global geopolitics

Amazon, Microsoft, Google, Meta and Oracle are borrowing at an unprecedented rate to fund AI infrastructure, even as memory prices climb and the revenue to justify the spending remains unproven

Microsoft, Amazon, Alphabet, Meta and Oracle built their market value over two decades on a specific financial identity, cash-generative software businesses that required comparatively little capital to scale. That identity has changed within roughly two years. Epoch AI’s research, published in June 2026, found that aggregate hyperscaler cash capital expenditure has been growing at approximately 70% a year against operating cash flow growth of about 23%, with the two lines projected to cross around the third quarter of 2026, the point at which combined free cash flow across the group reaches zero. Oracle had already crossed that threshold by the time of the report, and Amazon was approaching it. PIMCO reached a similar conclusion, estimating that capital expenditure would consume 94% of hyperscaler operating cash flow in 2026, compared with under 50% two years earlier.

Deutsche Bank’s global head of macro and thematic research, Jim Reid, has documented the same shift from a different angle, warning clients that continued economic growth tied to the AI investment cycle would require capital spending to remain what he called parabolic, a trajectory he described as highly unlikely to persist. His colleague George Saravelos went further, writing to clients that Nvidia was effectively carrying the weight of American GDP growth given the scale of capital goods it supplies to the AI investment cycle. These are not claims from AI sceptics operating outside financial institutions; they come from inside one of the world’s largest investment banks, tracking capital flows that are now large enough to distort national output figures.

The financing gap this creates has driven hyperscalers into corporate debt markets at a pace with no recent precedent among technology companies. Amazon’s attempt to raise $25 billion through a bond sale in July 2026 illustrates the shift in investor appetite. Peak demand for the offering reached $62 billion before underwriters trimmed the final spread, at which point orders settled at roughly $41 billion, or 1.6 times the size of the deal, according to Bloomberg’s reporting on the transaction. That figure compares unfavourably with Amazon’s own bond sale four months earlier, in March 2026, which was described at the time as heavily oversubscribed. Across the sector, index-eligible new debt issuance from hyperscalers had already crossed $136 billion by mid-2026 according to PIMCO, and Barclays projected total hyperscaler issuance would exceed $200 billion for the year, a figure that could climb again in 2027. Meta, Oracle, Alphabet and Amazon have each completed bond sales in excess of $15 billion since late 2025, with Meta’s $30 billion offering in October 2025 standing as the largest non-acquisition-related high-grade corporate bond sale on record at the time.

Rising financing costs have coincided with a separate, structurally distinct cost pressure inside the same supply chain: the price of high-bandwidth memory, the specialised chip component that sits directly on graphics processing units and determines how quickly an AI accelerator can move data. Three companies, Samsung, SK Hynix and Micron Technology, control effectively all global production of this component. TrendForce and the Taiwanese industry publication DigiTimes both reported in mid-2026 that high-bandwidth memory prices were on track to more than double by 2027, with next-generation HBM4 pricing potentially reaching $4 to $5 per gigabit, up from roughly $2 in the second half of 2026. Micron’s own fiscal third-quarter results, announced in June 2026, gave a direct financial expression of that pricing power: non-GAAP gross margin reached 84.9%, a company record, up from 39% in the same quarter a year earlier, with net income of roughly $28 billion driven substantially by data centre and cloud memory sales. Micron’s own guidance for the following quarter pointed toward margins expanding further still, to approximately 86%.

That combination, rapidly rising infrastructure costs financed increasingly by debt rather than operating cash, has not been matched by comparable clarity about where the corresponding revenue will come from. Ed Zitron, whose newsletter Where’s Your Ed At has tracked AI industry financials since 2023, has calculated that OpenAI and Anthropic together carry approximately $1.1 trillion in compute commitments contingent on continued growth in their customer base, a figure consistent with public statements from OpenAI chief executive Sam Altman, who told investors in late 2025 that the company was looking at approximately $1.4 trillion in infrastructure commitments over eight years, before that projection was reportedly revised down toward a $600 billion compute-spend target through 2030 following investor concern that the original figure outpaced realistic revenue growth. Anthropic, for its part, signed a 20-year lease with the former bitcoin mining company TeraWulf in July 2026, covering approximately 400 megawatts of capacity at a data centre campus in Hawesville, Kentucky, an agreement TeraWulf said could generate around $19 billion in contracted revenue over its term, a figure larger than TeraWulf’s own market capitalisation at the time the deal was announced.

Commitments on this scale rest on an assumption that computing demand will continue expanding at the rate hyperscalers have projected internally, an assumption that several of the companies’ own recent decisions appear to complicate. Microsoft, Google and Amazon have each moved to cap or reduce internal AI usage budgets for employees in 2026, a step that sits awkwardly alongside continued external capital commitments running into the hundreds of billions of dollars. Whether this reflects short-term cost discipline unrelated to underlying demand, or a genuine signal about the limits of current large language model usefulness inside the companies best positioned to judge it, remains a live and unresolved question rather than a settled one.

The comparison most often invoked when discussing a potential downturn in this sector, the 2008 financial crisis, does not map cleanly onto the present situation, whatever the surface similarities in scale of borrowing. The core of the 2008 intervention was not the Troubled Asset Relief Program’s purchase of distressed mortgage securities, politically prominent as that programme was, but a set of Federal Reserve facilities, including the Primary Dealer Credit Facility and the Term Securities Lending Facility, that existed to keep short-term interbank and money-market funding operating at all, because a collapse there threatened the basic mechanics through which cash moved through the American financial system. Data centre operators and AI laboratories, however large their debt loads have become, do not perform an equivalent systemic function; a default by a company such as OpenAI or Anthropic would not, on its own, threaten the plumbing of the wider financial system in the way a collapse at American International Group did in 2008. That distinction matters for how policymakers in Washington are likely to respond if hyperscaler or AI-lab financing strain intensifies, and it argues against assuming that large-scale federal intervention would follow automatically from a downturn in this sector.

Deutsche Bank’s own broader research, distinct from Reid’s warnings on capital intensity, offers a more measured framing than an outright collapse scenario. A separate Deutsche Bank analysis divides the AI boom into three components, a valuation bubble, an investment bubble, and a technology bubble, and finds that current public-market technology valuations, while elevated, remain less extreme than those recorded at the peak of the dot-com era in 2000, attributing much of the increase to genuine earnings growth rather than speculation in isolation. That assessment sits in tension with the financing and margin data set out above, and the two are not straightforwardly reconcilable, as rising debt, tightening cash-flow coverage and record supplier margins can coexist, for a period, with earnings growth that public-market investors judge to be fundamentally sound. Which of these dynamics proves decisive will depend on factors not yet resolved in the public record, including whether hyperscaler capital expenditure plans are revised downward before financing costs become unsustainable, and whether large language model deployment inside enterprise customers generates revenue at a pace commensurate with the infrastructure now being built to support it.

Authored By: Global GeoPolitics

If you prefer to make a one time donation in support of my work, you can do so by clicking any link below:

https://donate.stripe.com/7sY8wPfvi864don0k3awo07 |

https://buymeacoffee.com/ggtv |

https://ko-fi.com/globalgeopolitics |

Bitcoin: 3NiK8BoRZnkwJSHZSekuXKFizGPopkE7ns

References

Epoch AI (2026). “Hyperscaler Capex Is on Trend to Outpace Their Cash Inflows by the End of 2026.” Isabel Juniewicz, 16 June 2026.

PIMCO (2026). Research note on hyperscaler capital expenditure and operating cash flow, cited in Yahoo Finance, “AI Hyperscalers Are Taking on Debt, But the Broader Market Looks More Leveraged,” 30 May 2026.

Deutsche Bank Research (2025–2026). Notes by Jim Reid, Global Head of Macro and Thematic Research, and George Saravelos, on AI capital expenditure and GDP contribution, September 2025 and July 2026.

Bloomberg News (2026). “Amazon Fuels AI Debt Boom With Bond Sale of at Least $25 Billion.” 7 July 2026.

TrendForce and DigiTimes (2026). Industry reporting on high-bandwidth memory pricing forecasts for 2027, June-July 2026.

Micron Technology, Inc. (2026). Fiscal Third-Quarter 2026 Earnings Release, SEC Form 8-K, Exhibit 99.1, 24 June 2026.

Where’s Your Ed At (2026). Zitron, E. “The AI Industry Is Losing.” June 2026.

TechCrunch (2025). “Sam Altman Says OpenAI Has $20B ARR and About $1.4 Trillion in Data Center Commitments.” 6 November 2025.

TeraWulf Inc. (2026). Announcement of 20-year data centre lease with Anthropic, Hawesville, Kentucky, 6 July 2026, reported by CNBC and Coindesk.

United States Department of the Treasury and Federal Reserve (2008–2009). Records of the Troubled Asset Relief Program, the Primary Dealer Credit Facility, and the Term Securities Lending Facility.

Deutsche Bank Research (2026). “The AI Bubble: Who’s Swimming Naked?” Analysis of valuation, investment, and technology bubble components within the AI sector.