Search epoch.ai
Hyperscaler capex has quadrupled since GPT-4’s release, nearing half a trillion dollars in 2025
Driven by investments in AI infrastructure, the combined capital expenditures at Alphabet, Amazon, Meta, Microsoft, and Oracle have been growing at an average of 72% per year since the second quarter of 2023. If this trend continues, they will spend $770 billion in 2026.
Company statements and analyst projections anticipate continued rapid spending growth in 2026, though slightly slower than naive extrapolation. Our data is sourced from companies’ financial filings and includes cash spending and new finance leases. This growth rate is consistent with the spending growth we observe in our AI chip sales dataset.
Authors
Published
February 26, 2026
Learn more
Overview
We sum quarterly capital expenditure across Amazon, Microsoft, Alphabet, Meta, and Oracle using two components extracted from SEC filings: (1) cash PP&E (cash payments for property, plant, and equipment) and (2) finance lease ROU assets obtained (right-of-use assets obtained in exchange for finance lease liabilities, effectively new leases). Where these exact tags are unavailable, we use the closest available tag or combination of tags. We then fit an exponential model to the post-Q2 2023 data to characterize the growth rate.
Data
All data is extracted from SEC EDGAR 10-Q (quarterly) and 10-K (annual) filings of Amazon, Microsoft, Alphabet, Meta, and Oracle from Q1 2022 through Q4 2025. We parse the structured XBRL tags rather than relying on company-reported aggregate figures, since companies define “capital expenditure” differently on their earnings calls. For example, Microsoft includes finance lease ROU and accrued-but-unpaid PP&E, Meta includes finance lease principal payments instead of ROU, while Alphabet and Oracle report cash PP&E only.
Fiscal year alignment: Microsoft (fiscal year ending June) and Oracle (fiscal year ending May) report on non-calendar fiscal years. Microsoft fiscal quarters map directly to calendar quarters, while we map Oracle quarters to the calendar quarter with the most overlap (e.g., Oracle FY26Q2, Sep-Nov 2025, maps to CY2025 Q4).
Quarterly derivation: SEC filings may tag standalone 3-month values, year-to-date values, or both. Our extraction logic tries the standalone period first; if unavailable, it derives the quarter by subtracting the prior cumulative from the YTD figure. Q4 values are derived as the 10-K annual total minus the sum of Q1–Q3.
Cash PP&E: Microsoft, Alphabet, Meta and Oracle directly tag us-gaap:PaymentsToAcquirePropertyPlantAndEquipment.
Amazon uses the tag us-gaap:PaymentsToAcquireProductiveAssets.
This component is available for all companies and all periods.
New finance leases: Microsoft, Amazon and Alphabet directly tag us-gaap:RightOfUseAssetObtainedInExchangeForFinanceLeaseLiability. However, prior to 2024, Alphabet did not use this tag; their 2023 10-K filing noted that “finance lease costs were not material for the years ended December 31, 2022 and 2023.” For Meta, we infer new leases from changes in us-gaap:FinanceLeaseRightOfUseAssetBeforeAccumulatedAmortization. For Oracle, we reconstruct new leases as us-gaap:FinanceLeaseRightOfUseAsset (instant) plus us-gaap:FinanceLeaseRightOfUseAssetAmortization (duration).
The claim that capital expenditure growth is driven by AI infrastructure is broadly supported by company earnings calls and reflects their phrasing choices (Microsoft, Meta, Alphabet, Amazon). The exact portion going into AI infrastructure is not provided.
Analysis
We fit an exponential growth model to aggregate quarterly capex (cash PP&E + new finance leases) from Q2 2023 through Q4 2025. The fit is performed via ordinary least squares on log-transformed data, yielding an annualized growth rate of 72% (90% CI: 66% to 78%).
Limitations
Our estimation method for finance lease ROU assets at companies that don’t provide the exact XBRL tag is an imperfect approximation, which may understate true new finance leases. Additionally, operating lease ROU assets are excluded from this analysis, both because their inclusion is less standard and because less granular data is available, which understates total hyperscaler investment in productive capacity.
