The desk’s shelf on one screen. Pick a piece on the rail and it reads right
here — signatures, teardowns, working notes, and the discipline pages — every one
keeping its own citable page and PDF where one exists.
source: the research manifest · receipts ledger · every figure traces to a filing or is labeled
The 2026 AI-compute cycle is usually argued on valuation. The more decision-relevant question is structural fragility — and it is legible in the data these companies file themselves. This brief reads the six fragility indicators straight from the filings.
The dataset
Filing-sourced and reproducible
Six filing-sourced indicator tables sit underneath this brief, each computed from the source filings: an accounting table of useful-life changes, a capex-versus-demand table, the insider Form 4 record, the financing graph of the compute complex, the disclosed energy commitments, and a ground-truth deterioration series. The brief is built to be reproducible — every figure derives from those filing-sourced tables.
Download the tables · CSV, primary-source notes inline
Each table carries its 10-K / Form 4 accession numbers inline, and shows blanks (NOT_SOURCED) rather than imputing. The circular-financing edge ledger (indicator 04) is published above — revised 2026-07-02 with the Amazon–OpenAI equity legs and a funded_usd column; the energy indicator (05) rests on qualitative disclosures and is discussed in its section. Cite as: The Desk, “AI Fragility” dataset (2026), /brief.
One discipline runs through all of it: where a value cannot be sourced cleanly from a filing, it is shown blank rather than imputed. The point is to read the cycle in the numbers the companies publish themselves, not in estimates layered on top of them.
Indicator 01 · Depreciation integrity
Are the assets aging faster than the books admit?
Has a firm extended the useful life of its depreciable assets — converting paper income without a matching dollar of cash? A life shortened scores zero, regardless of size.
The first indicator asks a narrow accounting question with a wide reach. When a firm extends the useful life of its servers, the same hardware cost is spread over more years, annual depreciation falls, and reported operating income rises — on paper alone, with no extra cash, no new customer.
# §2.1 — paper benefit from a useful-life extension delta_dna = ppe_depreciable × ( 1/life_old − 1/life_new ) # hard rule: a life shortened scores 0, regardless of size
The direction of travel is uniform: every firm that touched a useful life lengthened it, and four did so while running the largest AI-capex programs on record. Amazon is the control — it moved the same lever the other way, six years to five, and absorbed a $1.4B charge against income, which is why it scores zero here despite carrying the heaviest depreciation line ($41.86B) in the set.
The signal is not the size of depreciation; it is the choice to make it smaller while everyone's assets are aging faster.
Indicator 02 · Capex vs demand gap
Is the spending outrunning the demand?
Is AI capital spending outpacing the revenue that would justify it? The break-even hurdle is set generously, so the firm gets credit for all segment revenue, not just AI lines.
# §2.2 — required revenue per $1 of capex per year factor = ( CoC + 1/L ) / m = ( 0.10 + 1/6 ) / 0.30 = 0.889 # fail when FY2025 segment revenue < capex × 0.889
One firm fails the break-even test on full segment revenue: Alphabet, where Google Cloud's $58.7B sits $22.6B below the $81.3B the capex requires — a 28% shortfall. Capex is also growing roughly 2–4× faster than the revenue lines it funds across the cohort, even where the level test still clears.
Firm
Capex / revenue growth
Meta
3.95×
Amazon
3.25×
Alphabet
2.07×
At the system level the aggregate gap widens from $78B to $90B over four quarters. Spending is being committed ahead of the demand — and the test is built to flatter the firms, not to indict them.
Indicator 03 · Insider selling intensity
What are the people who know most actually doing?
Two kinds of insider selling look identical on a tape and mean opposite things. Pre-scheduled 10b5-1 plan sales score low; the signal is discretionary selling — a sale an officer chose to make, in a window when they held material non-public information, with no 10b5-1 footnote on the Form 4.
The three compute leaders divide cleanly. The discretionary cluster — not the headline dollar — is what scores, which is why the largest sellers by dollar (both on 10b5-1 plans) are discounted while smaller discretionary clusters rate higher.
Firm
Discretionary
10b5-1 plan
Largest single seller
NVDA
$0.93B
$1.57B
Dir. Mark Stevens $802M discretionary
AVGO
$0.50B
—
Co-founder Samueli $749M plan
AMD
$0.02B
$0.29B
CEO Su plan
NVDA's $0.93B discretionary is led by director Mark Stevens at $802M with no detected plan, against $1.57B run through confirmed 10b5-1 plans — including CEO Huang's $1.05B, under 1% of his stake. AVGO's $0.50B discretionary is spread across the entire C-suite — CEO Tan, the CLO, the CFO, and two more officers, none with a detected plan. AMD is the quiet one.
Discretionary selling is not a one-quarter event. The universe-level Form 4 total rises every quarter across the window — from $0.85B in 2025Q3 to $1.10B in 2026Q2, a 29% increase — while the same names were guiding investors toward accelerating AI demand.
Indicator 04 · Circular financing
Is the money going in a circle?
The structure is a loop: an investor funds a lab, the lab commits to buy compute from the investor's cloud, that cloud revenue underwrites the investor's capex, and the capex buys the investor's own chips through the lab it funded.
The financing graph of the AI-compute complex is a directed multigraph over twelve principals and four edge types — invests · buys_compute · supplies · marks_up. The recycling ratio measures the loop's leverage: compute committed out of the core labs (OpenAI, Anthropic, xAI) divided by equity put in, across three provenance tiers.
The same dollar of disclosed equity supports roughly 15.5× committed compute on a funded-cash basis (revised 2026-07-02 from 26× — Amazon's Q1 2026 $15B funded OpenAI stake widened the equity base), easing to ~3.6× only when every reported secondary round is admitted as equity. Present-valued at 10% over each commitment's disclosed horizon, the funded-cash ratio is about 13× — nearer 11× if the undated Microsoft commitment is discounted over a typical cloud term. Provenance, not arithmetic, moves the number; stock or flow, discounted or not, the loop turns far above any arm's-length benchmark.
Recycling ratio by equity tier — funded cash → filed → +reported → PV-adjusted.
Two destinations carry the loop: of the labs' committed compute — the same $540B universe as the ratio — Microsoft and Amazon receive 96% (98% on the filing-grade subset). Mark-to-model gains booked on those same customer stakes total $18.2B (Microsoft +$5.9B — primarily the OpenAI recapitalization dilution gain — Amazon +$12.3B) — earnings recognized on the appreciation of the firms one funds. Eight directed cycles run through the cash-flow subgraph, and the largest single commitment — Nvidia's $6.3B backstop to CoreWeave — surfaced only in a September 2025 8-K (accession 0001769628), absent from the March 2025 IPO prospectus that first sold the relationship.
Indicator 05 · Energy & diminishing returns
Are physical limits starting to bind?
Are power, cooling, and chip economics beginning to cap capability gains? This is the thinnest-data indicator in the framework and carries the lowest weight (0.10) — we will not present estimate as measurement.
The firm-level cost-per-capability curve is largely proprietary, so this indicator does not try to measure it. What the filings do record, unambiguously, is the scale of power being committed — the appearance of gigawatt-scale capacity figures inside the same compute-purchase agreements that drive the circular-financing loop. The build stops being denominated in dollars and starts being denominated in power.
Power commitment
Capacity
Provenance
OpenAI → AMD
6 GW
Filing 8-K EX-99.1, 2025-10-06
Anthropic → Amazon
5 GW
Media not yet filed
Anthropic → Google
>1 GW
Media not yet filed
Three edges carry an explicit gigawatt figure — 12 GW in aggregate — but exactly one is filing-sourced. By the methodology's own rule, that single filing item is the floor under any elevated read: the indicator is directionally supportive, not independently load-bearing, and is flagged as such. The cost-per-capability curve that would let it stand on its own is deferred to Phase 2.
Indicator 06 · Organic end-user demand
Is the revenue real, or recycled?
Does reported AI revenue reflect genuine paid adoption by independent end-users — or is it recycled through the same ecosystem that funds the build, or rebranded from existing product lines?
The test is anchored on the MIT NANDA finding that roughly 95% of enterprise GenAI pilots show no measurable P&L impact (Fortune, August 2025). Headline growth in the 30–50%+ band scores well only when paired with demonstrated paid retention and pilot-to-production conversion above 50%; growth sourced from ecosystem participants scores worse, not better. The indicator scores the source of the growth, not its rate.
Revenue growth alone clears the headline band for most of the complex — CoreWeave at 168%, Broadcom at 64%, four firms clustered at 32–36%. CoreWeave is the limiting case: 67% of its FY2025 revenue is a single counterparty — Microsoft, "Customer A" in its 10-K — with the remainder committed by OpenAI, Meta, and Nvidia. Every named buyer is an investor in, or a lab funded by, the same circular structure.
That is growth from ecosystem participants rather than demonstrated independent end-user retention — the band the rubric reserves for recycled demand, and exactly what the NANDA anchor predicts: an "AI revenue" label growing fastest where the demand is most recycled, not where paid conversion is most proven.
The synthesis · Divergence gauge
The tape versus the filings
D(t) = M(t) − G(t) sets a market signal against a ground-truth signal. The market term M(t) is the equal-weight mean of three full-window z-scored components of SOXX price behaviour — 63-day momentum, price-to-trend overextension, and 20-day annualized instability. The ground-truth term G(t) is the negative mean of three deterioration z-scores — AI-layoff share, discretionary insider selling, and the capex gap. The gauge widens when momentum and overextension climb while the fundamentals erode.
Toggle between the composite (M, G, D) and the three ground-truth signals underneath G(t). Source: SOXX + ground-truth series.
Through 2025Q1 the two signals track close and D(t) sits below zero — price had not yet detached from fundamentals. In 2026Q2 the gap inverts hard: M(t) jumps to +2.83 as SOXX closes at 639.45 (63-day momentum +88.0%, instability +0.74 annualized) while G(t) falls to −1.23, dragged by the AI-layoff share and discretionary insider selling both reaching their window highs.
D(t) widens from −1.80 to +4.06, a +5.86 swing — the strongest move in this four-quarter series so far (n=4: descriptive, not a long-run signal).
Method & limitations
What would prove this wrong
This brief is built to be reproducible: every figure derives only from filing-sourced inputs. Each indicator is computed only from filing-sourced inputs; where a value cannot be sourced cleanly it is shown blank rather than imputed.
Two Phase-1 simplifications are stated plainly. The divergence gauge standardizes its components over the full window — it is descriptive, not real-time: it carries look-ahead bias and is not a tradeable signal, and an expanding-window version is deferred. It also weights its three market components equally; empirical calibration is future work. Indicator 05 (energy) rests on the thinnest data in the set and is weighted accordingly — directionally supportive, not independently load-bearing.
The falsifier is built in: if the ground-truth signal turns back up — demand converting, the capex gap closing, insider selling normalizing — the divergence closes and the boom earns its price. We publish the number either way.
Committed AI compute divided by outside funded cash — the desk's flagship measurement of circular financing, from primary filings. 15.5× est., 2026Q2.
The recycling ratio measures how much committed AI compute the core labs’ disclosed outside funding actually supports. The numerator is disclosed compute commitments — ~$539B in total, of which ~$395B walks to a primary SEC filing accession; the remaining ~$145B is corroborated from company disclosures and reporting and carries a lower confidence label in the ledger. On the strictly primary-filed numerator the ratio reads ~11.3×. The denominator is outside cash actually funded — not announced, not committed, funded. When the ratio is high, the build-out is financing itself in a loop: vendors funding customers whose commitments come back as the vendors’ own backlog.
The figure is labeled an estimate because the denominator depends on which instruments count as arm’s-length outside funding — the papers walk that judgment edge by edge, every accession number cited.
Not “is it a bubble” — whether the math closes. The full argument, torn down to its arithmetic: the spending is enormous and immediate, the payoff uncertain and lagged, and the race between them decides the repricing.
01 · The question
Not "is it a bubble." Whether the math closes.
Michael Burry's 2008 call was never that houses were expensive. It was the
math of the financing structure underneath them. The AI build-out deserves the
same treatment, and the same refusal to argue about vibes. We are not bears shouting "bubble." We
do the arithmetic the consensus is skipping, and we report what it says.
The arithmetic resolves to a single tension. The spending is enormous and immediate.
The payoff is uncertain and lagged. So the question becomes a race: does the
productivity arrive before the financing that funded it has to be repriced?
02 · Two clocks
The financing runway and the productivity lag.
Clock one — the runway. Can the spending survive? This is capital
expenditure versus demand, circular financing, and depreciation that hasn't been honestly marked.
The arithmetic of whether the capex is paid for by revenue that actually exists.
Clock two — the lag. Will the payoff arrive? Solow's computers took
the better part of a decade to show up in the aggregate productivity statistics — "everywhere
except the numbers." AI may earn its own lag-then-boom. Or the lag may outlast the money.
The bull case is a lag-then-boom. The fragility case is capex repriced before the
payoff ever lands. Others time one clock or the other; we hold both on a single scoreboard — with explicit falsifiers, revised in public.
03 · Clock one — the money carousel
The money circles home.
Capital recirculates among a tight loop — Nvidia, the hyperscalers, OpenAI, Oracle,
CoreWeave, Anthropic — where vendors finance the customers who buy their product. When a vendor
finances its customer to buy the vendor's product, revenue and demand stop meaning what they appear
to mean. The growth is real; what it measures is the open question.
Indicator I4 · circular leverage
04 · Clock one — the unmarked cost
The depreciation illusion.
Amazon shortened a subset of servers and networking from six years
to five, effective Jan 1 2025, citing AI/ML obsolescence verbatim — a change that raised 2025
depreciation by ~$1.4B (FY2025 10-K). Separately, it took ~$920M of
one-time accelerated depreciation retiring AI-exposed gear early in Q4 2024 (FY2024 10-K). That reversal is the canary:
the first time a hyperscaler names AI to shorten asset lives, and both charges sit in filed 10-Ks.
Meta extended useful life to about 5.5 years the same month — near
−$2.9B of FY2025 depreciation, which flatters earnings. Microsoft,
Google and Meta extended useful lives and held them; Amazon alone reversed, shortening for
AI-exposed gear — the contrarian tell.
Stretch the assumed life and today's earnings look better than the cash can support.
Michael Burry (Scion) estimates the understated depreciation near
$176B across 2026–2028 — enough to flatter the very earnings base
that justifies the capex. We carry it as his allegation: a disclosed short-seller's projection,
not an audited figure. The canary, by contrast, is filed. Naming the difference is the point.
Amazon FY2025 10-K Note 1 · Meta FY2025 · $176B — Burry/Scion allegation, Nov 2025
05 · Clock two — the demand problem
Record shipments, unproven payback.
~95%
enterprise GenAI pilots not yet at measurable P&L (MIT NANDA, Aug 2025)
≥30%
GenAI projects abandoned post-PoC by end-2025 (Gartner)
The infrastructure layer is booking real, record revenue. The end-user layer, where the
investment must ultimately be repaid, is not yet showing the productivity that would justify it.
That is exactly the shape of a lag — and a lag is survivable only for as long as the financing
holds. This is where clock two meets clock one.
MIT Project NANDA · Gartner
06 · Convergence
One red light is noise. Four is a structure.
We score six indicators — depreciation integrity, capex-vs-demand, insider selling,
circular financing, energy, and organic demand. Any one can be explained away. The signal is
corroboration: when several light at once, the explanations have to start
contradicting each other. They share a common driver (capex ahead of monetization), so we weight
their agreement for that overlap rather than as fully independent votes. On the compute bellwether,
four of six are elevated — not a coincidence you can narrate around.
The discipline cuts both ways. We deliberately hold two indicators — insider selling
and energy — as contained, because forcing them red to fit the story would forfeit the
only thing that makes the rest worth reading.
07 · The fork
Repriced, or propped.
Fast repricing
Demand fails to close the gap in time, depreciation is
marked honestly, and the financing structure reprices quickly. Painful, visible, and over
relatively fast — the market clears.
Administratively propped
The structure is held up — sovereign AI
subsidies, national-security compute mandates, too-big-to-fail support. The cost is not erased.
It is relocated, and stretched over a longer, quieter horizon.
Propping a structure never deletes the cost; it moves it. The
operational tell is the transition itself — the moment market repricing gives way to administrative
support. Watching for that handoff is part of the standing read.
08 · Falsifiers
What would prove us wrong.
A thesis that can't be killed isn't research, it's faith. Each pillar has an explicit
exit. If these print, we revise — in public.
Demand closes the gap. The MIT no-measurable-ROI rate falls below 60% for two
consecutive quarters. The lag was real and the payoff arrived.
Depreciation honesty spreads. Hyperscalers converge on shorter, marked useful lives and
absorb the hit. The earnings base stops being flattered; the illusion self-corrects.
The financing de-circularizes. External, end-customer revenue replaces vendor financing
as the marginal dollar. Then the demand is genuine and the carousel was a phase, not a flaw.
The spending is certain. The payoff is lagged. Everything
hangs on which clock finishes first.
This is a living document. Verified figures trace to primary
filings; attributed estimates are named as such; unverified figures are flagged, not dressed up.
As the data moves, so does the read.
Two clocks, one race: does the AI productivity payoff arrive before the build-out's financing has to be repriced? The combined scoreboard of the lag and the runway.
——
now · 2026
Financing runway
payoff
Financing runway (reprices ~2028)Productivity payoff landsThe fragility gap
Scenario, not forecast. The runway window (~2027–2029) is where hyperscaler free cash flow turns negative and the AI debt issued in 2025–26 comes up for refinancing. The payoff range (2027 fast → 2037 slow) is the Lag Index fork. Drag to test any arrival date; the verdict is the gap between them.
~6–10 q
Financing runway
before repricing (~2027–29)
2027
Fast payoff (bull)
lands inside the runway
2037
Slow payoff (bear)
decade after repricing
The gap
What we measure
runway-end → payoff
How to read it
The bull case and the fragility case are the same picture
Almost every AI-cycle argument is really a claim about timing. The optimist says the payoff is near; the pessimist says the bill comes first. Put both on one axis and the disagreement becomes measurable: does the productivity payoff land before the financing has to be repriced? If it does, the long depreciation schedules and the debt were justified — a supercycle. If it doesn't, the capex reprices into a payoff that hasn't arrived — the fragility case, and the shaded gap is its size.
Clock 1 · the runway
How long the money lasts
Capex vs demand, circular financing, depreciation pulled forward, debt issued against returns that must show up on schedule. Instrumented on Capex Watch and Stranded Compute.
Clock 2 · the payoff
When the lag closes
Adoption is ~78%; realized productivity is ~5% of potential. The J-curve payoff arrives in 2027 — or 2037. Instrumented on The Lag Index.
Research discipline · what would settle the race
The falsifiers
Payoff side: AI-attributable TFP turning up in the official series, broadening past a few industries — the runner gains.
Runway side: utilization climbing through 70%, or AI revenue growth overtaking capex growth — the runway lengthens, repricing recedes.
Against the bull: hyperscaler free cash flow turning negative on schedule and debt refinancing at higher rates — the runway shortens.
Against the bear: training silicon repurposing cleanly for inference, extending real asset life toward the booked schedule.
We track both clocks and revise in public when a falsifier moves.
Method & sources
The Race-o-meter is an interactive framing of two sourced instruments, not a forecast; the runway window and payoff range are scenarios you can move. Payoff fork (2027/2037) & the Lag Index: our Lag Index (Slok / Apollo; Kansas City Fed; BLS). Runway / financing pressure: Stranded Compute & Capex Watch (hyperscaler FCF and debt issuance via industry reporting; filings).
Not investment advice. The Desk is a research institution; nothing here is a recommendation to buy or sell any security.
The productivity lag as a number: AI is deployed everywhere and visible almost nowhere in the output statistics. The Lag Index tracks the gap — and the race against the financing runway.
5
Lag Index · 0–100
Deep in the trough. Inputs are nearly fully deployed; payoff has barely begun to register. The index is the share of the technology's productivity potential that has actually shown up in the aggregate statistics — and it is still in the single digits.
Inputs deployed
78%
Payoff realized
~5%
The lag (gap)
73pp
Lag Index = realized AI-attributable productivity as a share of full-diffusion potential. Inputs = enterprise adoption. Payoff = AI-attributable TFP (~0.07pp/yr) against a full-GPT-payoff benchmark (~1.5pp/yr, the late-1990s IT surge). Illustrative calibration; see sources.
~78%
Enterprise adoption
AI deployed somewhere
~90%
Report zero impact
on output in 3 yrs
+0.07pp
AI-attributable TFP
/yr · vs +40% self-reported
≥3 yr
Min. implementation delay
adoption → measured output
The shape · adoption vs payoff
Deployed everywhere, visible nowhere — yet
Plot the two curves on one scale — each as a share of full potential — and the lag is the white space between them. Adoption has raced to roughly 78%. Realized payoff has crawled to about 5%. Every prior general-purpose technology opened a gap like this; the question is only how long it stays open. Economist Torsten Slok frames the fork bluntly: the J-curve payoff could arrive in 2027 — or 2037.
Adoption and realized payoff as % of full-diffusion potential, 2018–2030. Solid = observed; dashed = the two scenarios from the trough. The shaded band is the lag. Adoption: enterprise surveys. Payoff: AI-attributable TFP vs the IT-surge benchmark. Projections are scenarios, not forecasts.
Calibration · the historical lags
Every engine waited
Electricity
~30 years
A three-decade productivity pause while factories redesigned around the motor — then manufacturing TFP ran +5%/yr through the 1920s.
IT / computers
~10 years
Solow's 1987 paradox — "computers everywhere except the statistics" — gave way to the 1995–2004 productivity surge.
AI
Year ~3
Adoption faster than either predecessor; payoff not yet in the data. Whether the lag is IT-short or electricity-long is the whole question.
Why the index matters: this is one of the two clocks we time. The Lag Index measures when the payoff arrives. Stranded Compute and Capex Watch measure how long the financing lasts. The fragility case is simply this: a lag that runs electricity-long against financing built for an IT-short wait. The bull case is the mirror — payoff by 2027, before the capex reprices.
Research discipline · what would move the index
The falsifiers
The index is built to move with the evidence. It climbs — and the bull case strengthens — if:
AI-attributable TFP turns up in the official series — the BLS/Fed aggregates, not self-reported surveys.
The gains broaden past a few industries into the whole nonfarm business sector.
Average usage deepens beyond the ~1.5 hours a week executives report today — the J-curve needs workflow redesign, not logins.
The implementation delay compresses below the three-year floor seen in firm-level data.
We track these on Capex Watch; when the official statistics move, the index moves with them.
Method & sources
The Lag Index is an illustrative calibration of published figures, not a proprietary econometric model; it expresses realized AI-attributable productivity as a share of full-diffusion potential. Figures are current as of mid-2026 and will move.
Adoption, "zero impact," and exec-usage figures: CEO survey via Fortune. AI-attributable TFP & the micro-macro gap: Kansas City Fed, BLS. The J-curve & "2027 or 2037": Torsten Slok / Apollo via industry reporting. Historical GPT lags (electricity, IT): general-purpose-technology literature (David; Brynjolfsson et al.). Cross-referenced with our own Capex Watch.
Not investment advice. The Desk is a research institution; nothing here is a recommendation to buy or sell any security.
One picture for the whole build-out: the forces of AI scaling, each on a single comparable log scale, indexed to their 2020 baseline. Three race upward — compute, context, and power. The cost to run a fixed-capability token is the one line collapsing.
Indexed to each series' baseline (≈2020 = 1×), log scale. The efficiencies overlays the compounding denominators; Absolute · units shows each force in its own real units. Solid = observed; dashed = scenario at the rates set below. Hover any point for the raw figure. Sourced anchors at the foot.
What-if · drag to bend the curves to 2030
What it says
The build accelerates; the unit cost caves
~5×/yr
Training compute
doubling ~every 5 months
~1000×
Context window
2020 → 2024
100MW → 2GW
Cluster power
per frontier site
~10×/yr
Inference cost ↓
cheaper, "LLMflation"
Read the engine first. Frontier training compute has grown about 5× a year since 2020 — doubling roughly every five months, an order of magnitude faster than the two-year cadence that defined classic Moore's Law. Context windows and cluster power ride the same exponential: a model's working memory went from 2,048 tokens to a million in four years, and a frontier training site went from a 100-megawatt "titan" to gigawatt campuses that draw as much power as a city.
Now read the counter-curve. The cost to run a model of fixed capability has fallen about 10× every year — a GPT-3-class model went from $60 to $0.06 per million tokens in three years. That is the scissors: the price of building the frontier explodes while the price of using it collapses. Switch to "The scissors" to watch the two curves cross, "The efficiencies" for the compounding denominators that make it possible, or "Absolute · units" for each force in its own real numbers.
The efficiency frontier
Why the build stays affordable
The cost collapse isn't magic — it rests on two compounding efficiencies underneath the price. Algorithmic efficiency: per Epoch AI, the compute needed to reach a fixed capability has halved roughly every eight months — about 3× a year. Hardware efficiency: the leading AI accelerators deliver about 34% more FLOP per watt each year. Stack those beneath the falling token price and you get the engine of "LLMflation" — the same intelligence, cheaper, every year. The "The efficiencies" view overlays all three rising denominators: smarter algorithms, more efficient silicon, and more tokens per dollar.
The fourth vector
Where scaling meets the workforce
The physical vectors are measurable. The fourth — adoption and labour — is where the scaling curves land in the real economy, and it resists a clean log line. So we mark it as the event it is: the capex pivot.
The crossover (2025–26): hyperscaler AI capital spending cleared $600B+ a year while 2026 tech layoffs neared 150,000 — spend surging as headcount falls. The build-out's bill is being paid, in part, by the workforce it is reorganizing. See it on Capex Watch and Layoffs.
Method & sources
Each series is indexed to its ~2020 baseline and drawn on a log scale so growth rates are directly comparable; the "Absolute · units" view shows the same data in native units on per-panel log axes. Hover a point for the raw figure. Anchors are rounded to convey orders of magnitude, not false precision; the dashed segments past 2026 are scenario extrapolations at the rates you set in the what-if panel, not forecasts.
Compute & efficiency trends (AlexNet→GPT-3→PaLM→GPT-4→Llama 3.1→frontier): Epoch AI. Context windows: model documentation (OpenAI, Google). Cluster power (Colossus 150MW→1.2GW→~2GW): SemiAnalysis / datacenters.com. Inference-cost decline ("LLMflation"): a16z. Algorithmic efficiency (~3×/yr, halving every ~8 months): Epoch AI. Hardware energy efficiency (~+34%/yr FLOP/watt): Epoch AI. Capex & layoffs: our own Capex Watch and Layoffs.
The overbuilt AI capacity that may never earn its capex — the stranded-asset thesis, quantified: the depreciation gap, the revenue lag, and the utilization fork.
$635–690B
2026 hyperscaler capex
~75% AI · >2× 2024
1–3 yr
Real GPU useful life
booked at 5–6 yr
~$176B
Understated depreciation
2026–28 (Burry est.)
40–60%
Cluster utilization
the zone of uncertainty
The mechanism · the depreciation gap
Booked over six years, dead in three
Hyperscalers depreciate AI accelerators on a five-to-six-year straight line. But Nvidia now ships a new architecture every year — Hopper (2022), Blackwell (2024), Rubin (2026) — so a frontier GPU's economic life is closer to two to three years. Microsoft's Satya Nadella put the worry plainly: he didn't want to "get stuck with four or five years of depreciation on one generation."
The chart shows annual depreciation on $100B of AI hardware under each assumption. In the early years the booked charge sits below the economic charge — so reported operating income runs ahead of economic reality. The bill doesn't disappear; it waits. Michael Burry estimates the gap understates depreciation by roughly $176B across 2026–28, leaving reported operating income at names like Oracle and Meta more than 20% above what he reads as economic truth.
Illustrative: annual depreciation per $100B of AI hardware. Booked = 6-year straight line (~$16.7B/yr). Economic = ~2.5-year decline reflecting annual obsolescence. The early-year gap is reported earnings borrowed from later writedowns. Schedules: Alphabet, Microsoft, Oracle, Meta filings.
The swing factor · utilization
The fork in the road
Whether the build-out is foresight or overbuild turns on one number almost no one discloses cleanly: how hard the clusters are actually running. Industry reads put current utilization in the 40–60% band — precisely the zone where both bulls and bears find evidence. Above 70%, the spend compounds; below 50%, it starts to look like the late-1990s telecom fibre glut.
> 70% · prescient
Supercycle
Demand fills the racks; the infrastructure compounds returns for a decade and the long depreciation schedule is justified.
40–60% · today
Uncertain
The zone of maximum ambiguity. Capex outruns AI revenue; the depreciation wave builds while the racks run half-full.
< 50% · stranded
Telecom replay
Demand never arrives at the booked scale. Writedowns begin in earnest; the stranded compute is realized as a loss.
The financing makes the clock real. AI revenue has not caught the depreciation wave from $380B+ in annual spend; Amazon's free cash flow is expected to turn negative in 2026, and hyperscaler debt issuance may exceed $400B. The capex is increasingly paid for by borrowing against a return that has to show up on schedule.
Research discipline · what would prove us wrong
The falsifiers
This is the fragility case, not a forecast. We hold it falsifiable. The thesis weakens — and the bull case strengthens — if:
Utilization climbs through 70% and stays there, on disclosed (not modelled) figures.
Old training silicon repurposes cleanly for inference at scale, genuinely extending economic life toward the booked schedule.
AI revenue growth overtakes capex growth for several consecutive quarters, closing the gap the depreciation wave opened.
Hyperscalers shorten useful-life assumptions themselves — taking the charge now, which removes the overstatement rather than proving it.
Each is observable. We track them on Capex Watch; when the evidence moves, the call moves with it.
Method & sources
The depreciation chart is illustrative of the mechanism, not a company-specific model; magnitudes follow disclosed schedules and the cited estimates. Figures are current as of mid-2026 and will move.
GPU useful-life debate & the ~$176B / 20% estimates: CNBC, Princeton CITP. 2026 capex scale ($635–690B, ~75% AI): hyperscaler guidance / AL Capital. Utilization fork & writedown timing: AL Capital / CFA analysis. Debt & cash-flow: Morgan Stanley via industry reporting. Cross-referenced with our own Capex Watch.
Not investment advice. The Desk is a research institution; nothing here is a recommendation to buy or sell any security.
Is the AI boom like the dot-com bubble? We aligned both cycles at their boom starts on the same NASDAQ series: the telecom era peaked at month 48 — the AI cycle is at month 44. The overlay shows shape, not prophecy. Sources: FRED.
The comparison everyone argues about
AI vs dot-com: the same chart, honestly aligned.
Every version of this argument uses a different index, a different window, and an
alignment chosen to make its point. So we did it the boring, checkable way: one index
(NASDAQ Composite, FRED), both eras, monthly closes, aligned at each boom's start — the
1996 Telecom Act and the November 2022 ChatGPT moment — indexed to 100. No peak-picking: aligning
at the start assumes nothing about where today's cycle tops.
Where we are on that curve
Reading:—
What this does and does not say:
the telecom-era tape peaked 48 months after its boom started; this cycle is at month 44. That is a
fact about shape, not a prediction — two cycles are a sample of two, the anchor choices are
judgments (stated above), and nothing here forecasts when or whether this cycle peaks. What the
overlay is for: making the last mania's shape visible while standing inside this one.
The deeper story is what the price tape can't show: in 2000 the "demand" was
vendor-financed — Lucent alone carried ~$15B of customer financing against ~$300M of operating cash
flow, and the overcapacity had a name: dark fiber. Today ~$540B of committed AI compute rests on
~$35B of filed outside equity (~15.5×), and a hyperscaler is renting out its "excess AI compute."
The structural analysis, edge by EDGAR-verified edge, is in the papers below.
The energy indicator, promoted to a first-class instrument: how power constraint scores across the AI build-out, read from the filings.
The Instruments · Energy & Power
The Power Draw
Energy is one of the six fragility indicators. It is also the one the market underweights and the build-out cannot escape — every token has a physical bind. This promotes it from a column on the scorecard to an instrument of its own, read the same way: from the filings.
The reading today
~40
MEAN ENERGY SCORE / 100
across 49 scored names · the calmest of the six indicators
75
HOTTEST · ADBE
the labs and the utility, not the chips, run hottest
10
CALMEST · CSCO
low physical-power exposure, per filings
Read straight, energy is quiet — a mean near 40 on a 0–100 fragility scale, calmer than capex-versus-demand (~65) or demand durability (~57). But the distribution is the story: the pressure concentrates exactly where power demand is most acute.
Where energy pressure concentrates
NextEra (NEE) · the utility64
OpenAI65
Anthropic63
xAI60
TSMC (TSM) · fabrication57
Nvidia (NVDA)50
Where it reads calm
Cisco (CSCO)10
Palantir (PLTR)12
Intel (INTC)18
Netflix (NFLX)19
UiPath (PATH)20
That split is the honest signal. The model labs and the one pure-play utility on the tape carry the highest energy readings; software and networking carry the lowest. Power constraint is not a market-wide fragility yet — it is a frontier fragility, sitting on the names actually pouring megawatts into the build-out.
What the indicator measures
Scored from each company’s filings, where disclosed: fab and data-center energy intensity, power draw and power-purchase exposure, energy as a share of cost of goods, and the diminishing-returns curve on efficiency (does each new generation of compute buy less output per watt?). Where a company does not disclose the underlying figure — and most do not disclose fab kWh-per-wafer or energy-as-%-of-COGS — the score carries a NOT SOURCED label rather than an invented number, exactly as every desk indicator does.
Why this is v1, and honest about it. Today’s energy score is company-level, read from SEC filings. That is real, but it is the shallow end. The deep signal in power lives one layer down — in the deals: power-purchase agreements, utility joint ventures, grid-interconnect queues, and the megawatt commitments behind each data center. That layer is not in a 10-K; it is in utility filings, project financings, and PPA disclosures. Building it is next week’s work — and it is the lane where the strongest independent energy-finance analysts already operate.
The build roadmap
Surface the existing energy indicator as its own instrument.DONE · this page The 49 filing-scored readings, the frontier-concentration finding, the methodology.
The deal layer. A data table of disclosed AI-power deals — PPAs, utility JVs, on-site generation, nuclear restarts — each tied to its primary source, the way the financing-edges ledger ties every recycling edge to an accession.
The megawatt tally. Committed data-center power against grid capacity, by region, from utility and interconnect filings — a sourced answer to “can the grid actually deliver the build-out the market has priced?”
The falsifier. State, in advance, the condition that would prove the power thesis wrong — e.g. interconnect queues clearing faster than capex grows — and put it on the receipts ledger like every other desk call.
Promote to first-class. Add to the nav, seal into the provenance manifest, and launch with a Dispatch piece and a flare card — a full instrument, not a scaffold.
Sources note (v1). Energy scores from the published company sheets on the 68-name tape (indicator 5 of 6), read from FY2023–FY2025 filings. Distribution figures computed from those scores. No power-deal or megawatt figures appear on this page yet — those arrive with step 2, each primary-sourced, or they do not appear at all.
The exits we publish in advance, and the revisions we make when the evidence moves. A live scoreboard of the thesis's falsifiers plus a dated revision ledger — credibility as an instrument.
The watch · what would change our mind
The exits, published in advance
Falsifier
Current reading
Status
Cluster utilization climbs above 70% on disclosed figuresweakens fragility · The Race · Stranded
~40–60% (industry estimates) — the zone of maximum ambiguity
Watching
AI-attributable TFP turns up in the official seriesweakens fragility · The Lag Index
~+0.07pp/yr (Kansas City Fed / BLS) — not yet in the aggregates
Not triggered
AI revenue growth overtakes capex growth for several quartersweakens fragility · Stranded · Capex Watch
Capex still materially outpacing AI-cloud revenue
Not triggered
Training silicon repurposes cleanly for inference at scale, extending real asset lifeweakens fragility · Stranded
Asserted by hyperscalers; no clean disclosed evidence yet
Watching
A hyperscaler shortens useful life citing AI obsolescenceconfirms fragility · Capex Watch
Hyperscaler free cash flow turns negative and AI debt refinances at higher ratesconfirms fragility · The Race
Amazon FCF expected negative in 2026; $400B+ AI debt issuance forecast
Watching
Average AI usage deepens well past ~1.5 hrs/week per knowledge workerweakens fragility · The Lag Index
~1.5 hrs/week (executive survey) — logins, not workflow redesign
Not triggered
When a reading crosses its threshold, the call moves with it — and the change is logged below. Sources are on each linked instrument.
The ledger · revise in public
What we've changed, and why
2026-07-02
Recycling ratio revised 26× → 15.5× — a missed Amazon–OpenAI equity edge, added
A line-item re-verification against EDGAR filing text found our edge ledger missing Amazon's Q1 2026 $15.0B funded OpenAI Series C investment and $35.0B commitment letter (accession 0001018724-26-000014). Both were added; the funded-cash denominator widened ~$21B → ~$35B and the headline ratio compressed to 15.5× (~13× present-valued). Why: the ratio must carry every filed edge, including the ones that cut our own headline. Note the falsifier distinction: this equity is intra-ring (a top cloud funding the lab committed to it), not the arm's-length capital that would weaken the thesis — the multiple fell while the circularity tightened.
2026-07-02
Microsoft's OpenAI gain relabeled: $5.9B investment gains, not $4.5B
The Q3 FY2026 10-Q (accession 0001193125-26-191507) reports $5.9B of net gains from OpenAI investments for the nine months — primarily the dilution gain from the OpenAI recapitalization. The $4.5B we carried was the same filing's after-tax net-income impact ($0.60 diluted EPS), mislabeled as the markup. Mark-to-model total is now $18.2B. Why: pre-tax investment gains and their after-tax income effect are different lines; the ledger should quote the one it names.
2026-07-02
Amazon's $920M one-time charge re-attributed to the FY2024 10-K
The ~$920M accelerated-depreciation charge (Q4 2024 early retirements) is verbatim in the FY2024 10-K filed 2025-02-07; we had cited it to the FY2025 10-K. The ~$1.4B actual-2025 depreciation step-up remains correctly cited to the FY2025 10-K. Why: filing-of-record accuracy — the quote must point at the document that contains it.
2026-06-26
Divergence gauge relabeled as a short-series, directional read
Flagged the +4.06 divergence as a four-quarter (n=4) reading — directional, not a long-run signal — on Capex Watch, the Brief, and the homepage. Why: a four-point series can't support two-decimal confidence; the precision implied more than the data holds.
2026-06-26
Amazon depreciation figure reconciled and sourced
Unified to a $1.4B run-rate depreciation step-up (6→5yr policy) plus a separate $920M one-time write-off, cited to Note 1 of the FY2025 10-K, across Capex Watch and the Brief. Why: the two pages had carried non-reconciling figures.
2026-06-26
Reproducibility wording corrected, and the tables published
The indicator pipeline is computed in Python, not "live through DuckDB"; we corrected the wording and published the underlying tables at /data with their accession numbers. Why: the original phrasing overstated the mechanism; the fix is to show the data.
2026-06-26
"Both clocks" claim softened to the defensible version
From "no one else is timing both clocks" to "others time one clock or the other; we hold both on a single scoreboard — with explicit falsifiers, revised in public." Why: the absolute claim contradicted our own sourcing.
This ledger is append-only. Each entry is a place we were less right than we wanted to be, fixed in the open. More on how we work
The desk measures one question — the race between the productivity lag and the financing runway — on two instruments. The Fragility Index is a structural read: the equal-weighted average of six filing-sourced indicators across the companies that are the build-out. The divergence gauge is a timing read: a market signal set against a ground-truth signal, quarter by quarter. The Index says how fragile the structure is; the gauge says whether the market's price has detached from the fundamentals. They sit side by side because either alone is half the picture.
The rubric
The six fragility indicators
Each company is scored 0–100 on six indicators (higher = more fragile). The weights are fixed and sum to 1.00. Two indicators have a published closed-form formula; the other four are computed from filings but described qualitatively rather than as a single equation.
Indicator
What it asks
Weight
01 · Depreciation integrity
Are the assets aging faster than the books admit? A life shortened scores zero, regardless of size.
0.20
02 · Capex vs demand gap
Is AI capital spending outpacing the revenue that would justify it? The break-even hurdle is set generously.
0.20
03 · Insider selling intensity
Pre-scheduled 10b5-1 sales score low; the signal is discretionary selling in a window holding material non-public information.
0.15
04 · Circular financing
Is the money going in a loop — investor funds a lab, lab buys the investor's compute, that revenue underwrites the capex?
0.20
05 · Energy & diminishing returns
Are power, cooling, and chip economics starting to cap capability gains? The thinnest-data indicator — lowest weight.
0.10
06 · Organic end-user demand
Is reported AI revenue genuine paid adoption by independent end-users — or recycled, or rebranded from existing lines?
0.15
The two published formulas
# 01 — paper benefit from a useful-life extension delta_dna = ppe_depreciable × ( 1/life_old − 1/life_new ) # hard rule: a life shortened scores 0, regardless of size
# 02 — required revenue per $1 of capex per year factor = ( CoC + 1/L ) / m = ( 0.10 + 1/6 ) / 0.30 = 0.889 # fail when FY2025 segment revenue < capex × 0.889
Honesty note
Indicators 03–06 are computed only from filing-sourced inputs but do not publish a single closed-form equation; they are read as standing analyst judgments on the 0–100 scale, reviewed each quarter. Indicator scores are internal judgment, read for convergence; figures are sourced or labelled.
Instrument one
The Fragility Index (0–100)
A company's composite is the weighted mean of its six indicator scores (missing scores are dropped and the weights renormalized). The Index is the equal-weighted average of those composites across the build-out core (Layers 1–4: compute & infrastructure, hyperscalers & cloud, model labs, AI software). The broader-market comparators in Layer 5 are excluded — they're the control group, not the build-out. It's reproducible from the published per-company scores.
composite(c) = Σ wᵢ·scoreᵢ / Σ wᵢ (over present scores) Index = mean( composite(c) ) for c in Layers 1–4
Band thresholds
Reading
Band
≥ 75
Severe
60–74
Elevated
45–59
Moderate
30–44
Contained
< 30
Low
Current reading: 49 (Moderate), 2026 Q2. The Index is recomputed from the published company scores; see the live Fragility Index.
Instrument two
The divergence gauge — D(t)
The gauge sets a market signal against a ground-truth signal:
D(t) = M(t) − G(t)
M(t), the market term, is the equal-weight mean of three full-window z-scored components of SOXX price behaviour — 63-day momentum, price-to-trend overextension, and 20-day annualized instability. G(t), the ground-truth term, is the negative mean of three deterioration z-scores — AI-layoff share, discretionary insider selling, and the capex gap. The gauge widens when momentum and overextension climb while the fundamentals erode.
The four-quarter series
Quarter
M(t)
G(t)
D(t)
2025 Q3
−0.82
0.98
−1.80
2025 Q4
−0.73
0.43
−1.16
2026 Q1
−0.50
−0.18
−0.32
2026 Q2
2.83
−1.23
+4.06
Stated limitation
The gauge standardizes its components over the full window — it is descriptive, not real-time: it carries look-ahead bias and is not a tradeable signal (an expanding-window version is deferred). The +4.06 reading is the strongest move in a four-quarter series (n=4: descriptive, not a long-run signal). It weights its three market components equally; empirical calibration is future work.
The set
The company universe — 43, 68, and the rings
Two counts appear across the site, and they measure different things:
Set
What it is
Names
Full board
All five layers of the build-out — the Markets tape.
68
Build-out core
Layers 1–4 (compute & infrastructure, hyperscalers & cloud, model labs, AI software). This is what the Fragility Index scores.
43
Comparators
Layer 5 — the broader-market control group, excluded from the Index.
25
So 68 = the full five-layer set; 43 = the build-out core (Layers 1–4); the remaining 25 are the Layer-5 comparators that act as a control group. The Fragility Index is computed on the 43, not the 68.
Separately, Explained maps the build-out as three rings — a core of 43, a supply chain of 39 that feeds it, and a demand ring of 21 industries that must pay it back. The ring counts are a different partition from the five-layer board; the enumerated name lists for the 39-name supply chain and 21-name demand ring are not yet published alongside the core.
Provenance
Data, sources & reproducibility
Every figure derives from filing-sourced inputs; where a value cannot be sourced cleanly from a filing, it is shown blank rather than imputed. Each table carries its 10-K / Form 4 accession numbers inline. The indicator pipeline is computed in Python, and the underlying tables are published open.
Named sources
SEC 10-K filings — accounting & capex tables (e.g. Amazon FY2025 10-K Note 1, accn 0001018724-26-000004). sourced SEC Form 4 — the insider-selling record (discretionary vs 10b5-1). sourced SEC 8-K — financing structure (e.g. Nvidia→CoreWeave backstop, 8-K accn 0001769628; OpenAI→AMD 6 GW, 8-K EX-99.1, 2025-10-06). sourced SOXX daily — iShares Semiconductor ETF price series, the market term for M(t) (soxx_daily.csv). sourced MIT NANDA / Fortune (Aug 2025) — ~95% of enterprise GenAI pilots show no measurable P&L impact. attributed Kansas City Fed / BLS — AI-attributable TFP (~+0.07pp/yr). attributed Michael Burry (Scion) — ~$176B understated depreciation 2026–2028; carried as his allegation, not an audited figure. attributed
Credibility is the only thing the desk sells, so every number carries its provenance. Three tags run through the site:
sourcedattributed / estimateeditorial read
Sourced figures trace to a primary filing with its accession number. Attributed figures are named as such — a short-seller's allegation, a survey, a Fed estimate — never dressed up as measurement. Editorial reads (the regime call, the loop framing) are labelled as The Desk's interpretation of the measures, not a measurement. Unverified figures are flagged or removed. A few standing examples:
· The regime call is The Desk's editorial read of the metrics, not a measurement.
· The recycling ratio moves with provenance, not arithmetic: ~15.5× on a funded-cash basis, ~13× present-valued at 10%, easing to ~3.6× when every reported secondary round is admitted as equity (revised 2026-07-02 — the Amazon–OpenAI equity legs entered the ledger).
· The convergence of indicators is weighted as corroboration, not four independent votes — they share a common driver (capex ahead of monetization).
· The self-measurement on the Instrument is n=1, illustrative — directional, not precision.
The exits
Falsifiers & revisions
The falsifier is built in: if the ground-truth signal turns back up — demand converting, the capex gap closing, insider selling normalizing — the divergence closes and the boom earns its price. We publish the number either way. The full ledger of what would prove us wrong, each with its current reading and status, is the Falsifier Watch — an append-only page that also records every revision we've made, dated. The most recent revisions reconciled the Amazon depreciation figure (the canary: a ~$1.4B actual-2025 depreciation step-up per the FY2025 10-K, plus a separate ~$920M one-time Q4 2024 charge per the FY2024 10-K), relabelled the divergence gauge as a short-series directional read, and corrected the reproducibility wording (Python pipeline, tables published at /data).
REVISION2026-07-02The call: the recycling ratio is ~26×
An EDGAR line-item re-verification (Amazon → OpenAI: $15B funded + a $35B commitment letter, accession on file) widened the outside-funded-cash denominator. Restated 26× → ~15.5×; the 26× is superseded and blocked at build time by the I4 out-of-band assert.
EDGAR: 0001018724-26-000014
REVISION2026-07The call: Amazon is extending server life to flatter earnings (I1)
Amazon in fact shortened useful life 6→5 years (~$1.4B added cost) — against the hyperscaler trend. Corrected: Amazon is the canary, not a culprit.
PENDINGopenThe refinancing wall bites in the next data-center debt-maturity window
Watching Q3–Q4 maturities; unresolved. Logged in advance so the call is checkable against what happens.
The desk’s claims are only worth something if the dates on them are real. So we hash every dated artifact — the papers, the receipts ledger, the Dispatch, the earnings watch list, the data files — into a manifest, and anchor that manifest to the Bitcoin blockchain via OpenTimestamps. Anyone can verify, cryptographically and without our cooperation, that an artifact existed in exactly its published form at the manifest’s time.
Why this exists
The desk publishes falsifiers before calls, watch lists before earnings, and corrections against its own headlines. All of that depends on one thing being true: that we wrote it when we say we wrote it. Platform timestamps require trusting the platform; our own dates require trusting us. A Bitcoin anchor requires trusting neither — the proof is in a chain nobody edits.
How to verify a manifest yourself
1 · Download a manifest and its proof file below (manifest-<date>.txt and .txt.ots).
2 · Install the open-source client: pip install opentimestamps-client
3 · Run ots verify manifest-<date>.txt.ots — it reports the Bitcoin block that anchors the manifest.
4 · Check any artifact against its line in the manifest: shasum -a 256 <file>.
If the hash matches, that exact file existed no later than the anchored time. Full stop.
Manifests
Newest first. Fresh proofs begin as calendar-server attestations and upgrade to full Bitcoin confirmations within hours; we re-upgrade proofs at every publish.
Loading…
What this proves — and what it doesn’t. An anchored manifest proves an artifact existed, in exactly its published byte-for-byte form, at a moment in time. It does not prove the artifact is correct — that burden stays on the sources and methods cited inside it, and on the receipts ledger where our record stands to be checked.
A working hub for the AI economy — the models, the tools, the data sources, and the reading we draw on. Click any card for the full read; the link out is inside.