AI Bubble Monitor

A one-page Next.js dashboard that tracks the signals behind one plausible AI-bubble-pop scenario: monetization stalls → price deflation → capex cuts → credit transmits the shock → circular deals unwind → macro spillover.

It refreshes once a day on Vercel and needs no API keys.

What's on the page

Section	What it shows	Source
Composite stress gauge	31 signals, each scaled 0 (calm) to 100 (crisis), then averaged	derived
AI chips, hyperscalers, neoclouds, private credit, bank lenders, DC REITs & power, on-prem AI infrastructure, broad market	Relative-performance charts vs. S&P 500 and a scoreboard (1D/1M/3M/YTD/off 52-week high)	Yahoo Finance
Hyperscaler capex vs. revenue	Trailing-twelve-month capex, revenue, capex/revenue and YoY growth for MSFT, GOOGL, AMZN, META, ORCL	SEC EDGAR (XBRL)
AI adoption	Reported users per AI lab over time (ChatGPT weekly users, Gemini, Meta AI, Copilot, Grok, Qwen monthly users), paid subscribers and work seats, latest figure per lab with 1-year growth, plus marked third-party estimates. Curated in data/ai-users.json, reviewed weekly	Company announcements, earnings calls and filings; Sensor Tower, Similarweb, QuestMobile
Anthropic	Quarterly revenue vs operating costs, expense breakdown, operating vs net income, compute cost as % of revenue, compute commitments vs revenue, $100K+ customers, annual/half-year results and other S-1 figures. Curated in data/anthropic.json; after the IPO, each quarter is appended from the 10-Q	Anthropic S-1 (SEC EDGAR), then 10-Q/10-K
Circular financing	Nvidia days sales outstanding, days of inventory, and non-marketable stakes in private companies	SEC EDGAR (XBRL)
AI money web	Live recreation of Bloomberg's Nvidia/OpenAI deal web: companies sized by market value (price × SEC share count) or last private round, arrows for each deal by type, round-trip pairs, full deal list with sources	data/ai-deals.json (curated, reviewed weekly), Yahoo, SEC
SPV tracker	GPU-backed loan pricing over SOFR by date, USD.AI stablecoin GPU-lending deposits (DefiLlama, live), data-center and GPU-backed SPVs: tenant, equity holders, lenders, debt, rating, terms, reported loan prices and events; lender rollup with stock drawdowns; hyperscaler lease and purchase obligations vs. reported debt	data/ai-spvs.json (curated, reviewed weekly), SEC EDGAR (XBRL)
Software & private credit	Public SaaS stocks and the IGV ETF vs. the S&P 500; sector facts on private-credit lending to software; each public BDC's software share next to its price / NAV	Yahoo, data/bdc-software.json (curated from 10-Qs, reviewed weekly), BIS, Octus, Kamakura
BDCs & redemptions	Price / NAV for public BDCs; SC TO-I tender-offer filings from non-traded private-credit funds	Yahoo, SEC EDGAR
GPU rental market	$/GPU-hour by source (weekly market median across ~90 providers with spot and 12-month reserved, Vast.ai marketplace, RunPod and Lambda list prices); weekly history charts, archived to Blob so they grow past the 53 free weeks	GetDeploying open dataset (CC BY 4.0), Vast.ai, RunPod API, Lambda pricing page
Cost of intelligence	Best open-weight vs. best closed model on Epoch's Capabilities Index over time (gap and lag), API price vs. capability for the top models, cheapest near-frontier price, automatic detection of API price changes, share of US businesses using AI, electricity prices	Epoch AI benchmark data, OpenRouter API, Census BTOS, FRED
Classic bubble gauges	A scorecard of seven market-wide warning signs (normal / elevated / flashing) and their history: Shiller CAPE vs P/E, excess CAPE yield (Shiller and TIPS methods), the Buffett indicator, the 10y−2y yield curve, Fed hiking cycles, breadth and margin debt	multpl.com (Shiller data), FRED, FINRA
Prediction markets	Polymarket odds on an AI bubble bursting, CoreWeave/OpenAI/Anthropic bankruptcy, a US recession, Fed hikes and the Anthropic IPO, with daily history	Polymarket (Gamma and CLOB APIs)
Credit spreads & stress	AI borrowers' reported 5-year CDS spreads and bond prices with implied default odds (data/credit-stress.json, curated weekly), Fed loan-officer survey (banks tightening business-loan standards), C&I loans, HY, CCC and BBB OAS, 10-year yield, VIX, NFCI, St. Louis Fed stress index	FRED
Concentration & Fed	Equal-weight ÷ cap-weight S&P 500, Fed funds rate, 10y−3m curve, recession probability	Yahoo, FRED
AI in GDP	US data center facility counts by stage (dcmap.us and Kadoa, saved daily), US data center capacity operating vs. under construction (CBRE, data/datacenters.json), how much of GDP growth comes from IT equipment & software investment (vs. the dot-com era), IT investment share of GDP vs. the dot-com peak, monthly data-center construction, GDP final sales of computers, business investment in computers and software, power construction	FRED (BEA, Census), Our World in Data
Consumers & household debt	Consumer sentiment, saving rate, household debt % of GDP, debt payments % of income, consumer credit, auto loans, credit card and mortgage delinquency, 30-year mortgage rate	FRED
Labor	Information and professional-services unemployment, claims, job openings	FRED
Adoption & compute	Gen-AI adoption share, hosting PPI, semiconductor production	FRED
Google Trends	Search interest for AI assistants (ChatGPT, Claude, Gemini, ...), "AI", bubble-fear terms, and Claude vs. peers. The last good copy is saved to Blob in case Trends rate-limits a refresh	Google Trends (unofficial endpoints)
News	Eleven Google News searches over the last 7 days	Google News RSS

All tickers, FRED series, news queries and scoring thresholds live in lib/config.ts and lib/signals.ts.

Updating the deal web

The deals live in data/ai-deals.json. Arrows point the way money flows (payer → recipient). Each deal needs a source URL and a status: announced (company statement, drawn solid), reported (press reports) or loi (letter of intent), both drawn dashed. data/ai-spvs.json holds the SPV tracker in the same way; each vehicle can carry a stress price (cents on the dollar) from a dated report. data/bdc-software.json holds each BDC's software share with the filing it came from. A weekly scheduled Claude Code session reviews the news, edits these files and pushes; Vercel redeploys on push. Market values update daily without any edits.

How the daily refresh works

• The page is statically generated with revalidate = 86400, and every upstream fetch is cached under one tag.
• vercel.json schedules a cron job at 21:30 UTC daily (after the US close) that calls /api/refresh. The job clears the cache tag, re-fetches everything, saves a snapshot, and rebuilds the page.
• Feeds fail independently. If one is down, its card says so and the rest of the page still renders. The failures are listed at the bottom of the page.

Optional: history charts

Prices and FRED series have their own history. The GPU rental rates, BDC price/NAV and composite score don't, so the refresh job stores a daily snapshot of them in Vercel Blob when a store is connected:

1. In the Vercel project, go to Storage → Create → Blob. Pick a private store and connect it to the project. This sets BLOB_READ_WRITE_TOKEN.
2. Redeploy. The score-history and H100-history charts appear after the second daily run.

Optional environment variables

• CRON_SECRET: if set, /api/refresh only accepts requests carrying it. Vercel Cron sends it automatically.
• SEC_USER_AGENT: SEC EDGAR requires a descriptive User-Agent with a contact email, for example "Your Name [email protected]".

Develop

npm install
npm run dev        # http://localhost:3000
npm run typecheck
npm run build

Not investment advice. The composite score is a heuristic, not a forecast.