Hyperscalers are spreading GPU costs over more years on the books than the hardware actually lasts — not a phase they'll grow out of, but a lasting feature of this AI buildout cycle. A dual-life EVA framework for bounding the debate, and a staged positioning framework for reading what comes next.
Hyperscalers are spreading GPU costs over more years on the books than the hardware actually lasts — not a phase they'll grow out of, but a lasting feature of the AI buildout cycle they've entered. Three independent lines of evidence converge on that conclusion, and none can be dismissed as a personality-driven bear thesis.
First, the depreciation debate is no longer academic: Michael Burry's understated-depreciation estimate ($176B across 2026–2028), footnotebrief's parallel reconstruction (~$228B rounded to ~$200B), and Goldman Sachs' finding that a 5→3-year life shift raises implied annual compute depreciation by nearly $1 trillion (from ~$3T to nearly $4T) all triangulate the same order of magnitude. Second, the engineering reality — Meta's own Llama 3 study implying a ~9% annualized H100 failure rate, GPU rental rates that fell 64–75% from their 2024 peak, and Nvidia's confirmed annual architecture cadence — is inconsistent with the 5–6 year lives most hyperscalers book. Third, the physical bottlenecks that would let hyperscalers "cascade" old chips into productive second lives (power availability, data-center shell reusability) are 5–10 years behind the compute-deployment curve.
The investment implication is not that the hyperscalers are uninvestable — Microsoft, Alphabet, Amazon, and Meta remain among the highest-return franchises in market history, and their core businesses subsidize the AI build. The implication is that single-point ROIC estimates on the AI infrastructure increment are analytically inadequate, and that the marginal 2026 capex dollar is being deployed at a return that is far more scenario-dependent than the sell-side constructive frame admits. Readers new to the ROIC–WACC framework will find a plain-terms primer below.
A single parameter — the assumed useful life of a GPU — swings the ecosystem depreciation burden by hundreds of billions to more than a trillion dollars. That is the size of the analytical gap this report exists to bound.
