Apr 30, 2026

The AI Bubble: Two Mechanisms Inflating Hyperscaler Earnings

Markets · AI Bubble · Accounting Quality 2026-04-30
Q1 2026 Earnings · Hyperscaler Balance Sheet · Earnings Quality

The AI Bubble: Big Tech's "Record" Quarter Lives In A Spreadsheet

This week (April 29–30, 2026), Alphabet, Amazon, and Meta reported net income spiking: +81%, +77%, +61%; Microsoft reported Q3 FY2026 the same day, but with a different structure. Read the reports closely, and two inflation mechanisms are running in parallel. Earnings: ~60% of GOOGL's NI ($37.7B) and ~55% of AMZN's ($16.8B) is a paper markup on its Anthropic stake - not one cent of it hit the bank. MSFT had a similar mechanism (~20%, $7.583B) one quarter earlier, from OpenAI's restructuring into a for-profit entity. META is a different case: ~30% of NI ($8.03B) is a one-time tax benefit. Strip out these add-ons, and operating income at all four rose evenly, 20–30% YoY - good, not a "record." Revenue: AI cash flows through a closed loop - a hyperscaler invests in a startup → the startup signs a cloud contract → the cloud spends on Nvidia → Nvidia reinvests in the startup. The same dollar gets counted multiple times at different nodes.

$GOOGL · Q1 2026
~60%
$37.7B paper markup on Anthropic / $62.6B net income
$AMZN · Q1 2026
~55%
$16.8B pre-tax gain on Anthropic / $30.3B net income
$MSFT · Q2 FY2026 (Oct–Dec 2025) · reported Jan 28
~20%
$7.583B one-time gain from OpenAI recapitalization / $38.5B net income
$META · Q1 2026
~30%

Before diving into the details, it's worth clearly separating the two inflation mechanisms running in parallel across hyperscaler reports. They target two different lines of the income statement, require two different reading skills, and reverse under two different scenarios:

Mechanism 1 · Top of P&L

Inflating revenue

Cash flows in a closed loop: a hyperscaler invests in a startup → the startup signs a cloud contract (with the same hyperscaler) → the cloud spends on Nvidia → Nvidia reinvests in the startup. The same dollar can be counted multiple times as "revenue" at different nodes in the network - creating the illusion of larger end-demand.

Shows up in: the Revenue line (AWS / Google Cloud / Azure).
Reverses when: the startup has no real end-customer to fund its take-or-pay obligations.
VS
Mechanism 2 · Bottom of P&L

Inflating earnings

As AMZN and GOOGL did in Q1 2026: ~55–60% of net income came from a paper markup on their Anthropic stake as its private valuation rose from $183B to $350B. No real transaction occurred; just an "unrealized gain" line in Other Income flowing straight to the bottom line. When Anthropic has a down round, that line reverses by the same magnitude.

Shows up in: Other Income (Expense), net.
Reverses when: Anthropic / OpenAI has a down round at a lower valuation.

Both are direct consequences of public filings and U.S. accounting standards - no independent analysis needs to be taken on faith. Parts II and III unpack each mechanism; the appendix discusses Michael Burry's GPU-lifespan thesis separately (it's more contested).

Scope: This analysis is based on Alphabet's and Amazon's Q1 2026 earnings releases (published April 29, 2026), their latest 10-Ks, cash-flow estimates from Cognitive Credit, MetricDuck, and Ferrante Capital, and Michael Burry's public essays on X from November 2025 through Q1 2026. All figures are from public filings - no inside information.

Note: This is not investment advice, still less a recommendation to short anything. This piece dissects earnings quality - a 10-Q/10-K reading skill every hyperscaler shareholder should have. Burry may be right about the structure but wrong about the timing - he was right about 2008 too, but had to wait three years.

On the percentages: $37.7B (GOOGL) and $16.8B (AMZN) are pre-tax gains in Other Income; the "60%" and "55%" figures are computed as pre-tax gain / after-tax NI - a common shorthand in the press, but one that overstates the effect since it doesn't net out deferred tax. Subtracting ~21% deferred tax (the federal corporate rate, used as a reference), the after-tax impact on NI falls to ~48% (GOOGL: $37.7B × 0.79 = $29.8B / $62.6B) and ~44% (AMZN: $16.8B × 0.79 = $13.3B / $30.3B). MSFT's $7.583B (SEC EX-99.1) is a figure the company itself disclosed as a non-GAAP adjustment - stripping it out, Q2 FY26 NI falls from $38.5B to $30.9B (+23% instead of +60%). META's $8.03B is different in nature: it's an income tax benefit sitting directly in the tax line - already after-tax by definition, so the "30%" figure compares apples to apples. This piece keeps the pre-tax/NI convention to make cross-referencing the press easier; readers should know the gap exists.

I. The First Observation - What Does "Net Income Up 81%" Actually Mean?

On the morning of April 30, 2026 (evening of April 29 US time), the market received two eye-catching numbers. Alphabet reported revenue of $109.9 billion (+22% YoY), operating income of $39.7 billion (+30%), and net income of $62.6 billion (+81%). Amazon reported net income of $30.3 billion - nearly double the $17.1 billion figure from the same period a year earlier (+77% YoY).

The headlines call this a "blowout quarter." But at GOOGL, where does the gap between net income of $62.6B and operating income of $39.7B - $22.9 billion - nearly the size of operating income itself - come from? The answer sits in the line "Other Income (Expense), net" - a line most retail investors skim past when reading the income statement.

Paper · GOOGL Q1 2026

Other Income - Paper Markup

Net unrealized gains, non-marketable equity+$37.7B
Main asset being marked upAnthropic, etc.
Share of net income ($62.6B)~60%
Corresponding cash flow$0

This is the measurement alternative mechanism (ASC 321) - when a new funding round establishes a higher price (an observable price event), the increase in the value of the non-public equity stake flows straight to the bottom line. No transaction happens; only a row in the books gets corrected.

Cash · GOOGL Q1 2026

Real Capex - Real Cash Going Out

FY2026 capex guidance (raised on Q1 ER)~$180–190B
vs. actual FY2024 ($53B)~3.4–3.6×
Actual FY2025 ($91B, +72% YoY)already rising sharply
Potential free cash flow↘ sharply*
Gap covered byDebt issuance + retained OCF

*Estimate from Cognitive Credit's and Burry's analysis (pre-Q1 ER): if capex reaches ~$175B, GOOGL's FCF could fall from ~$73B (FY2024) to ~$8B in 2026. After the April 29 Q1 ER, GOOGL raised guidance to $180–190B - pushing FCF toward zero or negative. Actual FY2025 capex was $91B; FY2025 FCF already narrowed considerably. Cash is flowing out at an unprecedented pace; paper profit is flowing in at an equally unprecedented pace. The two lines have nothing to do with each other.

The story at Amazon runs in parallel. AMZN reported net income of $30.3 billion. Read the footnote: $16.8 billion is pre-tax gains on its Anthropic investment (grafa). That means 55% of the profit in a "record" quarter lives entirely in a blue cell on a valuation spreadsheet. In that same quarter, AMZN's cash capex hit $43.2 billion - up 79% YoY from $24.1 billion in Q1 2025; actual FY2025 capex already reached $131.8 billion, with FY2026 guidance of ~$200 billion - the full-year increase of ~$68 billion alone is larger than Caterpillar's entire annual revenue.

AMZN's underlying business is actually fine: AWS reaccelerated to 28% (the fastest in 15 quarters), AI revenue run rate exceeds $15B, custom silicon (Trainium/Inferentia) surpassed a $20B run rate, and overall operating margin is 13.1%. This piece isn't arguing "AWS is dying." It's pointing to one simple fact: the shape of AMZN's 2026 cash-flow structure is unrecognizable next to AMZN's 2024. And the "blowout" part of the profit - the part pushing the stock price up - is the part with nothing to do with the core business.

A warning for 10-Q readers

"Other Income, net" is not a "minor miscellaneous" line. In hyperscaler 10-Qs for Q1 2026, this line determines 50–60% of net income. The footnote usually sits in Note 4 or Note 5 of the filing, under the heading "Investments in Equity Securities." Read that footnote before reading the headline. The first question to ask: "how much of net income is realized cash, and how much is a mark-to-model gain?"

II. Inflating Earnings - All Four Hyperscalers, Each In Its Own Way

To understand why a startup that hasn't gone public generated $37.7 billion in "profit" for Alphabet in a single quarter, you need to know two U.S. accounting standards:

  • ASC 321 (updated in 2018) - for non-public equity (like Anthropic), companies may use the measurement alternative: carry it at cost, then adjust up or down whenever a new funding round establishes an observable price. That adjustment flows straight to net income - no need to sell the stake, no transaction required. Before ASC 321, these holdings were typically carried at cost until sold or impaired.
  • "Observable price changes" - for non-public equity, a company can mark up its holding whenever a new funding round at a higher price establishes an "observable price." No sale is required; another investor simply has to pay a higher price for an equivalent block of shares.

Anthropic has had a notable 2025–2026 fundraising timeline:

Date Event Valuation
September 2025 Series F (ICONIQ Capital lead; Lightspeed co-investor) $183B
November 2025 MSFT + NVDA commit $15B ~$200B+
January 2026 Term sheet with Coatue + GIC $350B
April 2026 Alphabet commits $40B ($10B cash + $30B milestone) $350B

Every time Anthropic closes a new round at a higher price, every existing investor - including GOOGL, AMZN, MSFT, NVDA - who has elected the measurement alternative under ASC 321 (a common accounting policy for their non-public investments) is required to recognize the corresponding value increase in net income. Their non-public equity stake has just been "repainted" at the new round's price. Note: the measurement alternative is an election (chosen once per investment), not a mandatory default - but once elected, marking up on an observable price event is unavoidable.

Quarterly markup = (New round price − Prior round price) × Shares held

Anthropic's valuation rose from $183B (9/2025) to $350B (1/2026) - nearly doubling in four months. Combined with hyperscalers continually increasing their stakes (GOOGL committing another $40B; AMZN committing another $20B), the book value of these investments on GOOGL's/AMZN's balance sheets multiplies, and each time it does, an "unrealized gain" line appears in Other Income.

Why do I call this a "paper markup"?

Because until the equity is actually sold (or Anthropic IPOs and GOOGL/AMZN exit), this gain only exists on paper. If Anthropic's next round is priced lower - a down round - the entire markup gets reversed (marked down) through the same "Other Income, net" line, pulling that quarter's net income down by the same amount.

In other words: symmetry. Every $37.7B added to net income today could become −$37.7B subtracted from net income in some other quarter - depending on how private-market valuations move.

This isn't theoretical. In 2022, when AI startup valuations collapsed after the 2021 peak, a string of Big Tech companies had to write down the value of their venture investments in the corresponding 10-Qs. Markup is a two-way street - the downside just never makes headlines.

Why this matters - a Soros perspective

This is a textbook example of reflexivity: market belief pushes private valuations up, higher private valuations push hyperscaler accounting net income up, higher EPS pushes hyperscaler stock prices up, higher stock prices confirm "AI is the right investment" - and the cycle repeats. When it reverses, every link reverses at once.

2.1 - All four hyperscalers had an "add-on" inflating net income - three this quarter, one a quarter earlier

Beyond GOOGL/AMZN: META also had an add-on boosting Q1 2026 NI; MSFT's structure is different - Q3 FY2026, the same week ($31.8B NI), is purely core business, since its large add-on happened one quarter earlier, in Q2 FY2026 (reported January 28, 2026). The mechanisms differ, but the direction is the same: the bottom line at all four is dressed up by something that isn't the core business.

Note - Microsoft's fiscal year

Microsoft runs a fiscal year from July to June, unlike Alphabet, Amazon, and Meta, which follow the calendar year (Jan–Dec). Result: the same calendar period, Jan–Mar 2026, is called Q1 2026 by GOOGL/AMZN/META, but Q3 FY2026 by MSFT. Likewise, MSFT's Q2 FY2026 = Oct–Dec 2025 - the quarter with the $7.583B OpenAI gain, reported January 28, 2026, one quarter before this week.

Company (Quarter) Reported net income "Add-on" inside it Share Nature
Alphabet (Q1 2026) $62.6B $37.7B unrealized gains on Anthropic stake ~60% Paper markup (ASC 321)
Amazon (Q1 2026) $30.3B $16.8B pre-tax gains on Anthropic stake ~55% Paper markup (ASC 321)
Microsoft (Q2 FY2026, Oct–Dec 2025, reported Jan 28, 2026) $38.5B $7.583B one-time gain from OpenAI's shift to for-profit ~20% One-time structural gain (OpenAI recapitalization)
Meta (Q1 2026) $26.8B $8.03B income tax benefit (one-time) ~30% Tax reform (One Big Beautiful Bill Act)

Three notable findings from the table above:

  1. MSFT has a similar mechanism, but it happened one quarter earlier, for a different reason. MSFT's $7.583B sits in Q2 FY2026 - that is, Oct–Dec 2025, reported January 28, 2026 - not in this week's Q3 FY2026 (Jan–Mar 2026). The mechanism is also different: it's a one-time structural gain from OpenAI's legal restructuring into a for-profit entity in late 2025. When the structure changed, MSFT was allowed to revalue its 27% stake at the new value, generating a $7.583B gain in Other Income during Q2. MSFT's Q3 FY2026 (this week) is actually clean - net income $31.8B (+23% YoY), OI $38.4B (+20%), revenue $82.9B (+18%); the OpenAI item is almost negligible - this is core business. The earnings-quality effect is the same though: ~20% of Q2 FY2026 net income ($38.5B GAAP, $30.9B ex-OpenAI) didn't come from Office or Azure - it came from a legal event at an AI startup.
  2. META is a different case - this is tax reform, not an AI markup. Meta's $8.03B benefit comes from the "One Big Beautiful Bill Act" and Treasury Notice 2026-7 - a U.S. tax law revision from late 2025 that created a deferred tax asset reversal for Meta. It partially offsets the $15.93B non-cash tax charge Meta recorded in Q3 2025, also because of this law. In other words: two sides of the same accounting event, not an AI bubble story. Excluding it, real net income was $18.7B (EPS $7.31) - versus Q1 2025 NI of $16.6B, a "real" increase of only ~13%, not the +61% in the headlines.
  3. "Record increase" is a headline shared across companies, but the substance differs. GOOGL +81%, AMZN +77%, META +61%; MSFT Q2 FY2026 +60% (due to the OpenAI gain), Q3 FY2026 +23% (core business). But dig in, and only operating income reflects the real business. OI growth was, respectively, +30% (GOOGL: $39.7B), +30% (AMZN: $23.9B vs. $18.4B in Q1 2025), +21% (MSFT Q2 FY26: $38.3B) and +20% (MSFT Q3 FY26: $38.4B), +30% (META: $22.9B vs. $17.6B in Q1 2025) - good, but not the "record" level that the inflated net income implies.

The question that follows: is 20–30% OI growth good or bad in the hyperscaler context? Set next to the last two years, it's a decent level - not a breakthrough, not a slowdown. FY2024 was a rebound year after the 2023 cost-discipline period, so FY2024 OI growth was unusually high. FY2025 already decelerated considerably. Q1 2026 is really a mild reacceleration from the trajectory at the end of FY2025 - but still below the FY2024 peak:

Hyperscaler FY2023 FY2024 FY2025 Q1 2026 Q1 2026 assessment
Alphabet (GOOGL) +13% +33% +15% +30% A clear reacceleration. FY2025 had been decelerating quarter to quarter (Q1 +20%Q2 +14%Q3 +9% / +22% ex-EC fine); Q1 2026 reversed that trend.
Amazon (AMZN) +201% +86% +17% +30% A reversal of the slowdown. Q1 2025 was only +20%, all of FY2025 just +17%; Q1 2026 jumped to +30%, reflecting AWS's reacceleration to 28% and custom-silicon margin.
Microsoft (MSFT) · fiscal year offset - +24% +17% +20% Stable, slightly ahead of FY2025. FY2024's +24% was the cycle's highest "rebound" level; FY2025 slowed down, but Q3 FY26's +20% held above the 5-year average. (The Q1 2026 column = Q3 FY26, since MSFT follows a July–June fiscal year.)
Meta (META) +62% +48% +20% +30% A mild reacceleration from FY2025, but below the FY2024 peak. Q3 2025 OI was +18%, Q4 2025 only +6% (margin compressed by capex); Q1 2026's +30% is a recovery from the prior quarter's trough.
Summing up - in what sense is 20–30% "good"?

Good by S&P 500 standards. The S&P 500's historical average OI growth is only ~6–8%/year. At 20–30%, hyperscalers are still growing at an outstanding rate - nothing to fault about the absolute speed.

Decent against their own FY2025. All four accelerated slightly versus FY2025: GOOGL from +15% → +30%, AMZN +17% → +30%, MSFT +17% → +20%, META +20% → +30%. This signals AWS/Azure reacceleration and Google Cloud growth spreading into OI.

But below the FY2024 peak (for META, MSFT, AMZN). FY2024 was a "rebound" year after 2023's cost-cutting - META +48%, AMZN +86%, MSFT +24%. Against that peak, Q1 2026 is merely "good," not a "record."

Most important - capex is growing many times faster than OI. OI growth is +20–30% while capex growth is +50–80% YoY (AMZN's Q1 2026 capex +79%, GOOGL's FY26 guidance ~3.4× FY24, META's capex ~+85%). The capex/OI ratio is deteriorating fast - meaning every additional dollar of OI now requires an ever-larger number of capex dollars to sustain. This is the signal actually worth watching: 20–30% OI growth isn't bad on its own, but paired with surging capex intensity, the quality of that growth is declining.

2.2 - Cash flow: where the real story shows

OI is an accounting line - affected by depreciation schedules, accruals, stock-based comp. Cash flow is the real money line: OCF (Operating Cash Flow) is the cash actually collected from operations, and FCF (Free Cash Flow = OCF − capex) is what remains after investing in assets. This is the line that actually separates "genuinely good" from "good on paper." Set next to Q1 of two years ago, Q1 2026 cash flow shows the OI–FCF gap cracking wide open:

Hyperscaler Q1 2024 Q1 2026 FCF Δ
(2 years)
OCF FCF Capex OCF FCF Capex
GOOGL $28.8B $16.8B $12.0B $45.8B $10.1B $35.7B −40%
AMZN (TTM) $99.1B $50.1B ~$49B $148.5B $1.2B $44.2B* −98%
MSFT (Q3 FY) $31.9B ~$21.0B $10.9B $46.7B $15.8B $31.9B† −25%
META ~$19.4B $12.5B $6.7B ~$32.2B $12.4B $19.8B −1%

Note: AMZN uses TTM (Trailing Twelve Months) figures since that's how AMZN reports FCF in its earnings release. *AMZN's Q1 capex is $44.2B (a single quarter only). †MSFT's Q3 FY26 capex of $31.9B doesn't include a $4.7B finance lease - the real total commitment is ~$36.6B; 2/3 goes to "short-lived assets" (GPU/CPU). META's FY26 capex guidance is $125–145B (raised) versus actual FY24 of ~$39B → ~3.5×.

Reading the table above: three findings

1. OCF at all four is still growing strongly (+30 to +66% over two years). The underlying business - search, e-commerce, cloud, ads - is still healthy and generating good cash. That's good news.

2. But FCF has turned in the opposite direction. AMZN −98% (near zero), GOOGL −40%, MSFT −25%, META flat. The gap comes entirely from capex: over the same period, Q1 capex multiplied 2.5–3× across the board. That means cash is still coming in - but it's flowing out even faster. This is the core difference from 2024: back then, every additional OCF dollar left ~50–60 cents dropping to FCF; today, almost 0 cents drops through.

3. AMZN is the early warning. TTM FCF fell to $1.2B from $50.1B two years earlier - nearly zero. If Q2 2026 capex holds at the Q1 pace ($44.2B), AMZN could become the first hyperscaler to report negative TTM FCF since the 2022 COVID cycle. At that point, the old question returns: who pays for buybacks, who pays down debt, who covers stock-based comp when the share price slides? The practical answer: issue more debt - and the "Sign (b) already happening" item in Part III describes exactly that.

A cross-check: the OCF/NI ratio - an earnings-quality indicator

A classic earnings-quality measure: OCF / Net Income. Ideally ~1; below 0.7, there's likely a large paper markup or accrual sitting between NI and OCF.

Q1 2026 ratios (estimated from public figures): GOOGL ~$45.8B / $62.6B = 0.73; AMZN's Q1 OCF is ~$33B / $30.3B NI = ~1.1 (good, thanks to working capital), but excluding the $16.8B Anthropic gain from NI: ~$33B / $13.5B core NI = 2.4 (very healthy - showing the core business is strong and NI is just noisy); META ~$32.2B / $26.8B = 1.20 (good, mostly thanks to a non-cash tax benefit); MSFT Q3 FY26 $46.7B / $31.8B = 1.47 (the cleanest). GOOGL has the lowest ratio at 0.73 - precisely the "paper markup" signal Section II pointed to.

2.3 - The pushback: weak FCF isn't necessarily a disaster - this is an AI arms race

Before concluding "earnings quality is deteriorating," it's worth taking the opposing view seriously. Weak FCF from high capex doesn't automatically mean a company is in trouble - it can be a rational strategic decision in a winner-take-most race. The bull case has three strong arguments, all grounded in public data rather than promises:

The bull case · High capex is rational

"This is CapEx with visibility, not a blind bet"

Backlogs are exploding. AWS backlog hit $364B in Q1 2026 (not including Anthropic's $100B commitment announced after the quarter), Google Cloud backlog is $462B (nearly doubling sequentially), Microsoft's commercial RPO is $625B in Q2 FY26 (+110% YoY). Combined, that's $1.63 trillion in signed commitments - capex is being poured into compute that already has take-or-pay customers, not built speculatively and hoped for. - Q1 2026 ER calls; Cloud Wars 2026
Compute-constrained - limited by supply, not demand. Sundar Pichai said on the Q1 2026 ER call: "what keeps us up at night is compute capacity - power, land, supply chain". Cloud revenue could be higher if capacity existed. In other words: capex is rising to meet demand that's already here, not to create demand that doesn't exist yet. Google Cloud customers are spending 30% faster than their commitments - a real sign of "consumed capacity." - Pichai, Q1 2026 ER call
Core-business AI monetization already has proof. Meta's Advantage+ AI ad suite has hit a $60B run rate; Google Search is +17% YoY in Q4 2025 (AI Mode queries are 3× longer); AWS's AI revenue run rate exceeds $15B in Q1 2026; Microsoft Copilot has permeated Office 365. This is AI revenue already being generated on existing distribution - no need to wait for some future product before the cash flow shows up. - Q1 2026 results · UC Intelligence 2026
Financing is still well within reach. All four still show strong OCF growth (+30 to +66% over two years), and positive net-cash balance sheets (META: $81B cash − $59B debt = +$22B; GOOGL: ~$95B net cash; MSFT: ~$60B). Issuing more debt is choosing cheaper capital structure, not a sign of being forced to by a cash shortage. T. Rowe Price calls this a "rational Nash equilibrium": invest heavily or lose monopoly economics - choosing not to invest is strategic suicide. - Q1 2026 balance sheets · T. Rowe Price 2026
The bear case · Three weaknesses remain

"Visibility ≠ profitability"

The loop is still a loop. A large chunk of the $1.63 trillion backlog is from the very startups hyperscalers invested in - Anthropic signing with AWS, OpenAI signing with Azure, OpenAI signing with Oracle's Stargate. This is a commitment, not end-customer demand. If OpenAI or Anthropic can't generate enough revenue to cover their take-or-pay obligations, that backlog turns into a write-down. - Section III, this article
Inference costs falling 50–200×/year could strand GPUs. If each new chip generation (Hopper → Blackwell → Rubin) pushes the prior generation's inference pricing down a notch or two, economic useful life ends up shorter than accounting useful life - exactly what AMZN admitted in early 2025 (see appendix). Long-term signed backlogs don't help if contract prices have to be renegotiated downward. - Burry thesis · appendix
A straight assessment - what's the reasonable conclusion?

The bull case is largely right about the present moment. A $1.63T backlog and strong OCF growth are real data - not promises. Compute-constrained is a genuine demand signal. High capex plus weak FCF in a winner-take-most race, backed by a still-thick balance sheet, is a reasonable strategic decision, not a sign of decline. If Q1 2026 is where this stops, this is the highest-ROI CapEx cycle in tech history.

But the question isn't "is today fine" - it's "when the cycle reverses, how fast does it reverse." Two bear-case risks still stand: (a) a large share of the backlog is an internal AI-ecosystem loop, not end-customer demand; (b) current ROI of $30–35B revenue / $332B capex = ~10% gross - enough to justify itself if growth holds pace, but no margin of safety if it slows.

The pragmatic conclusion: weak FCF alone isn't a red flag - it's an investment decision. But weak FCF combined with a paper markup inflating EPS combined with a backlog dependent on the AI-startup loop is a combination worth watching. This piece isn't saying "sell" - it's saying "know what you're holding." At this point, the bull and bear cases aren't mutually exclusive; they're just betting on two different outcomes for 2027–2028.

III. Inflating Revenue - Capex Circling Among A Handful Of Addresses

The second mechanism hits the top line: the circularity of AI cash flow means the same dollar gets counted multiple times as "revenue" at different nodes in the network. This isn't fraud - every transaction is real and gets audited. But it creates an illusion of end demand many times larger than reality.

Circular capex - AI money circling among a handful of addresses
1. Hyperscaler invests GOOGL/AMZN/MSFT/NVDA pump capital into OpenAI, Anthropic, xAI...
4. Nvidia reinvests NVDA signs a $100B LOI with OpenAI (Jensen Huang later said it's "never a commitment"), invests $2B in CoreWeave, signs with other hyperscalers

The natural question: when will real end-customer demand (businesses using AI to cut costs or generate more revenue) be large enough to justify $650B in annual capex? Combined revenue from Amazon, Apple, Alphabet, Microsoft, Meta, and Nvidia doesn't even reach $2 trillion - yet they're pursuing an AI buildout on the order of $2.9 trillion in data-center capex through 2028 (Morgan Stanley); Bain estimates $2 trillion in annual revenue is needed to fund AI compute through 2030, and the world will fall $800B short. This equation only balances if (a) AI generates enormous end-customer value within 3–5 years, or (b) a large chunk of capex gets funded with debt.

Sign (b) is already happening

Meta issued $30B in bonds (November 2025) - the company's largest bond sale ever - to fund a data-center project called Hyperion in Louisiana with Blue Owl Capital (~$30B total). Oracle committed $300B to OpenAI's Stargate over five years while its net debt has already risen sharply. Microsoft disclosed $58.5B in uncommenced data-center leases (not yet on the balance sheet, expected to commence 2026–2031) alongside $4.8B in finance-lease assets already recorded. Amazon's capex booked via finance leases keeps growing - not fully reflected on a traditional balance sheet. The funding structure for the AI buildout is shifting away from retained earnings and toward debt and leases. This isn't the sign of a cash-rich industry self-financing; it's the sign of an industry borrowing to spend.

The weakest link - Oracle is the "canary in the coal mine"

Among the "addresses" in the capex loop, Oracle is the one most worth watching - and the reason credit markets call ORCL the "canary in the coal mine" for the debt-fueled AI spending race. Unlike other hyperscalers (cash-rich, A+/AA-), Oracle sits at the edge of investment-grade, and its dependence on a single customer is unprecedented:

Why this matters - a cross-default scenario: If OpenAI can't generate enough revenue to pay Oracle's $60B/year (current run rate ~$15B), Oracle would have to write down that portion of its backlog → breach a debt/EBITDA covenant → be forced to raise equity or cut capex sharply. The shockwaves would flow back to (i) Microsoft - OpenAI's main cloud partner, which would have to absorb extra workload if Stargate slows; (ii) Nvidia - Oracle is one of the top customers for Blackwell/Rubin, so GPU orders would drop; (iii) private credit funds holding Oracle debt (Blue Owl, Apollo, Ares all already have data-center exposure) - exactly the scenario described in the Private Credit Crisis piece. Oracle isn't the biggest hyperscaler, but it's the first fracture point worth watching - and an Anthropic/OpenAI down round from Section II plus an Oracle covenant breach here would trigger at the same time.

IV. The 2027–2028 Conversation - When The Cycle Reverses

The two mechanisms above are two faces of the same phenomenon: earnings quality is deteriorating quickly, while the headlines and stock prices say the opposite. Four risks converge on 2027–2028, visible from today's numbers:

RISK 1

An Anthropic / OpenAI down round

If enterprise AI demand doesn't grow fast enough by 2027, Anthropic's next round at a valuation lower than $350B would kick off a cascade of mark-downs at GOOGL/AMZN/MSFT/NVDA - reversing tens of billions of dollars in "profit" from 2025–2026. In that same quarter, EPS could fall 30–50% purely from a fair-value adjustment. This is the risk of the earnings mechanism.

RISK 2

Take-or-pay contracts coming due

Most AI revenue commitments (Anthropic's $100B with AWS; OpenAI's multi-year deal with Azure) are multi-year contracts. If a startup can't generate enough end-customer revenue to pay them, the hyperscaler - playing both investor and supplier - would have to write down both lines at once. This is the risk of the revenue mechanism. One startup going bankrupt: markup + revenue + receivable vanish in a single quarter.

RISK 3

Hyperscaler FCF approaching zero

If capex holds at $180–190B (GOOGL's Q1 2026 guidance), ~$200B (AMZN's FY2026 guidance, +52% from actual FY2025 of $131.8B), $125–145B (META, raised at Q1 ER) in 2026 while the underlying business only grows 10–15%, FCF could fall to near zero or negative. At that point the question becomes: who pays the dividend / buyback? More debt? Issuing shares? Cutting capex (and admitting AI isn't an arms race that must be won)?

RISK 4

The GPU useful-life debate erupts

If Blackwell → Rubin makes H100 inference pricing collapse faster than expected, hyperscalers could be forced to pull useful life down to 4–5 years - creating tens of billions in impairment. AMZN already did this in early 2025 (reversing from 6 years back to 5 for part of its server fleet; see appendix) - the first empirical evidence. MSFT, GOOGL, META haven't followed yet. The rest is contested: Burry believes it will spread; Microsoft's CFO, CoreWeave, and Bernstein don't.

"AMZN's 2026 cash-flow profile is unrecognizable next to 2024. Q1 2026 capex of $43.2B (+79% YoY) and FY2026 guidance of $200B confirm it: hyperscalers are no longer cash-rich self-financiers - and the long, uncomfortable conversations about depreciation schedules in 2027–2028 are coming, whether anyone wants them or not." - Synthesized from Q1 2026 ER releases + Burry's essays (1–4/2026), author's interpretation
What does "depreciation schedules in 2028" mean, and why 2028?

What a depreciation schedule is. When a company buys a long-lived asset - GPUs, servers, a data center - it doesn't record the entire cost on the P&L at once. Instead, the cost is "spread out" over its estimated years of use (useful life). Example: a $1B GPU with a 6-year useful life → $167M in depreciation expense recorded each year. A "depreciation schedule" is the table determining how much gets recorded in which year - and it directly affects operating income, EPS, and period-to-period comparisons.

Why this is a problem. As the appendix table shows, during 2023–2024, hyperscalers extended GPU useful life from 3–4 years to 6 years. The result: quarterly depreciation expense fell, EPS rose - but this is EPS "borrowed from the future." The real cost doesn't disappear; it just gets pushed into later quarters. If GPUs actually become obsolete faster than the 6-year schedule assumes, a company must either (a) shorten useful life - causing a sudden jump in quarterly depreciation expense, or (b) record a one-time impairment - a large write-down on the P&L.

Why 2027–2028. Hyperscalers' peak GPU buildout happened in 2023–2024 (the big H100/H200 cohort). If real economic useful life is 4–5 years: 2023 + 4 = 2027; 2024 + 4 = 2028. The older cohort (2022 - mostly A100) already came due in 2026 but at a much smaller scale, while the Hopper peak fell in 2023–2024 - so impairment pressure concentrates in 2027–2028, not 2026. By then, the books still carry 1–2 years of "cushion" under the 6-year useful life while real market value may have already collapsed. If the fair value of H100/H200 falls below book value, accounting standards require recognizing impairment. That's why it's an "uncomfortable conversation" - not because anyone wants to admit $100–200B evaporating off the balance sheet.

An important addition: depreciation is a non-cash expense - it doesn't directly affect FCF. But it does reduce operating income, EPS, and net income. That's why this piece places this risk fourth and splits it into an appendix: the mechanism is real, but its timing and scale depend on how fast GPUs actually become obsolete in practice - which is still contested.

V. A Checklist For Hyperscaler Shareholders

This piece isn't saying "sell hyperscalers." The underlying business - search, e-commerce, cloud, ads - is still very good. AWS reaccelerating to 28% is a real signal. A $20B custom-chip run rate is a real signal. The issue isn't the companies themselves; it's that the gap between reported profit and high-quality profit keeps widening. When you read next quarter's 10-Q, here are six things to check:

  1. "Other Income" as a share of net income - if it's above 30%, dig into the footnote. How much is a realized gain (already sold), and how much is an unrealized markup?
  2. Useful life of PP&E - look for the phrase "estimated useful life" in the "Property and Equipment" note. Has it changed over the past two years? How much does that affect depreciation expense?
  3. Operating cash flow vs. net income - the OCF/NI ratio should be close to 1. If NI rose 80% while OCF only rose 20%, there's a large paper markup in between.
  4. Free cash flow - and its trend - capex minus OCF. Is it positive or heading toward zero? How has it changed versus last year?
  5. Total off-balance-sheet commitments - the "Commitments and Contingencies" note. Lease obligations, take-or-pay contracts, guarantees. The AI funding structure is shifting off the traditional balance sheet - read this note.
  6. Debt structure - net debt / EBITDA, maturity schedule. Hyperscalers used to be cash-rich; if net debt is rising fast, ask where that money is going.
Methodology reference

This entire checklist is a basic skill already covered in Reading U.S. Financial Statements By Industry. Applying it to tech - especially hyperscalers spending heavily on AI infrastructure - is just taking seriously warnings already sitting in accounting textbooks: revenue ≠ cash, net income ≠ FCF, fair value ≠ realized.

VI. Four Final Lessons

LESSON 1
Read Other Income before you read the headline

A headline saying "net income up 81%" can be true under GAAP and meaningless economically. An "unrealized gain on non-marketable equity" line larger than operating income isn't a good signal - it's a signal that most of that quarter's profit didn't come from the business.

LESSON 2
Separate structural facts from future theses

The Anthropic markup and circular capex are structural facts readable directly from the 10-Q - no need to believe any thesis, just to know how to read the footnote. Future-facing theses (like GPU depreciation) are worth reading but shouldn't be bet on one-sidedly. Weigh each observation by its own level of certainty; don't buy or sell the whole package at once.

LESSON 3
Two-way reflexivity

The paper markup is pushing net income up, pushing hyperscaler stock prices up, further justifying AI capex, further validating AI startup valuations. When the cycle reverses - and history says it always does - the entire chain reverses at once. This is systemic risk, not idiosyncratic risk.

LESSON 4
Being "structurally right" doesn't mean being "right on timing"

The history of every financial bubble: the risky structure is usually visible years before it bursts. Subprime was visible from 2005, burst in 2008. Dot-com was visible from 1998, burst in 2000. The lesson: recognizing a structure and betting against it are two different things. This piece recommends recognizing it, not shorting it. The best defense is knowing what you hold, and not being surprised when the cycle turns.

Appendix - Michael Burry's Depreciation Thesis (Contested)

This section is split off because it differs from the two main mechanisms above: those two are facts readable directly from this quarter's 10-Q; the depreciation thesis has one empirical data point (AMZN reversed its server useful life in early 2025, recorded in its FY24 10-K - see below) but is mostly still a prediction about the future that remains contested for the other three hyperscalers. Included so readers can judge for themselves - not as a recommendation.

In November 2025, Michael Burry (who shorted subprime before 2008) announced that Scion Asset Management had bought ~$1.1 billion in put options on Nvidia and Palantir. In subsequent essays (michael-burry.com, 1–4/2026), he targeted a purely accounting point: GPU useful life. His argument: hyperscalers quietly extended useful life to 6 years during 2023–2024, reducing depreciation expense and inflating EPS by ~24%.

Hyperscaler Old useful life Current Disclosed impact
Amazon (AWS)4 years5 (2022) → 6 (2024) → 5 (2025, reversed)+$3.6B NI FY24, then −$0.7B op income FY25
Microsoft (Azure)4 years6 years (FY2023)↓ ~33% depreciation; +$3.7B NI FY23
Google Cloud4 years6 years (2023)↓ ~33% depreciation
Meta (servers)4–5 years5.5 years (2025)~$2.9B reduction in FY25 depreciation

The useful-life change is real and disclosed in the 10-Ks. Meta's ~$2.9 billion reduction in FY25 depreciation is the figure Meta itself reported. The contested part isn't "did the change happen", but "was the change justified." Two sides:

Burry & the critics

"GPUs really only last 3–5 years"

Nvidia's ~18-month cycle (Hopper → Blackwell → Rubin) makes H100s obsolete fast. Accounting useful life is drifting away from economic useful life. - Burry, michael-burry.com 1/2026
"Rapid AI innovation cycles could make GPUs obsolete within 3–4 years." Asset valuations are inflated, future costs understated. - Jim Chanos, 2026
Cumulative total for 2026–2028: an estimated ~$176B in depreciation expense has been "pushed out" - it will have to show up later as impairment or accelerated depreciation. - Burry estimate
The opposing side

"GPUs are already contracted for nearly their full useful life"

"Most of the GPUs we're buying are already contracted for nearly their full useful life. The risk he [Burry] is pointing to largely doesn't exist - it's already been sold." - Amy Hood, Microsoft CFO, Q2 FY2026 ER call
"A100s shipped six years ago are still running at full utilization." CUDA lets new software run on old hardware. - Colette Kress, Nvidia CFO, 2026
Customers recontract H100s within 5% of the original contract price when they come due - real-world evidence that value holds up across multiple cycles. - CoreWeave CEO, CNBC, 11/14/2025
Burry's track record since 2008: shorted Tesla in 2021 at a loss, closed the Scion fund in 11/2025 (citing "valuations not in sync with markets"). Being "right" about 2008 doesn't guarantee being "right" about AI in 2026. - Reuters 11/2021, Zacks/TradingView 2026
2025 empirical evidence - AMZN already reversed course

In early 2025, AMZN became the first hyperscaler to formally shorten server useful life. Its 4Q24 report (2/2025) showed AMZN reversing from 6 years back to 5 years for part of its server fleet - along with $920M in accelerated depreciation in Q4 2024 and an admission that 2025 operating income would fall ~$0.7B because of this adjustment. AMZN's own management concluded that part of its prior-generation server fleet wouldn't last the full 6 years assumed when it extended useful life to 6 in 2024.

This is a data point favoring Burry's side - not theory, but an actual accounting decision made. However: (a) AMZN reversed only part of its fleet, not all of it; (b) MSFT, GOOGL, META haven't followed; (c) AMZN may have reversed because of its custom-chip strategy (Trainium) replacing hardware faster, not because NVDA GPUs are obsolete. This evidence is necessary but not yet sufficient to confirm an industry-wide pattern.

A straight assessment

Amy Hood's (MSFT CFO) strongest counterargument shifts the risk from "GPUs becoming obsolete" to "can AI customers actually pay their take-or-pay contracts" - which actually overlaps with Risk 2 in Part IV. That means the risk hasn't disappeared, just changed shape. Combined with AMZN's early-2025 useful-life reversal, the balance has tilted slightly more toward Burry than it was at the end of 2025 - but with MSFT, GOOGL, META not yet following, the overall conclusion remains: the thesis could be right, and it could be wrong; it will take another 2–3 years of Blackwell-utilization data to know. The two main mechanisms in Parts II and III stand on their own without needing this thesis - which is why it's split into an appendix.

Read next

More from the shelf

Jul 24, 2026Why Losses Lead to Riskier BetsJun 4, 2026Nassim Nicholas Taleb: Panic Early Is Better Than Panic LateApr 29, 2026Tenbagger: Hunting X10 Stocks The Peter Lynch WayApr 28, 2026Margin of Safety: Why The Richest Man On Wall Street Fears Losing Money The Most

Pass it on

If it found you, share it kindly

XEmail

03 Discussion

Leave a note

A considered space for questions, counterpoints, and useful additions. Civil, on-topic, signed.

Reader notes

...

Loading notes...