What "Capex" Actually Means

Capital expenditure is the cash a company spends to acquire, build, or upgrade long-lived physical and intangible assets expected to produce economic benefits over multiple years. For hyperscalers, this translates into a relatively concentrated set of categories: land and shell for data-center campuses; power distribution, transformers, and backup generation; cooling and HVAC systems; racks and physical security; servers, storage arrays, and network switches; GPUs and custom AI accelerators; the fiber and long-haul networks that interconnect regions; and a portion of capitalized software and internal tooling that meets accounting criteria for long-lived assets. Dell'Oro estimates that GPUs and custom AI accelerators now account for roughly one-third of total data-center capex — the single largest growth driver in the segment.

The accounting treatment is consistent across the major hyperscalers. Capitalized outlays hit the balance sheet as property, plant, and equipment (PP&E) and are then expensed gradually through depreciation and amortization, typically over five to fifteen years depending on asset class. From an income-statement perspective, this smooths the impact: a $200 billion AI build does not show up as a single year's expense. From a cash perspective, however, the impact is immediate and visible on the cash flow statement under "purchases of property, plant, and equipment" (and increasingly under finance-lease right-of-use additions, which Epoch AI aggregates alongside cash PP&E to capture the true scale of new asset commitments).

Two practical points follow.

• Earnings hide cash stress. A company can keep reporting rising GAAP earnings while free cash flow turns sharply negative, because depreciation does not require cash today even though building the asset did.
• Working capital and finance leases matter. Reported "capex" on the cash flow statement is the cleanest number to track, but power-purchase agreements, GPU-as-a-service contracts, and large equipment-financing facilities can move a meaningful share of capex off the cash statement. Cross-checking reported capex against balance-sheet PP&E growth and finance-lease additions gives a fuller picture.

Capex Versus Opex — Why the Cloud Changed the Math

Historically, enterprise IT departments bore a significant share of data-center and server capex. The rise of public cloud shifted that burden onto hyperscalers and let customers convert what had been fixed capital outlays into flexible operating expenses billed by the hour or the gigabyte. The cloud model — IaaS, PaaS, and increasingly AI-as-a-service — therefore concentrates capex on a small number of operators while distributing the resulting compute and storage capacity across a much larger enterprise customer base.

The economics of this shift are well documented. McKinsey's cloud-ecosystems analysis notes that the migration to consumption-based cloud services has been one of the largest reallocations of corporate IT spending in decades, replacing internal capex-heavy data centers with on-demand compute that scales elastically with workload. From an investor's standpoint this matters because hyperscaler capex is not just a cost — it is also the productive asset base that earns the cloud and AI revenue. The relevant question is not whether the absolute dollars are large but whether the revenue and free-cash-flow generated on the installed base justify the capital tied up.

What Hyperscaler Capex Looks Like in Practice

Translating the general definition into the specifics of a hyperscale data-center build yields a relatively standard cost stack, even as unit costs vary substantially by region, power source, and design.

• Land and shell. Site acquisition, grading, the building itself, and physical security. Costs vary widely by metropolitan area; large campuses can require hundreds of acres.
• Power infrastructure. Substations, transformers, switchgear, on-site generation (often natural gas turbines or fuel cells for backup), and increasingly co-located renewable generation or PPAs. Power is the binding constraint on AI capacity and the largest source of permitting and grid-interconnection delay.
• Cooling and HVAC. Liquid cooling and rear-door heat exchangers are now standard at AI cluster densities; air cooling is being phased out for accelerated compute.
• IT equipment. Servers, GPUs, custom accelerators (such as Google's TPU, Amazon's Trainium, and Microsoft's Maia), high-bandwidth memory, networking switches, and fiber interconnect. This is the line item that has driven the most striking capex acceleration since 2023.
• Network and interconnect. Long-haul fiber, optical transceivers, and the proprietary interconnect fabrics that tie accelerators into large training clusters.
• Capitalized software and internal tools. A smaller but non-trivial share of capex is internal-use software and platform tooling that meets the accounting criteria for long-lived assets.

Dell'Oro's tracking breaks hyperscale customers into four buckets — the Top 4 U.S. cloud (Amazon, Google, Meta, Microsoft), Top 4 China cloud (Alibaba, Baidu, ByteDance, Tencent), Top 4 Tier 2 (Apple, CoreWeave, IBM, Oracle), and a "Rest-of-Cloud" segment of neo-clouds and GPU-as-a-service providers — plus colocation, telco, and traditional enterprise. The Top 4 U.S. cloud alone accounted for nearly half of global data-center capex in 2025.

From Cloud Build-Out to AI Supercycle: How This Cycle Differs

The current hyperscaler capex supercycle did not begin with AI. It builds on more than a decade of cloud and data-center investment that funded the basic capabilities — global storage, elastic compute, content delivery, mobile backends — that the AI build now relies on.

In the mid-2010s, the largest hyperscale cloud and internet companies were already spending tens of billions per year on capex; by 2017 that group spent roughly $75 billion for the year and about $22 billion in the fourth quarter alone. Through the late 2010s and early 2020s, capex grew alongside SaaS adoption, video streaming, and digital services, but spending remained relatively balanced between general compute and emerging specialized workloads. In 2023, Amazon, Google, and Microsoft collectively spent more than $127 billion on capex, roughly flat versus 2022, reflecting a mature yet still-expanding cloud market that had built the "table-stakes" infrastructure needed to pivot when generative AI began scaling.

The AI inflection changed the slope. Since the second quarter of 2023, combined capex at Alphabet, Amazon, Meta, Microsoft, and Oracle has been compounding at roughly 72 percent per year, and forward 2026 guidance from the major hyperscalers suggests combined capex could approach or exceed $600 billion in 2026 (Epoch AI, Penn Capital). This is materially faster than the prior cloud cycle and concentrated in a smaller number of asset categories — accelerators, networking, and power infrastructure — than the broader mid-2010s build.

Three structural differences are worth highlighting.

• Concentration in accelerators. Roughly one-third of hyperscaler data-center capex is now GPUs and custom AI chips, an asset class that did not meaningfully exist before 2017 and that depreciates more rapidly than traditional servers.
• Power as the binding constraint. The capital intensity of AI has shifted the bottleneck from silicon supply to power generation, transmission, and grid interconnection. Permitting timelines for new generation and transmission now routinely gate the pace of capacity additions.
• External financing. Whereas the 2010s cloud build was largely funded from operating cash flow, the AI build has pushed hyperscalers deeper into investment-grade debt, equity issuance, joint ventures with private infrastructure funds, and structured GPU-financing facilities.

The AI Capex Supercycle in Numbers

Several independent research houses now converge on the picture of a multi-year AI capex supercycle.

• Dell'Oro Group projects worldwide data-center capex to grow at a 21 percent CAGR through 2029, reaching roughly $1.2 trillion annually, with hyperscale cloud service providers accounting for about half of that total. GPUs and custom AI accelerators represent approximately one-third of total data-center capex, the single largest growth driver. Dell'Oro also expects accelerated servers for AI training and domain-specific workloads to represent about half of data-center infrastructure spending by 2029.
• Goldman Sachs Research estimates combined capex at the large technology companies leading the AI build-out (primarily the major hyperscalers) of $5.3 trillion from 2025 through 2030, up from a prior estimate of $4.5 trillion before first-quarter 2026 earnings reports.
• Epoch AI finds combined capex at Amazon, Microsoft, Alphabet, Meta, and Oracle has roughly quadrupled since the release of GPT-4 in early 2023, with combined quarterly capex running above $200 billion annualized.
• Penn Capital estimates 2026 combined hyperscaler capex guidance at nearly $690 billion, up roughly 81 percent versus 2025 and more than 200 percent versus 2024.
• 650 Group documents record data-center capex guidance from U.S. hyperscalers in 2025, with continued upward revisions through 2026.
• KPMG provides a global benchmarking view of capex and opex in data centers, noting that AI workloads are reshaping the relative weight of compute, power, and cooling within the cost stack.
• State Street Global Advisors (SSGA) frames the AI capex cycle as having more staying power than prior tech build-outs, citing the breadth of demand drivers and the need for recurring model retraining across enterprise customers.
• Apollo Academy's Hyperscaler Capex note lays out the corporate financing picture, including investment-grade debt issuance, private credit facilities, and structured GPU financing.
• CreditSights provides independent credit-research estimates of 2026 hyperscaler capex, including company-level forecasts and sensitivity to power and chip pricing.
• CoBank frames the AI infrastructure build as a capital supercycle with implications for credit, equities, and rural infrastructure investment.
• Ninety One places the AI capex wave within a broader macro view that also includes energy transition and nearshoring — together contributing an estimated $2.5 to $5 trillion of incremental annual global capex by 2030.
• MUFG Americas documents the financing side of the AI supercycle, including corporate bond issuance, syndicated loans, and private-credit capacity committed to data-center projects.

Where estimates diverge, they typically differ on (a) the share of capex that is incremental AI versus table-stakes cloud, (b) the speed of power and grid build-out, and (c) the durability of AI revenue growth that justifies the spend. They converge on the conclusion that the absolute size of hyperscaler capex is unprecedented relative to any prior corporate IT build-out.

What This Means for Free Cash Flow

The most direct implication of a capex supercycle is compression of free cash flow. Operating cash flow at the hyperscalers has continued to grow, but capex has grown faster. The result is that AI-related capex could consume over 90 percent of operating cash flow by 2026, up from roughly one-third in 2023, according to asset-management estimates from J.P. Morgan.

This matters for three reasons.

• Buybacks slow. Free cash flow is the engine of capital return. As more of it gets absorbed by data-center build-out, share repurchases and dividend growth slow even when reported earnings are healthy. Investors who model capital return must watch the FCF line, not the buyback announcement.
• Balance sheets grow. The cumulative stock of PP&E on hyperscaler balance sheets is set to roughly double over the next several years. Even after depreciation, the embedded capital base is materially larger, raising the bar for incremental returns.
• Financing cost enters the picture. As internal cash flow is insufficient to fund the build, hyperscalers are issuing debt at investment-grade levels. The cost of that debt — and the duration of maturities — becomes a non-trivial driver of equity returns.

The flip side is operating leverage. If AI and cloud demand scales as the hyperscalers expect, the marginal revenue on an installed base that already exists is high. Fixed infrastructure costs get spread across a larger usage base, and revenue growth can flow through to free cash flow faster than the early-cycle capex burden suggests. The investor question is whether the revenue line will materialize on a schedule that matches the depreciation curve — and what level of growth the market is implicitly assuming.

How Hyperscalers Are Paying for It

Three financing channels are now visible.

• Operating cash flow. Still the largest single source, even at the elevated capex run-rates. The investment-grade rating and the scale of the underlying cloud and advertising businesses give the major hyperscalers meaningful cash-generation capacity.
• Investment-grade debt. Hyperscalers now sit among the largest issuers of investment-grade corporate debt. Goldman Sachs Research notes that issuer concentration in widely tracked corporate bond indices is itself becoming a market-saturation concern, because the same handful of names are absorbing ever-larger debt allocations.
• Private capital and structured financing. Private infrastructure funds, sovereign wealth funds, and other institutional investors are increasingly co-investing in data-center campuses and GPU clusters. The Apollo Academy note documents structured GPU-financing facilities and project-level debt as a growing share of the funding mix.

The mix between these channels has equity implications. Heavy reliance on debt can pressure credit ratings over time, particularly if capex extends longer than expected. Equity issuance dilutes existing holders. Private-capital partnerships reduce the asset intensity on the hyperscaler balance sheet but also reduce the upside capture from the resulting compute capacity. Each approach involves tradeoffs that are not visible from the capex headline alone.

Returns: What ROI Looks Like and the Risks

Capital expenditure earns a return only if the assets it builds generate revenue and operating profit in excess of their cost of capital over their useful life. For hyperscaler AI and cloud build, three return measures are worth tracking.

• Return on invested capital (ROIC). Hyperscaler ROIC has historically been high because cloud has been a software-like business running on infrastructure with depreciation schedules measured in years. As AI capex grows and depreciates faster, the blended ROIC will compress unless revenue scales accordingly.
• Payback period. For AI accelerators in particular, the relevant question is how quickly training and inference revenue repays the chip and networking cost. Industry estimates commonly cite three-to-five-year useful lives for GPU clusters, but those estimates depend heavily on utilization rates, model demand, and the absence of faster obsolescence.
• Free-cash-flow conversion. The ratio of FCF to net income will likely remain depressed for several quarters as capex outpaces operating cash flow growth, even when reported earnings continue to grow.

The principal risks to the capex thesis are well documented but worth restating.

• Demand shortfall. If enterprise AI adoption disappoints or pricing erodes faster than expected, the installed base earns less than the capex implied. Asset-management research emphasizes that hyperscalers will need to demonstrate sufficient demand and pricing power quickly enough to avoid over-stretching cash flow and balance sheets.
• Supply chain and power delays. Permitting, grid interconnection, and equipment lead times can push revenue generation out beyond the depreciation curve, lengthening payback periods.
• Technology obsolescence. AI accelerator generations are improving rapidly. A cluster deployed in 2026 may be partially obsolete by 2028, raising the effective depreciation rate and compressing returns.
• Regulatory and political risk. Cross-border data flows, export controls on advanced chips, and energy and environmental rules can all affect both the cost of build-out and the addressable market.
• Financing-cost shock. If investment-grade spreads widen materially or the liquid credit market absorbs less issuance, the marginal cost of capex rises and the return profile falls.

What Investors Should Watch

Three indicators help separate the capex narrative from the capex reality in upcoming quarters.

• Capex as a share of operating cash flow. The cleanest single indicator of capex pressure. Sustained readings above 80 percent imply that capital return will remain compressed and financing needs will rise.
• Power capacity in service. Megawatts actually delivered to active data centers — not announced or under construction — is the leading indicator of revenue-generating compute. Industry trackers like Dell'Oro and 650 Group now publish multi-year capacity forecasts.
• Backlog and contracted compute. Long-dated AI and cloud commitments from enterprise customers, governments, and AI labs indicate whether demand is matching the build pace. Quarterly disclosures from hyperscalers increasingly break out remaining performance obligations and contracted cloud backlog.

The next twelve months of earnings will be especially informative. Hyperscalers have guided to materially higher 2026 capex, and the market will be watching whether the corresponding revenue ramp — AI services, cloud consumption, and enterprise commitments — follows the curve implied by the spend. The upside scenario is that AI workloads scale quickly enough to absorb the new compute, free cash flow re-expands, and the capex supercycle becomes a durable earnings story. The downside scenario is that revenue lags, FCF stays compressed for longer than expected, and the equity market re-prices the cost of the AI bet. Both scenarios are credible. The data in the next four quarters will narrow the range materially.

Strategic Takeaways

• Capex is the scoreboard, not the strategy. The dollar amount alone tells you little. The composition (AI accelerators versus general compute), the source of financing (cash flow versus debt versus private capital), and the resulting return profile are what matter.
• Free cash flow is the constraint that binds. When capex exceeds 80–90 percent of operating cash flow, capital return slows and equity returns become highly sensitive to the next year's revenue ramp.
• Power is the bottleneck. The pace of grid interconnection and generation build-out will determine how quickly committed capex translates into revenue-generating compute.
• Financing mix is a leading indicator. Shifts toward debt, private credit, and structured financing signal that internal cash flow is no longer sufficient to fund the build — useful information about how durable the supercycle is.
• The cycle is real, but not uniform. Hyperscaler dispersion is high, and the variance in execution, power access, and AI monetization will produce wide return dispersion even among the biggest names.

Related reading

• Investing in Hyperscalers: How the AI Infrastructure Supercycle Is Reshaping Cloud and Capital Markets — the investment case and portfolio construction around the buildout

Sources, methodology & compliance

Published: July 24, 2026. This piece draws on research from Dell'Oro Group, Goldman Sachs Research, Epoch AI, Penn Capital, 650 Group, KPMG, State Street Global Advisors, Apollo Academy, CreditSights, CoBank, Ninety One, MUFG Americas, J.P. Morgan Asset Management, McKinsey & Company, and Platformonomics. All options strategies described here are computed using the Black–Scholes–Merton framework.

— Dependability Research Desk

Disclaimer: This research is for informational purposes only and does not constitute investment advice. Options trading involves substantial risk of loss. Past performance is not indicative of future results.