Skip to content

KB/framework

AI capital cycle watch

Last verified

The thesis in one paragraph

Capital cycles run a repeatable shape: a flood of investment chases a narrative, a physical buildout follows (often years behind the capital commitment), output prices eventually reflect the new capacity, and when the financing character of the buildout was aggressive enough, the unwind is a credit event before it is an equity event. The AI buildout tracks this shape across three parallel cascade tracks — hardware supply, power/energy infrastructure, and direct demand signals — any one of which can arm a single-file sequence of downstream events. The watch maintains 23 explicit, admission-gated triggers across those tracks and reduces them to one composite state.

The four measurement altitudes

The watch reads the cycle at four altitudes, top of the funnel to bottom:

  1. Committed capital — capex level, funding source, and debt structure. The character of the money (cash-funded vs. debt-funded) matters more than the headline dollar amount.
  2. Physical inputs — the hard constraints capital has to clear before it becomes usable capacity: chip lead times, chip inventory, power-equipment lead times, interconnection queues.
  3. Capacity — the built, energized, priced infrastructure itself (datacenter capacity, colocation lease rates) — distinct from the compute sold out of that capacity.
  4. Output price — what the buildout ultimately sells for: compute rental rates, inference token prices. The verdict altitude — slowest to move, hardest to narrate around, because it’s a real transaction price, not a forecast.

A healthy boom shows output prices still rising while physical inputs stay scarce. The turn shows up first as physical constraints easing, and confirms once output prices roll over.

Three cascade tracks converging on one sequence

Three tracks run in parallel, any one of which can arm the converged sequence:

Any track arming the convergence point starts a single-file sequence: output prices roll over, credit reprices, collateral markdowns force a depreciation reckoning, capex guidance gets cut, an index-level guidance-miss event fires, the macro feedback loop engages, and — at the far end — a durable-asset inheritance phase opens for whoever is left with a shopping list. A behavioral basket (insider selling, IPO- window quality, private-credit inflows, secondary-market depth) runs alongside as a noisy, weeks-to-months-early leading layer.

Which track arms the sequence changes what kind of bust it is: a silicon- or power-led arming looks like a classic supply glut (slower, more wreckage to inherit); a demand-led arming is faster, gives less warning, and leaves less capacity behind to buy cheaply.

What the trigger table is

23 triggers (T1-T23) are declared as config-driven wires, one per observable in the watch doc’s founding table. Each wire carries a track, a cascade position, a threshold, a current status, and the evidence behind that status. A candidate only earns a wire when it has a public observable, a stated review cadence, a numeric or binary threshold, and a defined position in the causal chain — no vibes-based additions.

Most wires are manual: an operator sets the status, and a status change ships as a config-edit PR citing its evidence. Automated collectors land wire-by-wire as they ship; the wire shape stays the same either way. Statuses: armed (watch set, not fired), amber (partial progress — e.g. one of two required quarters), fired (cleared), no_data (honestly unmeasured, never a stand-in for “clear”), and calendar (one wire — a dated supply-mechanics event that is informational only and never escalates the state).

Auto wires

T13 (NVDA inventory growth exceeding revenue growth) is the first wire to flip auto: SEC EDGAR’s free quarterly financial filings give a real, audited read on both series, so a session-typed evidence string is no longer the best source. The read is fully mechanical — no narrative, no judgment call — and replaces the manual status once at least three overlapping quarters of inventory and revenue data have accrued; before that, the wire shows its founding evidence as an honest fallback rather than going blank. T10 (hyperscaler capex guide-down) stays manual — a guide-down is a spoken earnings-call statement no filing encodes directly — but gains an auto companion read: the capex-to-operating-cash-flow ratio per hyperscaler, computed from the same audited filings. That companion can only escalate the wire (never quietly downgrade a status an operator already set), and it never substitutes for the guide-down call itself.

T9 (useful-life shortening) flips auto next: SEC EDGAR’s full-text filing documents (not structured facts this time — a footnote’s plain prose) give a real read on each hyperscaler’s + CoreWeave’s stated server/GPU useful life. Each filer’s two most recent 10-K/10-Q filings are compared; a genuine shortening at any filer fires the wire, a disclosure that CHANGED but couldn’t be turned into a clean number escalates to an honest “manual read needed” rather than a fake fire, and a filer lengthening its schedule further stays armed with an earnings-quality note attached. Same founding-evidence fallback as T13 — until at least one filer has two filings on record, the wire shows its founding read.

T16 (neocloud insider selling cluster — the behavioral basket’s leading tell) is the third auto wire: SEC EDGAR’s Form 4 filings give a real, timely read on which insiders are selling at CoreWeave, Nebius, and IREN. The wire counts distinct insiders with at least one open-market sale at a single covered name in a trailing 90-day window; three or more fires the wire, exactly two escalates to an honest partial read, and fewer stays armed. A per-name coverage state travels alongside the count — this matters because Nebius and IREN are foreign issuers, and a genuinely Section-16-exempt filer’s absence of Form 4s must never be read as “nobody’s selling, all clear.” Live-checked at build time: all three names currently file Form 4s, so today the wire is a genuine three-name read. Same founding-evidence fallback as T13/T9 — until at least one covered name has a Form 4 on record, the wire shows its founding read.

T14 (token-consumption growth decelerates — the demand track’s one direct observable) is the fourth auto wire and the first fed by a non-EDGAR source: OpenRouter’s own weekly aggregate cross-model token-volume read. The wire fires on two CONSECUTIVE quarters of decelerating growth — the growth rate slowing, not necessarily going negative — against the founding ~33T-tokens/week, 4-5x-year-over-year baseline; a single quarter of deceleration escalates to an honest partial read, and stable-or-accelerating growth stays armed. This source carries its own selection caveat, distinct from the exemption question T16 worries about: OpenRouter is one router among several and skews toward indie/developer traffic, so its number is read as a DIRECTION signal, not a precise census-level figure — that caveat rides every response alongside the computed read. Same founding-evidence fallback as the other three auto wires — until enough quarters of real observations have accrued, the wire shows its founding read.

T5 (H100 rental rate sustained below the Oct-2025 low — the “verdict altitude” output-price observable) is the fifth auto wire and the second fed by a non-EDGAR source: getdeploying.com’s structured per-model GPU-pricing listings, which turned out to carry something better than expected — a genuine 12-month reservation-rate tier across a broad panel of providers, a real free 1-year contract-rate proxy rather than only spot pricing. The wire fires on two CONSECUTIVE calendar months where that reservation-rate median sits below $1.70/hr (the research baseline’s dated low); a single month’s dip escalates to an honest partial read — matching the watch doc’s own framing that one dip is noise and a sustained lower low is the turn — and a reading at or above the threshold stays armed. Same founding-evidence fallback as the other four auto wires — until enough months of real observations have accrued, the wire shows its founding read.

T6 (frontier token pricing cut without a capability jump — the other “verdict altitude” observable, alongside T5’s rental rate) is a DIFFERENT shape from every other auto wire: it stays manual, never fires from automation. The threshold has two parts — a mechanical “frontier-tier price cut of more than 50% within roughly two quarters” and a judgment call, “without a capability-generation jump justifying it” — and only a human reviewing what actually shipped can answer the second part; a steep price cut can be entirely explained by a genuine new model generation, or it can be a price war, and no automated read can honestly tell those apart. So the automated half — tracking the newest-released model in each of Anthropic’s, OpenAI’s, and Google’s flagship lineage via OpenRouter’s public pricing catalog, and comparing today’s price against its own trailing six-month peak — can only DETECT a qualifying cut and escalate the wire to an honest partial read; the fire itself stays an operator call once the capability-jump question is answered. This is the same hybrid pattern the hyperscaler capex-guide-down wire (T10) uses: the wire is about the judgment, the automation is a same-PR pre-signal that can only escalate it, never decide it.

The composite read

The composite state mechanizes the watch doc’s own standing rules: one firing gets noted, two firings escalate to weekly full-pass monitoring, three or more behavioral-basket wires flashing tighten the review cadence even with no track firing, and any firing at or past the point where credit starts repricing puts the watch into event mode (same-day re-review). The state is one of none, watch, elevated, or high — how urgently the watch needs attention, never a directional call on price.

Where it surfaces