The useful life of a boom

My base case for the next twelve months is a volatile range: several drawdowns of 10% or more, and several new all-time highs, arriving in the same year and probably in the same quarter. Not a melt-up. Not a bust. A market that repeatedly frightens people out of positions it then rewards.

Last week was the rehearsal. The Nasdaq entered its second correction of 2026 in late July. Days later, on 4 August, the S&P 500 closed at a record 7,737 as oil eased on hopes of a deal to reopen the Strait of Hormuz. In between, the index had spent most of July below its 50-day moving average. That is the whole thesis compressed into ten trading sessions.

The reason the range holds is not sentiment. It is arithmetic. FactSet’s 31 July Earnings Insight has the S&P 500 reporting blended second-quarter earnings growth of 47.4% — the strongest since Q2 2021 — on revenue growth of 14.1%, with a blended net profit margin of 16.7%, the highest since FactSet began tracking the series in 2009. Analysts see 29.1% earnings growth for calendar 2026. Eighty-six per cent of reporters beat, against a ten-year average of 76%.

Against that, the forward twelve-month P/E is 19.6 — below its five-year average of 19.9, and only marginally above the ten-year average of 19.0. Multiples are elevated on a twenty-year lens. They are not elevated against this growth rate. Put earnings growth in the denominator and the index screens closer to cheap than to dear.

Where I disagree with my own bull case

The growth number deserves an asterisk that almost nobody applies, and applying it makes the thesis stronger rather than weaker.

Alphabet’s Q2 GAAP EPS of $9.11 against a $2.88 estimate included a roughly $98bn gain, largely unrealised marks on equity securities. Amazon’s $5.75 against $1.82 included $53.4bn of non-operating pre-tax other income, principally from its investment in Anthropic. Strip those two companies out and index earnings growth falls from 47.4% to 28.8%. The aggregate earnings surprise falls from 31.4% — which would be the largest since FactSet started measuring it in 2008 — to 9.2%.

Then note that the Energy sector grew earnings 135.3% because average crude in Q2 was $92.55 against $63.68 a year earlier, a 45% increase driven by a war. That is a windfall, not a franchise.

So the honest PEG is not 19.6 over 29.1. A cleaner denominator strips the valuation gains and the oil windfall and lands nearer the high teens — still strong, still supportive, but a very different quality of earnings. And here is the part that matters for everything that follows: the circular financing of the AI build is no longer a private-market curiosity. It is now inside the reported earnings of the S&P 500. When Amazon books a mark-up on its Anthropic stake as income, index EPS becomes a function of private AI valuations. Those marks go both ways.

The index is now long the private AI capital structure, through the income statements of two of its largest constituents.

The stress test already ran

2026 has not been a calm year. WTI opened near $57, printed $113 in April, and has traded above $100 since. The Fed under Kevin Warsh has held at 3.50–3.75% for five consecutive meetings with three dissents in favour of a hike at the July meeting, and markets now price one to two increases by year end rather than the cuts expected in January. The Treasury sold thirty-year paper at a 5% yield for the first time since 2007. Japanese and Chinese official holdings have been shrinking, with China’s at an eighteen-year low, while the yen broke 160 and JGB yields hit levels last seen in the 1990s.

That is a genuine exogenous shock — war, an energy spike, a hawkish repricing, and foreign demand for Treasuries wobbling — and equities made a new high through it. This is the single most useful piece of evidence available about what breaks the AI trade, and it is a negative result: discount-rate shocks alone do not do it. Not while the earnings are printing.

The shock that matters has to hit one of three specific transmission channels: the funding window that finances the build, the demand signal that justifies it, or the accounting that converts it into profit. Everything else is noise the market absorbs on its way to another record.

Base rates for a range

Coexisting drawdowns and records are the norm, not the exception. And this is a US midterm year: the widely-cited Longview Economics work on data back to 1926 puts the average drawdown in the twelve months preceding a midterm at roughly 18%, with a range from about 7% to 42%. Meanwhile power prices have become an election issue, which — as Note 04 argues — is not a coincidence.

Two mechanical shifts are inflating reported profit across the largest companies in the index. Neither is fraud. Both are reversible. Together they explain why margins are at records while cash generation is collapsing.

Shift one: operating expense becomes capital expense

Salaries are expensed the quarter they are paid. Servers are capitalised and released into the income statement over five or six years. Every dollar moved from the first bucket to the second improves this year’s operating margin, regardless of whether anything real improved.

That migration is now enormous. Through June 2026 the technology sector announced 139,156 job cuts, up 83% year on year and close to a third of all US layoffs; broader trackers run to 170,000–200,000 depending on methodology. Meta cut roughly 8,000 people — about 10% of staff — in May while guiding 2026 capex to $125–145bn. Cisco cut around 4,000 after beating estimates. Cloudflare cut 1,100 — a fifth of its workforce — in the same quarter it posted record revenue up 34%. Citigroup is targeting roughly 20,000 reductions by year end and has tied them publicly to AI adoption.

Gross margins improve. Headcount falls. The story is productivity. The evidence for the story is thin: a May 2026 Gartner survey of 350 executives at billion-dollar-plus firms actively deploying AI found that the companies cutting hardest showed essentially the same financial returns as those cutting least, and in some cases worse. Stanford HAI data shows employment of software developers under 26 down close to 20% since 2024 — a pipeline decision, not a productivity result.

Some of this is genuine automation. Some of it is a 2021 hiring error being re-labelled, and payroll dollars being reallocated to GPUs.

Shift two: the useful life

The second shift is the one that eventually forces the reckoning. Between 2020 and 2024 the major operators steadily extended the assumed lives of servers and network gear, each extension lowering annual depreciation and lifting reported profit. Then, in 2025, the trend split.

FIGURE 1Same hardware, opposite conclusions — disclosed changes to server / network useful life assumptions

Source: company SEC filings as compiled in public reporting. Amazon is the outlier — and the precedent.

The bear argument, made loudest by Michael Burry from November 2025, is that Nvidia’s roughly annual product cadence — Hopper to Blackwell to Rubin — means the economic life of frontier compute is nearer two to three years than five or six, and that the industry is therefore understating depreciation by about $176bn across 2026–2028, with Oracle’s and Meta’s operating earnings overstated by roughly 27% and 21% respectively by 2028. Independent work using a four-year counterfactual across a broader asset base gets to a larger number, around $228bn.

The bull rebuttal is real and should be stated plainly. Nvidia’s position is that customers observe four-to-six-year lives in practice. Goldman Sachs noted in April 2026 that A100 and H100 rental prices remain consistent with five-to-six-year economics, implying the secondary market does not believe in rapid obsolescence. Older silicon does not stop working when newer silicon ships; it cascades down to inference, to fine-tuning, to batch workloads. And a data-centre shell, its substation and its cooling plant genuinely do last twenty to thirty years.

This is also the argument we stopped being willing to conduct on anecdote. There is no clean tape for used GPUs the way there is for cars or aircraft — a Goldman note here, a broker listing there — so in August we built one: COMPUTE-RV, our GPU residual value index. It derives residual value as the present value of remaining rental income, caps it at the frontier replacement-cost ceiling, and referees the result against secondary-market prints — script-codified and re-run verbatim each period, so the number cannot drift with the narrative. The first print, 11 August 2026, lands closer to the bulls for now: composite retention across four generations is 74.4%, every observable secondary print sits inside its modelled band, and it is the replacement-cost ceiling — not cash flow — that binds on all four generations. But the same print quantifies the fragility: a single 1.7× frontier price-performance step — the Rubin volume ramp, roughly four months out — cuts composite retention from 74.4% to 43.7% at a stroke.

That last point is the one most bears skip, so let me do the arithmetic honestly.

FIGURE 2What the assumption is worth — annual depreciation on the 2026 capex vintage, equipment portion only (illustrative)

Author’s illustrative calculation. Takes 2026 guided capex for the four largest US hyperscalers of ~$725bn, assumes 65% is compute, server and network equipment (~$471bn) with the balance shell, land and power infrastructure left on long lives, and applies straight-line depreciation with no salvage value or part-year convention. One vintage only. The 2027 vintage — consensus points to $1.1–1.2 trillion — layers on top of this, and the 2025 vintage sits underneath it.

Note what this arithmetic does and does not say. It does not say anyone is lying. It says that a three-year change in one estimate, on one year’s spending, on the equipment portion only, is worth roughly $78bn of annual pre-tax profit across four companies. Stack three vintages and you are in the range of the sector’s entire earnings growth.

One further wrinkle worth knowing: US tax rules already permit heavily accelerated write-off, so cash taxes broadly reflect a faster life. The extended life flatters the book number — the one the multiple is applied to — not the cash. Which brings us to the tell.

The tell: earnings up, cash down

If the profit were real in an owner’s sense, cash would follow it. It is doing the opposite, and the divergence is now historic.

The company-level detail is starker than the aggregate. Alphabet’s free cash flow turned negative in Q2 2026 — the first negative quarter since it listed in 2004 — with $44.9bn of quarterly capex exceeding operating cash flow. Amazon’s trailing-twelve-month free cash flow has gone to roughly negative $7.6bn from $25.9bn a year earlier. Meta’s quarterly free cash flow fell about 91% year on year, it has issued $55bn of debt in six months and paused buybacks. Oracle’s fiscal 2026 capex ran at roughly 174% of operating cash flow. Sector-wide, capex is now around 100% of operating cash flow against a ten-year average near 40%.

Alphabet’s return on invested capital has fallen from roughly 45% to around 28% as the invested capital base nearly doubled. This is the number that has to normalise, and normalising it is what the eventual re-rating is about.

There is thirty years of finance literature on exactly this configuration. Sloan’s 1996 accruals work established that when reported earnings are driven by accruals rather than cash, subsequent returns disappoint — not because the accounting is wrong, but because investors capitalise the accrual as though it were cash. Right now the largest weights in the index are running the widest earnings-to-cash gap of the modern era, and the multiple is being applied to the earnings.

The build has outgrown operating cash flow. What filled the gap is a structure the market has not priced because, technically, it is not on the balance sheets the market looks at.

More than $120bn of data-centre spending has been moved off technology-company balance sheets in roughly eighteen months. The template is Meta’s Hyperion campus in Louisiana: a special purpose vehicle called Beignet, 80% owned by funds managed by Blue Owl and 20% by Meta, funded with about $27bn of debt arranged by Morgan Stanley and roughly $2.5bn of equity, anchored by PIMCO with BlackRock among the buyers. The bonds were rated A+ and priced around 6.58%, running to 2049. Meta leases the finished capacity back on four-year initial terms with a residual value guarantee. S&P confirmed it would not consolidate the vehicle’s debt with Meta’s.

It was the largest private credit transaction on record when it closed, and it was immediately copied. Meta is reported to be arranging a further vehicle of over $13bn for El Paso. Oracle’s OpenAI facility at Abilene sits in an SPV with roughly $13bn from Blue Owl and JPMorgan; a $38bn debt package funds sites in Texas and Wisconsin; an $18bn loan funds New Mexico.

Control without consolidation. The obligation is a lease footnote; the leverage sits in someone else's fund.

The strain is already visible where it always shows first — in credit, not equity. Oracle’s total debt exceeds $130bn alongside roughly $248bn of new lease commitments. Both S&P and Moody’s have had it on negative watch. Its five-year CDS has widened around 310% to a sixteen-year high and has become the market’s liquid proxy for AI capex risk generally. In January 2026 bondholders led by the Ohio Carpenters’ Pension Plan sued over the $18bn of notes issued in September 2025, alleging the offering documents did not disclose that Oracle would return seven weeks later for $38bn of loans. CoreWeave shareholders filed their own class action two days earlier. Blue Owl walked away from a $10bn Michigan project before PIMCO and Blackstone stepped in — and a specialist private lender declining fee income of that size is a more informative signal than any sell-side note.

The circle

Layered on top is the circularity, which by 2026 estimates spans north of $800bn of arrangements. OpenAI carries infrastructure commitments in the order of $1.15–1.4 trillion — roughly $350bn to Broadcom, $300bn to Oracle, $250bn to Microsoft, $100bn to Nvidia, $90bn to AMD, $38bn to AWS, $22bn to CoreWeave — against revenue around $13bn and a projected 2026 loss near $14bn. Microsoft owns about 27% of it. Amazon has put more than $83bn into Anthropic and OpenAI, which have committed over $238bn of compute back to AWS.

In a single week in late July, Nvidia announced or opened talks on more than $750bn of deals: a $500bn-plus letter of intent with SK Group, $5bn into a pre-revenue lab in exchange for it buying ten times that in Nvidia compute, and — the striking one — a proposal to guarantee up to $250bn of OpenAI’s data-centre lease payments plus finance $350bn of its chip purchases. A supplier underwriting its customer’s ability to buy from it is not an analogy for vendor financing. It is vendor financing.

Lucent and Nortel did this in 1999. The revenue was real when it was booked and worthless when the customer failed.

The bubble does not deflate on its own logic. It needs a catalyst, and the catalyst does not have to be about AI at all — it only has to close the funding window, break the demand signal, or force the accounting. Ranked by what they actually transmit, not by how alarming they sound.

FIGURE 4Candidate catalysts and their transmission channel — ordered from mainstream to tail

The three left-hand entries are not tail events. They are already happening, at a pace the market currently reads as background noise.

Once the first crack appears, expect the write-downs to be larger and faster than the underlying deterioration justifies. That is not cynicism. It is the documented behaviour of management teams under exactly these conditions.

The accounting literature is consistent. Jordan and Clark found that goodwill impairments cluster in periods when earnings are already depressed, because the marginal market punishment for an additional loss is small when the base is bad. Van de Poel and co-authors showed impairments are recognised alongside other losses and when pre-impairment earnings are low. Work on CEO turnover finds that a change at the top materially raises the probability of a big bath, since responsibility can be assigned to the predecessor. The common thread: the optimal time to recognise every bad thing at once is when the market has already stopped paying for good news.

Apply that to a hyperscaler in the quarter after a genuine shock. Management is looking at capacity commissioned into a demand air-pocket, an assumed useful life it can no longer defend to its auditor, non-cancellable purchase obligations, and a share price already 25% below the high. The rational move is to do all of it in one quarter: shorten lives, impair stranded assets, restructure the workforce, take the residual value guarantee to account. Blame the environment. Reset the base.

The write-down is not the accident. The write-down is the plan, executed at the moment it costs least.

Mechanically, the effect on reported earnings is violent. Depreciation, accelerated impairment and restructuring hit the same line at once, against a cost base already stripped of the headcount that would normally cushion it. Absolute earnings capitulate. Prices follow, because at that point the market is marking to the printed number rather than to the cash — the mirror image of what it is doing today.

And the trailing P/E goes vertical, which is exactly when it becomes useless. A denominator collapsing faster than a numerator produces an optically terrifying multiple at the precise moment the asset is cheapest. That is the trap in both directions: the multiple looks cheap at the top because earnings are inflated, and looks insane at the bottom because earnings are crushed.

What to watch, in order

• 01
  Any peer matching Amazon's five-year lifeThe single highest-information disclosure available. It arrives in a 10-Q footnote, not a press release, and it makes every remaining six-year assumption indefensible.
• 02
  Critical audit matters citing useful lifeWhen an auditor elevates the estimate to a CAM, the internal argument has already been lost.
• 03
  Lease and purchase-obligation footnotesNon-cancellable commitments are where the off-balance-sheet leverage actually resides. Track the growth rate, not the level.
• 04
  A syndication that does not clearOr a private lender walking, as Blue Owl did in Michigan. Credit tells you before equity does.
• 05
  Oracle five-year CDSThe market's chosen proxy for the whole trade. It moves first.
• 06
  Capex guidance that flattens rather than fallsThe first flat guide is the top. Falls come later and are already priced by then.
• 07
  Marks on private AI stakesNow a direct input to index EPS. A single down-round reverses billions of reported earnings at two index heavyweights.
• 08
  A COMPUTE-RV print breaking bandOur GPU residual value index is the live read on whether the resale market still underwrites the assumed life. Retention rolling over, or a secondary print landing below its modelled band, is the tape confirming what the footnotes will only admit later. The Rubin volume ramp is the scheduled test.

The reason to map the downside carefully is not to sell everything. It is that this particular downside has an unusually well-understood recovery mechanism, and most participants will be looking at the wrong metric when it arrives.

A capital cycle ends when spending stops. Depreciation, being non-cash, keeps running through the income statement while capex falls away — so reported earnings stay depressed while cash generation inflects hard. The cost cuts made under duress do not reverse. Utilisation of the installed base rises because nobody is adding to it. Free cash flow returns violently, ahead of accounting earnings, and the market re-rates on the cash long before the E in P/E recovers.

We have the template within living memory. Meta’s 2022: capex peaking into a demand slowdown, a stock down roughly three-quarters, then a declared year of efficiency, a capex plateau and a cost reset — and one of the great re-ratings of the decade, driven by cash conversion rather than revenue acceleration.

And we have the deeper analogue. Between 1996 and 2001 the telecoms industry laid over 80 million miles of fibre, close to a trillion dollars of capital, on WorldCom’s claim that internet traffic was doubling every hundred days. It was doubling annually. Four years after the bust, 85–95% of that fibre was still dark. The Nasdaq fell 78%; roughly $5tn of market value and more than 500,000 telecom jobs went with it. Corning went from near $100 to about $1. Ciena’s revenue fell from $1.6bn to $300m.

And then that fibre carried the entire cloud era, YouTube, streaming, and eventually the data centres now under discussion. The infrastructure was not wasted. It was simply owned by the wrong people, at the wrong price, on the wrong capital structure. The capital was destroyed; the capacity was not.

The asset gets used. The question is only who owns it when it does, and what they paid.

What that means practically

Through the range, valuation discipline matters more than direction. Owning the trend is fine; owning it with leverage in the wrong part of the stack is not. The differentiation that matters is between operators funding the build from operating cash flow and equity — Alphabet raised $80bn in stock in June, including $10bn from Berkshire — and those funding it from debt against contracted revenue from a counterparty that is itself loss-making.

Through the shock, the screen changes. Earnings quality — cash conversion, accrual ratios, the gap between reported profit and owner earnings — stops being an academic exercise and becomes the primary filter. Balance sheet capacity to keep spending through a downturn becomes the durable competitive advantage, exactly as it was for Amazon in 2001.

Through the recovery, throw out the P/E. At the trough the useful measures are EV to sales, EV to gross profit, and free cash flow yield calculated on maintenance capex rather than the growth capex that has just ended. That is where the generational entry point is priced, and it is why the best buying will look statistically expensive on the metric most people default to.

What would prove me wrong

I want to be explicit about the falsifiers, because a thesis you cannot lose is not a thesis.

Microsoft’s commercial remaining performance obligations stand near $627bn and Google Cloud’s backlog around $460bn — contracted, not hoped for. Microsoft has disclosed an Azure backlog it cannot fulfil for want of power. Google Cloud grew 63% year on year with operating margin above 30%, from a loss three years ago. If inference demand compounds faster than capacity through 2027 and utilisation stays high, the six-year life is simply correct and the depreciation debate resolves in the bulls’ favour. If secondary rental prices for prior-generation silicon hold — as Goldman observed in April, and as COMPUTE-RV’s first print corroborates, with all three observable secondary prints inside their modelled bands — obsolescence is a story rather than a fact. And if aggregate free cash flow troughs in 2027 and inflects with revenue attached rather than with capex withdrawal, then this was a J-curve, not a bubble, and the write-down chapter never gets written.

Watch the cash. It resolves the argument earlier and more honestly than the earnings will.

1. FactSet, Earnings Insight, 31 July 2026 — Q2 blended growth, margins, forward P/E, CY26/CY27 estimates, Alphabet and Amazon surprises, oil comparison. factset.com/earningsinsight
2. FactSet Insight, S&P 500 earnings season updates, April–July 2026. insight.factset.com
3. Trading Economics — US equity index levels, 4 August 2026. tradingeconomics.com
4. Advisor Perspectives / dshort — S&P 500 snapshots and drawdown tracking, June–July 2026. advisorperspectives.com
5. Tom’s Hardware / Financial Times — $725bn 2026 hyperscaler capex, +77% y/y. tomshardware.com
6. TrendForce — top-nine CSP capex ~US$830bn in 2026, +79%. via Yahoo Finance
7. Yahoo Finance — Bank of America aggregate free cash flow forecasts, 2025–2028. finance.yahoo.com
8. CNBC — hyperscaler capex scrutiny and free cash flow, 28 July 2026. cnbc.com
9. Decoding Discontinuity — Q2 2026 hyperscaler FCF dispersion. decodingdiscontinuity.com
10. ProCap Insights — capex as a share of operating cash flow; Alphabet ROIC. procapinsights.com
11. Yahoo Finance / Motley Fool — useful-life disclosures by company from SEC filings. finance.yahoo.com
12. Deep Quarry / National Law Review — useful lives of GPUs, key considerations. natlawreview.com
13. Footnote Brief — depreciation and circularity, alternative counterfactual estimate. footnotebrief.com
14. El Contenido — Goldman Sachs April 2026 note on A100/H100 rental economics. elcontenido.substack.com
15. Quinn Emanuel — AI data centre financing and litigation risks (SPV structures, Oracle metrics, off-balance-sheet totals). quinnemanuel.com
16. Bloomberg / Business Standard — Meta–Blue Owl Hyperion financing. business-standard.com
17. Forbes — analysis of the Meta SPV structure and pricing. forbes.com
18. Reuters / CNBC — Oracle bondholder class action, 14 January 2026. cnbc.com
19. CNBC — bond market anxiety over AI capex budgets, July 2026. cnbc.com
20. Bloomberg — AI circular deals interactive. bloomberg.com
21. Axios — Nvidia’s proposed OpenAI financing guarantee. axios.com
22. AI Circular Economy — deal map and commitment tracker. ai-circular-economy.com
23. US Bank / Trading Economics / CNBC — FOMC July 2026 decision, dissents, oil path, CPI. usbank.com
24. Fortune — Japanese repatriation risk and the 5% thirty-year auction. fortune.com
25. CNBC — foreign official Treasury holdings, China at eighteen-year low. cnbc.com
26. Wolf Street — TIC data, record $9.49tn foreign holdings. wolfstreet.com
27. Mapshock / Bloomberg Economics — Taiwan scenario modelling and GDP impact. mapshock.com
28. Atlantic Council — data centre backlash, moratoria and cancelled projects. atlanticcouncil.org
29. MultiState — state data centre legislation, 2026 session. multistate.us
30. Brookings — rising electricity rates and the 2026 midterms. brookings.edu
31. TechCrunch / Founder Reports / Tech Times — 2026 layoff tallies, Gartner survey, Stanford HAI data. techcrunch.com
32. Fortune — dot-com fibre buildout, dark fibre and equipment-vendor outcomes. fortune.com
33. Bridgewater — macro implications of the AI capex boom (GDP contribution). bridgewater.com
34. ANTIKYTHERA — COMPUTE-RV, GPU residual value index — all index prints on the research register (first print 11 August 2026, spec 0.1.0: composite retention, print-band validation, replacement-cost ceiling, Rubin-roll stress). antikythera.com.au/research
35. Academic: R. Sloan (1996) on the accruals anomaly; Jordan & Clark, and Van de Poel et al., on impairment timing and big-bath behaviour; Daniele et al. (2024) on big baths around CEO turnover.

This document sets out a personal market view for discussion. It is general commentary only, prepared without regard to any person’s objectives, financial situation or needs, and it is not a recommendation to buy or sell any security. Figures are drawn from the sources listed and are current to 4 August 2026; forecasts are estimates and will change. Company examples are used to illustrate accounting and financing mechanics, not to allege wrongdoing by any party. Figure 2 is the author’s own illustrative arithmetic on stated assumptions, not a forecast of any company’s reported results. Past performance is not a guide to future performance.