The AI Capex Boom:
When Does Investment Become Return?

The scale of AI infrastructure spending is now large enough to matter well beyond technology stocks.

Artificial intelligence may prove to be one of the most important productivity technologies of our lifetime. But that does not mean every dollar invested in AI infrastructure will earn an attractive return. Those are two different questions, and investors increasingly need to separate them.

The scale of the buildout is extraordinary. The International Energy Agency reported that capital expenditure by five large technology companies exceeded $400 billion in 2025 and was expected to increase by another 75% in 2026. That implies spending of more than $700 billion across those companies alone, although the IEA notes that its 2026 figures are estimates.[1]

The investment question is therefore changing. The debate is no longer simply whether AI demand is real. The harder question is whether the revenue, margins, cash flow and useful lives of the assets being built will ultimately justify the capital being committed today.

How Much Are the Tech Giants Spending on AI?

Company disclosures reinforce the IEA's broader picture. Meta currently expects $130 billion to $145 billion of 2026 capital expenditures, including principal payments on finance leases.[1] Microsoft previously expected roughly $190 billion of calendar-year 2026 capex, but after an accounting-policy change shifted more future data-center leases from finance leases to operating leases, Microsoft revised the comparable capex expectation to approximately $175 billion. The company said the underlying investment expectations were otherwise unchanged.[2]

Oracle's fiscal 2026 capital expenditure reached $55.7 billion, up from $21.2 billion in fiscal 2025. Oracle said the increase reflected its expanding cloud infrastructure footprint. Its operating cash flow was about $32.0 billion, while free cash flow was negative $23.7 billion for the year.[3]

These figures are not perfectly comparable. Fiscal calendars differ, lease accounting differs, and not every dollar is exclusively AI. That distinction matters. Still, the direction is unmistakable: AI has moved from a research-and-development story into one of the world's largest capital-investment cycles.

*2026 bar illustrates a 75% increase from the IEA's stated 2025 level of more than $400 billion; therefore the actual total would exceed $700 billion.

AI Spending Is Moving Into the Physical Economy

AI can feel like a software story, but the current investment cycle is increasingly physical. It requires data centers, servers, advanced chips, networking equipment, cooling systems, transmission lines, substations and dependable power. That means the AI buildout is reaching into parts of the economy that investors may not immediately associate with technology.

Microsoft provides a useful example. In fiscal 2026, the company generated $182.9 billion of operating cash flow, up from $136.2 billion in fiscal 2025. Cash additions to property and equipment rose to $115.9 billion from $64.6 billion. Microsoft's net property and equipment also increased to $313.1 billion from $205.0 billion a year earlier.[1]

Source: Microsoft FY2026 fourth-quarter financial statements. This chart shows cash additions to property and equipment, not total capex including finance leases.

That distinction matters. A data-center shell may remain useful for decades, while the computing equipment inside it can become economically obsolete much faster. The faster the hardware replacement cycle, the more important utilization and pricing become. High demand can make an expensive asset productive. Underutilization can make the same asset a drag on returns.

Amazon offers another view of the same cycle. In its second-quarter 2026 results, AWS revenue rose 37% year over year to $42.2 billion. Yet Amazon's trailing-twelve-month free cash flow fell to an outflow of $7.6 billion, driven primarily by a $66.1 billion year-over-year increase in purchases of property and equipment, net of proceeds and incentives.[1] The contrast shows why free cash flow deserves as much attention as reported earnings during an intensive capital expansion.

AI's Growing Demand for Power

AI infrastructure is only valuable if there is enough electricity available to power it. The IEA estimates that global data-center electricity consumption was about 485 terawatt-hours in 2025 and could roughly double to 950 TWh by 2030. Electricity consumption from AI-focused data centers is expected to grow even faster, tripling over the same period.[2]

In the United States, Lawrence Berkeley National Laboratory's 2025 Update estimates that data centers could account for 11.8% of total U.S. electricity use by 2030, with scenarios ranging from 9.5% to 15.3%.[3]

Source: International Energy Agency, Key Questions on Energy and AI, 2026.

The scale is significant, but the more important issue may be location. Data centers are concentrated loads. A new facility can create enormous electricity demand in one market, often faster than generation, transmission and interconnection capacity can expand. The IEA reports that bottlenecks across the value chain are reducing the likelihood of more aggressive near-term data-center growth scenarios.[1]

The broader grid challenge is also substantial. The IEA estimates that more than 2,500 GW of generation, storage and large-load projects, including data centers, remain stalled in grid-connection queues worldwide. It estimates annual grid investment would need to rise by roughly 50% by 2030 from today's approximately $400 billion level to meet forecast electricity demand.[1]

This creates a second layer to the AI investment story. Beneficiaries may include electrical equipment, power generation, transmission infrastructure, engineering, construction and cooling technologies. At the same time, physical constraints can raise project costs and delay the point when a data center begins producing revenue.

Major U.S. transmission projects operate on a much longer timeline than many of the data centers they are intended to serve. The Department of Energy reports that federal permitting for a new electric transmission line has historically taken approximately four years.[2] The challenge extends well beyond permitting. A DOE review of more than 30 transmission projects initiated since 2005 found that new transmission projects take about 10 years, on average, to develop, permit and construct, with some requiring 15 years or more.[3] DOE has established a coordinated federal permitting process intended to shorten federal reviews for qualifying projects to a two-year timeline. Investors should be aware however, that states continue to hold primary authority over the siting of most transmission facilities and that policies and approval timelines can be extremely variable.[4]

Even with a shortened permitting process by regulators, this timing gap creates an important investment risk. A data center can be announced and constructed much faster than the transmission infrastructure needed to deliver electricity to it. Delays in transmission, substations or interconnections can therefore postpone the point when billions of dollars invested in AI infrastructure begin generating revenue and cash flow.

The Hidden Cost of the AI Buildout: Depreciation

One of the easiest mistakes during a capital-spending boom is to focus only on the year when the cash is spent. Buildings, servers and networking equipment are capitalized and then depreciated over their estimated useful lives. A large wave of spending today can therefore create a rising depreciation expense tomorrow.

Microsoft's fiscal 2026 cash-flow statement reported $38.5 billion of depreciation, amortization and other adjustments, compared with $29.4 billion in fiscal 2025, while its property-and-equipment base expanded sharply.[1] Alphabet has made the same economic point directly: as technical-infrastructure investment rises, depreciation and data-center operating costs can place increasing pressure on the income statement.[2]

Depreciation is not hidden in an accounting sense. The investment concern is the speed and magnitude of the asset build. If AI revenue and utilization rise rapidly, companies can absorb the depreciation burden through a larger stream of gross profit. If demand disappoints, pricing falls or hardware becomes obsolete faster than expected, the economics can look very different.

When Does AI Capex Become an Investment Risk?

The AI capex boom does not need to collapse for investors to experience disappointing returns. A technology can transform the economy while individual investments in that technology still fail to earn adequate returns. The central risk is not necessarily that companies build infrastructure nobody uses. A more subtle risk is that they build infrastructure that generates revenue but earns an inadequate return on the capital invested.

Utilization will matter, so will pricing. If demand for AI compute continues to exceed available supply, infrastructure owners may retain strong pricing power. If supply eventually catches demand, competition could pressure prices precisely when depreciation, energy and replacement costs begin to rise.

Hardware obsolescence adds another complication. An asset does not need to stop working to become obsolete. It only needs to become too expensive or inefficient compared with the next generation. Investors should therefore keep asking: How long will today's AI hardware remain economically competitive?

What Investors Should Watch Next

The outcome will probably be uneven. Some companies may earn excellent returns on their infrastructure. Others may earn adequate returns. Some projects may never justify their cost. The technology can succeed even while investment results vary widely.

Investors should watch whether cloud and AI revenue growth keeps pace with growth in the capital base. Utilization should remain high as new capacity enters service. Pricing will matter as competitors add compute. Depreciation and power expenses deserve increasing attention as the infrastructure base grows. Free cash flow may become particularly revealing because, eventually, new assets must generate enough incremental cash to justify their cost.

Customer commitments and backlog can also provide useful evidence that capacity is being built against identifiable demand; balance-sheet strength matters as well. The first wave of spending has been dominated by companies with extraordinary cash generation. As financing expands to infrastructure partnerships and outside capital, capital structure becomes part of the investment thesis.

Meta's July 2026 El Paso data-center venture illustrates this evolution. BlackRock-managed funds will own 80% of a venture expected to fund about $14 billion of development costs, while Meta retains 20%. The structure also includes debt financing and long-term lease arrangements with Meta.[1]

AI Capex and the Security of Everything™

AI is not simply a technology theme. It intersects with energy security, semiconductor supply chains, grid reliability, industrial capacity, critical infrastructure and capital-market discipline. In this sense, it is the Security of Everything.

A semiconductor shortage can limit computing capacity. A shortage of transformers can delay a data center. Insufficient electricity generation can determine where a project gets built. Grid congestion can affect economics after construction. Capital availability can determine which projects receive financing. These connections illustrate why investors increasingly need to look beyond traditional sector boundaries.

The AI investment cycle therefore extends well beyond the companies whose names appear most frequently in technology headlines. For global investors, that may create opportunities in less obvious places. It also reinforces the need for discipline: technological importance does not automatically equal investment return.

The Question Is No Longer Whether AI Will Matter

Artificial intelligence may become one of the defining technologies of this generation. The extraordinary level of investment underway suggests that some of the world's largest companies see a substantial opportunity.

But the investment question becomes more complicated as spending moves from software development into hundreds of billions of dollars of physical infrastructure. Investors will need to look beyond revenue growth alone and consider capital intensity, utilization, depreciation, power availability, financing costs, hardware replacement cycles and free cash flow.

The central question is therefore no longer simply, “Will AI matter?” It almost certainly will. The more useful investment question may be: “At what price, with what capital intensity, and for whose benefit?” In an investment cycle measured in hundreds of billions of dollars, the answer may become one of the defining investment questions of the next several years.

Hemispheres Investment Management

HIM is a wealth manager with a global value investment focus (domestic and international investments in the same portfolio). Because of HIM’s global investment experience, we are uniquely qualified to meet and benefit from this secular trend toward a polycentric investment world. Our team of seasoned professionals each has over 35 years of experience in research, strategy development, and management of investment portfolios, including deep proficiency in U.S., international, and emerging markets. Hemispheres can assist you in diversifying your portfolio globally. Global Equities is Hemispheres’ flagship investment product.

Please contact Hemispheres Investment Management for a free consultation. We provide guidance to assist you in optimizing your investment strategies and helping you achieve your investment goals. Contact us at [email protected].

[1] International Energy Agency (IEA), “Key Questions on Energy and AI,” Executive Summary, 2026. The IEA reports that capital expenditure by five large technology companies exceeded $400 billion in 2025 and was expected to rise another 75% in 2026. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary

[2] Meta Platforms, Inc., “Meta Reports Second Quarter 2026 Results,” July 29, 2026. Meta narrowed its 2026 capital-expenditure outlook to $130-$145 billion, including principal payments on finance leases. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Reports-Second-Quarter-2026-Results/

[3] Microsoft Corporation, “Fiscal Year 2026 Fourth Quarter Earnings Conference Call,” July 29, 2026. Microsoft stated that a change affecting lease classification adjusted its calendar-2026 capex expectation to approximately $175 billion, while underlying investment expectations remained unchanged. https://www.microsoft.com/en-us/investor/events/fy-2026/earnings-fy-2026-q4

[4] Oracle Corporation, “Oracle Announces Record Q4 and FY 2026 Results Driven by Cloud Infrastructure & Cloud Applications,” June 16, 2026. Fiscal 2026 capital expenditures were $55.663 billion; operating cash flow was $31.977 billion; free cash flow was negative $23.686 billion. https://investor.oracle.com/investor-news/news-details/2026/Oracle-Announces-Record-Q4-and-FY-2026-Results-Driven-by-Cloud-Infrastructure--Cloud-Applications/

[5] Microsoft Corporation, “Microsoft Cloud and AI Strength Fuels Fourth Quarter Results,” FY2026 Q4 financial statements, July 29, 2026. Net cash from operations was $182.935 billion; additions to property and equipment were $115.948 billion; net property and equipment was $313.076 billion. https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast

[6] Amazon.com, Inc., “Amazon.com Announces Second Quarter Results,” July 30, 2026. AWS Q2 2026 sales were $42.232 billion, up 37% year over year; trailing-twelve-month free cash flow was an outflow of $7.6 billion, driven primarily by a $66.1 billion year-over-year increase in purchases of property and equipment, net of proceeds and incentives. https://ir.aboutamazon.com/news-release/news-release-details/2026/Amazon-com-Announces-Second-Quarter-Results/default.aspx

[7] International Energy Agency (IEA), “Key Questions on Energy and AI,” Executive Summary, 2026. The IEA states that bottlenecks across the value chain are reducing the likelihood of more aggressive near-term data-center growth scenarios. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary

[8] U.S. Department of Energy, “Powering America’s AI Future—Data Center Resource Hub,” Data and Transparency, accessed September 21, 2026. DOE cites Lawrence Berkeley National Laboratory’s United States Data Center Energy Usage Report: 2025 Update, which estimates that data centers could account for 11.8% of total U.S. electricity use by 2030, with scenarios ranging from 9.5% to 15.3%. https://www.energy.gov/powering-americas-ai-future-data-center-resource-hub

[9] International Energy Agency (IEA), “Key Questions on Energy and AI,” Executive Summary, 2026. The IEA states that bottlenecks across the value chain are reducing the likelihood of more aggressive near-term data-center growth scenarios. https://www.iea.org/reports/key-questions-on-energy-and-ai/executive-summary

[10] International Energy Agency (IEA), “Electricity 2026,” Executive Summary, 2026. The IEA reports more than 2,500 GW of generation, storage and large-load projects stalled in grid-connection queues and estimates annual grid investment must rise roughly 50% by 2030 from about $400 billion today. https://www.iea.org/reports/electricity-2026/executive-summary

[11] U.S. Department of Energy, Office of Electricity, “Coordinated Interagency Transmission Authorizations and Permits Program,” accessed September 21, 2026. DOE reports that federal permitting for a new electric transmission line takes approximately four years on average and that CITAP establishes a two-year federal permitting timeline for qualifying projects. https://www.energy.gov/oe/coordinated-interagency-transmission-authorizations-and-permits-program

[12] U.S. Department of Energy, “Transmission Impact Assessment: Power Sector Infrastructure Deployment to Reduce Costs, Improve Reliability, and Lower Pollution,” 2024. DOE reports that a review of more than 30 transmission projects initiated since 2005 found an average of about 10 years to develop, permit and build new transmission, with some projects requiring 15 years or more. https://www.energy.gov/sites/default/files/2024-10/DOE_OP_2024_Report-Transmission_Impact_Assessment.pdf

[13] Federal Energy Regulatory Commission, “Electric Transmission Facilities Permit Process,” accessed September 21, 2026. FERC states that transmission siting authority principally resides with the states and that states retain jurisdiction over most electric transmission projects; federal siting authority applies in limited circumstances. https://www.ferc.gov/electric-transmission-facilities-permit-process

[14] Microsoft Corporation, “Microsoft Cloud and AI Strength Fuels Fourth Quarter Results,” FY2026 Q4 financial statements, July 29, 2026. Net cash from operations was $182.935 billion; additions to property and equipment were $115.948 billion; net property and equipment was $313.076 billion. https://www.microsoft.com/en-us/investor/earnings/fy-2026-q4/press-release-webcast

[15] Alphabet Inc., “2025 Q3 Earnings Call,” October 29, 2025. Alphabet stated that increased technical-infrastructure investment would continue to pressure the P&L through higher depreciation expense and related data-center operating costs, including energy. https://abc.xyz/investor/events/event-details/2025/2025-Q3-Earnings-Call-2025-4OI4Bac_Q9/default.aspx

[16] Meta Platforms, Inc., “Meta Announces New Strategic Venture with BlackRock to Develop Data Center in El Paso,” July 28, 2026. The venture contemplates approximately $14 billion in development costs, with BlackRock-managed funds owning 80% and Meta retaining 20%, supported in part by debt financing and lease agreements with Meta. https://investor.atmeta.com/investor-news/press-release-details/2026/Meta-Announces-New-Strategic-Venture-with-BlackRock-to-Develop-Data-Center-in-El-Paso/default.aspx