OMEN's indexes and gauge read what markets are pricing. This page reads the plumbing: who is lending, who is quietly switching to equity, how much of the economy the capex has eaten, and how many announced gigawatts actually get energized. Eight theses, each with the metric that would confirm or kill it – five from Paul Kedrosky's ROI teardown, three from Jim Chanos's telecom-collapse analogy. The structural risk underneath them is the same one: deflating token prices funding 10–15-year fixed-payment debt – a classic duration mismatch.
Frameworks: Paul Kedrosky on Better Offline – "Why AI Has No ROI" (Jun 2026) and Jim Chanos, "The AI Bubble Is Much Worse Than Dot-Com" (2026); the bull-side tape rows follow Gavin Baker on Invest Like The Best – "The AI Selloff Doesn't Match the Data" (Aug 2026), plus the reference series cited under each panel.
Four feeds the eight theses kept pointing at but never measured: realized hardware demand (TSMC's monthly tape, upstream of every hyperscaler press release), issuance velocity (EDGAR full-text counts of the actual debt and equity paperwork), paid adoption (Ramp's transaction data – the revenue-side answer to "does the capex earn its cost"), and the generator pipeline behind the stranded-GW thesis. Four more carry the bull case in the same auditable form – the counters Gavin Baker put on the record after the July selloff: cash-flow acceleration (hyperscaler OCF growth, XBRL), token demand (OpenRouter's platform-wide weekly tape), the GPU repricing gap (spot vs contract vintages), and the inference-margin floor (flagship token prices vs compute cost at spot rents).
Monthly revenue growth (TWSE filing, –). The buildout is real only while this number is.
Ramp AI Index (–) – share of ~70k firms with a paid AI transaction that month.
FWP + 424B filings by the six AI-capex issuers, trailing 90 days vs the 90 before (–).
Operating cash flow across the AI-capex filers, latest complete quarter (–) – audited XBRL, not guidance. Acceleration pending prior-quarter data.
OpenRouter platform-wide (wk of –) – the demand tape behind the capex. Growth loads with the live series.
Vs its own 10-year monthly range (percentile loading) – SEC XBRL EPS × Yahoo closes. The "decade-low multiple" claim, audited.
| Series | Reads on | Latest | Δ |
|---|---|---|---|
| TSMC monthly revenue (–) | Realized AI-silicon demand, ~10-day lag | – | – |
| TSMC revenue, YTD vs prior year | Trend, strips the monthly noise | – | |
| Korea semiconductor exports, 20-day | HBM/memory pulse every ~10 days – the classic cycle canary | manual |
| Paperwork | Reads on | 90d | Prior 90d |
|---|---|---|---|
| FWP + 424B, six AI-capex issuers | Thesis i in filing counts – the debt wave, mechanically | – | – |
| S-1s mentioning "artificial intelligence" | IPO pipeline – Kedrosky's "gonging of the bell" | – | – |
| Form Ds mentioning "artificial intelligence" | Private raises – the pre-IPO froth layer | – | – |
| Series | Reads on | Latest | Δ |
|---|---|---|---|
| Ramp AI Index (–) | Share of US businesses with paid AI subscriptions – transaction data, not surveys | – | – |
| Census BTOS estimate (contrast) | The survey number Ramp's transaction data corrects | – | |
| Anthropic Economic Index, latest release | Usage composition (augmentation 57/43 automation, hand-updated) – freshness tracked live | – | |
| Agent-stack installs, npm weekly (–) | Claude Code + Codex + Gemini CLI downloads – the agentic S-curve, read in package managers | – | – |
| Agent SDK installs, PyPI weekly | claude-agent-sdk + openai-agents – a hard undercount (OpenAI-compatible traffic rides the openai package) | – |
| Series | Reads on | Latest | Prior |
|---|---|---|---|
| Big-5 OCF, YoY (–) | Baker's master loop – contracts roll to spot, cash flow accelerates, ≈$700B of credit demand disappears | – | – |
| Acceleration, quarter over quarter | The second derivative is the claim: decelerating growth breaks the loop before levels do | – | |
| Newest quarter, cohort status | A 4-of-5 cohort never gets quoted against a 5-of-5 baseline – it would read as a collapse that never happened | – |
| Series | Reads on | Latest | Δ |
|---|---|---|---|
| Routed tokens, weekly (wk of –) | Aggregate inference demand in the price-sensitive routed segment – the mix-shift-vs-demand-shift test | – | – |
| 4-week growth | Momentum – Baker's "token growth is accelerating" claim, checkable weekly | – | |
| ~1-year growth | The scale of the demand curve the buildout is racing | – | |
| Chinese-lab share of the week | Open-weight substitution inside the same demand curve – margin dollars migrating, not demand leaving | – |
| Generation | Reads on | Spot | Contract vintage | Gap |
|---|---|---|---|---|
| H100 (–) | Legacy fleets – 2023-vintage LTAs sit far above today's spot; the renters' churn risk, the owners' cushion | – | ≈$4.40/hr 2023 LTA | – |
| B200 curated | Current generation – Dec-2025-vintage contracts sit under spot; the installed base is underearning, repricing upside on roll | ≈$3.90/hr Aug 2026 | ≈$2.50/hr Dec 2025 | – |
| Line | Reads on | Output $/M | Implied margin |
|---|---|---|---|
| Serving-cost floor at spot | H100 spot rent ÷ throughput, at 0.5–2.0M output tok/GPU-hr – the denominator every model shares | – | |
| US flagship (Claude Opus) | Frontier pricing power – Baker puts frontier token margins at 80–95% | – | – |
| CN flagship (DeepSeek pro) | Open-weight serving economics – Baker puts these near 30%; compression biting shows up here first | – | – |
| Status | Reads on | Nameplate GW |
|---|---|---|
| Operating | What the grid actually has | – |
| Under construction | Steel in the ground | – |
| Planned / approvals pending | The press-release layer | – |
| Canceled or postponed | Thesis v in federal data – capacity that stopped | – |
| US interconnection queue (LBNL, 2023-12) | Requests, most of which never build | ≈2,600 |
| PJM capacity clears (2025/26 → 2026/27) | Scarcity pricing the AI load already caused | $269.92 → $329.17/MW-day |
The buildout has migrated from cash flow to credit. When prime corporates saturate the IG market, the marginal buyer changes – recent books lean on European insurance funds and Middle East sovereign wealth, the buyers who historically show up at the end of a cycle. Watch issuance volume, book quality, and new-issue concessions.
Cash proceeds from long-term debt issuance across the AI-capex filers, trailing four reported quarters – audited XBRL, not a press tally.
The comparison group hyperscalers overtook in Q1 2026. Stays hand-curated on purpose: JPM and WFC tag no debt-issuance concept the XBRL API exposes, so a computed bank total would omit two of six and flatter this very comparison.
Oct 2025, the largest corporate bond deal of the year – demand at the top is not the question; capacity is.
| Date | Issuer | Vehicle | Size |
|---|---|---|---|
| Sep 2025 | Oracle | IG bonds, 5–40y | $18.0B |
| Oct 2025 | Meta | IG bonds – largest corporate deal of 2025 | $30.0B |
| Oct 2025 | Meta / Blue Owl | Hyperion SPV private placement (off balance sheet) | $27.3B |
| Nov 2025 | Alphabet | IG bonds, USD + EUR tranches | $25.0B + €6.75B |
| Dec 2025 | Amazon | IG bonds – first issuance since 2022 | $15.0B |
| H1 2026 | Complex-wide | Follow-on IG + SPV/private-credit vehicles | ≈$100B+ |
The gauge's credit family reads HYG drawdown and the HY/IG ratio – the whole market. This section watches the instruments that fund the buildout itself: neocloud high-yield, Oracle's curve, and the SPV paper. If the duration-mismatch thesis is right, this is where stress prints first, months before it reaches the index.
The purest listed neocloud credit. Widened ≈85bp over 90 days while generic HY was flat – early divergence.
Roughly double its pre-buildout level. Oracle carries the most leveraged AI capex program of the majors.
SPVs, vendor financing, and data-center lending sitting outside public marks – repricing arrives late and all at once.
| Instrument | Reads on | Level | Δ 90d |
|---|---|---|---|
| CoreWeave 9¼% 2030 | Neocloud funding cost – GPU-backed, contract-concentrated | ≈+560bp | +85bp |
| Oracle 10y benchmark | Most-leveraged major – capex running ahead of operating cash flow | ≈+130bp | +22bp |
| Meta Hyperion SPV 2049 (Blue Owl) | Off-balance-sheet data-center paper – the new marginal structure | ≈+240bp | +15bp |
| ICE BofA HY OAS (contrast) | Generic high yield – what the gauge already reads | – | – |
A mega-cap selling stock at the market is a tell, not a flex. Equity is the most expensive money a prime credit can raise – you only reach for it when cheaper channels (operating cash flow, IG bonds, SPVs) are saturated. Alphabet's $80B program is the first mega-cap ATM era in the modern market. Track who follows.
Cash actually raised, per Alphabet's XBRL cash-flow statement – against an $80B announced program ($10B Berkshire placement plus two ATM sale programs, Jul 2026). Alphabet had never tagged this concept before 2026; the tag appearing at all is the tell.
MSFT · ORCL · CRWV, trailing twelve months – and this is ordinary employee stock-plan flow, not secondary issuance. That is the comparison: one mega-cap's raise against the entire rest of the complex's routine share sales.
Stock comp forces buybacks, buybacks eat cash flow, capex eats the rest – then the SPVs and the ATM appear.
| Date | Issuer | Event | Size |
|---|---|---|---|
| Jul 2026 | Alphabet | Berkshire private placement | $10B |
| Jul 2026 | Alphabet | Two at-the-market sale programs | ≈$70B |
| Watch | Meta · Oracle | Next-most-stretched capex/OCF ratios – candidates to follow | – |
| Watch | OpenAI · Anthropic | IPO S-1s – check whether training costs get capitalized ("earnings before bad stuff") | – |
When one investment category carries a visible share of GDP growth, the macro cycle and the capex cycle become the same cycle – a slowdown in data-center spend reads as a recession print. The monitor already tracks audited capex vs operating cash flow from SEC XBRL; this section adds the macro layer on top.
MSFT + GOOGL + AMZN + META + ORCL, each filer's own trailing four quarters per XBRL filings – roughly triple the 2023 run rate.
Data-center construction (–, Census C30 SAAR) plus computer-equipment investment (–, BEA) over nominal GDP. Narrower than the monitor's big-5-capex/GDP ratio, which counts five companies' worldwide capex.
Big-5 capex growth as a share of the change in nominal GDP – strip it out and the economy is materially slower.
| Buildout | Peak share of GDP | What followed |
|---|---|---|
| Railroads, 1880s | ≈6% | Panics of 1873/1893 · ≈half of boom-era track eventually abandoned |
| Telecom fiber, 2000 | ≈1.2% | 2001–02 bust · fiber found reuse, but only after the equity was destroyed |
| AI data centers, 2026 | – | Open – this page exists to watch it |
Announced gigawatts are a press release; energized gigawatts are a utility interconnection. The gap between them is the stranded-asset pipeline: speculative shells, behind-the-meter gas plants with 30–40-year lives, and county budgets pre-spending tax revenue that never arrives. The metric is brutal and simple – announced vs under construction vs energized.
Dedicated AI data-center capacity announced to date – roughly the load of 35 million homes.
Steel in the ground per utility filings and interconnection queues – a quarter of the announcements.
Of 14 announced gigawatt-class projects, none is fully powered. Abilene is the closest, at partial load.
| Project | Sponsor | Target | Status |
|---|---|---|---|
| Stargate Abilene, TX | OpenAI · Oracle · Crusoe | 1.2 GW | Partially energized |
| Hyperion, Richland Parish, LA | Meta | 5 GW | Under construction |
| Prometheus, New Albany, OH | Meta | 1 GW | Under construction |
| Colossus 2, Memphis, TN | xAI | 1.5 GW | Under construction |
| Fairwater, Mt Pleasant, WI | Microsoft | ≈1 GW | Under construction |
| Wonder Valley, AB (Canada) | O'Leary / Greenview | 7.5 GW | Announced |
| Utah / New Mexico mega-sites | Various | 1–10 GW | Speculative |
| Orbital constellation (LEO) | SpaceX / Starlink | 1–2 GW (8 GW rumored) | Speculative |
| Starcloud orbital DC | Starcloud (NVIDIA-backed) | 5 GW target | Announced |
A buildout can be real and ruinous at once – the internet did need the fiber. What killed the telecom names was timing: capacity energized years before demand could absorb it, funded by debt that came due first. The AI tell is the same – falling unit economics on already-built compute. When rental prices for last-generation GPUs fall faster than the boxes depreciate, the market is telling you supply has outrun paying demand.
Of the strand-miles laid into the telecom bubble, most was never lit – ≈$5T of market value erased before it found reuse.
Live vast.ai median (–); p10 –. Down from ≈$8/hr in 2023 – depreciation on a ~2-year asset the hyperscalers still carry over 5–6. The spot tape is noisy day to day; the multi-year direction is the claim.
≈48 GW announced against a demand curve that plausibly absorbs a fraction on the stated timeline – the overbuild ratio the fiber map warned about.
| Mechanic | Telecom, 2000 | AI, 2026 |
|---|---|---|
| Asset built ahead of demand | Long-haul fiber | GPU clusters + power |
| Vendor financing | Lucent / Nortel to CLECs | Nvidia to neoclouds + labs |
| Depreciation vs reality | 20–25y book life on gear obsoleted in ~5 | 5–6y book life on GPUs obsoleted in ~2–3 |
| Utilization tell | % of fiber lit | GPU rental price · fleet utilization |
| Outcome | ≈$5T erased; capacity reused post-bankruptcy | Open |
Thesis four watched capex ÷ operating cash flow climb toward 1.0. Push one line further down the statement and you get the punchline: operating cash flow minus capex – true free cash flow – has fallen ~80% across the Big-5, and turned negative at Oracle and Amazon. GAAP earnings don't show it for exactly Chanos's reason: a ~2-year chip depreciated over 5–6 years keeps net income high while the cash walks out the door now.
OCF minus capex across MSFT · GOOGL · AMZN · META · ORCL – down from ≈$300B+ before the buildout, on roughly flat-to-higher net income.
The filers whose capex outruns what their operations generate in cash. Audited XBRL, each filer on its own trailing four quarters.
The gap between reported Big-5 net income and free cash flow – the non-cash cushion that lengthened depreciation keeps inflating.
This is the froth tell that sits next to the buildout, not inside it. A treasury vehicle trading at 1.6× the value of what it owns is priced as a perpetual-motion machine: premium funds purchases, purchases lift the asset, the higher asset "justifies" the premium. Chanos's point is that the mechanism is reflexive – it runs in reverse just as fast. The pattern started in crypto; watch it migrate to AI-compute and token treasuries.
Market cap ÷ net asset value of the largest crypto-treasury vehicle – you pay ≈$1.60 for $1.00 of the underlying it holds.
Public companies whose primary "operation" is holding a financial asset bought with issued equity – a category that barely existed three years ago.
At-the-market issuance funds asset purchases; purchases lift the mark; the mark justifies the next raise – until the premium inverts.
| Vehicle type | Underlying | mNAV | State |
|---|---|---|---|
| Flagship BTC treasury | Bitcoin | ≈1.6× | Premium |
| Second-wave BTC treasuries | Bitcoin | ≈1.0–1.2× | Compressing |
| ETH / SOL treasuries | Ether · Solana | ≈0.9–1.1× | At/near NAV |
| Emerging AI-compute treasuries | GPUs · token credits | – | Watch |
This page is part live, part curated, and every figure says which it is. Live from primary sources on the data cadence: the whole of the live tape (TWSE, EDGAR full-text, Ramp, Hugging Face, EIA-860M), plus every capex, cash-flow and issuance number in theses i, iii, iv and vii (audited SEC XBRL), the capex/GDP ratio (Census C30 + BEA), the GPU rental tape (vast.ai) and the HY OAS contrast (FRED). Still curated by hand and marked as such: the single-name credit spreads in thesis ii (no free per-CUSIP source exists), the bank issuance comparator, the gigawatt project tracker, and the treasury-vehicle watchlist. Levels marked ≈ are approximations assembled from filings, TRACE prints, and press, and can be stale or wrong; treat them as a reading list with numbers, not data. Unlike the indexes and the gauge, these are fundamentals, not market prices – they can't tell you when, only how much has been borrowed against the answer. The eight theses come from two aligned bearish worldviews – Kedrosky's ROI teardown and Chanos's telecom analogy; the bull rebuttal – that inference demand outruns the depreciation schedule – is exactly what the Bull index prices on the other side of the pair.