跳到正文
MarketHOT
English
原文
圣路易斯联储 · 经济分析·· 2 小时前精选AI 评分74

圣路易斯联储分析:超大规模科技企业如何为AI投资热潮融资

Hyperscalers’ Financing of the AI Investment Boom

AI 导读

圣路易斯联储分析显示,数据中心建设企业在2024年第四季度至2026年第二季度平均每季度投资196.38亿美元,接近经营现金流加研发支出的197.52亿美元,净债务发行升至30.17亿美元。

推荐理由

分析对比了数据中心建设企业与半导体企业在AI投资扩张中的现金流、债务和股东支付差异,并指出前者对外部融资的依赖有所上升。

正文 · 原文

KEY TAKEAWAYS

  • “Hyperscalers” — large tech firms involved in data centers and computer chips — are drawing from substantial financial resources to fund investment during the AI boom.
  • Hyperscalers’ high operating profitability has provided capacity to finance greater investment with internal funds.
  • Hyperscalers’ corporate debt relative to assets has fallen in recent years and remains below that of other publicly traded U.S. firms. Payments to shareholders have declined since 2022 and converged with those of other firms, suggesting hyperscalers are retaining money.
  • Data center builders lean more heavily on debt and have reduced shareholder payouts to fund growing investment, while computer chip suppliers cover more-modest levels of investment with operating cash flow.

In our previous blog post, we showed that, since the beginning of the artificial intelligence (AI) boom, an increase in capital expenditures and research and development (R&D) at major U.S. technology companies has corresponded with a decline in cash holdings. The decline in cash holdings, however, does not identify how these companies funded their expanding investments. Today, we will examine debt, shareholder payouts, operating income and cash holdings as possible sources of financing.

Our researchThis two-part series outlines findings from our Nov. 10, 2026, article. See Julian Kozlowski, Masataka Mori, Juan M. Sánchez and Nicholas Sullivan, “Financing Artificial Intelligence: Cash, Debt, and Corporate Investment,” Federal Reserve Bank of St. Louis Review, Fourth Quarter 2026, Vol. 108, No. 13, pp. 1-20. focuses on 10 large technology firms: Alphabet Inc., Amazon.com Inc., Apple Inc., Broadcom Inc., Intel Corp., International Business Machines Corp. (IBM), Meta Platforms Inc., Microsoft Corp., Nvidia Corp. and Oracle Corp. As before, we refer to this group as “hyperscalers,” using a broad definition that includes semiconductor suppliers alongside cloud and data center operators. We define the AI era as 2018 to the present, which corresponds to significant advancements in deep learning.

This analysis relies on current balance sheet and cash flow measures and does not capture off-balance-sheet commitments, such as future lease obligations, purchase obligations or remaining performance obligations, which may represent a significant share of data center builders’ forward-looking investment and financing plans.

Debt, Payouts, Profitability and Cash Holdings

Firms can obtain additional resources by increasing borrowing, retaining funds that would otherwise be distributed to shareholders, or relying on internally generated earnings and cash holdings. In the figures below, we show how these measures differ between hyperscalers and all other publicly traded U.S. firms.

The first figure reports debt-to-assets ratios from the first quarter of 2010 to the first quarter of 2026. Hyperscalers’ leverage rose substantially during the first half of the sample period, but it remained below that of other firms throughout. More recently, hyperscalers’ ratio fell from nearly 30% in 2023 to roughly 25% in early 2026. The ratio for other firms remained close to 34%. Thus, even if the dollar amount of hyperscalers’ debt has increased, it has grown more slowly than their total assets in recent years. The recent investment expansion has, therefore, not been accompanied by a comparable increase in hyperscalers’ balance sheet leverage.

Firms’ net payments to shareholders encompass dividends as well as share repurchases (net of shares issued). In the second figure, we look at net shareholder payouts scaled by firms’ total assets. Hyperscalers distributed more to shareholders than other firms did throughout most of the period. Their payout ratio reached approximately 4% in 2022 but then declined to about 2% by early 2026. Consequently, hyperscalers’ ratio is now similar to that of other firms. This decline indicates that hyperscalers have retained a larger share of their financial resources, rather than distributing them to shareholders.

The third figure reports operating income relative to total assets. Hyperscalers were substantially more profitable than other firms throughout the sample, with an operating income ratio that was generally about twice as large. After declining to approximately 21% in 2023, the ratio rose to 25% in early 2026. In contrast, the ratio for other firms remained near 13%. This high level of operating profitability has given hyperscalers considerable capacity to finance expenditures internally.

The figures above show that hyperscalers have leaned less on debt than other firms, pulled back on shareholder payouts, and maintained high operating profitability throughout this period. This combination suggests that hyperscalers had substantial internal resources to draw on as they fund their investment expansion.

Financing the Investment Expansion

We use the following cash flow identity to characterize the sources and uses of funds associated with investment:

inv t = OCF t + ΔD t - div t - ΔM t

In this equation, invt denotes net investing cash outflows plus R&D expenditures. The four terms on the right capture:

  1. Internally generated funds (OCFt): Cash generated by the firm’s operations.
  2. Net borrowing (ΔDt): Funds raised through debt issuance minus debt repayments. Positive net borrowing provides funds for investment; net repayments use funds.We treat R&D spending as investment. Because R&D is recorded as an operating expense, we add it back to reported operating cash flow. Thus, OCFt denotes operating cash flow plus R&D, and invt denotes net investing cash outflows plus R&D. The full identity also includes a small residual reflecting other financing flows and exchange rate effects, which we omit here for simplicity. See our Review article for details.
  3. Net payments to shareholders (divt): Dividends plus share repurchases minus proceeds from issuing equity. These payments use funds, so they enter with a minus sign. Lower payments leave more funds available for investment, while net equity issuance provides additional financing.
  4. Cash accumulation (ΔMt): The change in cash and cash equivalents. Accumulating cash uses funds that could otherwise finance investment, so it enters with a minus sign. Drawing down existing cash balances releases funds for investment.

The table below reports the average components of hyperscalers’ investment financing over two periods: the pre-ChatGPT period (the first quarter of 2012 to the fourth quarter of 2018) and the AI investment boom (the fourth quarter of 2024 to the second quarter of 2026). The top panel reports results for the data center builders (Alphabet, Amazon.com, IBM, Meta, Microsoft and Oracle), while the bottom panel reports results for the semiconductor producers (Broadcom, Intel and Nvidia). While Apple is included in the dataset for all hyperscalers, the company is excluded from the table because its cloud and chip activities are structured primarily around its own consumer products.

Average Quarterly Investment and Financing Flows among Hyperscalers
Component 2012:Q1-2018:Q4 (Billions of GDP-Deflated Dollars) 2024:Q4-2026:Q2 (Billions of GDP-Deflated Dollars) Percent Change between Period Averages
Data Center Builders
Investment (inv) $34.59 $196.38 467.8%
Operating cash flow plus R&D (OCF) $44.46 $197.52 344.3%
Net debt issuance (ΔD) $3.23 $30.17 833.3%
Net shareholder payouts (div) $12.04 $26.84 123.0%
Cash accumulation (ΔM) $1.21 $4.43 266.1%
Semiconductor Producers
Investment (inv) $8.54 $22.68 165.4%
Operating cash flow plus R&D (OCF) $10.55 $40.98 288.3%
Net debt issuance (ΔD) $1.10 -$0.79 -171.8%
Net shareholder payouts (div) $2.96 $14.10 375.9%
Cash accumulation (ΔM) $0.05
$2.34
4,580.0%
SOURCES: Compustat, the gross domestic product (GDP) deflator from FRED and authors’ calculations.
NOTES: Data center builders are Alphabet, Amazon.com, IBM, Meta, Microsoft and Oracle. Semiconductor producers are Broadcom, Intel and Nvidia.

Data center builders’ investment has grown so large that in the recent period it nearly matches their operating cash flow: $196.38 billion versus $197.52 billion, respectively, per quarter. Having largely exhausted what their operating cash flow can cover, data center builders are leaning more heavily on debt. Net debt issuance rose from $3.23 billion to $30.17 billion per quarter. Despite this thin margin, data center builders still increased both cash accumulation and shareholder payouts in dollar terms. However, both shareholder payouts and cash accumulation grew more slowly than operating cash flow and investment. As a consequence, they fell as a share of operating cash flow. Shareholder payouts fell from 27.1% to 13.6% of operating cash flow, and cash accumulation slipped from 2.7% to 2.2% of operating cash flow.

Semiconductor producers tell almost the reverse story. Their operating cash flow has outpaced investment, growing 288.3% and 165.4%, respectively. Operating cash flow reached $40.98 billion per quarter against $22.68 billion in investment. This leaves them with far more funds than they need to cover their own expansion. Semiconductor producers have paid down debt, shifting from net debt issuance of $1.10 billion to net repayments of $0.79 billion per quarter. At the same time, they have sharply increased shareholder payouts and cash accumulation, with payouts rising from 28.1% to 34.4% of operating cash flow and cash accumulation rising from 0.5% to 5.7%.

Conclusion

Together with the decline in cash holdings documented in our previous post, these financing patterns reveal a divide within the hyperscaler group. Data center builders have had to lean more heavily on debt and pull back on shareholder payouts and cash accumulation to fund their rapidly growing investment, while semiconductor producers have comfortably covered their more-modest investment growth with operating cash flow alone — leaving room to increase shareholder payouts, build up cash and pay down debt. Going forward, data center builders’ ability to sustain the current pace of investment will depend more heavily on continued profitability, retained earnings and access to external financing.

Notes

  1. This two-part series outlines findings from our Nov. 10, 2026, article. See Julian Kozlowski, Masataka Mori, Juan M. Sánchez and Nicholas Sullivan, “Financing Artificial Intelligence: Cash, Debt, and Corporate Investment,” Federal Reserve Bank of St. Louis Review, Fourth Quarter 2026, Vol. 108, No. 13, pp. 1-20.
  2. We treat R&D spending as investment. Because R&D is recorded as an operating expense, we add it back to reported operating cash flow. Thus, OCFt denotes operating cash flow plus R&D, and invt denotes net investing cash outflows plus R&D. The full identity also includes a small residual reflecting other financing flows and exchange rate effects, which we omit here for simplicity. See our Review article for details.

来源:圣路易斯联储 · 经济分析 · stlouisfed.org