• Goldman Sachs (GS) forecasts hyperscaler AI infrastructure spending will surge 54% to $1.2 trillion in 2027, above Wall Street consensus, with an upside case of $1.4 trillion.
  • Amazon (AMZN), Alphabet (GOOG), Microsoft (MSFT), Oracle (ORCL), and Meta (META) are already on track to spend $800 billion this year, but the break-even math requires roughly $300 billion in annual AI revenue.
  • The spending boom is increasingly reliant on debt financing, with an estimated $400 billion in investment-grade bond issuance expected in 2027 to fund the buildout.

Goldman Sachs Lifts AI Capex Forecast to $1.2 Trillion

Goldman Sachs has raised its outlook for AI infrastructure spending by the five largest U.S. hyperscalers—Amazon, Alphabet, Microsoft, Oracle, and Meta—projecting they could collectively deploy roughly $1.2 trillion in 2027, a 54% increase from the prior year and above prevailing Wall Street estimates of around $920 billion, according to people familiar with the analysis.

The forecast underscores an unusually capital-intensive phase of the AI race, where the central question is no longer whether demand exists but whether revenue, power supply, data-center capacity, financing, and regulation can scale quickly enough to earn an adequate return. In an upside scenario, Goldman projected spending could reach $1.4 trillion, contingent on cash flow and investment-grade debt-market capacity.

Recent company disclosures suggest the spending momentum is already strong. Amazon recorded about $44.2 billion in quarterly capital expenditures, Alphabet $35.67 billion, Microsoft $30.88 billion, and Meta raised its full-year capex guidance to $125–$145 billion, citing components and data-center costs. Microsoft’s fiscal third-quarter capex jumped 84% year over year, while its AI revenue surpassed a $37 billion annualized run rate.

The Break-Even Challenge and Rising Leverage

The massive outlay comes with a hefty price tag. Goldman estimates hyperscalers need roughly $300 billion in annual AI revenue to break even on the buildout, a figure that highlights the gap between technical adoption and economic success. To fund the investment, the group is expected to increasingly turn to debt markets, with one estimate cited by Goldman pointing to about $400 billion of global investment-grade bond issuance in 2027 against approximately $1.14 trillion in capex.

“The financing mix is shifting,” said an analyst briefed on the forecast, who requested anonymity to discuss internal deliberations. “Preserving cash for operations and acquisitions is a priority, but it makes borrowing costs and credit-market access far more important.”

Spending is concentrated in AI accelerators and servers, data-center buildings, networking, storage, cooling, and electricity infrastructure. IDC found that servers comprised 98% of AI-centric infrastructure spend in the second quarter of 2025, with accelerator-equipped servers accounting for 91.8% of total AI-server spending. The near-term constraint is increasingly physical rather than software-related: available power, transmission connections, data-center land and construction capacity, networking gear, advanced chips, and skilled labor.

Power Grid and Regulatory Hurdles

The electricity effect is particularly large. The U.S. Department of Energy cites an estimate that data centers could consume up to 9% of U.S. annual electricity generation by 2030, up from about 4% of load in 2023. A separate DOE/LBNL-based estimate places data centers at roughly 6.7%–12.0% of U.S. electricity consumption by 2028, versus 4.4% in 2023. For utilities, the boom creates a potential growth opportunity alongside reliability and cost-allocation challenges.

“We have a constant balance with the banks, which really we consider our partners and not only our binary competitors,” said a private credit executive involved in data-center financing, speaking on condition of anonymity. “It’s much more of a convergence between the two solutions.”

Policymakers are generally supportive. The White House’s AI Action Plan calls for more chip factories, data centers, and power generation, including identifying federal land suitable for such projects. But rapid buildout can still face local opposition, transmission bottlenecks, environmental review requirements, and limited generation capacity. Internationally, U.S. export controls on advanced chips to China continue to evolve, with the Commerce Department moving to case-by-case review for licenses involving Nvidia (NVDA) H200, AMD (AMD) MI325X, and similar chips in January 2026, subject to security conditions.

Market Implications and Historical Parallels

The market impact is shifting from software enthusiasm to an industrial buildout. AI is now driving spending in chips, electrical equipment, cooling, power generation, fiber, networking, and real estate. IDC reported that cloud and shared deployments made up 84.1% of AI infrastructure spending in the second quarter of 2025, with hyperscalers, cloud providers, and digital-service providers responsible for 86.7% of the quarter’s spending.

Goldman explicitly framed a high-capex AI scenario as comparable to historical investment waves in railroads and automobiles, suggesting incremental AI investment of about 2%–3% of GDP could support roughly $1.1 trillion in hyperscaler capex in 2027. The bullish outcome is that AI becomes embedded across enterprise software, search, advertising, coding, and scientific research, absorbing upfront costs through higher cloud revenue and productivity gains. The bearish outcome is an overbuild period: AI demand grows but does not monetize fast enough, capacity becomes underutilized, debt burdens rise, and investors force capex discipline.

The base case is likely between those extremes: sustained AI investment, but with periodic spending pauses and a growing focus on utilization, margins, power efficiency, and provable customer demand rather than raw model scale alone. For now, Goldman’s forecast signals that the spending race is far from over—and that the bill is increasingly being financed by borrowed money.

Correction: An earlier version of this article misstated the projected investment-grade bond issuance for 2027. It is approximately $400 billion, not $400 million.