Wall Street spent the past year debating when the AI spending spree will finally slow, but Big Tech may not be done yet.
Hyperscalers are likely to spend a lot more on AI than the market expects — even after a series of massive upward revisions, according to analysts at Goldman Sachs in a Wednesday note.
“Consensus 2027 hyperscaler capex estimates are too conservative,” they wrote.
Goldman’s analysts estimate that hyperscaler capital spending could reach roughly $1.1 trillion in 2027, compared with the roughly $920 billion expected by Wall Street.
In a bullish scenario, spending could climb to $1.4 trillion.
Goldman’s key assumption is that the demand for AI computing power is still in its early stages. The bank forecasts that token consumption will increase 24 times through 2030, largely led by the rise of enterprise agents.
More tokens require more computing power, which in turn fuels demand for data centers, chips, networking equipment, and power infrastructure.
“Higher input costs also put upward pressure on the nominal dollars of capex required to support a given amount of token consumption,” the analysts wrote.
Still, the note highlighted a risk: Several companies have recently flagged token expenses associated with AI tools, raising questions about whether productivity gains will ultimately exceed the cost of running increasingly sophisticated models.
The tension underscores a broader debate playing out across corporate America: Companies are spending aggressively on AI while still struggling to prove the return on investment.
One of the strongest signals supporting Goldman’s bullish view comes from cloud providers themselves.
Google Cloud and Amazon Web Services reported a combined backlog of $832 billion as of the first quarter, up from $358 billion just six months earlier, Goldman’s analysts noted.
But the analysts don’t expect AI supply and demand to reach balance until at least the second half of 2027, suggesting spending could remain elevated for longer than investors expect.
The firm also argues that history suggests investors may be underestimating how large the buildout could become.
AI-related investment amounted to roughly 1.5% of GDP in 2026, according to Goldman. By comparison, investment booms tied to railroads, electrification, and automobiles reached peaks of roughly 2% to 3% of GDP.
The real bottleneck isn’t money
Even so, financing is unlikely to be the main constraint on future spending.
Instead, the bigger bottlenecks may be physical.
“There are numerous delayed data center projects in the pipeline and memory, power, and labor have been flagged as constraints to the capex build-out,” they wrote.
Goldman said stronger-than-expected capex spending should continue to support earnings growth for companies tied to the AI buildout, including semiconductor, networking, cooling, and power suppliers.
But the bank also warns that parts of the trade are becoming increasingly crowded. Valuations for many AI infrastructure stocks have expanded rapidly in recent months, with share prices outpacing earnings revisions in several parts of the sector, increasing the risk of volatility.
At the same time, evidence of widespread AI-driven productivity gains remains limited.
While 54% of companies discussed AI productivity during first-quarter earnings calls, only 11% quantified specific productivity benefits and just 2% quantified an impact on earnings, according to Goldman’s analysis.
Goldman’s note came amid a tech sell-off in recent days amid concerns about the war in Iran and an uncertain outlook for interest rates.
The Nasdaq 100 closed 2% lower on Wednesday and has fallen 6% since the start of Friday’s sell-off. The index remains 13% higher this year.
Nasdaq futures were up 0.6% at 12:03 a.m. ET on Thursday.
