Goldman Sachs: Hyperscaler Capex Could Reach $1.1 Trillion by 2027
Alina Collins
Goldman Sachs estimates hyperscaler capital spending will hit $1.1 trillion in 2027, nearly 20% above Wall Street consensus; the bank argues AI compute demand is still in its early stages and the market is systematically underpricing both the scale and duration of this buildout.
Wall Street has already raised forecasts — why does Goldman say it's still not enough?
Goldman estimates 2027 hyperscaler (Google, Amazon, Microsoft and peers that build their own data centers) capex at roughly $1.1 trillion, rising to $1.4 trillion in a bull case.
Wall Street consensus sits at about $920 billion — already revised up multiple times, yet Goldman says it still trails the actual demand curve.
This means → even though the Street is chasing the numbers, it isn't chasing fast enough; the AI infrastructure spending cycle may last far longer than current stock prices imply.
How strong is demand, exactly?
By the end of Q1, Google Cloud and Amazon AWS had a combined backlog of unfulfilled orders worth $832 billion — up from just $358 billion six months earlier, more than doubling.
Goldman forecasts token consumption (the basic unit measuring how much work AI models do) will grow 24× by 2030, driven mainly by mass deployment of enterprise AI agents.
In plain terms = orders are piling up faster than capacity can be built; Goldman expects supply and demand won't reach balance until at least H2 2027.
How does this buildout compare to past tech revolutions?
AI-related investment in 2026 accounts for roughly 1.5% of U.S. GDP.
Historically, peak investment during the railroad, electrification, and automobile revolutions reached 2–3% of GDP.
This means → if AI's economic impact is comparable to those transformations, current spending levels may be only halfway to the peak, not near the top.
The real bottleneck isn't money — so what is it?
Goldman states explicitly: financing capacity is not the binding constraint.
The real bottlenecks are physical — data-center project delays, memory supply shortages, power capacity limits, and labor shortfalls are all flagged as material constraints on capex expansion.
In plain terms = the hyperscalers have the cash; what they lack are the physical inputs to deploy it — not enough power, not enough workers, not enough chips.
Valuations are running ahead of earnings — where's the risk?
Goldman warns that valuation expansion in AI infrastructure stocks has outpaced earnings revisions; crowding in semiconductors, networking, cooling, and power-supply names is rising, and so is volatility risk.
In Q1 earnings calls, 54% of companies mentioned AI productivity, but only 11% quantified the benefits and just 2% tied them to an earnings impact.
This reflects the central unresolved question of this AI investment boom: companies are spending heavily, but whether that capex converts into measurable returns remains far from answered.
Content is for reference only, not financial advice.