McKinsey: Global Data Center Spending Could Reach Nearly $7 Trillion by 2030
Consulting firm estimates that to keep up with computing power demand, global data centers may require approximately $6.7 trillion to $7 trillion in capital expenditures by 2030; of this, about $5.2 trillion is for AI workloads and approximately $1.5 trillion for traditional IT. Global installation demand may nearly quadruple, with about 70% coming from AI.
The U.S. is projected to be the largest recipient of this spending: McKinsey states that over 40% will be invested in the U.S., with domestic demand growing by about 20% to 25% annually. BloombergNEF has revised its forecast for U.S. operational capacity in 2030 to 118 GW, a 52% increase from its December 2025 prediction, and expects it to reach 194 GW by 2035; since December 2025, the U.S. has announced an additional 101 GW in the pipeline.
Six major operators—Amazon, Microsoft, Alphabet, Meta, Oracle, and SpaceX—are estimated by S&P Global to have a combined capital expenditure exceeding $1.3 trillion by 2027. The World Economic Forum cites that the five largest U.S. cloud and AI infrastructure companies have committed to capital expenditures of about $66 billion to $69 billion by 2026, nearly doubling and accounting for about 1.9% of GDP.
Electricity is a hard constraint. NEMA expects data centers to contribute 38% of the net electricity increase in the U.S. by 2037; Goldman Sachs estimates electricity demand will rise by about 165% by 2030, with grid upgrades costing about $720 billion. BNEF states that if only looking at chip shipments, the U.S. could theoretically add 207 GW by 2033, but due to energy constraints, there is a gap of about 63 GW between its capacity forecast and chip implied values. Bank of America expects data centers to add about 125 GW of electricity over the next five years, while utilities plan to add only about 93 GW of reliable capacity, resulting in a gap of over 100 GW, leading developers to turn to on-site gas and storage solutions.
Mechanically, the buyers are large-scale cloud and model companies, while the sellers are chip, power, cooling, and land developers. The event is driven by the rising power density of AI training and inference: cabinet power consumption is shifting from about 227 kW to around 400 kW. Funding is flowing from tech companies' operating cash flows, investment-grade bonds, infrastructure funds, and high-yield debt into cabinets; beneficiaries are owners who secure electricity and turbine scheduling first, while communities in queue with the grid and utilities still planning based on old loads are under pressure.
The publicly predicted range is quite large: in a constrained scenario, the increment is about 78 GW with spending of about $3.7 trillion; in an accelerated scenario, the increment is about 205 GW with spending of about $7.9 trillion. The $7 trillion figure does not represent signed orders.
Source: Public information
ABAB AI Insight
"Three times the U.S." is often used as a headline, but the original text from the institution indicates that global capacity is nearly three times, with the U.S. taking the largest share. BNEF has revised the U.S. figure for 2030 upwards by more than half, citing the expansion of announced pipelines, not the already connected capacity doubling. The upward revision itself indicates that projects have shifted from "planning posters" to "GW campuses," and it also shows that cancellations and delays have not been equally accounted for.
The capital path indicates that chips account for about 60% of AI-related spending, power and cooling about a quarter, and land and civil construction the remainder. Jensen Huang's quoted cost per GW is higher than most seller analyses; using his lower limit for reverse calculation, announced projects are close to several trillion dollars. The financing side draws in pension funds, direct lending funds, and large-scale operating cash flows into the same power supply curve.
Analogies are made to interstate highways and the Apollo program in terms of GDP share, as well as the electrification of railroads, which are built once and depreciated over decades; in contrast, data centers renew GPUs and power equipment every four to six years. In terms of industry positioning, cloud companies are moving from renting cabinets to building their own power plants: those who lock in turbines and substations first will collect payments first.
The structural judgment belongs to the concentration of capital overlapping with industrial chain reconstruction: pricing power is shifting from software licensing to electricity and cabinet density. The mechanism is that model capabilities are written as a power curve that must continuously expand; when the grid cannot keep up, on-site generation rewrites the energy companies' schedules into the capital expenditure schedules of tech companies.
ABAB News · Law of Cognition
The computing power bill will ultimately be written on the electricity grid.
Upward revisions in forecasts indicate that posters are faster than grid connections.
Equipment that is replaced every four years cannot support the recovery story of a century-old infrastructure.