The depreciation & duration-mismatch trap: useful-life as the fifth self-marked number

Compiled 2026-06-10, triggered by the justdario.com piece "The Unhappy Ending of the Whole AI Dream." Formalized in models/z3/depreciation_trap.py. Overlay claims graded fact | contested | weak | unsupported; the Z3 result is a proof. Cross-refs spec-sec-filings-primary, macro-bank-htm-marks, reflexive_marks, self_marked_value, spec-unwind-timing.

The article's checkable claims verify against primary data, and it supplies a mechanism this project had not yet formalized: AI compute depreciates faster (~2–3 yr) than the debt and leases financing it (5–19 yr). So even if the promised revenue arrives, the asset is gone before the loan is repaid — and the "useful life" assumption that hides this is the same defect as the machine-proven self-marked-value theorem, applied to depreciation. Useful life is the fifth self-marked number.

What the article argues (and how it grades)

justdario's thesis: the US AI buildout is debt-financed, fast-depreciating, winner-take-all capital waste with no graceful exit — abandoning triggers a Meta-metaverse-style stock crash, continuing demands perpetual capital, and (unlike Meta's cash-funded metaverse) the debt impairs the balance sheet and legacy operations when it unwinds; QE won't rescue it because debt service eats the cash flow. This is strongly aligned with the project's machine-proven circular core + insolvency-at-zero-inflow, and adds the depreciation and asset-light→asset-heavy angles (fact: the alignment). The "inevitable collapse" framing is the author's — this project proves structure, not date (spec-unwind-timing).

Verified primary claims

Burry's depreciation critique (fact, Nov 2025). Hyperscalers (Meta, Amazon, Microsoft, Google, Oracle) depreciate Nvidia GPUs over 5–6 years when the true economic life is closer to 2–3 years — ~$176B of understated depreciation / overstated profit 2026–2028, ~$50–60B/yr if the true life is 3 not 6. Concrete instances: Meta raised most server/network useful life to 5.5 yr, cutting depreciation ~$2.3B over 9 months of 2025; Microsoft's ~$17B GPU purchases booked over 6 yr instead of 3 → ~$2.9B/yr earnings overstatement. The consensus read: "probably not fraud, but almost certainly optimistic estimates that will require adjustment" — i.e., useful life is a discretionary assumption (that part is fact; the $176B is Burry's estimate, contested).

Oracle's asset-light → asset-heavy pivot (fact, FY2026 8-K). Free cash flow negative ~$23.7B, ~$50B capex, >$108B debt (plus $30B raised early 2026 in IG bonds + mandatory convertible preferred). The $523B RPO rests >half on one customer (OpenAI via the ~$300B Stargate deal); $75B of the big AI contracts are customer-prepaid or customer-supplied GPUs. And $248B of additional datacenter/cloud leases of 15–19-yr term, substantially off the balance sheet — long-tenor financing against ~3-yr-life GPUs. CNBC's framing: "Oracle is building yesterday's data centers with tomorrow's debt."

Geopolitical color (weaker). China's ~$295bn state-directed AI buildout (80%+ domestic semiconductors) — roughly one year of Google's spend (contested, single-source figure). US operational datacenters reportedly exceed all other countries combined / ~10× Germany (weak, magnitude unverified here).

The cluster's unified useful-life table + forward rate (added 2026-06-15, #54)

A per-company audit of the disclosed server/GPU useful life and its earnings effect:

• The Amazon reversal is the tell. Amazon is the only hyperscaler to publicly cut a useful-life category (6→5 yr, Jan 2025), its filing citing "an increased pace of technology development, particularly in … artificial intelligence and machine learning" — an inside-the-filings admission that the extend-life / harvest-profit trend is reversing.
• The CoreWeave-vs-Nebius natural experiment. Identical neocloud business models, GPUs over 6 yr (CoreWeave) vs 4 yr (Nebius) — a ~50% depreciation-rate gap from the useful-life choice alone; at 4 yr CoreWeave's reported operating profit flips negative. The cleanest demonstration that useful-life manufactures profitability.
• Burry/Scion (Nov 2025), graded. ~$176B understated depreciation 2026-2028; Oracle earnings overstated ~26.9%, Meta ~20.8% by 2028; ~$1.1B notional Nvidia+Palantir puts (later wound down, fund closed). Nvidia's rebuttal (Kress: six-year-old A100s "still running at full utilization") vs Burry (utilization ≠ value; older chips far less power-efficient → continued use reflects scarcity, not retained value). All changes are disclosed estimate changes (prospective) — "fraud" is rhetorical/contested, not adjudicated.
• The unified forward rate. The cluster sits at a ~6-yr accounting life for servers/GPUs while a defensible blended GPU economic life is ~3–4 yr (training-edge ~1.5–2 yr; inference tail ~3–4 yr) against Nvidia's ~annual cadence (Hopper→Blackwell→Rubin). Marking GPU-heavy fleets from ~6 yr to ~3–4 yr raises annual depreciation ~50–100% on the affected base. Key tension: the accounting life is set on blended fleets (incl. 15+-yr data-center shell/power); the GPUs are the fast-decaying subset the blended life masks — exactly the D2/D4 mechanism below.

The mechanism, formalized (models/z3/depreciation_trap.py)

• D1 (SAT) — a long assumed life makes the same asset print a profit: profit = revenue − opex − capex/life. The life is the free parameter.
• D2 (UNSAT) — the true, shorter economic life strictly lowers profit; the overstatement is structural, = capex·(1/Lₜᵣᵤₑ − 1/Lᵦₒₒₖ) ≈ $2.9B/yr for one MSFT-like GPU tranche, ~$176B industry-wide 2026–2028.
• D3 (UNSAT) — the duration mismatch: when asset life (~3-yr GPUs) < financing tenor (5–7-yr bonds, 15–19-yr leases), equity cannot stay whole — there is an interval with a live debt and a dead asset (Oracle's $248B of 15–19-yr leases).
• D4 (UNSAT) — depreciation is only timing: no life choice avoids both near-term losses and a retirement writedown; a stretched life just defers the loss and forces a catch-up when the asset is retired.

Why this matters to the map

Useful life joins HTM cost, AI fair-value marks, private-credit NAVs, and insurance captive marks as a chosen number held above realizable until a forcing event (self_marked_value U1–U4) — here the event is retirement / obsolescence, not a deposit run or an IPO. It is the cleanest answer to "but the AI firms have real assets backing the debt": the real assets age out faster than the debt amortizes, and the accounting that makes today's earnings look positive is borrowing those earnings from a future writedown. That is the formal spine under justdario's asset-heavy / no-graceful-exit thesis.

← Research index · structured data: fin-ai-depreciation-debttrap.json · fin-ai-depreciation-debttrap.md