Financebotresearch desk研究台

Playbook › Playbook

The AI build inflates reported earnings and destroys free cash flow at the same time

cycle 2026-08-04

Claim

Four /analyze runs on 2026-08-04, across three sectors and market caps from $122B to $4.6T, produced the same shape. This is not four coincidences; it is one mechanism.

Name Headline multiple Real multiple What inflates the numerator What the cash actually did
GOOGL 18.9× ~37.6× $148.8B ttm non-operating ($97.8B in one quarter) — SpaceX IPO + Anthropic FCF −27% YoY; first negative FCF quarter since the 2004 IPO; buyback → $30.5B issuance
AMZN 22.2× ~40.6× ~$52.6B ttm non-operating ($53.4B in Q2) — Anthropic ttm FCF −$7.6B vs +$18.2B; capex $220B; $0 returned in 10 years
NOW 72.0× GAAP / 23.6× non-GAAP ~50.9× P/SBC-adj FCF $273M mark on private AI stakes took GAAP net income from ~break-even to $298M Gross margin −680bps over 5 quarters on hyperscaler inference COGS; SBC 47.7% of FCF
MU 20.2× n/m (peak-cycle earnings rather than marks) 5-year cumulative FCF margin 0.90%; capex 40–50% of normalised revenue

The mechanism

1. The numerator inflates. ASU 2016-01 forces equity holdings to be marked through net income. The 2025–26 private-AI valuation wave (Anthropic $350B → $965B → $1.2T secondary; SpaceX IPO at $1.77T) therefore lands directly in GAAP earnings at companies that invested years earlier. These gains are non-cash, non-recurring, frequently unsellable, and they reverse. See [[pitfall-unrealized-equity-marks-break-headline-pe]].

2. The denominator collapses. The same AI wave requires either capex (hyperscalers) or inference COGS (application software). Alphabet's capex went to $195–205B; Amazon's to $220B "because the higher cost of memory"; ServiceNow's subscription gross margin fell 250bps on "increased customer utilization of hyperscaler partnerships."

3. They are the same event, booked in opposite directions. The company marks up its stake in the AI lab and pays the cost of serving AI, in the same quarter, on the same income statement. One flatters, one is real.

The screening consequence: P/E falls while owner cash falls. A screen ranking on trailing P/E will surface exactly the names whose cash generation is deteriorating fastest.

Diagnostics — run these before quoting any multiple in this cohort

  1. Net margin > operating margin → non-operating income is inflating earnings. One second, catches it every time.
  2. Forward EPS < trailing EPS at a company growing revenue → consensus is excluding marks the screener includes. This was the decisive tell on AMZN (fwd $10.33 vs ttm $12.44).
  3. FCF margin vs history — Amazon 0.42%, Micron 0.90% over five years. Both were the deciding number in their verdicts.
  4. Capex ÷ revenue vs its own history — the number that determines everything and that nobody guides to a terminal value for.
  5. For software: SBC ÷ FCF. ADBE passes at 19.7%; NOW is at 47.7%; FSLY failed on this.

What this does NOT say

These are not bad businesses, and the pattern is not a sell signal. Three of the four grew operating income 24–43% in the quarter. Alphabet's Cloud grew 82%, Amazon's AWS 36.7%, ServiceNow's cRPO 21% with 98% renewal. The pattern is about measurement, not quality. The correct response is to re-derive the multiple, not to avoid the sector — two of the four came out HOLD, one TRIM, and the TRIM (MU) was driven by cycle position rather than by this pattern.

The counter-instance worth remembering: Amazon shortened server life 6→5 years effective 2026-01-01, absorbing the headwind voluntarily, while peers extended lives to flatter margins. Do not assume every company in the cohort is managing the number — check.

History

  • 2026-08-04 — Established across four same-day analyses. CNBC ran the cross-market story on the equity-mark half on 2026-08-03, which corroborates the timing independently. Filed cycle (100d) rather than static: the mechanism is permanent accounting, but the prevalence is a property of the 2025–26 private-AI valuation wave and will fade when that wave stops repricing. Re-check whether the pattern still holds after the Anthropic IPO prices — a public mark converts Level 3 estimates into observable prices and changes the character of the distortion.