US hyperscalers clearly cooking their books
With US earnings season out of the way, there are a number of AI-related issues specific to hyperscalers which are the engine of data centre spending that keep us awake at night. Although we have addressed the rising pricing pressure on LLMs which has been sparked by Chinese labs which are dramatically discounting their products, we think the fast-deteriorating quality of financial reporting by this subset is becoming too obvious to ignore.
The current market valuation of AI mega-caps rests on an elaborate triad of accounting cosmetics designed to manufacture the appearance of sustainable profitability. On the balance sheet, hyperscalers are reportedly masking an estimated $1.6 trillion in shadow leverage through uncommenced data centre leases, take-or-pay purchase commitments, and off-balance-sheet special purpose vehicles, keeping massive construction liabilities out of plain sight while free cash flows plunge.
Meanwhile, headline earnings are heavily flattered by selective Non-GAAP methodology, where analysts routinely strip out recurring operating expenses like stock-based compensation schemes while waving through tens of billions of dollars in non-recurring, mark-to-market paper gains from private AI holdings like Anthropic and OpenAI.
We were shocked to learn that these paper gains made up over 70% of Alphabet’s (GOOG) quarterly profits and over 65% of Amazon’s (AMZN). For those old enough to remember the latter years of Japan’s financial bubble in 89/90, the bookings of similar gains under the practice dubbed ‘Zaiteku’ came immediately to mind seeing this.
Compounding this financial alchemy is the artificial suppression of operating costs through aggressively stretched depreciation timelines. By extending server and computing useful life assumptions from the normal practice of two to three years to five or six years, hyperscalers are erasing tens of billions in annual depreciation expenses from their income statements, despite the harsh physical reality that high-power AI silicon faces economic obsolescence within 24 to 36 months under relentless duty cycles and extreme thermal stress.
Between hidden shadow leverage, inflated venture markups, and an impending cliff of accelerated GPU write-downs, these manicured metrics create a dangerous veneer of quality that obscures severe structural margin drag and rising capital costs. The above scenario is exacerbated by the rising cost of compute—which, as we have long argued, is the secular inflation within the AI supply chain eating into spending budgets—and the falling price of compute as LLMs fight for market share.
We continue to believe that the canary in the coal mine for things going pear-shaped is the neocloud segment, given that these firms have been banking on compute demand continuing to rise without much consideration of the intensifying competition from Chinese labs. As we have argued, these challengers possess a natural advantage through cheaper access to electricity and water for their data centres. Thus, we keep a close eye on the share price performances of names like Oracle (ORCL), CoreWeave (CRWV), Nebius (NBIS), and IREN (IREN) as key indicators for sentiment in the AI space.
As long as their shares, which have bounced strongly in the past few weeks, remain at least stable in the near term, we think the prospects for other AI names, including those trading in Japan, should be okay. However, any notable deterioration in sentiment in this specific space could once again be a sign of trouble ahead.
