The 7.6 Trillion Dollar AI Infrastructure Buildout

By Nekud October 16, 2026 16 min read

In September 2026, the Goldman Sachs Global Institute and Global Investment Research teams published a research note that put forward a forecast whose sheer numerical magnitude has subsequently dominated the discussion of the AI infrastructure industry. The forecast was for 7.6 trillion dollars of cumulative AI infrastructure investment from 2026 through 2031, covering compute, data centres, and power. The forecast was for annual AI capex to reach approximately 1.6 trillion dollars by 2031. The forecast was for the five largest hyperscalers to spend approximately 800 billion dollars on capital expenditures in 2026 alone, up 94 percent from 2025. The forecast framed itself, modestly, as Goldman Sachs’ “baseline” estimate; the implied range of plausible outcomes was substantially wider, and the high end of the plausible range was higher still. The 7.6 trillion dollar number became, within days of its publication, the headline that financial-market analysts, AI-industry executives, and policy commentators were all working with.

The 7.6 trillion dollar forecast is, in numerical terms, comparable to the largest industrial-investment cycles of the post-war era. It is comparable to the U.S. interstate-highway investment of the late 1950s and the 1960s, to the global telecommunications-buildout investment of the late 1990s, and to the global energy-transition investment of the late 2010s and the early 2020s. The AI infrastructure buildout is, in scope, larger than any of those historical analogs when measured by the share of global GDP that it is projected to absorb: the Goldman Sachs forecast implies that AI capex as a share of global GDP will climb from approximately 0.4 percent in 2026 to approximately 1.6 percent by 2031, with the growth rate having no real precedent in modern industrial-investment history. The PwC team’s parallel forecast of 27.23 trillion euros of AI infrastructure investment through 2050 is, in numerical terms, three times larger than the Goldman Sachs forecast, with the difference reflecting the longer forecast horizon and the broader scope of investments included in the PwC forecast.

The most striking numerical fact in the Goldman Sachs forecast is the rapid acceleration in 2026. The five largest hyperscalers spent approximately 150 billion dollars in 2023, are estimated to spend approximately 800 billion dollars in 2026, and are projected to spend approximately 1.4 trillion dollars in 2028. The 800-billion-dollar figure for 2026 alone is larger than the entire 2018-vintage 5G infrastructure capex for the global telecommunications industry. The acceleration from 2023 to 2026 is, in numerical terms, a 5x increase in just three years. The acceleration from 2026 to 2028 is a further 1.75x increase in just two more years. The compound effect is that the AI infrastructure buildout is, in the late 2020s, the single largest industrial-investment cycle in the global economy, and the buildout is, in growth-rate terms, one of the fastest in modern industrial history.

This essay traces the 7.6 trillion dollar AI infrastructure buildout through three threads. The first is the dollar thread: how the investment is being allocated across the four principal categories of AI infrastructure (compute silicon, data-centre buildings, network and interconnect, and power). The second is the geography thread: how the buildout is being distributed across the U.S., Europe, and Asia-Pacific, and how the geographic distribution is being shaped by the power-wall constraint. The third is the financing thread: how the buildout is being paid for, what the financing constraints are, and what the policy environment is going to need to look like for the buildout to continue at the projected pace. The thesis the essay develops is the following. The 7.6 trillion dollar AI infrastructure buildout is the largest single industrial-investment cycle of the 2020s and is the cycle on which the rest of the AI industry’s growth depends. The buildout is, by late 2026, already substantially under way; its trajectory over the next several years is going to be one of the most consequential industrial-policy questions of the late 2020s.

Where the dollars go

The 7.6 trillion dollar AI infrastructure buildout is going to be allocated across four principal categories of investment, with the allocation shifting over the forecast horizon. The first category is compute silicon. The compute-silicon category includes the merchant AI accelerator (principally Nvidia’s GPUs), the hyperscaler custom silicon (Microsoft Maia, Google TPU, Amazon Trainium, and the long tail of independent accelerators), the HBM memory subsystem, the server CPU and the companion chips that support the AI compute accelerators, and the network and interconnect silicon that ties the AI compute accelerators into the rack and into the larger data-centre. The compute-silicon category is, in late 2026, the largest single category of AI infrastructure spending, with the Gartner 92 percent revenue growth in the global semiconductor industry to 1.56 trillion dollars in 2026 being substantially attributable to the AI compute silicon demand. The compute-silicon category is projected, on the Goldman Sachs forecast, to grow at approximately 25 percent per year through 2031, faster than any other category of AI infrastructure spending.

The second category is data-centre buildings. The data-centre-buildings category includes the land, the structural shell, the cooling system, the electrical distribution within the data centre, the security system, and the related infrastructure that turns a piece of land into a building that can host AI compute silicon. The data-centre-buildings category is the second-largest single category of AI infrastructure spending, with the colocation-builders (Equinix, Digital Realty), the hyperscaler captive-builders (Microsoft, Google, Amazon, Meta), and the neocloud captive-builders (CoreWeave, Lambda, Crusoe) all spending substantial amounts. The data-centre-buildings category is, in late 2026, the category in which the most rapid growth is occurring, with the largest AI data-centre campuses (Microsoft’s planned multi-gigawatt Maia deployment, the Anthropic 5-gigawatt TPU and AWS deployment, the OpenAI gigawatt-scale deployment of Jalapeño silicon) all requiring substantial new data-centre buildings. The PwC forecast implies that the data-centre-buildings category will double in nominal size between 2026 and 2031.

The third category is network and interconnect. The network-and-interconnect category includes the optical fibre and the optical transceivers that connect the AI compute data centres to each other and to the internet, the high-bandwidth networking that connects the AI compute accelerators within a data centre, and the long-haul fibre backbone that connects the data centres to the user traffic. The network-and-interconnect category has been, in late 2026, the slowest-growing of the four categories, with the AI compute buildout’s network requirements being more modest, in absolute terms, than the AI compute silicon or the data-centre buildings. The network-and-interconnect category is, however, becoming increasingly important as the agentic AI workload drives the demand for low-latency connectivity between the AI compute data centres and the user devices, and as the buildout of new AI data centres drives the need for substantial new fibre connectivity.

The fourth category is power. The power category includes the new power generation (natural gas, nuclear, renewables), the new transmission and distribution capacity (HVDC and AC transmission lines, substations, transformers), the new on-site power systems (gas turbines, fuel cells, battery storage), and the new grid-interconnection capacity that allows the AI data centres to draw power from the grid. The power category has been, in late 2026, the smallest of the four categories, with the existing grid capacity supplying the bulk of the AI compute power. The power category is, however, becoming increasingly important as the AI compute buildout pushes against the power-wall constraint, and the power category is projected to be the second-fastest-growing of the four categories through 2031. The IDC forecast that data-centre semiconductors will account for nearly half of the total semiconductor market by 2030 is, in substantial part, a reflection of the AI compute silicon demand that the power-wall constraint is driving.

How the buildout is distributed geographically

The 7.6 trillion dollar AI infrastructure buildout is being distributed across the world in proportions that are shaped by both the existing AI compute customer base and the power-availability constraint. The PwC forecast attributes approximately 48 percent of the 2026-2050 AI infrastructure investment to the U.S., with the Asia-Pacific region receiving approximately 26 percent and Europe receiving approximately 18 percent (the balance going to other regions). The U.S. share is, in absolute terms, by far the largest: the PwC forecast implies approximately 13 trillion dollars of AI infrastructure investment in the U.S. over the next 24 years, with the bulk of the investment concentrated in the late 2020s and the early 2030s. The U.S. share reflects the concentration of the AI compute customers (Microsoft, Google, Amazon, Meta, Oracle, and the long tail of AI-cloud companies) in the U.S., and the U.S. share is, accordingly, substantially larger than the U.S. share of the global AI service market would suggest.

The Asia-Pacific region’s 26 percent share is, in absolute terms, substantial, with the bulk of the share going to China, Japan, South Korea, Singapore, and Australia. China’s AI infrastructure investment is, on the PwC forecast, the largest single component of the Asia-Pacific share, with China’s investment driven by the domestic AI compute customer base and by the strategic-autonomy considerations that have been a persistent feature of China’s AI policy since the late 2010s. Japan’s AI infrastructure investment is driven principally by the AI compute customers that operate substantial Japan-based AI compute (notably the Japanese telecommunications companies and the Japanese AI-cloud companies). South Korea’s AI infrastructure investment is driven principally by Samsung and SK Hynix’s substantial HBM production for the global AI compute market, which is requiring substantial AI compute capacity to be deployed in South Korea to support the HBM production testing. Singapore’s and Australia’s AI infrastructure investments are driven principally by the international-AI-compute-customer buildout that is going on in Southeast Asia and in the Asia-Pacific region more broadly.

Europe’s 18 percent share is, in absolute terms, smaller than the U.S. share and the Asia-Pacific share, with the European share driven by the European AI compute customers (principally Mistral AI, the European telecommunications operators, and the European sovereign-AI-cloud customers). The European share is, however, growing faster than the global average, in part because of the European Union’s substantial industrial-policy efforts in the AI infrastructure space. The European share is also being driven by the substantial investments of the U.S. hyperscalers in European data-centre capacity, with Microsoft, Google, and Amazon all having substantially expanded their European data-centre footprints through 2025 and 2026 in response to the European customer demand.

The geographic distribution of the AI infrastructure buildout is being reshaped, in real time, by the power-wall constraint. The historical concentration of the AI compute buildout in the U.S. Pacific Northwest, northern Virginia, Texas, and the Phoenix corridor is being supplemented by substantial new buildouts in regions with abundant power supply: the nuclear-power-adjacent sites, the regions with substantial new renewable generation, and the regions with substantial new transmission capacity. The geographic distribution is also being reshaped by the international AI compute buildout, with several of the largest AI compute customers (Anthropic, OpenAI, Microsoft, Google) all deploying substantial new AI compute capacity outside the U.S. in response to the power-wall constraint and to the strategic-diversification considerations.

How the buildout is being financed

The 7.6 trillion dollar AI infrastructure buildout is being financed through a combination of corporate-balance-sheet investment, debt issuance, and various forms of project finance. The largest single source of financing is the corporate-balance-sheet investment of the hyperscalers and the AI compute customers: the five largest hyperscalers spent approximately 800 billion dollars on capital expenditures in 2026, with the bulk of the capex funded from operating cash flow and from balance-sheet cash. The hyperscaler balance-sheet capex is, in late 2026, the most significant single source of financing for the AI infrastructure buildout, and it is the source that has the smallest exposure to the debt-market dynamics that are going to shape the other sources of financing.

The second largest source of financing is debt issuance. The AI compute customers have, in 2024 and 2025, become some of the largest single issuers of corporate debt in the global capital markets, with several of the largest AI compute customers (Meta, Alphabet, Amazon, Microsoft) issuing substantial multi-billion-dollar bond offerings to fund the AI infrastructure buildout. The debt issuance has been well received by the bond markets, with the AI compute customers’ bonds trading at premium spreads reflecting the strength of the AI compute customers’ balance sheets and the visibility of the AI compute customers’ cash flows. The Data Center Frontier reporting on the AI infrastructure financing indicates, however, that the financing terms are tightening as the size of the buildout grows: the September 2026 Anthropic-Lambda $35 billion cloud deal and the September 2026 Anthropic-Akamai $11.6 billion (with a $9 billion extension option) CPU cloud deal are both substantially larger than the deals of the previous year, and the financing of the deals requires more sophisticated financing structures than the simpler deals of the previous year.

The third source of financing is project finance. Project-finance deals are typically used for the largest individual AI infrastructure projects, with the debt issued against the cash flows of the individual project rather than against the balance sheet of the AI compute customer. Project-finance deals have been, in late 2026, an increasingly important part of the AI infrastructure financing landscape, with several of the largest individual AI data-centre projects being financed through project-finance deals rather than through corporate-balance-sheet investment. The project-finance deals are typically structured with a longer tenor than the corporate-balance-sheet deals (often 15 to 20 years rather than 5 to 10 years), with the longer tenor reflecting the longer useful life of the AI infrastructure assets. The project-finance deals are also typically structured with the substantial equity contribution from the AI compute customer providing the credit enhancement for the debt.

The fourth source of financing is sovereign-AI-cloud and government-backed financing. Several countries have, in 2025 and 2026, established substantial sovereign-AI-cloud financing programs, with the funds typically used to support the construction of AI data-centre capacity within the country. The sovereign-AI-cloud financing programs are typically structured as a combination of equity investment from the government, debt issuance guaranteed by the government, and partnerships with private-sector AI compute customers who would deploy their AI compute capacity in the country in exchange for the favorable financing. The sovereign-AI-cloud financing programs are a substantial source of financing for the AI infrastructure buildout, and they are also a substantial source of the international AI infrastructure buildout that the geographic distribution of the buildout has implied.

The strategic implications of the 7.6 trillion buildout

The 7.6 trillion dollar AI infrastructure buildout is, in strategic terms, the largest single industrial-investment cycle of the 2020s, and the strategic implications of the buildout are correspondingly large. The first strategic implication is that the AI infrastructure buildout is, by late 2026, becoming a major determinant of the global semiconductor industry’s revenue trajectory. The Gartner forecast of 92 percent semiconductor industry revenue growth in 2026, the IDC forecast of data-centre semiconductors accounting for nearly half of the semiconductor market by 2030, and the Makreo Research forecast of 22 percent CAGR in the AI chip market through 2032 are all, in substantial part, reflections of the AI infrastructure buildout’s demand for the AI compute silicon, the HBM memory subsystem, and the support chips. The semiconductor industry’s strategic position is, in 2026, the strongest it has been in the history of the industry.

The second strategic implication is that the AI infrastructure buildout is, in late 2026, becoming a major determinant of the global electric-utility industry’s revenue trajectory. The Morgan Stanley 57 gigawatt supply gap by 2029, the Goldman Sachs 7.6 trillion dollar investment forecast’s substantial allocation to the power category, and the substantial new power-generation investment that the AI compute customers are committing to are all, in substantial part, reflections of the AI infrastructure buildout’s demand for the new power generation and the new transmission capacity. The electric-utility industry’s strategic position is, in 2026, the strongest it has been in the history of the industry, and the substantial new investment in the power category is reshaping the industry’s competitive landscape.

The third strategic implication is that the AI infrastructure buildout is, in late 2026, becoming a major determinant of the global data-centre industry’s revenue trajectory. The colocation-builders (Equinix, Digital Realty), the hyperscaler captive-builders, and the neocloud captive-builders are all investing at unprecedented rates, with the colocation-builders having substantially expanded their land banks, their pipeline of new buildings, and their customer relationships through 2025 and 2026. The data-centre industry’s strategic position is, in 2026, the strongest it has been in the history of the industry.

The fourth strategic implication is that the AI infrastructure buildout is, in late 2026, becoming a major determinant of the global financial-services industry’s revenue trajectory. The substantial new debt issuance, the substantial new project-finance deals, the substantial new equity investment in the AI compute customers, and the substantial new advisory work on the AI infrastructure buildout are all, in substantial part, reflections of the AI infrastructure buildout’s demand for the global financial-services industry’s services. The financial-services industry’s strategic position is, in 2026, being substantially strengthened by the AI infrastructure buildout.

The longer-term trajectory

The 7.6 trillion dollar AI infrastructure buildout is, in late 2026, on track to proceed at the pace that the Goldman Sachs forecast implies. The early-2026 corporate disclosure from the five largest hyperscalers, the Anthropic-Lambda and Anthropic-Akamai deal disclosures, the OpenAI Stargate joint-venture disclosures, and the Microsoft Maia 300 multi-gigawatt deployment plan all indicate that the buildout is proceeding at a pace that is consistent with the Goldman Sachs forecast’s 2026 capex forecast of 800 billion dollars and its 2028 capex forecast of 1.4 trillion dollars. The buildout is, however, on a trajectory that is going to be increasingly constrained by the power-wall constraint, and the principal uncertainty in the forecast is the pace at which the power-wall constraint is going to limit the buildout.

The principal policy implication of the 7.6 trillion dollar AI infrastructure buildout is that the AI infrastructure industry is, by late 2026, one of the largest single industrial-policy challenges of the late 2020s. The buildout requires substantial regulatory innovation in the U.S. electric-grid context, substantial regulatory innovation in the international trade context (with the export-control regime on advanced AI silicon being a substantial determinant of the international AI infrastructure buildout), and substantial regulatory innovation in the financial-services context (with the substantial new debt and equity issuance required to fund the buildout requiring substantial regulatory support). The policy environment in which the AI infrastructure buildout proceeds is, accordingly, going to be one of the most important strategic variables of the late 2020s.

The principal corporate-strategy implication of the 7.6 trillion dollar AI infrastructure buildout is that the AI compute customers that are best positioned to capture the buildout are the ones with the largest balance sheets, the longest track records of infrastructure investment, and the most sophisticated financing capabilities. The five largest hyperscalers — Microsoft, Google, Amazon, Meta, and Oracle — are the corporate beneficiaries of the buildout. The neoclouds — CoreWeave, Lambda, Crusoe — are also substantial beneficiaries, with their strategic positioning dependent on the buildout’s pace. The corporate beneficiaries are, in late 2026, also the principal corporate competitors in the AI service market, and the AI infrastructure buildout is, accordingly, both a source of competitive advantage and a source of competitive pressure for the same companies.

The 7.6 trillion dollar AI infrastructure buildout is, in summary, the largest single industrial-investment cycle of the 2020s and is the cycle on which the rest of the AI industry’s growth depends. The buildout is being driven by the AI compute customers’ demand for the new compute silicon, the new data-centre buildings, the new network and interconnect, and the new power generation and transmission. The buildout is being financed through a combination of corporate-balance-sheet investment, debt issuance, project finance, and sovereign-AI-cloud financing. The buildout is being distributed geographically in proportions that reflect the existing AI compute customer base and the power-wall constraint. The buildout is reshaping the strategic positions of the global semiconductor industry, the global electric-utility industry, the global data-centre industry, and the global financial-services industry. The buildout’s trajectory over the next several years is going to be one of the most consequential industrial-policy questions of the late 2020s, and the trajectory is going to be determined by the pace at which the power-wall constraint and the regulatory environment allow the buildout to proceed.