How to Read the AI Market Watch

Every week someone declares that the AI bubble has burst, and every week Nvidia ships more chips. Both camps are mostly watching share prices, and share prices are the least informative place to look. A 6% fall in the semiconductor index tells you investors are nervous. It doesn’t tell you whether that nervousness is about to become somebody else’s problem.

The AI Market Watch exists to answer a narrower question, and to answer it the same way every time: is stress in the AI investment cycle spreading into the wider financial system, and is the investment itself paying its way? Most of what happens on any given day is noise around those two questions. The job of each report is to say which bits weren’t.

This page sets out the rules. If you read one report, you should be able to read the next one and see exactly what moved, in which direction, and why it did or didn’t change the overall judgement.

Why the format matters more than the opinion

The temptation with a series like this is to write a fresh essay every day. That produces good prose and bad tracking. Indicators appear when they’re dramatic and vanish when they’re dull, the language of each verdict drifts, and after a month every report says “meaningful deterioration” because nobody defined what an improvement would look like.

A watch that can only ratchet upwards is an alarm, and alarms are easy to build. The harder and more useful thing is a fixed set of measures, a fixed scale, and published thresholds that work in both directions. When the situation gets better, the report should say so in the same words, in the same place, with the same weight as when it gets worse.

The firewall

The single most important idea in the series is the firewall: the boundary between stress that belongs to AI and stress that belongs to everyone.

AI companies and their financiers can have a terrible month without it mattering much to anyone else. CoreWeave bonds can trade at 77 cents on the dollar, a data-centre loan can fail to syndicate, and a leveraged AI hedge fund can blow up, and the rest of the corporate world can still borrow at normal prices. That is painful repricing inside a sector. It becomes a different kind of event when ordinary companies with no AI exposure start paying more to borrow, or can’t borrow at all, because lenders are now pulling back from everything.

The cleanest public measure of that boundary is the spread on ordinary corporate bonds: the extra a company pays to borrow over what the US government pays for a loan of the same length. The ICE BofA indices for investment-grade, high-yield and CCC-rated US bonds are published daily and free on FRED, which makes them the backbone of the broad-credit pillar. When AI-specific measures are deteriorating and those spreads are flat, investors are discriminating between borrowers. When those spreads move too, and stay moved, the stress is spreading.

The eight pillars

Every report scores the same eight pillars. Each one has a fixed set of indicators that appear in every report, whether or not they moved, so readers can see the quiet days as well as the loud ones.

Central-bank and regulator warnings from the BIS, IMF, Bank of England, Bank of Japan and others appear in the narrative as context. They aren’t scored. A central bank saying a risk has risen is useful information about what supervisors are seeing, but it isn’t itself a market event, and scoring it would double-count the evidence they’re reacting to.

The status scale

Each pillar gets one of five statuses, plus a trend arrow comparing it with the previous report.

The trend arrow is worse, unchanged or better. A pillar can be Stressed and improving, and that combination matters: it is often the most important thing a report says. Worse is shown as a red arrow pointing down, better as a green arrow pointing up, and unchanged as a grey square, so the direction reads the same way whether you see the colour or not.

For the market-priced pillars, the statuses are anchored to numbers so they can’t drift with the mood of the day. These are starting calibrations and will be reviewed as the series builds a history.

Where the two credit measures disagree, the worse one sets the status. The CCC spread is reported alongside as an early warning, because the weakest borrowers feel tightening first, but it doesn’t set the pillar on its own.

The rates status does not come from the latest close. It moves only when the 10-year has finished five consecutive trading days in the new band, or further from Calm than that band: five closes at or above 4.75% to reach Elevated, five closes above 5.25% to reach Stressed, and five closes back on the easier side of the line to move down. The same count applies at 4.25%. A single close through a threshold is reported in the change section and leaves the status where it is. Critical is not a yield level. It needs a failed auction or official intervention.

The non-market pillars (financing flow, leverage, private credit and AI economics) are scored against the definitions in the first table rather than numeric bands, because the evidence arrives as events and filings rather than as a daily price. Each report states the specific evidence behind those scores.

Carrying state forward

A scorecard is only useful if the next one can be compared with it, so every report records its scores in a machine-readable block in its frontmatter:

previous_report: 2026-10-30
scorecard:
  as_of: 2026-10-30
  stage: 3
  stage_entry_evidence: 2026-10-23
  economics: stressed
  pillars:
    ai_equities: elevated
    rates_macro: elevated
    broad_credit: calm
    ai_credit: stressed
    financing_flow: elevated
    leverage_liquidity: elevated
    private_credit: elevated
    ai_economics: stressed

The next report reads that block and the comparison is mechanical. The trend arrow on each pillar is the difference between this report’s status and the last one’s, not a feeling about the week. The “change since last report” section is the full diff: stage, economics status and all eight pillars, previous against current, with the date of the report being compared against stated at the top. A pillar whose status hasn’t moved but whose underlying indicators have is marked unchanged in status and the movement goes in the reason column, which is how the reader sees pressure building before a threshold is crossed.

The same block is what makes the symmetric persistence rule enforceable. The stage can’t move down until the entry evidence has been absent for the stated period, so the block records the date that evidence was last confirmed, and the report can say exactly when a downgrade becomes possible rather than leaving it to judgement.

Because every report keeps its block, the series also has a history. Below the diff, the change section shows a strip for the stage and for each pillar across the last 20 reports, with Calm at the top and Critical at the bottom so that worse always points down, as the trend arrows do. A flat line means nothing has crossed a threshold; a step down in red is a pillar getting worse, a step up in green is one getting better. Faint vertical lines mark each report, so uneven gaps between reports are visible. One week’s diff tells you what just moved. The strips tell you whether it is part of a pattern.

Where no previous report exists in this format, the baseline is back-scored from the most recent report in the earlier archive and labelled as such, so the first comparison is still against something rather than against nothing.

The whole archive, from 4 September 2026 onwards, has since been back-scored the same way and restructured into this format, so the strips have a history from the start of the series. The market-priced pillars and the stage 1 tripwire were scored mechanically from FRED and Yahoo Finance closes. The other pillars, and the evidence for stages 2 and 3, were read from each report’s own text, using only what that report knew on the day. Every back-scored report says so in its change section.

The transmission chain

The scorecard tells you where the pressure is. The transmission chain tells you how far it has travelled.

Market crashes that matter beyond the sector they start in tend to spread through the same mechanisms, in roughly the same order. Prices fall first. Then lenders start charging the sector more. Then investors who borrowed to hold those assets are forced to sell, and their selling pushes prices down further. Then the losses and the caution spread to lending in general, and companies with nothing to do with the original story pay more to borrow. Finally, in the worst cases, the plumbing of the financial system itself stops working.

Each of those steps needs a mechanism, not just a mood. A share-price fall becomes a credit problem only if someone has lent against the shares, or against the revenue the shares were priced on. Credit stress becomes forced selling only if someone is leveraged. That is why the chain is useful: at each stage it says what has to be true for the next one to happen, and what evidence would show it happening. Most corrections stop at stage 1 or 2. The chain exists to tell you when one hasn’t.

What each stage looks like

Stage 0: expensive but stable. Valuations are stretched and a handful of companies dominate the indices, but money is still easy to raise and nobody is selling because they have to. Being at stage 0 is not a prediction that nothing will happen. It means nothing has happened yet that would make the high prices somebody else’s problem.

Stage 1: AI equity correction. Investors are revising down what AI companies will earn. It is painful for shareholders, but on its own it is contained: equity holders absorb the losses, and a falling share price doesn’t stop a company paying its debts. The dot-com bust of 2000 to 2002 was overwhelmingly a stage 1 event for most of its length: share prices fell by more than three-quarters while most of the financial system carried on.

Stage 2: AI-specific credit stress. Lenders, not just shareholders, now think AI is riskier. AI borrowers pay more than comparable non-AI borrowers, their existing bonds fall below face value, and new deals are harder to place. This matters because debt has deadlines. A company whose shares fall can wait; a company that can’t refinance can’t. The telecom debt crisis that followed the dot-com crash, ending in the defaults of WorldCom and others in 2002, is the precedent: the equity story turned into a debt story when companies that had borrowed to build capacity couldn’t fill it.

Stage 3: forced deleveraging. Investors who borrowed to hold AI assets are being made to sell, by margin calls, by their lenders cutting credit lines, or by losses breaching their risk limits. Forced selling is different in kind from ordinary selling: it happens regardless of price, and it hits whatever the seller can sell, not just what has gone wrong. That is how stress jumps from one market to another. The collapse of Long-Term Capital Management in 1998 and the “quant quake” of August 2007, when many hedge funds running similar strategies were forced to cut positions in the same week, are both stage 3 events.

Stage 4: firewall breached. The stress has escaped AI. Ordinary companies with no AI exposure are paying more to borrow, or can’t borrow, because lenders are pulling back from everything. This is the point at which an AI downturn becomes an economic event: higher borrowing costs and less available credit slow hiring and investment across the economy. The widening of corporate spreads through 2008 is the clearest example.

Stage 5: systemic liquidity event. The markets that the financial system uses to fund itself stop working properly: banks stop lending to each other, the market for US government debt becomes disorderly, funds stop letting investors withdraw, and central banks step in. September 2008, March 2020 (when even the Treasury market seized up and the Federal Reserve intervened on a vast scale) and the UK gilt crisis of September 2022 (when pension funds’ forced selling led the Bank of England to buy gilts) are all stage 5 events.

The stages aren’t a staircase

The chain describes the usual order, not a required one. The UK gilt crisis went from forced selling straight to official intervention without broad corporate spreads breaching anything like our stage 4 tripwire. Evidence for a higher stage can appear while a lower one is still forming.

So the report marks each stage separately as confirmed, partial or not present, and the headline is always the highest stage whose entry evidence is confirmed. Partial evidence for the next stage up is shown in the chain diagram and discussed in the judgement, but it doesn’t move the headline.

How the stages relate to the pillars

The stages aren’t scored independently of the pillars. Each stage reads from the pillar where its evidence lives:

A Stressed main pillar is usually what puts a stage on watch. The entry evidence in the table is what confirms it.

Reading the chain diagram

Each report draws the chain as a line of five numbered stages. Stages whose entry evidence is confirmed are filled black. The headline stage is filled red, and the red route runs from stage 1 to it, with a dotted arrow pointing at the next stage up. Stages with partial evidence have a dashed outline. Stages not present are left as outlines. Under each stage is the date its evidence was confirmed, and under the headline stage the earliest date it could move down.

Tripwires, and the rule that makes them work in both directions

The stage moves on published tripwires, chosen in advance so that the judgement can be checked against the evidence rather than against the tone of the day’s news. Every stage has an entry condition and an exit condition, and both carry a persistence period: the condition has to hold for that long before the stage moves.

The persistence condition does most of the work. A single day of spread widening after a bad earnings print is common and usually reverses. Five days of it means lenders have changed their minds. For the event-driven stages, a month is long enough for a one-off to be seen as a one-off, and short enough that the chain doesn’t lag reality by a quarter.

The same rule runs in reverse, and that is deliberate. A stage moves down when its entry evidence has been absent for the persistence period. Because each report records the date the evidence was last confirmed, it can say exactly when a downgrade becomes possible, and readers can check it. Downgrades are reported with the same prominence as upgrades. Without that symmetry, the series would only ever record bad news, and readers would be right to stop trusting it.

Worked example

Stage 3 was confirmed on 23 October 2026, from prime-brokerage data showing synchronised hedge-fund losses and deliberate deleveraging. A week later there was no new evidence, but the stage stayed at 3: the exit condition is a month without new evidence, so the earliest possible downgrade is 23 November. If new forced selling is reported before then, the clock restarts from that date. If not, the report on or after 23 November moves the headline to stage 2, provided stage 2’s own evidence is still confirmed, and says so in the headline.

The second axis: AI economics

The transmission chain asks whether stress is spreading. It doesn’t ask whether there’s a hole underneath it in the first place, and that turns out to be a separate question.

AI financing can be under pressure because rates are high and lenders are cautious, even while the underlying investment is perfectly sound. The reverse is also possible: the money can keep flowing while the returns quietly fail to show up. Alphabet’s second consecutive quarter of negative free cash flow, with search advertising partly cannibalised by its own AI Mode, is that second kind of signal. The economics of the investment are getting worse, while nothing in the credit market has yet reacted.

So each report carries a separate AI economics status on the same five-level scale. It draws on the AI economics pillar: whether hyperscaler capex is being funded from operating cash or from outside money, whether cloud revenue growth is keeping pace with spending, whether contracted capacity is holding up, and whether projects are being delayed or cancelled.

The two axes together are the judgement. A crash needs both a hole and a way for the hole to spread. Weak economics with an intact firewall usually means a slow, grinding deflation of expectations. Strong economics with a breached firewall means a financing shock that the underlying business can probably survive. The dangerous combination is both at once.

How evidence is weighed

Every claim in a report carries one of three labels, so readers can see how much weight it can bear:

• Confirmed means primary data or a company or official filing: an ICE BofA index value, a quarterly report, a central-bank publication.
• Reported means credible journalism from a named outlet, typically a single source that hasn’t yet been confirmed by the parties involved.
• Estimate means a projection or a management claim that hasn’t been independently checked, such as an analyst forecast or a CEO’s view of industry economics.

Market data taken from FRED or from Yahoo Finance’s index history counts as confirmed; in the source list it is labelled official-data-series or market-data-series.

Only confirmed evidence can move a stage on its own. Reported evidence can move a pillar status, and can move a stage when several independent reports point the same way. Estimates go in the narrative and nowhere else.

Each report also has to fill both halves of its headlines: the bad and the good. If there is no good news on a given day, the report says “none” rather than leaving the heading out. That small rule forces the question to be asked every time, and it’s the main defence against a series that only notices bad news.

Discrimination and contagion are kept separate throughout. Investors demanding 115bp from AI issuers against 78bp for the wider investment-grade market is discrimination: lenders are pricing a specific risk, which is what lenders are supposed to do. Contagion is when the 78bp moves too.

What a report looks like

Each report follows the same order:

1. Headlines. The stage and the AI economics status, then the week’s events in two halves: the bad and the good. Anything that looks like a signal but isn’t goes underneath, as “not what it looks like”.
2. Scorecard. All eight pillars, each with its status, trend and a one-line reason.
3. Change since last report. The diff against the previous report’s scorecard block: stage, economics status and every pillar, previous against current, with the date being compared against. This is where the arrows in the scorecard are accounted for. Below it, the history strips show the stage and each pillar across the last 20 reports.
4. Transmission chain. Stages 1 to 5 marked confirmed, partial or not present, with the date the entry evidence was last confirmed and the earliest date a downgrade becomes possible.
5. Judgement. A few paragraphs on how the evidence was weighed against the existing picture. The headlines say what happened; the judgement says what it means.
6. Key indicators. The fixed indicator set, grouped by rates and macro, broad credit, AI-specific credit and quarterly financials, with latest value, previous value, change, tripwire and source. Below the table, a 20-report history of the market-priced indicators against their thresholds. Capex guidance and free cash flow are reported on each company’s own quarterly cadence rather than the report date, so they stay in the table, carried forward from the report that last cited them, rather than being drawn on the same daily axis. Then what would change our view: the specific evidence that would make things worse, and the evidence that would make them better.
7. Sources. Numbered, dated and linked.

When the week’s news depends on a concept that needs explaining, an explainer follows the sources, or links to a standing one.

New reports start from a blank copy of this structure, the report template, and are checked against it before publication: every section in order, a scorecard block that matches the tables, trend arrows that match the previous report, the fixed indicator set, and a numbered source for every footnote.

Each numbered section carries a “What does this mean?” link to the part of this page that explains it. Underlined terms throughout the report, such as five-year CDS or free cash flow, link to the AI Market Watch explainers.

Reports are published on trading days when there’s something to score, and not otherwise. A quiet day with an unchanged scorecard is still worth publishing if a reader would otherwise wonder whether the silence meant something.

What this isn’t

This is not investment advice, a price forecast or a call to buy or sell anything. It tracks the mechanisms through which an AI investment downturn could spread into the wider financial system, and reports honestly when those mechanisms are and aren’t operating. The thresholds are judgements, published so you can disagree with them on specific grounds rather than on vibes.

When a report gets something wrong, through a figure later revised or a story that turns out to be inaccurate, the correction goes at the top of the next report and the original is annotated rather than quietly edited.

Next: all AI Market Watch reports