Amazon, Alphabet, Microsoft, Meta, and Oracle will spend over $600 billion on infrastructure in 2026 — a 36% increase over 2025’s already historic levels. Roughly 75% targets AI infrastructure. Each of the four largest hyperscalers now exceeds $100 billion in annual capex, with capital intensity reaching 45–57% of revenue. US data centre power demand is projected to nearly triple to 134 GW by 2030. Transformer lead times have stretched to 2–4 years. Power constraints are extending construction timelines by 24 to 72 months. The bottleneck constraining AI infrastructure is no longer the chip or the software. It is the electricity.
The AI industry is attempting the largest infrastructure buildout in technology history, and the physical world is pushing back. The five largest hyperscalers are collectively spending over $600 billion in 2026 on data centre infrastructure — more than the GDP of Sweden. Capital intensity has reached levels historically associated with utilities and industrial companies, not technology firms. Amazon is projected to have negative free cash flow of $17–28 billion in 2026 as it invests $200 billion in infrastructure. Alphabet held a $25 billion bond sale and quadrupled its long-term debt to fund the buildout.[1][2]
The origin of this cascade is D6 — operational. The physical infrastructure required to run AI at scale (data centres, power systems, cooling, networking) is being built at a pace that exceeds the capacity of the supply chains that serve it. The bottleneck has migrated: in 2021–2024, the primary constraint was manufacturing AI chips at scale. By 2025–2026, the constraint has shifted to the power infrastructure that runs the data centres housing those chips. AI accelerators and liquid cooling systems advance on monthly cycles. The lead times for transformers, switchgear, and gas turbines stretch two to four years.[3]
$600B+ in committed 2026 capex. Every major hyperscaler individually exceeding $100B. Goldman Sachs projects $1.15 trillion in hyperscaler capex from 2025–2027.
Transformer lead times: 2–4 years. Power constraints delaying construction 24–72 months. US needs 100+ GW of new data centre capacity by 2035 — 10× NYC peak demand.
The diagnostic signal is the mismatch between the speed of capital deployment and the speed of physical infrastructure. Money moves at the speed of a wire transfer. Electrical transformers move at the speed of grain-oriented electrical steel procurement, factory assembly, and heavy-haul trucking. Power plants move at the speed of permitting, construction, and interconnection. The $600 billion is real. The power to use it doesn’t yet exist at the locations and timescales the capital requires.[4]
US data centre grid-power demand stood at approximately 62 GW in 2025. S&P Global’s 451 Research projects this will rise to 75.8 GW in 2026 and nearly triple to 134.4 GW by 2030. For context, 100 GW of new data centre demand by 2035 is roughly ten times New York City’s summer peak demand.[5]
Data centre demand in 2025 (up from 9.3 GW in 2024). Over 90% of projected new power demand in PJM is data centres. Dominion Energy running out of capacity in Loudoun County.
Data centre utility demand in 2025. ERCOT passed SB-6 requiring new interconnection processes for loads exceeding 75 MW. Grid stress events increasing in frequency.
Capacity market clearing price for 2026–2027 delivery year — over 10× the $29/MW price two years prior. Data centre growth is the primary driver.
The transformer supply chain is the sharpest bottleneck. Since 2019, demand for generator step-up transformers has grown 274%. Lead times for large power transformers average 128 weeks; generator step-up transformers average 144 weeks. Prices have increased 77% for power transformers and 45% for GSUs since 2019. Only 20% of large transformers used in the US are manufactured domestically. Nearly $1.8 billion in new North American manufacturing capacity has been announced but will take years to come online.[6]
The response is restructuring how data centres get built. Developers no longer buy land first and coordinate power second — they secure power first and find land around it. Hyperscalers are investing directly in energy generation: Meta’s Hyperion project in Louisiana involves a $3.2 billion investment in a 2 GW combined-cycle gas plant. Microsoft has signed nuclear PPAs. Amazon is quadrupling US data centre capacity from 3 GW to 12 GW, including a 2.2 GW campus in Indiana that would consume roughly half the electricity of all Indiana households combined.[4]
Individual rack-level power density has surged from 10–14 kW to over 100 kW for AI workloads — a tenfold increase requiring fundamental redesigns of electrical distribution, cooling systems, and building infrastructure. Liquid cooling is no longer optional for AI-density deployments; it is a prerequisite. The data centre of 2026 bears little physical resemblance to the data centre of 2020.[4]
Origin: D6 (Operational). The physical infrastructure required for AI at scale — power, cooling, networking, facilities — is being built at a pace that the electrical grid, the construction supply chain, and the workforce cannot match. The cascade radiates outward from this operational mismatch.
| Dimension | Score | Diagnostic Evidence |
|---|---|---|
| Operational (D6)Origin — 80 | 80 | Power is the binding constraint on AI infrastructure. US data centre demand: 62 GW (2025) → 75.8 GW (2026) → 134.4 GW (2030). Transformer lead times: 2–4 years. Power constraints extending construction timelines 24–72 months. Rack density surging from 10–14 kW to 100+ kW, requiring liquid cooling. PJM capacity market prices increased 10× in two years. Amazon building 2.2 GW Indiana campus (half state’s household electricity). Microsoft turned away customers due to power shortages. Land without power is worthless to hyperscalers. The bottleneck migrated from chips to packaging (UC-219) to electricity — each layer deeper into physical infrastructure.[3][4][5][9][12] Power-First Infrastructure |
| Revenue (D3)L1 — 75 | 75 | $600B+ hyperscaler capex is the largest single-year infrastructure investment in technology history. Amazon $200B, Alphabet $175–185B, Meta $115–135B, Microsoft $120B+, Oracle $50B. Capital intensity at 45–57% of revenue. Hyperscalers raised $108B in debt in 2025; projected $1.5T over coming years. Amazon facing negative free cash flow of $17–28B in 2026. Equinix investing $4–5B/year through 2029, doubling capacity. Digital Realty leading in US leased power (15% market share). Colocation market: $104B (2025) → $204B (2030). Record annualised gross bookings across the sector.[1][2][7][8][10][11] Historic Capital Deployment |
| Quality (D5)L1 — 62 | 62 | Facility design evolving rapidly under competitive pressure. Liquid cooling transitioning from optional to mandatory for AI workloads — air cooling insufficient above 40 kW per rack. Direct-to-chip liquid cooling and full immersion cooling becoming standard. Data centres moving to locations with cheap power and water rather than network proximity — a fundamental shift in site selection logic. Geographic diversification: Iowa, Ohio, Nordic countries, Middle East. Google disclosed 6.1 billion gallons of water usage across its data centre portfolio in 2023. A 100 MW facility consumes ~300,000 gallons per day, equivalent to 2,600 households. Quality means sustainability as much as uptime.[4] Facility Design Revolution |
| Customer (D1)L2 — 58 | 58 | Enterprises and AI companies constrained by data centre availability. Colocation wait times at 18–24 months in key markets (Northern Virginia, Dallas, Phoenix). Microsoft has acknowledged turning away customers due to power shortages. Cloud GPU pricing remains elevated. Cloud region selection is becoming an infrastructure strategy decision, not just a latency decision. Enterprises unable to secure sufficient compute face competitive disadvantage in AI deployment. The capacity constraint is democratic — it affects startups and Fortune 500 companies alike.[3][4] Capacity Constraint |
| Employee / Workforce (D2)L2 — 55 | 55 | Data centre construction and operations workforce under extreme demand pressure. Electricians, mechanical engineers, power systems engineers, and data centre operations staff are among the scarcest technical roles. The US would need to purchase 90% of global semiconductor manufacturing output to support announced data centre load through 2030 — a staffing and training challenge at every level. Equinix aiming to double capacity in 5 years (matching 27 years of prior buildout). The pace implies a workforce buildout that the trade school and engineering pipeline was not designed to produce.[3][7][9] Workforce Bottleneck |
| Regulatory (D4)L2 — 48 | 48 | Local zoning and permitting are the binding constraints on where and how fast data centres can be built. Texas SB-6 redefines interconnection for large electrical loads (>75 MW) in ERCOT, requiring disclosure, financial commitments, and transmission cost coverage. PJM interconnection queue delays. Northern Virginia data centre moratorium discussions. Environmental review for new transmission lines can take a decade. Tariffs on copper (up to 50%), steel, and aluminium raise transformer manufacturing costs. Cities increasingly creating data centre-specific regulatory frameworks as economic development tools. The regulatory dimension is local and fragmented, not federal and monolithic.[5][6] Zoning & Permitting |
-- The Data Center Gold Rush: $600B Chases Power That Doesn't Exist (Diagnostic)
-- Sense -> Analyze -> Measure -> Decide -> Act
FORAGE data_center_gold_rush
WHERE hyperscaler_capex_2026 > 600_000_000_000
AND capex_yoy_growth > 0.35
AND ai_infrastructure_share > 0.70
AND transformer_lead_time_weeks > 100
AND power_constraint_binding = true
AND rack_density_increase > 5x
ACROSS D6, D3, D5, D1, D2, D4
DEPTH 3
SURFACE the_data_center_gold_rush
DIVE INTO infrastructure_mismatch
WHEN capital_speed > infrastructure_speed -- money moves faster than concrete and steel
AND grid_demand_exceeds_capacity = true
AND liquid_cooling_mandatory = true
AND capex_exceeds_free_cash_flow = true -- Amazon negative FCF
TRACE the_data_center_gold_rush -- D6 -> D3+D5 -> D1+D2+D4
EMIT diagnostic_cascade_analysis
DRIFT the_data_center_gold_rush
METHODOLOGY 85 -- capacity planning is a mature discipline
PERFORMANCE 35 -- AI demand broke every forecast model
FETCH the_data_center_gold_rush
THRESHOLD 1000
ON EXECUTE CHIRP critical "6/6 dimensions, $600B capex, power-constrained buildout, transformer bottleneck"
SURFACE analysis AS jsonRuntime: @stratiqx/cal-runtime · Spec: cal.cormorantforaging.dev · DOI: 10.5281/zenodo.18905193
UC-219 showed that the AI chip supply chain is constrained by CoWoS packaging. UC-220 shows that even if packaging resolves, the data centres housing those chips are constrained by electrical power. Each bottleneck sits deeper in the physical stack and takes longer to resolve. Chips: 12–18 months. Packaging: 2–3 years. Power infrastructure: 3–10 years. The AI industry is discovering that digital products rest on an analogue foundation.
Hyperscaler capital intensity at 45–57% of revenue resembles utilities and industrial companies, not technology firms. Amazon projected to have negative free cash flow. $108 billion in debt raised in 2025, with $1.5 trillion projected. The hyperscalers are becoming infrastructure companies that happen to write software, not software companies that happen to own infrastructure. This is a structural shift in how the industry is funded and valued.
Data centre site selection has inverted. In the old model: find land, then coordinate power. In 2026: secure power, then find land around it. Developers invest directly in generation. Meta builds a $3.2B gas plant. Microsoft signs nuclear PPAs. Amazon quadruples capacity to 12 GW. The company with secured power wins. The company without it waits 2–4 years for a transformer and loses the AI infrastructure race while it waits.
Goldman Sachs projects $1.15 trillion in hyperscaler capex from 2025–2027. Google co-founder Larry Page reportedly stated he would rather go bankrupt than lose this race. The 2000s telecom overbuild and the 1860s railroad overbuilding are the historical templates. Whether the AI demand curve justifies the capital deployed is the prognostic question this cluster’s capstone (UC-223) will address. For now, the buildout is a measured reality, not a prediction.
One conversation. We’ll tell you if the six-dimensional view adds something new — or confirm your current tools have it covered.