As of early October 2026, the AI boom is confronting the cost of turning demand into buildings, chips, power contracts and cooling systems.
A recent paper by Columbia Business School professor Stijn Van Nieuwerburgh frames AI as a physical capital boom. A revised October 2 version estimates that a 188-gigawatt U.S. AI buildout completed by 2032 would require nearly $9 trillion in investment, equal to about 3.2% of U.S. GDP per year from 2025 through 2032. The estimate is scenario-based rather than a forecast, but it illustrates the physical scale behind the AI expansion.
The revision itself is a useful reminder that the precise number can change as assumptions about data center capacity and construction costs move. Even at nearly $9 trillion, the buildout makes power, chips, networking, cooling and financing central to the economics of AI.
Goldman Sachs Research arrived at the scale from another angle. In August, it estimated that global AI-related investment would exceed $1 trillion in 2026, including about $581 billion in the United States. Goldman also cautioned that any estimate requires assumptions about which capital expenditures are truly AI-related, because hyperscaler capex includes non-AI spending and global companies do not always disclose the location or use of every investment.
Those qualifiers matter. The precise number can move. The direction is harder to dismiss. AI is consuming more capital, more electricity and more financing capacity than the market expected when generative AI first moved from demo to mass adoption.
Earnings reports show the bill in real time
Company reports from the second quarter of 2026 show the infrastructure bill turning up line by line.
Alphabet said capital expenditures were $44.9 billion in the quarter, with the vast majority tied to technical infrastructure for AI. The company also reported negative free cash flow of $5.9 billion in the quarter and raised its 2026 capex guidance to a range of $195 billion to $205 billion, up from $180 billion to $190 billion.
Microsoft reported $41 billion in capital expenditures for its fiscal fourth quarter, with roughly two-thirds going to short-lived assets, primarily CPUs and GPUs. It also reported $5.6 billion in finance leases, mainly for large data center sites, and $35.8 billion in cash paid for property and equipment.
Meta reported $31.08 billion in second-quarter capital expenditures, including principal payments on finance leases. The company narrowed its full-year 2026 capex outlook to $130 billion to $145 billion, with spending tied to AI efforts and the core business.
Amazon’s Form 10-Q shows cash capital expenditures of $53.1 billion in the second quarter and $96.3 billion for the first six months of 2026. Amazon said those expenditures primarily reflected technology infrastructure, the majority of which supported AWS business growth, and additional capacity for its fulfillment network.
These figures are not all pure AI spend. Fulfillment, non-AI cloud workloads, networking, offices and other long-lived assets still sit inside companywide capex. But the latest disclosures make one point clear: AI is pulling the largest technology companies into a heavier, more capital-intensive operating model.
Revenue is growing, but the payback clock is tight
The case for continued spending is not imaginary. Alphabet reported Google Cloud revenue growth of 82% in the second quarter and a cloud backlog of $514 billion. Microsoft reported Microsoft Cloud revenue of $59.3 billion for the quarter and commercial remaining performance obligation of $678 billion. Amazon’s AWS operating income rose in the second quarter of 2026, even with added spending on technology infrastructure.
The tension is timing. Infrastructure spending happens before the revenue fully appears. Data centers require land, interconnection, construction, power arrangements and equipment. Accelerators and servers then depreciate much faster than the buildings around them. Brookings notes that financing GPU and related hardware is different from financing conventional real estate because the equipment has a shorter economic life and faces rapid obsolescence risk.
Bain & Company’s 2026 Global Technology Report put the revenue challenge in starker terms. The firm said the global AI industry would need about $6 trillion in annual revenue by 2031 to justify the capital being deployed for data centers and related infrastructure. Bain estimated existing consumer and enterprise AI services may generate as much as $1.8 trillion of that amount, leaving about $4.2 trillion in new annual revenue to be created.
That is not a prediction of failure. It is a measure of how much new value must be captured for the infrastructure race to stay financially comfortable. The AI sector does not only need users. It needs paying customers at a scale that can support construction, electricity, chips, networking, leases, debt and replacement cycles.
Power is becoming a financing risk
The International Energy Agency estimated that data centers consumed about 415 terawatt hours of electricity in 2024, equal to roughly 1.5% of global electricity consumption. In its Base Case, the IEA projects data center electricity use to roughly double to around 945 terawatt hours by 2030, just under 3% of global electricity consumption.
The global percentage can make the issue look manageable. The local concentration makes it harder. The IEA notes that data centers tend to cluster in specific locations, which can make grid integration more difficult even when the global share of electricity use remains limited.
That local constraint is already showing up in financing. Reuters reported in September that Oracle issued a force majeure notice connected to Project Jupiter, a New Mexico data center campus being developed by a Blue Owl unit to support OpenAI. Reuters reported that Oracle invoked the clause because of delays securing power to the site, citing a person familiar with the matter. Blue Owl said the notice did not change the parties’ financial commitments to the multiyear project.
The same Reuters report said at least 45 projects worth $68 billion faced opposition from community groups in the second quarter of 2026, citing Data Center Watch, after 75 projects worth about $130 billion were disrupted in the first quarter.
That is the point where AI stops looking like a software upgrade and starts looking like an energy, permitting and project-finance problem. Even strong end-user demand does not eliminate the risk of a delayed power connection, a contested permit or a financing structure that has to absorb a year of slippage.
Private capital is moving in because balance sheets are not enough
Goldman Sachs Research said in June that private markets are expected to play a larger role in data center financing as hyperscalers are expected to spend $5.3 trillion on AI and data centers by 2030. Goldman pointed to private infrastructure funds, real estate funds, debt markets and other structures as increasingly important sources of capital.
Brookings described a similar shift. As hyperscaler capital expenditures grow beyond internal cash flow, financing is moving toward leases, joint ventures, project debt, private credit, securitization and special-purpose vehicles. Those structures can expand funding capacity, but they can also make risks harder to see because ownership, leverage and obligations sit across multiple entities.
The risk is not simply that large technology companies will run out of money. Many still have enormous revenue bases, investment-grade credit profiles and dominant platforms. The more immediate risk is that the marginal dollar becomes more expensive, more conditional or harder to place. If investors demand better terms, if lenders grow cautious about tenant concentration, or if local power constraints delay projects, the pace of the buildout can change even while AI demand keeps growing.
The pressure will reach the rest of the market
The infrastructure bill will not remain a Big Tech accounting issue. If capacity stays tight, the effects are likely to appear in cloud pricing, access to frontier models, regional availability, minimum commitments and the economics of AI products built on top of rented compute.
Higher inference costs can show up as usage limits. Capacity shortages can slow product rollouts. Expensive compute can push vendors toward smaller models, narrower features or pricing plans that recover more of the infrastructure cost from customers.
The strongest use cases will receive capital first because they can support clear revenue, productivity gains or customer retention. Speculative products that rely on cheap frontier-model access may face a harder path if compute remains expensive.
The next phase of AI may be decided as much by power contracts, depreciation schedules and financing terms as by model benchmarks.
I can suggest related Tech Help Canada articles based on the topic you’re reading now.
Tech Help Canada Staff researches, writes, and reviews practical content for business owners and professionals. Our coverage spans business, marketing, SEO, technology, and the tools and systems people use to grow and operate online. We focus on clear, useful information backed by research, hands-on experience, and editorial review. Learn more about our team and editorial standards. Need help with something? Contact Us
