Hyperscaler Capital Spending Nears $785 Billion as the BIS Warns of a Bust

By The Edge Research Team

Moody’s puts hyperscaler capital expenditure at 785 billion dollars in 2026, a figure the agency describes as hyperscaler capital spending rather than as an AI-specific total, and its forecast of 11 May 2026 has 2027 approaching one trillion dollars. Forbes separately reports a further 821 billion dollars of what it calls leases and expenses that have not started yet, across six companies: Alphabet, Amazon, CoreWeave, Meta, Microsoft and Oracle. The rating agency’s view, reported on 23 July, is that heavy capital spending relative to revenue will drive free cash flow down and in some cases negative, and that a material shift in the structure of these companies’ balance sheets is becoming evident. Three of the four largest spenders raised their guidance in the fortnight that followed.

The quarterly figures are the starting point. Alphabet spent 44.92 billion dollars on property and equipment in the second quarter, reported 22 July. Microsoft spent 35.80 billion dollars in cash additions, or about 41 billion including finance leases, in the quarter to 30 June, reported 29 July. Meta spent 31.08 billion dollars including finance lease principal, reported the same day. Amazon spent 54.21 billion dollars gross, reported 30 July, taking its trailing twelve-month gross figure to 173.03 billion dollars.

The guidance moved up, not down. Alphabet raised its 2026 range to 195 to 205 billion dollars from 180 to 190 billion, having started the year at 175 to 185 billion, and said capital expenditure would increase significantly again in 2027. Meta narrowed to 130 to 145 billion dollars from 125 to 145 billion, lifting the floor, against an original guide of 115 to 135 billion. CNBC reported on 30 July that Amazon had raised its 2026 figure to 220 billion dollars from about 200 billion, attributing the increase to higher memory and component costs. Microsoft’s calendar-2026 number fell to about 175 billion dollars from the about 190 billion it gave on its fiscal third-quarter call of 29 April, but this is an accounting effect and not a retreat: the company is extending the estimated useful life of data centres and office buildings from 15 to 25 years from the start of its 2027 fiscal year, which reclassifies future data centre leases from finance to operating, and it stated that outside this useful-life impact its calendar-2026 expectations are unchanged. It guided the first quarter of fiscal 2027 to more than 50 billion dollars.

What has changed most is not the size of the spending but its circularity, and the disclosures are now explicit enough to trace. Microsoft’s commercial remaining performance obligation rose 84 percent to 678 billion dollars, but 25 percent excluding a single counterparty. Amazon booked 53.4 billion dollars of non-operating pre-tax other income in the quarter, primarily from its investment in Anthropic, whose carrying value was about 74.2 billion dollars at the end of March before a further 5.0 billion dollar investment and a financing facility not to exceed 20.0 billion dollars. Alphabet’s non-marketable securities rose to 131.5 billion dollars from 68.7 billion in December, with a 99.03 billion dollar gain on equity securities in the quarter. Nvidia’s non-marketable equity holdings nearly doubled to 42.3 billion dollars in a single quarter. AMD agreed on 22 July to invest up to 5 billion dollars in Anthropic alongside a commitment of up to two gigawatts of its accelerators. Anthropic itself raised a 65 billion dollar round on 28 May at a 965 billion dollar post-money valuation and said run-rate revenue had crossed 47 billion dollars. Oracle’s remaining performance obligation reached 638 billion dollars, up 363 percent, of which 75 billion dollars is prepaid or customer-supplied hardware, while free cash flow ran at minus 23.7 billion dollars.

Because operating cash flow no longer covers it, the funding has moved to debt. Some 236 billion dollars of AI-related debt had been priced globally by 31 May, roughly four times the prior-year pace, with Morgan Stanley projecting 570 billion dollars for the full year. Bond order books thinned from around five times covered in February to below two times by July. Data centre securitisation topped 30 billion dollars in 2025, about three times the 2024 level, and S&P Global Ratings notes technology’s share of global non-financial corporate bond issuance rose to 16.7 percent in 2025 from 11.6 percent. Oracle announced a 45 to 50 billion dollar debt-and-equity financing plan for calendar 2026, Meta placed roughly 27 billion dollars of development cost into a joint venture with Blue Owl for a single campus, and CoreWeave issued 3.5 billion dollars of senior notes in June.

The physical constraint is electricity. The International Energy Agency puts global data centre consumption at 485 terawatt hours in 2025, rising 17 percent on the year, and projects 950 terawatt hours by 2030, around 3 percent of global electricity demand. Within that total, the Agency says electricity demand from AI-focused data centres is set to triple by 2030. Data centres account for roughly half of all United States electricity demand growth to 2030.

Official institutions have moved from observation to warning. The Bank for International Settlements, in its annual report published at the end of June, noted that the five largest hyperscalers are set to spend over a trillion dollars on AI-related capital expenditure across 2025 and 2026 together, warned that disappointment in returns could trigger a sudden pullback in financing and turn the capital expenditure boom into a protracted investment bust, and observed that the terms of these deals are typically poorly disclosed, with the risk of the same asset being pledged multiple times. That is the BIS identifying a risk that arises from weak disclosure, not a finding that double-pledging has occurred. It added that direct lending funds have quadrupled their exposure to AI and information technology in five years, to about 15 percent of portfolios. The Bank of England’s July financial stability report records that AI-related companies now account for roughly half of S&P 500 market capitalisation against about a quarter in 2022, and models a scenario — not a forecast — in which United States equities fall 45 percent over six quarters alongside a 350 basis point widening in credit spreads, which it estimates would take 2.2 percentage points off United Kingdom GDP. The IMF’s July update warns that frothy equity valuations in technology-concentrated markets could correct sharply, and the European Central Bank’s May review flags rising concentration and interconnection among a handful of large United States technology firms. The Federal Reserve’s own May stability report, in the survey exhibit on its second page, found AI cited as a salient risk by half of surveyed market contacts, against none in the autumn of 2025. Markets have not yet taken the point: the Nasdaq Composite closed at 25,192.82 on 31 July.

Why it matters: The Gulf has positioned itself on the capital side of this rather than the model side, and the distinction is the whole argument. Kuwait’s Investment Authority is a founding investor in Helix Digital Infrastructure, launched on 11 June with KKR as anchor and Nvidia as strategic partner and more than 10 billion dollars of committed long-duration capital, and it was the first non-founder financial anchor investor in the AI Infrastructure Partnership, which targets 30 billion dollars of equity and up to 100 billion including debt. The UAE’s MGX closed its first fund at 49 billion dollars on 1 July, above target, and completed the acquisition of Aligned Data Centers at roughly 40 billion dollars enterprise value with 5 billion dollars of growth capital on 21 July. Qatar’s Qai formed a 20 billion dollar partnership with Brookfield in December, and AMD’s collaboration with Saudi Arabia’s HUMAIN runs to 10 billion dollars over five years. These are infrastructure positions with long asset lives, which is a materially different risk profile from equity in model developers. Whether each vehicle carries contracted revenue is not disclosed vehicle by vehicle, so the point stands on asset duration rather than on offtake. The BIS caution about the same asset being pledged more than once is nonetheless the right question for any investor in this chain to ask of its own documentation.

Outlook: Two near-term markers. Nvidia reports its second fiscal quarter in late August, having guided to 91 billion dollars in revenue on 20 May with no data centre compute revenue assumed from China; that print is the cleanest read on whether demand is still ahead of supply. The second is the bond market, where order-book cover has fallen from roughly five times to below two in five months. Capital expenditure guidance can be raised on a call; the debt to fund it has to clear at a price, and that is where the first real constraint will show.

The four largest spenders

The financing and the warnings

Sources: company earnings releases and SEC filings for Alphabet, Amazon, Meta, Microsoft, Nvidia, Oracle and CoreWeave; Moody’s; S&P Global Ratings; Bank for International Settlements; Bank of England; International Monetary Fund; European Central Bank; Federal Reserve; International Energy Agency; CNBC; Forbes; MGX; BlackRock; Brookfield; AMD.

Get Complimentary Insights

Follow us