Anbarasu: Nalla, what the hell is happening with AI stocks?! A few weeks ago, everything looked fine. Now every stock I see is down 10 per cent, 20 per cent…
Nallasivam: Try worse than that! From their 52-week highs, CoreWeave is down 60 per cent, Super Micro Computer about 50 per cent, Oracle some 40 per cent and how can I forget SoftBank, which is investing heavily in OpenAI? Its shares are down about 25 per cent.
Anbarasu: What’s causing this bloodbath? It’s not like these companies suddenly did something terrible together.
Nallasivam: Exactly – same companies, same business models. But recent events have flipped market sentiment upside down.
Anbarasu: Like what?
Nallasivam: Many bankers and analysts have become sceptical as to whether these stocks would make money for investors that justify their massive capex in high-end chips and data centres. This while valuations were through the roof. CoreWeave for one, was valued at about $90 billion a few months ago. This for a loss-making company that made around $2 billion in revenue in 2024. Look at the price-to-sales multiple on that!
Anbarasu: It’s bonkers!
Nallasivam: Something even more interesting happened this week. Have you watched ‘The Big Short’?
Anbarasu: Of course, I have.
Nallasivam: Then I assume you know who Michael Burry is. He posted on X that hyperscalers such as Amazon, Microsoft and Oracle are inflating profit by understating depreciation on chips and servers.
Anbarasu: How so?
Nallasivam: Burry argues these companies depreciate chips over five to six years of useful life, when two to three would be more realistic. By stretching it out, he claims hyperscalers will collectively understate depreciation by $176 billion from 2026-28.
Anbarasu: Wow! Aren’t there regulations against this?
Nallasivam: Well legally, companies are in the clear here. Under US GAAP, useful life is an accounting estimate left to the judgment of companies – unlike India, where the Companies Act prescribes it for each asset category.
Anbarasu: So companies can just pick whatever they want? And these estimates vary between companies?
Nallasivam: Companies are mandated to disclose useful life assumptions in financial statements, which keeps them in check. They should be fine as long as their judgement is reasonable. But yes, it creates inconsistency. Alphabet depreciates servers over six years; Nebius depreciates GPUs over four.
Anbarasu: Hold on. Are these chips really that fragile? They don’t just stop working after three years, do they?
Nallasivam: Of course, they’ll keep functioning, but that’s not Burry’s point. He opines that given the rapid advancement in tech and hence shorter product life cycles, the older chips are prone to a great deal of obsolescence. For example, in 2024, Nvidia announced that it will design new AI chips every year, instead of every two years. These older chips might not earn top dollar, forcing hyperscalers to constantly upgrade, in order to stay competitive.
Anbarasu: Just like how my iPhone 16 lost value when the 17 was launched.
Nallasivam: That’s right. But imagine the 17 to have 2x, 3x performance upgrade over the 16. Here’s a telling example.
Amazon increased the useful life of servers from five to six years in January 2024, boosting earnings by $2.5 billion for 2024. However, from January 2025, it backtracked on certain servers, reverting to five years. It even accelerated depreciation on some as they were retired from use due to ‘increased pace of technology development’. Amazon expects both these to cost the company $1.3 billion of operating income in 2025.
Depreciation is just a non-cash expense until it is not. The void left by such retired servers might need to be replaced, costing actual cash and draining free cash flows.
Anbarasu: So Burry thinks such retirements will become the norm?
Nallasivam: It might, it might not. But he’s certainly given us food for thought. Would be worth watching how this plays out.
Published on November 15, 2025
