What happens to old AI chips? They’re still put to good use and don’t depreciate that fast, analyst says

Concerns that hyperscalers are understating depreciation on older AI chips—highlighted by Michael Burry’s estimate that Big Tech could understate depreciation by $176 billion between 2026 and 2028—have pressured AI stocks, but Alpine Macro argues those fears are overstated. It notes three offsets: software advances can materially boost performance of older processors (Nvidia’s five‑year‑old A100 can be 2–3x faster than at launch), inference workloads will outpace training and can be met by greater chip quantity rather than only cutting‑edge quality, and China’s persistent demand (including secondary channels) keeps older hardware utilized. Yardeni Research also points to 'distributed AI' that taps idle GPUs in consoles, mining rigs and institutions, implying lower write‑down risk and greater resilience in chip inventories than some headline estimates suggest.

Analysis

AI sector momentum has softened amid concerns that hyperscalers may be understating depreciation on legacy AI chips; Michael Burry specifically estimated Big Tech could understate depreciation by $176 billion between 2026 and 2028, which has pressured AI-related equities in recent weeks. The headline risk centers on potential large future write-downs as newer-generation accelerators enter the market faster than anticipated.

Alpine Macro counters that these depreciation fears are likely overstated for three reasons: software enhancements can materially raise the effective performance of older processors (the note cites Nvidia’s five-year-old A100 improving two- to threefold versus its initial release), inference workloads—where quantity can substitute for bleeding-edge quality—are expected to outpace training demand, and Google continues to run seven- to eight-year-old TPUs at full utilization. The note also highlights persistent Chinese demand and secondary channels that keep older hardware employed despite export controls.

Yardeni Research’s observation of “distributed AI” — leveraging idle GPUs in consoles, mining rigs, offices and universities — adds another demand-side absorber that could reduce the scale of hyperscaler write-downs, though it may be slower and less centralized. Net market impact is mildly positive but cautious: NVDA and Google-related exposures get relatively higher sentiment scores, yet regulatory, disclosure and secondary-market risks remain material and should be monitored closely.

Key Decisions for Investors

• Reduce binary exposure to a near-term broad AI sell-off but consider selective accumulation of NVDA given software-led longevity for older GPUs and the higher per-ticker sentiment score
• Monitor hyperscaler capex and depreciation policies and 10-K/10-Q disclosures for signs of accelerated write-downs through 2026–2028 and use that data to size positions or activate hedges
• Increase attention to firms and strategies that benefit from inference growth or distributed-AI architectures and to secondary-market channels servicing China as potential demand backstops