Reference

Glossary

Every term this site defines, in plain language and alphabetical order. Dotted-underline words across the site link back to these entries.

Updated July 19, 2026

ARR (annualized run rate) #: Annualized run rate, or ARR. Take the latest month (or quarter) of revenue and multiply it out to a full year. Startups and AI companies quote ARR because it flatters fast growth: a company that made $100 million last month can claim a '$1.2 billion run rate' even if it earned far less than that over the actual past year. ARR is a useful speedometer but a poor odometer. When you see 'annualized revenue' on this site, that is what it means.
Basis points #: A basis point is one hundredth of a percentage point. If a bond yield rises from 5.00% to 5.50%, it rose 50 basis points. Finance uses basis points because small moves in rates matter a lot when trillions of dollars are involved.
Capex (capital expenditure) #: Capital expenditure, or capex, is spending on long-lived assets: factories, railway track, fiber-optic cable, data centers, computer chips. Unlike day-to-day expenses, capex is a bet that the asset will earn its cost back over many years. Every infrastructure bubble in history shows the same signature: capex racing far ahead of the revenue the assets actually produce. That gap, and how long it can persist, is a central question of this site.
Circular financing #: Circular (or reciprocal) financing is when participants in a boom fund each other's demand. A chip maker invests in a startup, which uses the money to buy the chip maker's chips; the chip maker then reports booming sales. It is like a shop lending its customers money to buy its own products and then reporting record revenue. The sales are real transactions, but the demand behind them partly originates inside the loop rather than from independent end users. History's versions include vendor financing (telecom, 1990s), share calls (railways, 1840s), and cross-shareholding (Japan, 1980s).
Collateral #: Collateral is the asset a borrower pledges against a loan. If the borrower defaults, the lender seizes and sells the collateral. The riskiness of a loan depends heavily on how well the collateral holds its value. A mortgage is backed by a house, which usually lasts decades. Some AI-era loans are backed by GPUs, chips that lose most of their market value within a few years. That difference matters a great deal if the loans ever need to be collected.
Credit spread #: The credit spread is the gap between what a company pays to borrow and what the US government pays. If Treasuries yield 4% and a data-center company's bonds yield 7%, its spread is 3 percentage points (300 basis points). Spreads are a fear gauge: when lenders start doubting they will be repaid, spreads widen. Historically, spreads on a bubble sector's debt often started widening months before stock prices peaked, which makes them one of the more useful warning lights.
Credit-to-GDP gap #: The credit-to-GDP gap measures whether borrowing across a whole economy is growing faster than its long-run trend. The Bank for International Settlements (the central bank for central banks) tracks it because readings of roughly +10 percentage points or more above trend preceded many historical banking crises, including 2008. It is one of the best-tested early-warning indicators, though it speaks to whole economies rather than single sectors.
CRPO (contracted backlog) #: Commercial remaining performance obligations, or CRPO, is accounting language for contracted future business: services customers have signed up to buy that have not yet been delivered or paid for. Cloud companies report huge CRPO figures as evidence of demand. The catch: a backlog is only as good as the customer's ability to pay. If one heavily-funded startup accounts for a large slice of a backlog, that backlog inherits the startup's risk.
Depreciation #: Depreciation is how accounting spreads the cost of a long-lived asset over the years it is used. Buy a $30,000 server and depreciate it over three years, and profits are reduced by $10,000 a year; assume six years instead, and only $5,000 a year. The asset and the cash spent are identical, but reported profit is higher under the longer assumption. That is why the question of how long AI chips really stay valuable, three years or six, moves billions of dollars of reported earnings and is one of the sharpest controversies in the AI boom.
Drawdown #: A drawdown is the fall from a peak to a low, expressed as a percentage. If an index goes from 5,000 to 1,100, that is a 78% drawdown, which is what the Nasdaq actually did between March 2000 and October 2002. Peak-to-trough drawdown is the standard way to measure how badly a bubble ended.
Equal-weight index #: The regular S&P 500 is cap-weighted: bigger companies move the index more. An equal-weight version counts each of the 500 companies identically. When the cap-weighted index races ahead of the equal-weight one, it means a handful of giants are doing the lifting while the average stock lags. That divergence is a classic sign of narrow market breadth, which has historically appeared in the late stages of bubbles.
Hyperscaler #: The hyperscalers are the handful of companies operating cloud computing at planetary scale: Microsoft, Amazon, Google, Meta, and increasingly Oracle. They matter to the bubble question because they are doing most of the AI spending, and unlike the startups of 1999 they are hugely profitable businesses funding the build-out largely from their own cash.
Inference #: AI has two kinds of computing cost. Training is the one-time (enormous) cost of building a model. Inference is the ongoing cost of using it: every chat reply, every generated image burns computing power. Inference is where AI's unit economics live. The cost of a fixed amount of inference has been collapsing, roughly ten times cheaper per year for equivalent capability, which is good news for users and complicated news for anyone who bought chips expecting yesterday's prices.
Insider selling #: Insider selling is when a company's own executives, founders, and directors sell their shares, which US law requires them to disclose. Clusters of heavy insider selling have shown up before many market tops. But it is a noisy signal: insiders also sell to diversify, pay taxes, or buy houses, and much of it happens on fixed schedules set in advance. Historians treat it as meaningful mainly when it coincides with extreme valuations and other warning signs.
IPO (initial public offering) #: An initial public offering is a private company's first sale of shares to the public. IPO waves are a late-cycle signal: promoters rush to sell while enthusiasm is high. The months around the 1929 peak, the 2000 peak, and the 2021 peak all saw record issuance of new investment vehicles. In the AI era, most of the equivalent action has happened in private funding rounds rather than public listings, which is one of the genuine differences from 1999.
Margin debt #: Margin debt is money borrowed from a broker to buy securities, with the securities as collateral. It cuts both ways: leverage amplifies gains in a rising market, but when prices fall, brokers demand repayment (a margin call), forcing investors to sell into the decline. Heavy margin debt helped turn the 1929 crash into a rout. US margin debt crossed $1 trillion for the first time in 2025.
Market breadth #: Market breadth measures how widely shared a rally is. Healthy bull markets lift most stocks; late-stage bubbles narrow to a handful of leaders while the average stock stalls. Breadth deteriorated 6 to 18 months before the 1929 and 2000 peaks. In 2025 the ten biggest S&P 500 companies reached a record share of the index, above 40%, which is why breadth is one of the loudest warning lights on this site's dashboard.
Market concentration #: Market concentration is the share of an index's total value held by its largest members. In 2025 the top ten S&P 500 companies reached about 40% of the entire index, a record, with AI-linked companies dominating that group. High concentration is not automatically fatal, but it means anyone who owns 'the market' through an index fund is, in practice, making a concentrated bet on a handful of AI-exposed giants.
Neocloud #: Neoclouds are the new generation of specialized cloud providers, CoreWeave, Lambda, Crusoe, Nebius and peers, built to rent out GPU computing for AI. They grew explosively by borrowing heavily, often using the GPUs themselves as collateral, and often depend on a small number of huge customers. They sit at the fragile center of the AI financing web: the fastest-growing, most leveraged, and most concentrated players in the ecosystem.
New issuance #: New issuance is Wall Street's supply response to enthusiasm: new stock listings, new bonds, new funds created to absorb investor demand. It is a reliable late-cycle indicator because promoters sell hardest when prices are highest. Investment-trust issuance peaked months before the 1929 crash; dot-com IPOs peaked with the 2000 top; SPACs peaked in early 2021. Today's equivalent runs heavily through private markets: venture mega-rounds and private credit for data centers.
Off-balance-sheet #: An off-balance-sheet arrangement is one where a company carries real economic exposure but does not record the assets and debt on its own balance sheet. This is not automatically improper: accounting rules decide who consolidates what, and the answer often turns on who controls the entity rather than who is exposed to it. The exposure is usually disclosed in the footnotes instead, which is where the interesting numbers tend to live. Enron is the notorious abuse. Meta's Hyperion joint venture is a current, disclosed and audited example: Meta reports a 20% stake and a stated maximum loss exposure rather than the project's assets and debt. See also VIE and special purpose vehicle.
Offtake agreement #: An offtake agreement is a long-term commitment to buy what a project will produce, signed before construction: a utility agreeing to buy a wind farm's power, or an AI lab agreeing to rent a data center's computing for 10 years. Lenders finance projects on the strength of these contracts. The catch in the AI build-out: many offtakers are startups whose own funding comes from the same ecosystem, so the 'guaranteed' revenue is only as solid as the buyer.
P/E ratio #: The price-to-earnings ratio divides a company's share price by its yearly profit per share. A P/E of 20 means you pay $20 for each $1 of current annual profit. High P/Es can be justified by fast growth, but extremes have marked every equity bubble: around 30 for the market in 1929, over 60 in Japan in 1989, and over 100 for many dot-com leaders in 2000. A 'forward P/E' uses next year's forecast profit instead of last year's actual.
Peak-to-trough #: Peak-to-trough describes the full extent of a decline: from the top (peak) to the bottom (trough), in both price and time. The Nikkei's peak-to-trough was about 80% over 13 years. The measure matters because partial declines can look survivable while the full journey was not, for anyone who had borrowed money along the way.
Price-to-sales ratio #: The price-to-sales ratio compares a company's total market value to its annual revenue, useful when profits are thin or absent (which is why it was the metric of choice in 1999). As a rule of thumb, above 10 times sales is historically expensive territory: the buyer is paying a decade of total revenue, not profit, up front. Several AI-era leaders trade well above that line.
Private credit #: Private credit is lending done outside the banking system, by investment funds such as Blue Owl, Apollo, and KKR. It has boomed since 2010 because it faces lighter regulation and discloses little. In the AI build-out, private credit funds tens of billions of dollars of data-center construction, often through separate legal vehicles that keep the debt off the big tech companies' own balance sheets, which makes the system's true leverage hard to see from public filings.
Project finance #: Project finance funds a single asset, a power plant, a toll road, a data center, on the strength of that asset's own expected cash flow. The debt is typically non-recourse, meaning lenders can seize the project but not the sponsor's other assets. It is an ordinary, century-old technique for building long-lived infrastructure, and there are good reasons to use it: it matches long debt to long assets and it stops one failed project from sinking the sponsor. The same features also mean the sponsor's reported debt understates how much building is going on in its name.
Rate base #: A regulated utility does not simply charge what it likes. A commission approves the assets it may recover from customers, the rate base, and the return it may earn on them. Bills then repay that investment over decades. This is why who pays for new power infrastructure is a regulatory question rather than a market one: if generation or transmission built for a large new customer enters the rate base and that customer's load does not arrive, the cost does not disappear, it is spread across everyone else's bills. Several states have written large-load tariffs specifically to stop that happening.
Residual value guarantee #: A residual value guarantee, or RVG, is a promise by a lessee that an asset will be worth at least a stated amount when the lease ends, and that the lessee will cover any shortfall. Because payment depends on a future condition, an RVG is disclosed as a contingent obligation rather than booked as debt. That makes it invisible in the headline leverage numbers and highly visible in a downturn, since the condition that triggers it, the asset being worth less than expected, is the same condition that hurts everything else at once. Meta has provided RVGs on the Hyperion joint venture with an aggregate threshold of roughly $28 billion that decreases over time.
Sale-leaseback #: In a sale-leaseback, an owner sells an asset to an investor and simultaneously signs a lease to keep using it. The seller converts a building into cash and a rent obligation. It is a standard corporate real-estate technique and is not inherently aggressive. What it changes is the shape of the balance sheet: a large owned asset and its debt are replaced by a stream of future lease payments, which are disclosed but sit outside the headline debt figure. In the AI buildout the same logic applies to whole data-center campuses.
Securitization #: Securitization pools income-producing contracts, mortgages, car loans, or data-center leases, into a separate entity that issues bonds backed by those payments. Rating agencies grade the bonds on the quality of the underlying tenants and the length of the leases, not on the developer's own credit. Data-center operators use it routinely: KBRA counted $48.69 billion of US data-center asset-backed and commercial-mortgage-backed issuance across 88 transactions from 2018 through May 2025. Securitization is well understood and heavily regulated. Its risk is concentration: the bonds are only as durable as the tenants' willingness to keep paying rent.
SPV (special purpose vehicle) #: A special purpose vehicle is a company created for a single purpose, typically to own one project and the debt that finances it. Used honestly, SPVs isolate risk (a failed project doesn't sink the parent). But they also keep debt off the parent's balance sheet, making the parent look less indebted than it functionally is. Enron's abuse of SPVs is the cautionary tale. Meta's 'Hyperion' data-center financing is a current example of the structure, disclosed but still off-balance-sheet.
Tightening (monetary policy) #: Central banks 'tighten' when they raise interest rates or withdraw money from the financial system, usually to fight inflation. Tightening makes borrowing costlier and speculation harder to sustain. It is the single most consistent bubble-killer in the historical record: in at least six of eight major bubbles since the 1840s, the peak came during or within roughly 18 months of a tightening cycle. The reverse, 'easing,' is when rates are cut.
Utilization #: Utilization is the share of time an asset is actually working. A GPU that runs jobs 5% of the time is 5% utilized, and 95% of the money spent on it is earning nothing. Surveys in 2026 found most enterprises running their own AI chips at 50% utilization or less, and one large study of cloud clusters measured average GPU utilization around 5%. Hyperscalers almost certainly do much better, but publish no audited numbers.
Valuation #: Valuation is the relationship between a company's price and its fundamentals (profits, revenue, assets). Metrics like the P/E ratio and price-to-sales ratio are valuation measures. High valuations are not a timing signal, expensive markets can get more expensive for years, but they set the stakes: the higher the valuation, the more future success is already paid for, and the further prices can fall when expectations reset.
Vendor financing #: Vendor financing is when a supplier lends its customers the money to buy its products. Telecom equipment makers Lucent, Nortel, and Cisco lent billions to shaky startups in the late 1990s so those startups could buy their gear; the vendors booked the sales as revenue immediately. Nine big vendors had at least $25.6 billion of such loans outstanding by end-1999, and consultants judged 30 to 40% of it 'at risk.' When the customers failed, the revenue proved partly fictitious and the vendors' stocks fell 90% or more. It is the closest historical template for today's AI financing loops, with important differences argued both ways.
VIE (variable interest entity) #: A variable interest entity is one that does not have enough equity of its own to fund its activities without outside support, so control cannot be read off the share register. US accounting rules therefore ask a different question: who has the power to direct the activities that most significantly affect its economic performance, and who absorbs its losses. That party, the primary beneficiary, consolidates it. Everyone else discloses their interest and their maximum exposure to loss instead. The judgement about who holds that power is where a great deal rides on relatively few words.