Why Global AI Investment Is Closer to $1 Trillion Than $800 Billion

The most widely cited benchmark for global AI investment is built from the public earnings disclosures of four US technology companies. According to Goldman Sachs Research, that proxy is missing more than US$200 billion in 2026 spending alone. As the AI infrastructure build-out accelerates into one of the largest capital deployment events in modern economic history, the figure most analysts reference, approximately US$800 billion, is constructed from a narrow and structurally incomplete measurement base. Goldman Sachs economist Joseph Briggs has published analysis arguing the true global AI investment figure exceeds US$1 trillion for 2026, with conventional methodology systematically excluding categories of spending that are material at a macroeconomic level. What follows is an examination of exactly where the US$800 billion consensus breaks down, what Goldman’s adjusted methodology captures instead, and why the gap between these two figures carries direct implications for energy infrastructure, data centre power demand, and critical minerals markets over a multi-year horizon.

Goldman Sachs Research on global AI investment details the augmented methodology developed by economist Joseph Briggs, identifying the three structural exclusions that cause the standard hyperscaler capex proxy to produce a reading approximately US$250 billion below the adjusted estimate.

Why the US$800 billion consensus is built on a narrow foundation

The prevailing US$800 billion figure for global AI investment in 2026 is not a comprehensive tally of capital deployed. It is an extrapolation from the aggregated capital expenditure disclosures of a handful of listed US hyperscalers, effectively a proxy inferred from a small and structurally unrepresentative sample.

Goldman Sachs’s own baseline AI infrastructure capex model, covering compute, data centres, and power, yields approximately US$765 billion for 2026. That figure sits within the same range as the consensus and confirms the starting point is not in dispute.

The AI investment cycle has historically been tracked through hyperscaler earnings disclosures, a methodology that made reasonable sense when the largest public cloud platforms dominated total spend, but which has become structurally inadequate as private capital, non-US firms, and cross-border deployments now account for a material share of global outlays.

The dispute is what happens next. Goldman’s adjusted methodology, which corrects for three structural exclusions in the standard benchmark, produces an estimate of US$1.019 trillion for the same year. The gap between US$765 billion and US$1.019 trillion, roughly US$250 billion, is the methodological question this analysis sets out to answer.

Goldman Sachs Research has noted that Wall Street analysts have repeatedly underestimated AI spending by focusing mainly on the largest public cloud platforms, a measurement approach that structurally excludes significant categories of capital deployment.

Investors who rely on hyperscaler capex as their proxy for total AI investment are pricing in a structurally incomplete picture. The benchmark’s construction is the first place to look.

The geography of a trillion-dollar build-out

The AI investment story is widely framed as a US technology sector story. Goldman’s adjusted estimates reframe it as a global capital deployment event with a geographic distribution that most market participants are not actively tracking.

The numbers split as follows:

• US AI investment in 2026: approximately US$581 billion, representing roughly 57% of the global total
• Non-US AI investment in 2026: approximately US$440 billion, representing roughly 43% of the global total
• A significant portion of US hyperscaler capital is physically deployed outside the United States, through international data centre construction and network infrastructure, meaning the geographic origin of spending and the geographic destination of spending are not the same

Just over half of global AI capital lands inside the United States. The remaining US$440 billion is distributed across other jurisdictions, creating demand for energy infrastructure, grid equipment, and materials in regions that US-centric capex models do not capture. For global resources investors, that non-US figure is not a residual. It is a structural demand signal in its own right.

Non-US AI infrastructure investment is no longer a residual category driven solely by sovereign digital policy ambitions; it increasingly reflects strategic commitments from the largest US AI developers who are deploying capital internationally to access power, land, and regulatory environments that support large-scale data centre construction outside the continental United States.

Systematic exclusions in AI investment benchmarks

The gap between US$800 billion and US$1 trillion is not the product of a single modelling assumption. It accumulates from three structural exclusions, each of which represents a category of real-world spending with a physical presence somewhere in the global economy.

1. Exclusion of private companies. The standard benchmark captures only publicly listed firms. Goldman’s adjusted methodology explicitly incorporates spending by private AI developers, infrastructure providers, and other firms outside the major US hyperscalers, none of which appear in public earnings-based capex tallies.
2. Geography distortions. The standard approach excludes spending by organisations headquartered outside the United States. It also misrepresents where US hyperscaler capital is actually deployed, since these platforms build data centres and networks globally, meaning a portion of their reported capex physically lands in non-US jurisdictions.
3. Non-AI capex bundled in hyperscaler totals. Total reported capex for large platforms includes general cloud infrastructure, content delivery networks, logistics technology, and other non-AI assets. Isolating AI-specific spending from this broader technology expenditure is essential to producing an accurate AI investment figure.

How the exclusions compound each other

These three gaps do not operate in isolation. A private, non-US AI infrastructure provider building data centres in Southeast Asia is invisible to the standard benchmark on all three counts simultaneously. The compounding effect means the consensus figure does not merely undercount by one category; it structurally misses entire segments of global capital deployment.

Each exclusion creates real-world demand for power, land, cooling, and materials. Recognising them individually clarifies why the US$800 billion figure produces a systematically low reading of the AI investment cycle’s true scale.

What global AI investment actually looks like when measured correctly

Goldman Sachs’s primary adjusted estimate for 2026 global AI investment is US$1.019 trillion, derived from an enhanced hyperscaler capex methodology that corrects for the three exclusions identified above.

The figure does not rest on a single model. Goldman validates it with two independent cross-check methodologies, and all three converge within a narrow band above US$1 trillion.

Three independent methodologies, starting from different data sources and using different analytical frameworks, all converge above US$1 trillion for 2026 global AI investment.

That convergence is the analytical foundation of Goldman’s argument. A single model can be debated on its assumptions. Three models that arrive at materially the same conclusion from different starting points shift the burden of proof onto anyone defending the US$800 billion figure. The adjusted estimate sits approximately US$220-250 billion above the consensus benchmark, a gap too large to attribute to rounding or methodology preference.

Understanding AI investment at macro scale: what 1% of world GDP means

A trillion dollars is a large number. Without a frame of reference, it remains an abstraction. Measured against global economic output, the AI investment cycle comes into sharper focus.

Global AI investment in 2026 is expected to represent approximately 0.9% of world GDP, according to Goldman Sachs Research. That share is projected to rise to approximately 1.3-1.4% by 2027-2028. For the United States specifically, AI investment is trending toward 2-3% of US GDP over the same window.

This is not a single-year event. Cumulative global AI investment since 2022 is expected to reach approximately US$1.8 trillion by the end of 2026, establishing a growing stock of deployed capital. Looking further ahead, Goldman estimates approximately US$7.6 trillion of cumulative global AI infrastructure investment between 2026 and 2031 across compute, data centres, and power.

Goldman Sachs estimates approximately US$7.6 trillion of cumulative global AI infrastructure investment between 2026 and 2031, spanning compute, data centres, and power generation.

GDP-relative framing and cumulative figures turn a large nominal number into a structural economic force with a multi-year demand curve (the total value of goods and services an economy produces, measured annually, against which investment intensity can be assessed as a proportion). For energy and resources investors, this is the frame that separates a cyclical capital expenditure spike from a sustained structural demand shift.

What the revised figure means for energy, grid, and critical minerals demand

The difference between US$800 billion and US$1.019 trillion is not a rounding error. It is a demand revision large enough to alter long-term capital allocation decisions across multiple upstream sectors.

The gap of approximately US$220-250 billion above consensus represents a structurally higher baseline for physical demand across:

• Electricity generation capacity
• Grid transmission and distribution equipment
• Data centre cooling infrastructure
• Copper and aluminium wiring
• Rare earth elements for magnets and semiconductors
• Semiconductor fabrication materials

Every additional data centre requires power, cooling, and cabling. Every additional gigawatt of generation capacity requires copper, steel, and grid connection equipment. The scale difference between the two estimates feeds directly into upstream supply chain pressure.

AI infrastructure metals demand for copper and aluminium wiring, steel for data centre structures, and rare earth elements for magnets and power electronics scales roughly in proportion to total invested capital, meaning the gap between an US$800 billion and a US$1 trillion investment figure translates into meaningfully different multi-year procurement volumes for upstream suppliers.

The non-US component of approximately US$440 billion distributes that pressure across multiple jurisdictions simultaneously. Demand for grid equipment and generation capacity is not concentrated in a single geography but spread across regions building out AI infrastructure in parallel. For materials suppliers, this geographic diversification broadens the demand base rather than creating a single-point dependency.

The cumulative figure of US$7.6 trillion through 2031 reinforces the duration of the demand signal. This is not a two-year procurement cycle. It is a multi-year, globally distributed capital deployment event that supports a sustained structural demand thesis for power infrastructure and critical minerals rather than a cyclical upswing.

A measurement problem with trillion-dollar consequences

The US$800 billion consensus benchmark for global AI investment is not a reasonable approximation. It is a structurally incomplete proxy that systematically excludes private companies, non-US firms, and non-AI capital expenditure bundled inside hyperscaler reports.

Three independent methodologies, each starting from different data sources, converge on a figure above US$1 trillion for 2026. The geographic distribution reveals that approximately 43% of this capital lands outside the United States, a share most market participants are not tracking and one that carries material implications for energy and materials demand in non-US jurisdictions.

The GDP and cumulative figures frame what follows: the AI infrastructure build-out is a multi-year, globally distributed capital event approaching 1% of world GDP in 2026 and projected to exceed US$7.6 trillion cumulatively through 2031. Any investment framework that prices it at US$800 billion is working from an incomplete map.

For readers wanting to model the specific electricity generation and data centre capacity requirements implied by a sustained multi-year build-out at this scale, our dedicated guide to AI power demand and data centre investment walks through the power intensity of different data centre configurations, projected gigawatt requirements by region, and the infrastructure investment categories most directly exposed to the spending surge.

This article is for informational purposes only and should not be considered financial advice. Investors should conduct their own research and consult with financial professionals before making investment decisions. Financial projections referenced in this analysis are subject to market conditions and various risk factors. Past performance does not guarantee future results.