NextFin News - Wall Street is bracing for an abrupt slowdown in artificial-intelligence infrastructure spending, but Goldman Sachs says the slowdown is a mirage. In a research note published Friday, the bank estimates hyperscaler capital expenditure will climb roughly 50% to about $1.2 trillion in 2027 - far above the roughly $920 billion consensus that would imply growth decelerating from 84% this year to just 22% next year. The call reframes the AI buildout not as a cycle nearing its peak, but as a structural investment regime comparable to the railroad and automobile buildouts of earlier eras, with spending still far short of the 2% to 3% of GDP those transformations absorbed.
The Consensus That Goldman Is Calling Wrong
The tension at the heart of the note is a numbers gap, not a disagreement about direction. Analysts covering the hyperscalers - Amazon, Alphabet, Microsoft, Meta and the other cloud builders - have already lifted their 2027 capital-spending estimates to about $920 billion. That figure sounds enormous, and it is. But it embeds a sharp deceleration: growth would fall from an estimated 84% in 2026 to 22% in 2027.
Goldman's strategists, led by economist Ryan Hammond, reject that inflection point. Their base case puts 2027 hyperscaler capex at roughly $1.1 trillion, implying 45% growth. In a more extreme upside scenario, the bank estimates that hyperscaler cash-flow generation plus investment-grade credit-market capacity could support as much as $1.4 trillion in capex, representing 89% growth. The headline $1.2 trillion figure sits between those two reference points and captures the bank's central message: the market's linear extrapolation is the error.
The historical analogy is the backbone of the argument. Goldman estimates AI-related investment will account for about 1.5% of GDP in 2026. The investment peaks driven by railroads, electrification and automobiles reached 2% to 3% of GDP. If incremental AI investment converges toward that historical range, the spending path does not flatten in 2027 - it steepens. A separate Goldman Sachs Research estimate, published on the bank's website, projects global AI investment at roughly $1 trillion in 2026, with AI capex rising from 1.8% of US GDP this year to 2.5% in 2027 and 2.8% in 2028.
The upward revision is not an isolated event. Goldman's equity analysts now expect a combined $5.3 trillion in capital spending from 2025 through 2030 by the large technology companies leading the buildout, up from $4.5 trillion before first-quarter earnings. Consensus 2026 capex for the group has climbed to $527 billion from $465 billion at the start of the third-quarter earnings season. Hyperscalers spent $106 billion in capex in the third quarter alone, a 75% year-over-year increase, though that growth rate is expected to slow to 49% in the fourth quarter and to 25% by the end of 2026.
Even the widely cited hyperscaler capex figure of $794 billion for 2026 understates the true scope, in Goldman's view. The bank's augmented measure - which adds non-hyperscaler AI investment and adjusts for the geographic split of spending - indicates the $794 billion number likely understates total global AI capex by around $200 billion, while simultaneously overstating the amount of US investment in AI by a similar margin. That adjustment matters for the GDP-share math: a global AI-investment base closer to $1 trillion in 2026 leaves more runway before the 2% to 3% historical ceiling is reached.
Why the Spending May Not Have a Cycle to Mean-Revert To
The cyclical read is seductive because it fits a familiar pattern: a capital-spending boom, an overbuild, a hangover. Under that view, today's hyperscalers are laying fiber and pouring concrete for demand that will not materialize, and the reckoning arrives when utilization falls short. The counter-argument - and Goldman's - is that this cycle has a different engine.
The mechanism runs through token consumption, not just capital budgets. Goldman forecasts global token consumption will surge 24-fold by 2030, led by the rise of enterprise agents. More tokens require more computing power, which in turn drives demand for data centers, high-performance chips, networking equipment and power infrastructure. That chain matters because it converts an abstract spending figure into a physical throughput requirement: the buildout is the supply side of a demand curve that the bank sees as still in its early stages.
There is also a funding argument. The bank's bull case - up to $1.4 trillion in capex - rests on the observation that hyperscalers combine large internal cash generation with access to investment-grade credit markets. A Goldman Sachs Research piece on data-center financing noted that hyperscalers already account for a large share of new corporate borrowing and may face market-saturation constraints as public-debt investors absorb ever-larger amounts from the same handful of issuers. That is the constraint Goldman is betting will not bind soon: balance sheets, not lenders, set the pace.
"The strong balance sheets of the largest hyperscalers - and their recent willingness to employ those balance sheets - support the outlook for continued capex spending growth," Hammond and his team wrote.
Private markets are being drafted in to close the gap between what hyperscalers can fund on their own books and what the buildout requires. The same Goldman research on data-center financing argues that private-market capital will play a growing role in funding AI infrastructure, precisely because the scale of planned spending exceeds what any single balance sheet can absorb efficiently. That layering of funding sources - corporate cash flow, investment-grade bonds, and private capital - is what allows the $1.4 trillion upside scenario to be financially supportable rather than aspirational.
This is why the railroad and automobile analogy carries weight. Those were not inventory cycles that reverted; they were general-purpose technology buildouts that re-priced the cost of moving goods and people across an economy. If AI is the same class of technology, then the relevant comparison is not last year's capex growth rate but the total investment a new infrastructure layer absorbs before it saturates. On that measure, 1.5% of GDP looks like a midpoint, not a peak.
The Second-Order Transmission: From Capex to Credit Spreads to Equity Dispersion
The first-order effect of Goldman's call is straightforward: more capex means more revenue for the infrastructure complex. The second-order effect is where the story gets interesting, and it runs through the credit market before it reaches equities.
As hyperscalers fund a $1.2 trillion buildout, their share of new corporate bond issuance rises, and the concentration risk that Goldman flags - public-debt investors absorbing ever-larger amounts from the same handful of issuers - begins to price into spreads. The Oracle episode is the warning flare: the company recently reported higher-than-expected quarterly capex, primarily for data-center construction, and its stock fell more than 8% in a single day - a reminder that investors now penalize capex that outruns visible revenue before the earnings can catch up.
The equity-market transmission is dispersion, not a uniform rally. The median AI infrastructure stock now trades at 26 times earnings, the highest multiple since the launch of ChatGPT, and Goldman cautions that "recent valuation expansion and positioning dynamics suggest additional volatility ahead." Hyperscaler stock-price correlation has fractured as investors separate spending from revenue outcomes - a sign that the market is no longer rewarding capex for its own sake. The beneficiaries are the picks-and-shovels layer: chipmakers, data-center builders, power and cooling providers, and networking-equipment vendors whose order books are visible quarters ahead. The exposed are the hyperscalers themselves if capital intensity rises faster than pricing power, and the software layer where Goldman expects the "AI-enabled versus AI-disrupted debate" to drive return dispersion.
The Fault Line: Spending Is Visible, Monetization Is Not
Here is the question the market is not asking loudly enough: even if Goldman is right about the size of the buildout, who captures the equity value? Capex growth and shareholder returns are not the same variable, and the gap between them is where the volatility lives.
Goldman's own enterprise-adoption data exposes the gap. Roughly 54% of companies discussed AI in the context of productivity during first-quarter earnings calls, but only 2% quantified the impact of AI productivity on earnings. The bank described enterprise adoption as "nascent." In other words, the demand side of the ledger is still mostly narrative. If hyperscalers spend $1.2 trillion in 2027 and enterprises cannot yet measure the profit impact of the tools they are buying, the utilization math becomes a bet on future adoption rather than a reflection of current returns.
The bank added that the token-price war among large-language-model providers such as OpenAI and Anthropic is intensifying, which could pressure revenue and cash flow at those companies and test their ability to sustain massive computing investments. Falling token prices are good for adoption and bad for the revenue that justifies the buildout - a tension that will only resolve when enterprise usage scales faster than price compression.
The Strongest Case Against the Bull Thesis
The bear case deserves its full weight, because it attacks the core of Goldman's argument rather than its edges. The strongest counter-thesis is not that AI demand is fake - it is that the financing capacity and physical supply chain cannot scale at the pace the model assumes, and that the equity market will discount the spend long before the infrastructure is complete.
On the physical side, Goldman itself identifies the bottlenecks: power supply, memory shortages, labor shortfalls and data-center project delays. These are not demand problems; they are supply constraints that can cap growth regardless of how much capital is available. A $1.2 trillion plan that cannot be physically deployed is not a $1.2 trillion outcome. Power in particular is the hardest constraint to finance away: data centers need grid connections that can take years to permit and build, and no amount of balance-sheet capacity shortens a substation queue.
On the financing side, Morgan Stanley analysts have described the concentration of index-level investment within AI as unprecedented, estimating that capex-to-sales ratios could run as high as 38%, 44% and 45% in 2026 through 2028, versus a roughly 32% peak in the dot-com era. That comparison is the bear's best evidence: the dot-com buildout ended in a collapse of valuations precisely because capital intensity outran revenue. If hyperscalers are walking the same path, the 26x multiple on AI infrastructure stocks is not a sign of confidence but of complacency.
The answer to the bear case rests on two distinctions. First, the dot-com builders were largely unprofitable companies raising capital in speculative markets; today's hyperscalers are cash-generative incumbents funding spend from operations and investment-grade debt. Second, the dot-com demand was consumer internet adoption, which did saturate; the AI demand driver is enterprise automation of cognitive work, which Goldman's 24-fold token-consumption forecast treats as still in its first inning. Neither distinction guarantees success, but both argue that the analogy cuts both ways.
The falsifying signal is concrete: if core hyperscaler capex growth in 2027 comes in at or below the consensus 22% - roughly $920 billion - and at least two of the four largest builders guide to flat or declining spending for 2028, the structural-supercycle thesis is wrong and the cyclical-overbuild read takes over. A secondary signal is enterprise monetization: if the share of companies quantifying AI's earnings impact remains in the low single digits through the next two earnings seasons, the demand narrative loses its anchor.
What to Watch and What It Means
The near-term read is volatility, not direction. Goldman's own warning about valuation expansion and positioning means the infrastructure complex can rise on the capex narrative and fall on the monetization narrative in the same quarter. The medium-term read depends on the earnings-impact data: the 2% of companies that currently quantify AI productivity gains must broaden, or the spending will begin to look like capacity ahead of demand. The long-term read is the structural one: if AI is a general-purpose technology on the railroad-and-automobile trajectory, then the total addressable investment is measured in trillions over a decade, not in single-year growth rates.
Scenarios, by trigger:
- Base case: 2027 hyperscaler capex lands between $1.1 trillion and $1.2 trillion (roughly 45% to 50% growth), enterprise AI monetization slowly broadens beyond the current 2%, and infrastructure beneficiaries outperform while hyperscaler returns disperse.
- Upside case: incremental AI investment reaches the 2% to 3% of GDP range Goldman compares to historical buildouts, pushing 2027 capex toward the $1.4 trillion bull scenario; power and memory constraints ease, and the equity multiple on the infrastructure complex expands further.
- Downside case: capex growth decelerates to the consensus 22% or below, two or more hyperscalers cut 2028 guidance, and enterprise monetization fails to broaden; the 26x multiple compresses and the overbuild narrative returns.
For investors, the practical implication is selectivity. The capex number is a rising tide for the picks-and-shovels layer, but it is also a warning that the hyperscalers are entering a period where every dollar of spending will be scrutinized for its revenue conversion. The market is about to stop rewarding spending for its own sake.
The central judgment: Goldman is likely right that the AI buildout has not peaked - the GDP-share math and the token-consumption trajectory argue against a 2027 inflection. But the bank's own data on enterprise monetization is the weaker link in the chain. The $1.2 trillion is the easy part; the hard part is proving that the spending earns its keep. This is not the market pricing a cyclical peak - it is the market pricing a buildout whose returns have not yet been measured, and that is a far more volatile proposition.
Explore more exclusive insights at nextfin.ai.
