The $1.4 Trillion AI Gamble: Inside Goldman Sachs’ Bull Case and the Hidden Debt Risks

Why Goldman Thinks the Consensus Is Too Low

The core of Goldman’s argument is that current Wall Street estimates imply an implausible deceleration in spending growth. The consensus for 2027 hyperscaler capex sits at roughly $920 billion, which would represent a sharp slowdown from the breakneck pace of 2025 and 2026 . Goldman challenges that assumption by modeling what happens if AI investment continues to consume 2% to 3% of GDP—a scenario that pushes annual spending toward the $1.1 trillion baseline and as high as $1.4 trillion in an upside case .

Behind the model is a bet on agentic AI. Unlike simple chatbots that answer a prompt and stop, AI agents operate continuously—executing multi-step tasks, calling APIs, and reasoning across extended chains of thought. Goldman expects this always-on behavior to drive a 24-fold increase in token consumption by 2030 . Each agent interaction consumes dramatically more compute, and with enterprises beginning to deploy agents at scale, the demand trajectory looks nothing like the linear growth curves that underpin consensus models.

The Real Bottlenecks: Power, Labor, and Physical Constraints

Goldman Sachs is remarkably blunt about where the true limits lie. In its report on powering the AI era, the bank states plainly: “a lack of capital is not the most pressing bottleneck—it’s the power needed to fuel it” . After a decade of flat electricity demand, global data center power consumption is projected to surge 160% by 2030 . The United States alone faces an estimated 45-gigawatt power shortfall for data centers by 2028, requiring 72 gigawatts of new capacity through 2030—the equivalent of roughly 72 large nuclear power plants .

The grid was not designed for this future. Transmission and permitting timelines for new natural-gas plants stretch five to seven years, wind and solar provide only intermittent supply, and nuclear is a longer-term solution . New gas combustion turbines, the workhorses of reliable power generation, are effectively sold out until 2030 .

Labor may prove to be the hardest constraint of all. Goldman estimates that approximately 760,000 additional electricians, linemen, and tradespeople are required to build the physical infrastructure AI demands, including 207,000 specialized roles that need three to four years of training . These are not jobs Silicon Valley can automate or offshore—they require boots on the ground, and the shortage means project timelines stretch further with every gigawatt of new demand .

The bank’s “Tracking Trillions” paper introduces the concept of “elongation risk”: power interconnection queues, permitting delays, and shortages of critical equipment like transformers and switchgear can extend build timelines well beyond initial plans. In stress scenarios, these delays feed back into demand-side doubt, creating a self-reinforcing cycle where projects take longer and the case for building more weakens . Even so, Goldman’s baseline estimate anticipates roughly $7.6 trillion in cumulative AI capital expenditures between 2026 and 2031 .

How Morgan Stanley’s Estimates Compare

Morgan Stanley’s projections have undergone their own dramatic upward revision. A year ago, the firm estimated combined hyperscaler capex at roughly $450 billion for both 2026 and 2027. After first-quarter earnings reports in 2026, analysts led by Brian Nowak raised those figures to approximately $800 billion for 2026 and $1.2 trillion for 2027 .

Morgan Stanley now forecasts $1.16 trillion in hyperscaler capex for 2027, a figure that exceeds Goldman’s baseline of roughly $1.1 trillion but falls short of Goldman’s $1.4 trillion upper bound . Through 2028, Morgan Stanley expects $2.9 trillion in global data center capital expenditures, with $1.4 trillion funded by hyperscaler cash flows and a $1.5 trillion financing gap that must be filled by debt, leases, and joint ventures .

Both banks agree that capex-to-sales ratios have entered uncharted territory. Morgan Stanley projects ratios of 34% to 39% from 2026 to 2028, exceeding the roughly 32% peak recorded during the dot-com era. When lease-adjusted figures are included, the ratios could climb as high as 44% to 45% .

The Shadow Ledger: Off-Balance-Sheet Leases and SPV Debt

Beneath the headline spending numbers lies a more troubling layer of financial engineering. Moody’s Ratings has estimated that the five largest U.S. hyperscalers—Amazon, Meta, Alphabet, Microsoft, and Oracle—hold $662 billion in future data center lease commitments that have not yet begun . Under Generally Accepted Accounting Principles, these obligations do not appear as current liabilities because services have not commenced. They sit off the balance sheet, visible mainly in footnotes .

When all undiscounted future lease commitments are tallied, the figure reaches an estimated $969 billion—roughly 113% of the combined adjusted debt of these five companies . As these leases commence over the coming years, they will begin to flow through income statements as operating expenses, potentially compressing free cash flow and limiting capacity for share buybacks that investors have long depended on .

A parallel concern is the growing use of special purpose vehicles to fund AI infrastructure. Major technology companies have structured more than $120 billion in data center debt through bankruptcy-remote SPVs that sit outside consolidated balance sheets . Morgan Stanley projects this off-balance-sheet funding mechanism could reach $800 billion by 2028 . These vehicles typically operate with thin equity cushions of 8% to 10%, rely on GPU collateral that depreciates rapidly, and involve lease terms as short as four years compared with the traditional ten-plus years .

The Oracle Warning and Widening Credit Spreads

Oracle has become a case study in how quickly AI financing assumptions can unravel. In late 2025, the company split with Blue Owl Capital over funding for a Michigan data center, exposing the fragility of the off-balance-sheet model. Oracle carries $124 billion in debt and $248 billion in lease commitments, and the market’s response was swift—credit was repriced “with brutal speed,” even for an investment-grade issuer .

The Bank for International Settlements has observed that credit default swap spreads for hyperscalers with lower credit ratings have already risen, reflecting both the sheer volume of debt supply and growing uncertainty about whether AI projects will generate adequate returns . The Financial Stability Oversight Council and the Bank of England have explicitly flagged the accumulation of off-balance-sheet AI infrastructure debt as a potential systemic vulnerability .

Concentration risk compounds the problem. Much of the SPV-based debt is tied to single-asset or single-tenant data centers. If the tenant fails or demand softens, the SPV structure offers limited recourse to the parent company’s balance sheet, creating the potential for cascading losses . PIMCO has also flagged the circular nature of AI financing, where suppliers such as GPU manufacturers extend credit or take equity stakes in the same SPVs they supply, exposing themselves to refinancing risks if capital markets tighten .

A Boom With Two Narratives

The AI infrastructure buildout is unprecedented in scale and speed. The five largest hyperscalers are on track to spend a combined $755 billion in 2026 alone—an 83% year-over-year increase . Morgan Stanley notes that the $800 billion figure for 2026 roughly matches what the entire non-tech group in the S&P 500 spent on capex the prior year .

Yet the financing structures that make this possible are generating their own risks. The bullish narrative rests on a 24-fold increase in token demand from AI agents that have not yet been deployed at scale. The bearish counterpoint, articulated within Goldman Sachs itself, is that the returns to date do not justify the investment . Between these two poles sit the physical realities: a power grid that cannot keep pace, a skilled workforce that does not exist in sufficient numbers, and a shadow ledger of nearly a trillion dollars in obligations that will soon come due, with consequences that extend well beyond the technology sector.