Lifespan of AI Chips: The $300B Question (blog.citp.princeton.edu)

OpenAI’s new multibillion-dollar Broadcom chip deal (following a separate AMD pact) is part of a massive industry buildout—eight major players are slated to spend over $300 billion on AI infrastructure in 2025. Mihir Kshirsagar highlights a critical technical-financial mismatch: AI inference/training hardware typically endures heavy thermal/electrical stress and becomes technologically obsolete in about 1–3 years, yet firms are depreciating these assets over 5–6 years. Rapid chip advances (e.g., Nvidia’s GB200 delivering ~4–5x faster inference than H100) amplify economic obsolescence. McKinsey-style cost breakdowns suggest roughly half of near-term infrastructure spend goes to computing hardware, so overstretched depreciation effectively halves reported annual replacement costs. That accounting “subsidy” creates a 3–6 year window during which hyperscaler-model coalitions (Microsoft–OpenAI, Amazon–Anthropic, Google, Meta) can underprice or absorb infrastructure costs, lock in enterprise integrations, and scale capacity faster than true economics justify. Example: if Microsoft’s $80B AI spend implies ~$13B/year real replacement costs but is depreciated over six years, reported depreciation falls to ~$6.5B, leaving a multi‑billion-dollar cushion to subsidize partners. The result: distorted unit economics, daunting capital gaps (Bain estimates an $800B annual shortfall by 2030; Kshirsagar’s TCO-based calc exceeds $1.5T), and a credible risk that temporary accounting advantages harden into durable market power—regardless of whether the investment model is ultimately sustainable.