AI chip depreciation poses a risk worth $176 billion. Tech giants suggest a 5-7 year depreciation period for a $40,000 AI chip, which is misleading compared to the real 24-month depreciation period.

The discrepancies between reported profits and real expenses stem from this depreciation strategy. Purchases of billions in hardware intended for a 2-3 year product cycle are depreciated over 5-6 years, inflating earnings artificially. This method extends the apparent economic life of assets, distorting profits.

AI chips have a 24-month functional lifespan. In contrast, physical failures occur within 1-3 years due to continuous operations. The unrealistic 5-7 year schedule leads to inflated financial expectations.

Value diminishes rapidly when new chip versions are introduced, offering superior features. The transition from A100 to H100 demonstrates a shift to quicker performance, rendering older models obsolete almost instantly. Yet, obsolete doesn’t mean worthless.

Despite obsolescence, chips find residual value. An A100 from 2020 retains some worth, marked down by 50-70% of its peak value. With projected overstated earnings and adjustments by financial institutions, the AI boom bears resemblance to past overinvestment trends. The reality contradicts elongated depreciation, echoing past economic bubbles.

We are looking at a market that is ignoring the true, rapid obsolescence of AI hardware. This isn’t just a technical detail; it’s a multi-billion dollar earnings risk hiding on the balance sheets of the largest tech companies. The 24-month hardware cliff is real, but the accounting schedules are still lagging dangerously behind.

The latest Q3 2025 earnings reports from the big cloud providers confirmed this spending frenzy, with combined capital expenditures soaring to over $60 billion for the quarter. A recent report from Gartner now estimates that over 75% of this spend is on AI servers with a functional lifespan of less than three years. We see this massive investment as a future liability, not the asset investors believe it is.

This situation suggests looking at long-dated put options on the major cloud players like Microsoft, Amazon, and Meta. We believe the market has not priced in the potential for significant earnings writedowns in 2026 and 2027. Options expiring in late 2026 could offer a direct way to position for this accounting correction.

The catalyst for this repricing is likely NVIDIA’s next-generation chip architecture, codenamed after another famous scientist and expected in late 2026. Looking back, the launch of the Blackwell B200 in late 2024 made the H100 functionally obsolete for top-tier training overnight. The same cliff is coming for the Blackwell chips, and the balance sheets are not prepared for that value collapse.

We’ve seen this playbook before in market history, particularly during the telecom boom of the late 1990s. Companies spent billions laying fiber optic cable, depreciating it over decades, only to have much of it become “dark fiber” after the 2000 crash. The subsequent asset writedowns were devastating for shareholder returns.

A more market-neutral strategy involves a pairs trade, going long on the chip designer like NVIDIA and short a basket of the hyperscalers. This position profits from the dynamic of the “shovel seller” getting paid regardless, while the “miners” are stuck with rapidly depreciating equipment. It isolates the accounting risk from broader market movements.