What it is. The rate at which an actor commits capital to positions that didn’t exist last year. Not reserves, not revenue, not spending in general — new-position formation per year. Free cash above a certain size is a failure signal; thrust is the opposite signal: capital that has already bought a position.

How we calculate it

Included: growth capex · M&A · strategic stakes · net new fund deployments · capitalized bets (a frontier training run is capex in everything but accounting) · take-or-pay compute and power reservations (rating agencies already treat these as debt-equivalent capital commitments).

Excluded: COGS, SG&A, utilities, debt service, dividends — and critically, maintenance capex. Buybacks are excluded here and tracked as their own signed channel: capital returned.

Depreciation, never amortization. The netting term is depreciation only — the proxy for physical wear-and-replacement. Amortization of acquisition intangibles is excluded: writing down a past deal’s goodwill has nothing to do with maintaining machines, and netting it produces artifacts (Broadcom’s capex minus its VMware amortization once printed a absurd negative thrust on this board — corrected under this rule).

Guarantees are not thrust. A financing guarantee (e.g. a vendor backstopping a customer’s debt) is contingent exposure — quasi-debt in rating-agency terms — not deployed capital. It enters the receipt as a contingent commitment and converts to thrust only as it actually funds. Equity stakes, by contrast, are cash deployed into new positions and count in full.

The maintenance split, honestly. The computable heuristic is maintenance ≈ depreciation, so thrust-capex ≈ capex − depreciation. It’s what makes Intel legible — capex ~$12.1B, D&A ~$12.4B, thrust ≈ zero — and it’s also known to distort during a buildout: depreciation reflects a small historical base while current spend is huge, and if AI hardware truly lives 2–3 years (the live depreciation-schedule debate), part of what looks like growth is replacement. We use the heuristic anyway, rough-is-fine, and flag hardware-heavy actors with an economic-depreciation caveat on their claims.

Prior art. Damodaran’s reinvestment rate already makes every extension we make — acquisitions counted, R&D capitalized, leases capitalized. Ghemawat adds the quality dimension we tag but don’t yet score: a dollar sunk into sticky, untradeable factors is more committed than a dollar in fungible form. Penrose adds the warning: thrust has a managerial speed limit.