The First Invoice
Two weeks ago we published the arithmetic of Big Tech's stretched depreciation schedules and said the bill was coming. Q2 earnings just delivered the first installment. Here we score our own note, line by line — and post the balance still outstanding.
Every figure is drawn from the companies' own SEC filings released this earnings cycle and is as-of dated. Fact ("the filing says…") is separated from estimate ("our model implies…") throughout. This note adjudicates a prior published claim out-of-sample; the tally in §2 states plainly what is confirmed, what is open, and what would falsify us. Independently red-teamed against primary filings before release; §7 documents the sourcing standard.
§0 How to read this invoice
This section is the reader's guide. If you already live in 10-Ks, skip to §1 — nothing here is new to you. If you don't, five minutes here makes the rest of the note self-explanatory.
Depreciation is how a company spreads a purchase over the years it expects to use it. Buy $12bn of servers and you don't expense $12bn this year; you expense a slice of it each year across the machine's "useful life." The size of that slice is a choice. On a three-year life the slice is $4bn a year. On a six-year life it is $2bn a year. Same machines, same cash out the door — but the six-year version reports $2bn more profit this year, every year, until the schedules catch up.
The useful life is an estimate management selects, and changing it is legal. It is disclosed, justified to auditors, and applied prospectively — the past is never restated; the benefit lands entirely in current and future quarters. Between 2020 and 2025 the companies in this note disclosed thirteen separate server-life extensions worth roughly $24bn of first-year earnings benefit — and exactly one change in the other direction. Nothing in this note alleges fraud. The question is not legality; it is whether the reported profits mean what a reader assumes they mean.
One ratio does most of the work in this series: capex divided by net income. Call it the dose. It measures how much a company spends on hardware per dollar it earns. When the dose is low — Apple's is 0.07× — the depreciation schedule barely matters; earnings are what they look like. When the dose is high — Meta now spends about $2 for every $1 it earns — the schedule is the earnings, and the choice of clock decides what the bottom line says. Our July finding: rank companies by dose and you have ranked how overstated their earnings are, nine concordant pairs out of ten.
Why the bill is already fixed. A "vintage" is one year's purchases. The record hardware vintages of 2024–26 are already bought; their depreciation through 2028 is baked in the moment the machines land. The only question the schedule answers is when the expense shows up. That is why we call it an invoice: the amount exists; only the due date was chosen.
§1 The entry we opened in July
The July note made four claims a reader could hold us to. Stated as ledger entries:
- The bill accelerates. As the record 2024–25 vintages land, reported depreciation must grow sharply at every high-dose name — schedule or no schedule.
- The spend continues. Guided capex holds or rises; the model assumed Microsoft would spend ~$115bn in FY2026 and the four hyperscalers roughly $630bn (model basis) in calendar 2026.
- The clocks stay long. Companies keep five-to-six-year server schedules through the fastest hardware-obsolescence cycle in computing history — and the pressure is to extend, not shorten. Amazon's February 2025 cut was the lone exception; we said watch whether anyone follows it down.
- The dose keeps rising, so the gap widens. Capex grows faster than income at every hyperscaler, pushing each name up the distortion curve toward Oracle's end of it — and the reported-versus-economic earnings gap widens through 2028.
Claims 1, 2 and 4 are testable every quarter. Claim 3 is testable every time a 10-K or 10-Q appears. This earnings season produced a Microsoft 10-K (filed the same day it reported — a full year-end filing, not a press release), an Alphabet 10-Q, and releases from Meta, Amazon and Apple whose 10-Qs all followed within a day. Every document needed to post a complete first tally is on file.
§2 The tally
| Claim (July 17) | What Q2 delivered | Stamp |
|---|---|---|
| 1 · The bill accelerates | Depreciation growth at the four: Microsoft +56% (FY26, 10-K), Meta +46% (Q2 D&A), Alphabet +43% (H1, 10-Q), Amazon +42% (Q2 PP&E depreciation, 10-Q segment table — AWS alone +67%). Control: Apple +16% (9M). | CONFIRMED |
| 2 · The spend continues | Microsoft actual FY26 capex $115.9bn vs the model's $115.0bn assumption — a 0.8% miss. Alphabet H1 capex +103% (~$161bn pace). Amazon TTM capex $169.0bn, +64%. Meta raised the floor of its guide to $130–145bn. | CONFIRMED |
| 3 · The clocks stay long | All five checks closed on filings: Microsoft's 10-K, no server-life change + a disclosure upgrade (§4); Alphabet's 10-Q, no change, six years holds; Meta's 10-Q, 5.5 years holds; Amazon's 10-Q, no further change; Apple's 10-Q, none disclosed. No company followed Amazon down — and per Microsoft's own call, one clock stretched further: buildings, 15→25 years eff. FY2027 (press-sourced; filing check FQ1-27). | CONFIRMED |
| 4 · The dose keeps rising | Microsoft 0.63× → 0.87× (FY26). Alphabet 0.69× → ~1.19× (H1, ex-gains). Meta 1.15× → ~1.96× (Q2). Amazon 1.70× → ~2.34× (TTM, ex-gains). Every arrow points right. Control: Apple 0.07×, capex falling. | CONFIRMED |
Measurement windows differ by company (fiscal years and quarter-ends differ); each figure's window and source filing are stated in §7. Doses marked "ex-gains" strip disclosed non-operating investment gains — Alphabet's and Amazon's second quarters both contain large ones (§5) — at the 21% statutory rate. Every claim is scored against documents on file; the Meta, Amazon and Apple 10-Qs all landed within a day of their releases and are scored here.
Four confirmations, zero refutations, no open entries — the tally closed when the Meta, Amazon and Apple 10-Qs landed within a day of their releases and every policy note held. On the evidence a single season can produce, the July arithmetic is running ahead of schedule — and the strongest single item is the one we least expected to land within one per cent.
§3 The bill is arriving
The core forward claim of the July note was mechanical, not directional: the vintages are bought, so the depreciation must come. One season in, it is coming at every name.
Sources: Microsoft FY2026 10-K; Meta Q2-26 8-K Ex-99.1; Alphabet Q2-26 10-Q; Amazon Q2-26 10-Q segment table (0001018724-26-000026 — PP&E depreciation ex-corporate, $9.77bn→$13.87bn; the release's bundled D&A line grew +31%); Apple FQ3-26 8-K Ex-99.1.
Microsoft is the cleanest read because its fiscal year just closed and the full 10-K is on file. Depreciation expense ran $15.2bn → $22.0bn → $34.3bn across FY2024–26 — 56% growth in the latest year. The servers-and-network gross book grew $83bn in one year (to $215.9bn), and another $26.7bn of equipment purchases sat unpaid in accounts payable at year-end, nearly four times the prior year's balance. The machine is being bought faster than it is being paid for, let alone expensed.
Source: Microsoft Forms 10-K, FY2019–FY2026 (depreciation expense per PP&E note); house-clock marker from the July model (v3), server-only delta, pre-tax.
Two honest cautions on this exhibit. The red marker is a model figure — the July note's estimate of what FY2026 depreciation would be on a three-year server clock — and it is server-only, so it is added to the reported total rather than replacing it. And 56% growth under the current schedule partially reflects Microsoft recognising its bill faster as huge vintages land even on six-year math: acceleration was always going to happen; the dispute is about the size of the remaining deferral, which the model puts at roughly $9bn pre-tax for FY2026 alone and growing through 2028.
§4 The server clocks held. A building clock stretched.
Claim 3 said the pressure runs one way: toward longer schedules. This season delivered both halves of the proof — no server clock moved anywhere in the group, and the one clock that did move moved <em>longer</em>, in an asset class the July note never even covered. The details deserve to be reported straight, because two of them count in the companies' favour.
Microsoft held its schedule and improved its disclosure. For years Microsoft was the least transparent name in the group: no server-specific line anywhere, server hardware buried in "computer equipment" with a 2-to-6-year range. The FY2026 10-K changes that — the policy note now reads "servers and network equipment, two to six years," and the PP&E table breaks out a "Servers, network equipment, and software" line for the first time ($215.9bn gross). The range is unchanged and there is still no point estimate, but naming the category is a real step toward the disclosure standard Amazon and Meta already meet. Credit where due.
Alphabet held six years, again. The Q2 10-Q discloses no change in estimate — and under GAAP a change must be disclosed in the quarter it occurs, so silence is confirmation. Its 60%-servers footnote, the disclosure our Alphabet share estimate rests on, is restated as of June 30 unchanged.
And on the same call, Microsoft stretched a clock the July note never covered — buildings. Per its earnings call (press reports; Microsoft publishes no transcript we cite from primary), Microsoft extended the useful life of datacenter and office buildings from 15 to 25 years, effective FY2027 — and because longer building lives push more future datacenter leases from finance leases (which count in Microsoft's capex-inclusive guidance) to operating leases (which don't), the calendar-2026 capex optic fell from roughly $190bn to roughly $175bn. CFO Amy Hood was explicit that spending itself is unchanged: "outside of this useful-life impact, our calendar year 2026 capex investment expectations remain unchanged." Nothing here touches servers, and buildings genuinely last decades — but the ledger should record what happened: the group's fourteenth disclosed life extension against one reversal, this one announced the same week the depreciation bill visibly arrived, with the side effect of making guided capex look $15bn smaller while every dollar still goes out the door. Sourced from three named press carriers pending the FQ1-2027 filing, and labeled accordingly.
Amazon's reversal remains the only reversal. Eighteen months after cutting a subset of servers back to five years — and realising double the cost it guided for — Amazon still stands alone. The remaining checks closed as this note went to press: Meta's 10-Q (filed July 30) holds the 5.5-year schedule with no change in estimate — and files a sharper figure than its release carried, servers-and-network depreciation of $4.62bn for the quarter, $9.01bn for the half; Amazon's 10-Q (July 31) discloses no further change ("we review the useful lives of equipment on an ongoing basis" is all it says); and Apple's (July 31) discloses none, which under the disclosure rules means none occurred.
The scoreboard read: on servers, the ratchet has paused, not reversed — and off servers, it is still ratcheting. No server schedule stretched this season, which removes the easiest fuel for the cynical version of this thesis — but no company moved toward the hardware cycle either, and the one life change announced anywhere in the group went the familiar direction. Holding a six-year clock while NVIDIA ships a new architecture every two years is a choice that compounds: the July note's arithmetic assumed exactly this stasis, and the arithmetic is what Q2 confirmed.
§5 The dose keeps rising
The July note's cleanest finding was that a single ratio — capex divided by net income — ranks the earnings distortion across companies. Q2 moved every company's ratio in the same direction: up.
Filled markers: FY2025 dose vs FY2025 overstatement, per the July note (v3 model). Open markers: latest disclosed dose (Microsoft FY26 annual; Alphabet H1-26 and Amazon TTM ex- non-operating gains at 21% statutory; Meta Q2-26). Vertical position is held at the FY2025 overstatement; the re-mark of overstatement itself requires full-year filings. Apple from FQ3-26 9-month figures.
Walk the arrows. Microsoft moved from 0.63× to 0.87× on full-year numbers — the gentlest move, on the strongest income growth (net income +31% to $133.7bn), and still a one-third jump in dose. Alphabet roughly doubled its H1 spend and its dose rose from 0.69× toward ~1.2×. Meta now spends about $1.96 per dollar earned in the latest quarter — near double its full-year 2025 ratio, with net income falling (Q2: $15.8bn vs $18.3bn, weighed by $3.6bn of legal and severance charges) while the floor of its capex guide rose for the second time this year: $115bn in January, $125bn in April, $130bn in July, per its own successive releases. Amazon reached ~2.34× on trailing numbers. In July we wrote that Oracle — at 3.26× and 25.8% overstatement — is what a hyperscaler looks like when capex outruns earnings. The group is walking toward it.
Two companies' headline profits this quarter illustrate why the dose must be computed with care — and both cut the same way. Alphabet's H1 net income of $174.8bn includes $135.8bn of gains on equity securities (primarily unrealised, per its own MD&A); Amazon's Q2 net income of $62.6bn includes $53.4bn of pre-tax non-operating income, "primarily from our investments in Anthropic." Strip the gains and the underlying engines earned about $67bn and $72bn respectively on the windows shown. We flag it because a casual reader of either release would conclude earnings comfortably cover the capex. On operating earnings, they don't: that is precisely what a negative free cash flow means.
Source: Amazon Q2-2026 earnings release (8-K Ex-99.1, 0001018724-26-000024), supplemental TTM series. Free cash flow figures as defined and reported by Amazon (Ex-99.2).
Amazon is the starkest exhibit in the series. AWS grew 37% — its fastest in eighteen quarters — and operating income rose 43%. The boom is real. And trailing free cash flow still went from +$18.2bn to −$7.6bn, because trailing capex reached $169bn, up 64% — an increase that, in the release's own words, "primarily reflects investments in artificial intelligence." The best quarter AWS has had in four years did not cover the invoice for building it.
Meanwhile the queue of hardware that hasn't even begun depreciating keeps growing:
Sources: Alphabet Q2-26 10-Q (assets not yet in service, PP&E note); Microsoft FY2024–26 10-Ks (purchases of PP&E in accounts payable).
Alphabet's "assets not yet in service" reached $122.8bn — up from $78.6bn six months earlier. That is hardware bought and owned on which the depreciation clock has not started. It is the purest picture of the series' central idea: the bill exists; only its due date was chosen. When those assets enter service, the depreciation arrives at whatever schedule is then in force — and it arrives on top of the acceleration already visible in §3.
And the control group behaved exactly as a control should. Apple posted a record June quarter — revenue $109.4bn, up 16%; EPS $2.02, up 29% (of which about $0.11 came from tariff refunds, worth noting) — while its nine-month capex fell, from $9.5bn to $6.8bn. Dose: 0.07×. Apple's depreciation schedule could halve or double and its earnings would barely notice. That is what an income statement looks like when the accounting choice doesn't matter — and it is the cleanest evidence that for the other four, it does.
§6 Balance carried forward
What would change our mind. The tally stays honest only if the falsifiers stay posted. Four things that would score against the July note in coming quarters: a second company following Amazon's schedule cut (would confirm the hardware clock but shrink the forward deferral we compute); depreciation growth decelerating at high-dose names while capex still grows (would suggest the vintage model mis-times recognition); a sustained capex decline group-wide (would shrink the 2027–28 gap directly — watch Meta's next guide); or disclosed server-life studies with utilisation data justifying six years (would weaken the two-year-architecture argument at its root). None appeared this season. All four remain open entries.
The balance outstanding. Nothing this quarter reduced the July estimate: roughly $221bn of depreciation deferred out of 2026–28 earnings across the five profitable names (band $192–240bn), with the FY2025 overstatement standing at 4.4–17.2% per company. The acceleration now visible is the early portion of that number arriving — on schedule, on the companies' own filings.
Where this series goes next. Each company in the scoreboard now gets its own valuation note, in which the depreciation re-mark meets a fair-value model:
The scoreboard itself will be re-scored each earnings season, same format, same falsifiers, until the 2026–28 window closes. The next hard checkpoints: Microsoft's FQ1-2027 10-Q, where the building-life change enters the filing record, and October's Q3 prints.
§7 Sources & Method
Every figure in this note is drawn from SEC filings via EDGAR, pulled and verified on July 29–30, 2026 (the evenings each document was filed), or from the July note's published model. No figure is sourced from press coverage, transcripts, or third-party aggregators.
Filings relied upon (this note). Microsoft FY2026 Form 10-K, filed 29 Jul 2026 (0001193125-26-323660) — useful-life policy, PP&E components, depreciation expense, cash-flow capex (XBRL PaymentsToAcquirePropertyPlantAndEquipment), net income, EPS; FY2025 10-K (0000950170-25-100235) for the prior-year policy wording. Alphabet Q2-2026 Form 10-Q, filed 23 Jul 2026 (0001652044-26-000071) — H1 depreciation and capex, PP&E note incl. assets not yet in service and the 60%-servers footnote. Meta Q2-2026 Form 8-K Ex-99.1, filed 29 Jul 2026 (0001628280-26-050596) — Q2 capex, D&A, net income, charges, FY2026 capex and expense guidance. Amazon Q2-2026 Form 8-K Ex-99.1/99.2, filed 30 Jul 2026 (0001018724-26-000024) — segment results, D&A, TTM capex series, TTM free cash flow as Amazon defines it, Q3 guidance. Apple FQ3-2026 Form 8-K Ex-99.1, filed 30 Jul 2026 (0000320193-26-000018) — quarterly results and nine-month cash-flow figures.
Model figures (the $221bn balance, the 4.4–17.2% overstatements, the $115.0bn Microsoft capex assumption, the FY2026 server-depreciation delta, and the FY2025 dose/overstatement pairs in the dose chart) are from the July note's v3 vintage model, published 17 Jul 2026 and unchanged since; its construction, validation against five disclosed life-change benefits, and limitations are documented in that note's §2 and §7. Nothing in the model was refit for this scoreboard.
Formulas. Depreciation growth = current-period reported depreciation (or D&A where that is what the release discloses; the difference is stated per chart) ÷ prior comparable period − 1. Dose = capex ÷ net income over the same window, using each company's freshest disclosed window: Microsoft FY2026 (annual, cash capex); Alphabet H1-2026; Meta Q2-2026 (capex incl. finance-lease principal, as Meta guides); Amazon TTM (capex net of proceeds, as Amazon reports); Apple nine-month FY2026. Where a window contains disclosed non-operating investment gains (Alphabet H1: ~$135.8bn unrealised; Amazon Q2/TTM: $53.4bn/$79.8bn pre-tax), net income is adjusted by the pre-tax gain × (1 − 21% statutory rate) before the ratio is taken; both adjusted figures are estimates and are labelled as such. Measurement windows differ across companies because fiscal calendars differ; no cross-company figure in this note nets windows against each other.
Late filings, incorporated. The Meta 10-Q (0001628280-26-050705, filed 30 Jul), Amazon 10-Q (0001018724-26-000026, 31 Jul) and Apple 10-Q (0000320193-26-000020, 31 Jul) all landed while this note was in preparation and are scored in the tally: no useful-life change is disclosed in any of the three, which closes claim 3's checks (a change in estimate must be disclosed in the period it occurs, so silence is confirmation). Meta's 10-Q additionally discloses servers-and-network depreciation of $4.62bn (Q2) / $9.01bn (H1) — a directly-filed figure sharper than the release's blended D&A. Amazon's release-basis D&A line includes content and lease amortisation; its 10-Q allocates PP&E depreciation by segment. Microsoft's earnings-call items — FY2027 capex guidance, and the datacenter/office building life extension (15→25 years, effective FY2027) with its finance-to-operating lease reclassification (~$190bn→~$175bn calendar-2026 capex-inclusive optic, spend unchanged per CFO Amy Hood) — are press-sourced (Benzinga, CFO Dive; Microsoft publishes its webcast but we cite no third-party transcript as primary) and are labeled as such wherever they appear; the building change enters the filing record with the FQ1-2027 10-Q, which is its verification checkpoint. The 10-K's commitments note ($34.6bn contracted construction) is the only issuer-document forward capex figure used. This note carries no target price and takes no position in any name; reference prices are therefore not cited. Each chart carries its source beneath it.
§8 Disclosures
Information only. TON618 Capital. This report is for information purposes only. Nothing here is an offer to sell or a solicitation of an offer to buy any security, fund interest, or digital asset, and nothing here is personalized investment advice or a recommendation regarding any instrument.
Publisher's exclusion. All research is published solely as general, impersonal information of regular circulation. It is not tailored to the objectives or circumstances of any individual and is not issued in connection with compensation from any client. The Fund has no clients and distributes all research free of charge. On that basis it publishes in reliance on the publisher's exclusion from the definition of "investment adviser" under the Investment Advisers Act of 1940 (§202(a)(11)(D); cf. Lowe v. SEC, 472 U.S. 181 (1985)).
Registration & conflicts. TON618 Capital is not registered as an investment adviser or broker-dealer in any capacity. The Fund is a Bitcoin fund and may hold or transact in the securities or digital assets it discusses. This report discusses Microsoft (MSFT), Alphabet (GOOGL), Amazon (AMZN), Meta Platforms (META), Apple (AAPL), Oracle (ORCL), CoreWeave (CRWV) and NVIDIA (NVDA); the Fund holds no position, long or short, in any of them, and has no economic interest in the price of any security named here. The Fund receives no compensation from any party in connection with its research.
Use of AI. Artificial intelligence is used in the creation of this research. The AI used is developed by Anthropic; this note separately reports Amazon's disclosed gain on its investments in Anthropic, and readers should weigh that connection — the Fund has no economic interest in Anthropic and the figure is quoted directly from Amazon's release. All methodology and data integrity are reviewed and approved before publication by TON618 Capital's Chief Investment Officer, Keyth Beck; errors may nonetheless occur, and readers should verify independently.
CFA. This report was prepared to align with CFA Institute analytical standards (methodology only). CFA® and Chartered Financial Analyst® are registered trademarks owned by CFA Institute. That reference describes the analytical framework applied; it does not imply the report was prepared, reviewed, or authored by a CFA charterholder, and the report is not issued, reviewed, endorsed, certified, or approved by — nor affiliated with — CFA Institute.
Risk & feedback. Past performance is not indicative of future results. Digital assets and equities are volatile and may result in total loss of capital. Corrections and feedback are welcome — please direct them to CIO Keyth Beck at keyth@ton618capital.com.