A Number So Large It Bends the Economy Around It
Goldman Sachs has put a price tag on the artificial-intelligence era, and it is staggering: roughly $7.6 trillion in cumulative capital expenditure on AI infrastructure between 2026 and 2031. That is not annual spending. It is the six-year running total the bank's Global Research division expects the industry to pour into chips, data centers and electricity to keep the AI boom alive — a sum equivalent to nearly one-quarter of annual U.S. GDP and roughly 1.4 times the entire annual output of Germany, the world's third-largest economy.
The figure, laid out in a Goldman report titled "Tracking Trillions: The Assumptions Shaping the Scale of the AI Build-Out," has become the new shorthand for the sheer financial gravity of the moment. It surfaced again this month when Anthropic president and co-founder Daniela Amodei cited the scale of the buildout at Bloomberg Tech, framing it as the unavoidable cost of frontier AI. "It's a really big upfront cost to train the models and to serve inference on them," Amodei said, speaking as her company's valuation approached the trillion-dollar mark.
Where the $7.6 Trillion Actually Goes
Goldman's model is not a single guess but a layered map. The bank breaks the total into three buckets: roughly $5.1 trillion for compute — the chips and servers themselves — $2.15 trillion for data centers, and about $358 billion for power generation and grid connection. Within the compute layer, the bank expects Nvidia to capture the lion's share, projecting the chipmaker will take roughly 75 percent of that $5.1 trillion over the six-year window.
The run-rate implied by the model is its own kind of vertigo. Goldman's baseline has annual AI capex rising from about $765 billion in 2026 to roughly $1.6 trillion by 2031 — more than doubling in five years. For context, the largest hyperscalers — Microsoft, Amazon, Alphabet and Meta, plus Oracle — have already guided toward a combined $635 billion to $690 billion in capital spending for 2026 alone, a 67 to 74 percent jump over the prior year.
Crucially, Goldman stresses that the $7.6 trillion is not a prophecy but the output of a handful of supply-side assumptions, any one of which could move the number by hundreds of billions. The most volatile, the bank argues, is how long an AI chip stays economically useful.
The Depreciation Wild Card
Today, hyperscalers typically depreciate their GPU servers over four to six years. But Nvidia has shifted to an annual product cadence, with each generation delivering order-of-magnitude gains in performance and energy efficiency — which makes hardware bought today look obsolete far faster than the accounting assumes. Goldman's sensitivity analysis is blunt: shortening the assumed chip lifespan from five years to three would push implied annual depreciation from 2026–2031 from roughly $3 trillion to nearly $4 trillion. Stretch it to seven years and the figure falls to about $2.2 trillion.
That single accounting choice — buried in footnotes today — could swing the economics of the entire buildout, and with it the reported profitability of the companies financing it.
What $7.6 Trillion Means for the Real World
A commitment of this magnitude does not stay on a spreadsheet. It lands on power grids first. Even Goldman's relatively modest $358 billion power allocation assumes a wave of new generation, transmission and substation capacity that utilities across the U.S. are only beginning to plan for. Data-center electricity demand is already reshaping regional grids, delaying coal-plant retirements and reviving interest in nuclear — and the build is still in its early innings.
It lands on chip demand second. If Nvidia is to absorb three-quarters of a $5.1 trillion compute bill, the entire semiconductor supply chain — from TSMC's advanced packaging to high-bandwidth memory suppliers like Micron and SK Hynix — must scale in lockstep. Any bottleneck ripples outward.
And it lands squarely in the middle of the returns-on-investment debate that has shadowed the AI trade for two years. The skeptics' case is not subtle. The MIT Project NANDA study found that 95 percent of enterprise generative-AI pilots produced zero measurable impact on profit and loss. Sequoia's David Cahn has calculated a roughly $600 billion annual revenue gap between what the industry is spending on infrastructure and what AI products are actually earning — a gap that, by his accounting, is widening rather than closing into 2026.
That is the crux of the bubble question. Capex is growing materially faster than the cloud revenue it is supposed to generate. If the productivity payoff arrives — and bulls point to early signals, such as Meta crediting AI recommendations for measurable gains in user engagement — then $7.6 trillion looks like the foundation of a supercycle. If it does not, it looks like the most expensive overbuild in corporate history, with a depreciation cliff waiting underneath.
What to Watch Next
The tell will be in the quarterly numbers. Watch whether hyperscalers begin to quietly extend chip depreciation schedules — a maneuver that flatters near-term earnings while masking the real pace of obsolescence — and whether any of them blink on their 2026 capex guidance. Watch the power-purchase agreements, because grid capacity, not capital, may become the binding constraint first. And watch the revenue line: the gap between Cahn's $600 billion and the industry's actual AI earnings is the single clearest gauge of whether Goldman's $7.6 trillion is an investment thesis or a warning. For now, the world's biggest companies are betting a sum the size of a major nation's economy that it is the former.
"It's a really big upfront cost to train the models and to serve inference on them."- Daniela Amodei, President and Co-Founder, Anthropic