Based on Goldman Sachs Global Investment Research; reporting via Yahoo Finance and Seeking Alpha.
Goldman Sachs Global Investment Research maps a 6-7x US spending surge — from $156B in 2022 to a projected $1T+ in 2027 — and argues the Street is still underestimating the scale.
What Happened
Goldman Sachs Global Investment Research released a chart and accompanying note arguing that Wall Street consensus for 2027 hyperscaler capital expenditure is too low. The chart tracks combined capex for Amazon, Microsoft, Google, Meta, and Oracle/OpenAI-tier spenders: $156B in 2022, $254B in 2024, $443B in 2025, a projected $764B in 2026E, and roughly $1,018B — more than one trillion dollars — in 2027E. Goldman’s own base case for 2027 sits at approximately $1.1 trillion; its bull case reaches $1.4 trillion. The range itself signals how genuinely uncertain — and how genuinely large — the outcome could be.
On the China side, the data is equally striking in its relative scale. Goldman tracks Alibaba, Tencent, ByteDance, and Baidu going from roughly $8B in 2022 to a projected $123B in 2027E — a 15x surge in absolute terms. But the absolute gap is the story: at those projections, US hyperscalers would be spending approximately eight times what China’s four largest AI infrastructure investors spend in the same year. That is compute bifurcation rendered as a single ratio.
Two critical hedges belong upfront. First, the 2026E and 2027E figures are estimates — capital budgets that can expand, compress, or shift timing. Second, ByteDance is explicitly listed as “not covered” in Goldman’s research; its figure is extrapolated, not directly sourced from company disclosures. The base-to-bull spread of $1.1T–$1.4T for 2027 reflects that uncertainty honestly. These are projections of committed intent, not spend already booked.
The key insight: This is not a chart about individual company budgets. It is a chart about a new category of sovereign-scale infrastructure commitment — one that, if Goldman’s projections hold, will consume a share of US GDP last seen during railroad construction and the electrification of the 20th century. The number is not the headline. The trajectory is.
The Structural Read
Understand this chart and you understand why every other scarcity in the AI economy exists. A trillion dollars of annual US hyperscaler capex is not a downstream effect of the AI boom — it is the primary cause of nearly every supply crunch, power constraint, and stack consolidation happening simultaneously. Memory is short because demand from this capex pipeline overwhelms fabs: Micron’s $250B AI memory buildout and SK Hynix’s record earnings are the supply side racing to keep pace. Power is the binding constraint because a data center campus sized to absorb this spend requires gigawatts, not megawatts — which is exactly why Meta is building gas-fired generation in Alberta rather than waiting for the grid. The entire AI stack is being verticalized because at this capex scale, dependency on external suppliers is an existential risk — the Great AI Verticalization is the hyperscalers’ rational response to owning the spend but not the supply chain.
The US-China gap is equally structural. An ~8x spending differential by 2027E is not a policy gap or a talent gap — it is a compute gap made quantitative. China does not attempt to match US hyperscaler spend; it rations foreign chips and optimizes for efficiency at a lower absolute base. DeepSeek’s architectural innovations are in part a product of scarcity: when you cannot buy your way to scale, you engineer around it. Two AI economies are forming — one that competes on raw capital deployment, one that competes on compute efficiency per dollar. These are not converging trajectories.
Goldman’s own framing reaches for the longest historical analogies: railroads in the 1870s, electrification in the early 20th century, the internet in the late 1990s. Each consumed 2–3% of GDP at peak investment. Each was eventually justified by the productivity it unlocked — but each also produced a brutal overbuild phase before monetization caught up. The comparison is Goldman’s, not a rhetorical flourish: at $1T+ in annual capex, AI infrastructure is approaching that GDP share threshold. The question is not whether this is the defining infrastructure buildout of the era. The question is whether AI monetization scales fast enough to validate the return on that capital before the cycle turns.
Goldman Sachs Global Investment Research
“Consensus 2027 hyperscaler capex estimates are too low. Our base case sits at approximately $1.1 trillion, with a bull case reaching $1.4 trillion — a range that reflects genuine uncertainty about AI monetization speed, not uncertainty about the direction of spend.”
Map of AI — Layer 0: Infrastructure Capex
The Engine Beneath Every Layer
The Map of AI framework identifies nine layers of the AI stack. Goldman’s chart is the financial representation of Layer 0: the raw infrastructure capital that makes every layer above it possible — or scarce. When this layer compounds at 6-7x over five years, every layer above it simultaneously becomes more valuable (more demand) and more constrained (more competition for the physical inputs). That is the core structural dynamic of the current AI cycle.
Three Implications
IMPLICATION 1 — SCARCITY IS STRUCTURAL, NOT CYCLICAL
Every physical bottleneck in AI — HBM memory, advanced packaging, grid power, data center land — is a direct output of this capex trajectory. These are not temporary supply mismatches. At $443B in 2025 accelerating toward $1T in 2027E, the demand signal to every supplier is clear and durable. Companies positioned in the supply chain for these physical inputs — memory, power infrastructure, networking silicon — are not riding a wave. They are inside a multi-year structural demand commitment.
IMPLICATION 2 — THE US-CHINA GAP IS THE GEOPOLITICAL FACT OF THE DECADE
An ~8x infrastructure spending gap by 2027E is not a technology lead — it is a capital moat. China’s response (chip rationing, efficiency-first architecture, domestic fab acceleration) is rational given the constraint. But it produces a structurally different AI economy: one optimized for doing more with less, not one that closes the absolute compute gap. Two AI paradigms — capital-intensive scale versus constraint-driven efficiency — will produce meaningfully different model capabilities, deployment economics, and geopolitical leverage over the next five years.
IMPLICATION 3 — A TRILLION DOLLARS A YEAR HAS TO EARN A RETURN
Goldman’s railroad/electrification comparison is the most honest framing available. Both analogies validate the long-run thesis and warn of the short-run overbuild. If AI application-layer monetization — autonomous agents, enterprise workflow replacement, inference-at-scale — does not accelerate to match this capital deployment by 2027–2028, the financial pressure on hyperscaler balance sheets will be acute. The bull case ($1.4T) requires not just spending the capital, but earning it back. That monetization race is the central unresolved question in the AI economy.
