuring. Neither framework is adequate for the speed at which the infrastructure sprint is moving.
What to Watch
Five indicators mark the next phase of this story.
First, utility power purchase agreements at scale. When hyperscalers sign long-term power contracts, particularly for dedicated generation capacity, it confirms infrastructure commitments that capex guidance alone doesn’t fully establish.
Second, grid policy developments in Texas and Virginia. State-level decisions on data center power access will shape where the next phase of the build-out concentrates.
Third, EU data center energy regulation proposals. The bloc has political motivation and existing industrial policy frameworks to move here. A draft framework would represent a significant escalation of regulatory risk for infrastructure plans already committed.
Fourth, additional hyperscaler capex announcements in Q1 2026 earnings. Meta’s guidance is public. Microsoft, Google, and Amazon’s capex trajectories will become clearer in earnings calls over the next 60 days.
Fifth, Q1 2026 employment data from BLS, particularly in technology support and administrative categories. If the NBER projections are accurate, the first statistical evidence should start appearing in official data within the next quarter.
TJS Synthesis
The $750 billion figure and the 502,000 projected job cuts are the same story told from different directions. Both reflect the operating logic of an economy reorganizing around AI infrastructure: capital concentrating at the compute layer, support functions contracting as automation absorbs a larger share of throughput. What’s happening isn’t a technology transition – it’s a capital allocation transition, with energy, labor, and regulatory systems that weren’t designed for its speed or scale. Investors, infrastructure planners, and workforce leaders who treat the data center build-out and the workforce restructuring as separate stories are missing the connective tissue between them. The connective tissue is AI’s actual deployment logic: build the infrastructure, reshape the workforce to match it, generate the revenue to justify the capital. That sequence is now visible in a single company’s week.