Goldman Sachs says AI spending could hit $1.4T by 2027

Goldman Sachs Says Wall Street Is Underestimating the AI Spending Boom

Goldman Sachs issued a research note on Wednesday arguing that consensus forecasts for hyperscaler capital expenditures are too conservative, projecting that spending could reach $1.1 trillion in 2027 — well above the $920 billion that analysts currently expect.yahoo

The Bull Case for $1.4 Trillion

The bank’s portfolio strategy research team drew on historical precedents from previous technology build-outs, including the development of railroads and the automobile industry, to argue that AI infrastructure investment is likely to follow a steeper and longer trajectory than markets are pricing in. In a more extreme upside scenario, Goldman estimated spending could reach $1.4 trillion in 2027, which would represent a growth rate similar to this year’s 84% surge.morningstar

Goldman projects hyperscaler capital expenditures will hit $757 billion in 2026, an 84% increase from 2025. The current consensus of $920 billion for 2027 implies a deceleration to 22% growth — a slowdown the bank views as unlikely given the dynamics of AI adoption.gurufocus

Enterprise Agents and the Token Explosion

Underpinning Goldman’s bullish outlook is its forecast that global token consumption will multiply 24 times by 2030, driven largely by the rise of enterprise AI agents rather than traditional human-prompted interactions. The bank expects total AI queries to climb from 5 billion in 2025 to 23 billion by 2030, with roughly 30% of those coming from agentic use cases. By 2040, if enterprise agents reach full-scale adoption, token consumption could hit 55 times current levels.duckittech

Physical Constraints, Not Capital

Goldman has previously noted that the real bottlenecks constraining AI infrastructure are physical — power supply, memory, and labor — rather than financing. The bank has estimated that global data center electricity demand will increase 50% by 2027, with 60% of that growth requiring entirely new power capacity. Supply and demand for AI compute are not expected to reach equilibrium until at least the second half of 2027, according to the firm’s earlier research, which projected datacenter occupancy peaking at around 93% before constraints begin to ease.theregister

Morgan Stanley has echoed a similar view, estimating U.S. hyperscalers alone will spend more than $800 billion on capex in 2026 — roughly matching what the entire non-tech group in the S&P 500 spent in the prior year.vtmarkets