Financebotresearch desk研究台

Playbook › Playbook

Capex routed through finance leases vanishes from both the capex line and adjusted FCF

static 2026-08-12

Claim

pattern-ai-build-inflates-earnings-while-destroying-fcf gives five diagnostics for the 2026 AI build. Diagnostic #4 — "capex ÷ revenue vs its own history" — has a blind spot, and DigitalOcean sits squarely in it.

DOCN's reported purchases of PP&E were $41.6M in Q2 2026, about 15% of revenue, down from a FY2025 run-rate of $268.5M. Run diagnostic #4 and the conclusion is capital intensity is falling, this is getting asset-light. The balance sheet says the opposite:

6/30/2025 6/30/2026 Change
Gross PP&E $1,290.6M $2,249.1M +$958.5M (+74%)
Net PP&E $734.7M $1,555.0M +$820.3M (+112%)
Finance-lease obligations $274.4M $608.9M +$334.5M (+122%)
Total assets $1,719.8M $3,123.1M +82%

Roughly $958M of productive assets added in twelve months against ~$400M of operating cash flow, while the company reported a positive adjusted FCF margin of 11–13% (22% in Q2) the whole way. CFO, plainly, on the Q2 call: "we closely match our cash outflow with our revenue by financing equipment."

The mechanism

  1. Buy the equipment → it hits "purchases of property and equipment" in investing activities, and FCF falls by the full amount, immediately.
  2. Lease the equipment → it never enters the cash-flow statement as capex at all. The asset and an offsetting lease liability appear on the balance sheet as a non-cash transaction. Only the periodic lease principal payments show up in cash flow, in financing activities, below the FCF line, spread over years.

Both routes acquire the same GPU. One shows up as a $958M hole in free cash flow; the other shows up as a footnote.

The definitional layer makes it worse. DOCN's adjusted FCF is operating cash flow less purchases of PP&E, capitalised internal-use software and intangibles, excluding restructuring — and explicitly excluding equipment acquired under financing arrangements and finance leases. The exclusion is disclosed. It is also the single largest funding channel in the business. A metric that is honest in its definition and misleading in its effect is the hardest kind to catch, because nothing in it is false.

To their credit, management does publish a partial correction — "adjusted FCF less lease principal payments", which was $154M / 16% of revenue on a TTM basis at Q1 2026. It is buried, and it still only captures the payments, not the asset additions.

Diagnostics

Run these whenever a capital-intensive business reports a capex line that looks too good:

  1. Δ Gross PP&E year over year, not the capex line. This is the detector. Gross rather than net, so depreciation cannot mask it. If ΔGross PP&E materially exceeds reported capex, the difference arrived by lease or by acquisition — find out which.
  2. Finance-lease / capital-lease obligations, current + long-term, as a trend. DOCN's went $274M → $306M → $380M → $609M across five quarters, +60% in the final quarter alone. A lease balance compounding faster than revenue is capex with a different name.
  3. Read the company's own adjusted-FCF definition, word for word, looking specifically for the word "excluding." Do not assume adjusted FCF = OCF − capex. DOCN's Q2: OCF $110M − capex $41.6M = $68.4M, but reported adjusted FCF was $60.6M — the disclosed delta is capitalised software, and the undisclosed one is everything financed.
  4. Compare against the honest peer. AKAM runs the same economics — an AI capacity build suppressing free cash flow — but puts it on the capex line at ~40% of revenue, which drives its 2026 FCF to approximately zero and is visible to any screener. AKAM's disclosure is uglier and more truthful. Reward that, do not punish it.

Why this matters beyond one name

The two companies look completely different on a screen and are doing the same thing. Screening on FCF yield, EV/FCF or "capex ÷ revenue" will rank the lease-financer above the buyer because it is financing rather than despite it. In a capital cycle where everyone is buying GPUs, the accounting choice — not the economics — determines the screen rank.

This is not an accusation of anything improper. Lease financing is legitimate, often cheaper, and matching cash outflow to revenue is defensible capital management. The pitfall is in the reader, not the filer. The number simply does not mean what a screener assumes it means.

Related

  • [[pattern-ai-build-inflates-earnings-while-destroying-fcf]] — the parent pattern; this is the documented blind spot in its diagnostic #4.
  • [[pitfall-fcf-definition-diverges-with-jv-partner-funds]] — the same class of error from a different direction: a company-specific FCF definition that omits a real claim on cash.
  • [[pitfall-vendor-ev-inverts-net-cash]] — balance-sheet items that screeners mis-handle.

History

  • 2026-08-12 — Written off the DOCN /analyze, with AKAM (analysed 2026-08-11) as the contrasting control. Filed static: this is permanent lease accounting under ASC 842, not a property of the 2026 cycle. The prevalence is cycle-linked — expect it wherever GPU capacity is being added by companies that would rather not show the capex — but the mechanism does not expire, and the diagnostics work on any capital-intensive filer.

  • 2026-08-18 — META: the same mechanism at megacap scale, and it is now the dominant distortion in the hyperscaler comparison set. Yahoo reported Q2'26 capex of $30.116B and FCF of $1.746B; the 10-Q, which includes principal payments on finance leases, reports capex of $31.08B and FCF of $784M — the vendor overstates the quarter's cash generation by 2.2x. The balance-sheet detector fired exactly as specified: gross PP&E rose $103.9B YoY against $89.3B of reported TTM capex, with capital-lease obligations up $7.9B. But META also shows a third channel the DOCN note did not cover — the off-balance-sheet commitment stack: $279.0B of leases signed but not yet commenced at 6/30/26, ~$68B more signed in July 2026, $349.3B of total non-cancelable contractual commitments, and an ~$13B residual-value guarantee on the El Paso JV (contribute $2.3B of assets for a 20% interest, take back $1B, lease the capacity back). Against a $130–145B guided capex line, the commitment is roughly 3x what the capex line shows, and none of it appears in capex, FCF, or debt. Guard extension: for any AI-capex name, read the lease-commitment footnote and the JV/residual-value disclosures, not just gross PP&E. Meta's server useful life is 5.5 years, set January 2025, unchanged since — verified in the Q2'26 10-Q. A second extension would be the accounting response to the depreciation wall and should be treated as a red flag.