For several months, investors have been questioning whether hyperscalers can earn an adequate return on the hundreds of billions of dollars being poured into artificial intelligence. However, the bill could be even larger than published capex alone suggests.
According to MacroMicro, nine tech giants now collectively have more than $3.5tn in off-balance-sheet AI-related commitments, mainly in the form of future purchases and lease contracts, but also strategic investments, guarantees and financing housed in special-purpose vehicles.
That eye-catching figure does not mean the tech groups have piled up $3.5tn of hidden debt. Most of it, about 85%, consists of spending that has already been contractually committed but will not show up in the financial statements until equipment is delivered or leases begin. Nearly $1.9tn (about 54%) relates to future purchase commitments, notably for semiconductors and data center equipment, plus roughly $1.1tn (about 31%) in lease contracts that have not yet started.
In other words, a substantial portion of the bill is already locked in, but it should hit the balance sheets later. As data centers are built and chips are delivered, these commitments will gradually turn into fixed assets, rent expense, lease liabilities and depreciation, with a more direct impact on cash flows and returns on invested capital.
A second category involves investments in companies such as OpenAI or Anthropic. Hyperscalers fund these players while often tying their investment to the use of their own cloud services or compute capacity to generate revenue. That model allows them both to support demand for their infrastructure and to benefit from rising valuations in these private companies. A meaningful share of Amazon's and Alphabet's earnings growth in the second quarter came from mark-up gains on stakes in unlisted or recently listed companies, notably Anthropic for Amazon and SpaceX for Alphabet.
The most sensitive structures, however, are those relying on special-purpose vehicles, residual value guarantees or other contingent commitments. In such cases, the debt and assets are legally held by an entity outside the group, limiting the immediate impact on leverage ratios without necessarily eliminating the economic risk. Nvidia, for example, invested in a Valor-led special-purpose vehicle that bought $5.4bn of GB200 chips to be leased to xAI. The vehicle is not consolidated in Nvidia's financial statements, but a default by xAI or a faster-than-expected decline in GPU values could still weigh on the investment's value and therefore on Nvidia's earnings.
In the same vein, Meta owns 20% of a data center project and uses, among other tools, short-term leases in order not to consolidate the entire vehicle. It does, however, provide a residual value guarantee that can extend up to 16 years, designed to help finance the project with an investment-grade rating, but if the value of the infrastructure ultimately falls below the guaranteed level, Meta could be required to cover part of that difference.
MacroMicro flags three main downside risks: US models losing pricing power to open source, faster-than-expected depreciation of older GPU generations, and tighter financing conditions.
A drop in lease rates for H100s or B200s could, in particular, undermine assumptions about chip useful lives and reduce the value of the assets used as collateral. Finally, a rise in the cost of capital or wider credit spreads would weigh most heavily on the most leveraged players, such as CoreWeave or Oracle, even as depreciation and interest expense keep accruing.
No comments:
Post a Comment
Note: Only a member of this blog may post a comment.