Writing · AI / Automation / Tech

2025-12-02
Five Accounting Tactics Hyperscalers Use to Make AI Capex Look Cheaper, Margins Look Fatter, and Leverage Look Low Munger called EBITDA “BS earnings”. I started thinking more about Munger’s jab at EBITDA. So I went down the accounting rabbit hole of the hyperscalers. We’re not in Enron territory. But there’s plenty of apple-polishing. Here are the pressure points analysts keep circling. 1. Meta’s Hyperion: the $27–29B asset that “isn’t an asset.” Meta designs it. Meta builds it. Meta operates it. Meta guarantees the debt. Yet Hyperion lives in a joint-venture SPV and stays off Meta’s balance sheet. This keeps tens of billions of capex and debt out of leverage ratios, ROIC math, and credit ratings. Structured finance if you’re being polite. “Come on… it’s yours” if you’re not. 2. Depreciation magic: turning 24–36-month hardware into 6-year assets Amazon, Microsoft, Alphabet, and Meta all stretched the useful life of servers and networking to roughly six years. In an industry where chips age like fruit, not steel. Depreciation falls. Earnings rise. Economics don’t budge. Michael Burry thinks this move inflates future earnings by over 170 billion dollars. Jim Chanos thinks the asset lives are wishful thinking. 3. Capitalizing labor to smooth margins Big Tech can shift engineering payroll from expense today to “capitalized software” amortized tomorrow. Perfectly legal. Heavily judgment-driven. Spikes in capitalization often show up right when management promises margin expansion. Not fraud. Just cosmetic tuning. 4. Stock-based compensation: the vanishing expense SBC is a real cost. GAAP counts it. Adjusted earnings make it disappear. Amazon alone hands out over 20 billion dollars a year. Across the hyperscalers, dilution is massive. You can’t give away 1–2 percent of the company every year and pretend it’s free. 5. Non-GAAP metrics: the “one-time” charges that never end Every quarter, companies strip out: • SBC • Layoffs • Restructuring • Impairments • Failed bets Yet many AI expenses labeled “non-recurring” show up year after year. Slide decks sparkle. Income statements groan. Three patterns jump out. 1. AI capex looks cheaper on paper. Longer asset lives Off-balance-sheet SPVs Write-down risk later 2. Labor looks cheaper on paper. SBC ignored Software development capitalized 3. Earnings look smoother on paper. Adjusted metrics Hidden leverage Narrative over economics No one is yelling “fraud.” But we’re staring at a cluster of yellow flags in the middle of a once-in-a-generation capex boom
AI / Automation / TechCapital / Finance / InvestingOperations / Property ManagementHiring / People / LeadershipMindset / Mental Models / Decision MakingReal Estate (general)

View original on LinkedIn

← Back to writing