Quick answer: Global AI spending is running into the trillions this year — the post puts it around $2.5 trillion — with much of it flowing into data centers and chips. The Bank of England has warned of a possible "sharp market correction"; Jeff Bezos has called it an "industrial bubble" — the kind that can still leave useful things behind when it pops. But the overlooked cost isn't the buildings, it's the silicon: hyperscalers are taking on hundreds of billions in debt to keep buying chips. Michael Burry's thesis is that these chips really last 2–3 years, not the 5–6 companies depreciate them over — an estimated ~$176 billion of understated depreciation between 2026 and 2028.
The shallow lesson: "it's all hype, get out" (or the mirror image, "it's fine, ignore it")
Both are vibes, not analysis. The story isn't the risk.
The real lesson: the risk lives in the depreciation schedule, not the headline
If a GPU's real useful life is 2–3 years but it's booked over 5–6, a lot of today's reported earnings are quietly borrowed from tomorrow. That's a math problem hiding inside an accounting choice — far more concrete than "is this a bubble."
The sobering part
An "industrial" bubble (Bezos's framing) can still leave real value behind — data centers, chips, inventions society keeps using. But the individual balance sheets built on optimistic depreciation are where the pain lands first. Build on economics you can defend if the useful life turns out to be short.
Full disclosure
Azati builds production AI systems, so our incentive is a healthy, durable AI market — and we'd rather clients plan around realistic hardware economics than headline capex. See our engineering work.
FAQ
Adapted from a LinkedIn post by Baryslau Yaravy, AI Engineering Practitioner at Azati – read the original. Sources: CNBC (Michael Burry; Bank of England); remarks by Jeff Bezos, 2025–2026.
