The AI bubble debate tends to get emotional. I think we should look at individual metrics more objectively instead. One useful example is GPU depreciation periods.

A useful life assumption is not a fact. It is just management's estimate. The longer that estimate, the lower this year's depreciation expense. That makes operating profit look higher than it might otherwise.

In November 2025, investor Michael Burry brought this issue into the public eye. He argued that hyperscalers depreciate their Nvidia chips over four to six years, even though the real technology cycle is closer to two or three years. He called this an artificial boost to earnings. He then took short positions against Nvidia and Palantir.

My starting point is a bit different from his. I do not treat a long useful life as a problem by itself. What matters more is whether that assumption is backed by real evidence. A longer estimate can also reflect genuine operational efficiency. Burry focuses on the idea that profits are already inflated today. What I want to focus on instead is the quality of the balance sheet. If the accounting estimate turns out to be longer than the real economic life, that gap will eventually show up as an impairment. When that happens, it hits earnings and the balance sheet at the same time, and it can also drag down how the market values the company. Burry's conclusion is to act now, through a short position. My conclusion is more modest. I think this is something worth watching closely over time.

I want to apply this same lens to Japanese stocks.

Sakura Internet (3778) and SoftBank (9434) both run AI businesses on the latest Nvidia GPUs, including the H100, B200, and GB200. That puts them on the same technological footing as the US hyperscalers. But when I checked their securities reports, neither company discloses a GPU specific useful life on its own.

Sakura Internet mostly follows the standard tax based useful life. SoftBank discloses a wide range, something like two to twenty years, without breaking out GPUs specifically. Neither shows the kind of clear, numeric changes that US companies disclose. In short, there simply is not enough information to judge this properly yet.

That is exactly why I think it matters to keep watching. Not just the headline numbers in earnings reports, but the footnotes on useful life assumptions, and how they compare across companies over time.