Vol. I · No. 001Tuesday, June 9, 2026Byline: Waggs
Token rate (frontier, per M):$1.10▼38% y/y
PJM power (avg):$58/MWh▲21% y/y
Hyperscaler capex (TTM):$412B
HBM lead time:41 wks
Who eats the loss on the $2 trillion compute buildout?
Mostly the equity holders of the hyperscalers — quietly, through depreciation.
The chips are being written off over six years while their economic life runs closer to three.
The gap is roughly $40B a year of earnings that exist on paper and nowhere else.
The loss has already happened; it just hasn't been booked.
7 min read · 3 exhibits · 3 signals
A railroad in 1873 and a GPU cluster in 2026 share one accounting problem:
the asset outlives the bond only if the traffic shows up. Everyone is arguing about whether the traffic
(AI demand) shows up. Almost no one is reading the depreciation footnotes, where the answer to
who pays if it doesn't is already typed out in 8-point font.
The framework for this issue: the scissors — the spread between the
stated life of an asset and its economic life. When the scissors open, earnings are borrowed
from the future. Every great capex bubble dies by scissors, not by demand.
Exhibit 1
The token rate collapses faster than any price in industrial history
Frontier-grade intelligence has fallen ~97% in three years. Your asset earns tokens; tokens are deflating; your depreciation schedule assumes they aren't.
Four companies, one bet — flick through the buildout
—
'26 AI capex
—
Implied 3-yr dep.
—
Earnings at risk
Stated depreciation (gold) vs. what a 3-year chip life would charge (dark blue). The white space between the bars is reported profit that may not exist.
The scissors, projected: when paper life meets real life
On a 6-year schedule the write-downs arrive in 2027–28 — precisely when the 2024–25 vintage of chips becomes uncompetitive. The bill and the obsolescence are synchronized.
Illustrative projection. Real issue: vintage-level capex model, auditor comment letters.
The other side — steelmanned
The bulls' best case is not "demand grows" — it's that old chips keep earning. An H100 at 70% utilization
running cheap inference still covers its marginal cost even three generations behind, the way a 20-year-old 737
still flies profitable routes. If utilization stays above ~60% across vintages, the six-year schedule is honest,
the scissors never open, and this issue is wrong. That is a real scenario. Watch utilization, not narrative.
Signals to watch
Useful-life disclosures
Any hyperscaler quietly shortening server depreciation from 6 years back toward 4 in a 10-Q footnote.
Trigger: first mover cuts ≥1 yr
Old-vintage rental rates
H100 hourly rates on the secondary clouds. The 737 thesis lives or dies here.
Trigger: <$1.00/hr sustained
Token rate vs. demand
If token prices fall 10x and inference revenue still grows >2x, deflation is being outrun. Bulls win.
Trigger: revenue growth < price decline
The coda
No CFO models it this way, but the buildout is not really priced off cash flows — it is priced off the fear of
watching a rival become a god while you held the dividend. That is mimetic desire wearing a hard hat. The railroads
were the same: the tracks were real, the traffic eventually came, and the men who laid them still died broke,
because wanting what your rival wants is not a discount rate. The chips will be useful. The question was never
the chips. It is who gets to be wrong with other people's money.