Will the global AI demand justify investment required to power the infrastructure boom? In its latest report titled, ''Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out,'' global brokerage and investment bank Goldman Sachs seeks to shifts the debate from whether AI demand will justify investment to the supply-side assumptions that dictate how much capital is actually needed. The report establishes a baseline estimate of $7.6 trillion in cumulative AI capital expenditure (capex) between 2026 and 2031. This includes spending on specialized chips (XPUs), data centers, and power infrastructure.
It explores the supply-side variables that could dramatically shift the estimated $7.6 trillion investment in AI infrastructure through 2031. Much of the market focus is on returns (demand), but this report argues that uncertainty in supply-side costs is an equally large risk for investors to manage, according to Goldman Sachs. The report identifies four variables that determine whether the final investment figure will be significantly higher or lower than the multi-trillion-dollar baseline which are as follows:
1. Economic Useful Life of AI Silicon
According to the report, this is the most influential variable. While buildings last 20+ years, chips typically have a useful life of 4-6 years. The single most influential variable is how long AI chips remain economically viable. Moving the replacement cycle from four years to six years could reduce cumulative spend by hundreds of billions.
The Risk: If rapid innovation (such as NVIDIA's annual release cycle) makes chips obsolete in three years instead of five, cumulative spend and depreciation costs would skyrocket.
The Mitigant: A "tiered model" where older chips are reused for less intensive tasks (like inference or edge computing) could stabilize their useful life and protect investment returns.
ALSO READ: India's AI Edge Is Real - Execution Discipline Will Decide The Winners
2. Data Center Cost and Complexity
AI workloads require far more power density than traditional cloud computing. Modern data centers are no longer just "warehouses for servers"; they are integrated high-density systems. Rising costs for advanced liquid cooling and power delivery are significantly driving up the price per megawatt.
Increased Density: Modern AI racks (sich as NVIDIA's GB300) generate immense heat, requiring liquid cooling and sophisticated power delivery.
System-Level Design: Data centers are no longer just "warehouses with servers"; they are tightly integrated "systems" where compute, cooling, and power are co-designed. This pushes the cost per Megawatt (MW) significantly higher.
ALSO READ: Claude's New AI-IT Services Bet: Should Infosys, TCS Be Worried? Jefferies Says Yes
3. Chip and Architecture Mix
Whether companies choose highly specialized ASICs or flexible GPUs affects whether compute demand is "elastic" (price-sensitive) or "inelastic" (performance-driven). The total cost depends on how compute demand behaves:
Elastic Demand: If users buy more compute just because it gets cheaper, margins might shrink, but the total capital spent remains high.
Inelastic Demand: If there is a fixed "budget" for compute, advancements in chip efficiency would actually reduce the total capital needed for the build-out.
ALSO READ: Anthropic Signs Computing Deal With SpaceX to Meet AI Demand
4. Physical Bottlenecks (Power, Labor, and Equipment)
Shortages in power grid capacity, specialized labor, and electrical equipment (transformers/switchgear) are "elongating" the build-out, potentially dampening the pace of investment regardless of demand. The speed of the build-out is currently constrained by the "physicality" of the infrastructure.
Power Grids: Connecting massive data centers to the grid can take years.
Labor & Parts: Shortages in specialized labor and long lead times for equipment (like transformers) create "elongation."
Feedback Loop: If these bottlenecks persist too long, they may cause investors to doubt the ROI, potentially cooling the demand for the very infrastructure being built. The report concludes that we are building the most expensive and physically demanding "computer" in human history. To understand if this is a bubble or a sustainable shift, investors must stop looking solely at ChatGPT's revenue and start monitoring the depreciation cycles of chips and the capacity of the electrical grid.
Hence, the multi-trillion-dollar scale of the AI build-out is not a fixed destination, but a highly conditional forecast that depends on a few "swing" assumptions rather than just market demand, according to Goldman Sachs. While the market often asks, "Will AI make enough money to justify this?" (the demand side), the report concludes that the supply side is equally volatile and carries its own set of risks
Essential Business Intelligence, Continuous LIVE TV, Sharp Market Insights, Practical Personal Finance Advice and Latest Stories — On NDTV Profit.