Hyperscaler AI Capex: A GAAP vs. Cash Flow Analysis

$500B in 2026 capex across Amazon, Alphabet, and Microsoft — modelled step by step, with assumptions explained.
Inspired by @deadneurons.

Hover dotted numbers for derivation
A1 = Assumption
SRC = Sourced fact

Part 1: What Just Happened

In early 2026, the three largest cloud companies guided to a combined ~$500 billionAmazon $200B (explicit guidance, Q4 2025 earnings) + Alphabet $175–185B (explicit guidance, Q4 2025 earnings) + Microsoft ~$125B (estimated midpoint of $105–145B analyst range; MSFT did not give full-year guidance, Q2 capex was $37.5B, +66% YoY but guided sequential decreases). Combined range: $480–530B.All from Q4 2025 / Q2 FY2026 earnings calls in capital expenditure for the year:

Company2026 CapexCertainty
Amazon$200BExplicit guidance
Alphabet$175–185BExplicit guidance
Microsoft~$125BEstimatedMicrosoft's fiscal year ends June 30. It did not give full-year guidance. Q2 capex was $37.5B (+66% YoY), but management guided to sequential decreases. Analyst estimates range $105–145B; we use the $125B midpoint.Q2 FY2026 earnings, Jan 28 2026
Combined~$500B

For scale: ~$500B is roughly Argentina's entire GDPArgentina GDP ~$530B (2024, World Bank). A G20 country's annual output, spent in one year by three companies. — a G20 country's annual output, spent in one year by three companies.

Market reaction

The market sold everything. Microsoft dropped ~10% on January 29-$357B in market cap, the second-largest single-day dollar loss ever, behind Nvidia's -$593B on Jan 27, 2025.CNBC. Amazon fell 8–10% intradayFeb 6, 2026. Closed down ~6% after partial recovery.Q4 2025 earnings reaction on February 6. Alphabet, cushioned by a strong earnings beat, closed essentially flat (-0.54%)Alphabet paired its capex guidance with a large earnings beat, absorbing the shock.Q4 2025 earnings, Feb 4 2026.

The market's message: $500 billion is too much. The rest of this post argues that the market may be over-weighting the lens that makes this investment look worst.


Part 2: The Bear Case — What GAAP Earnings Say

Here's how a sell-side analyst models this investment. Every assumption is labelled.

Step 1: Split the $500B

Data centre capex splits into short-lived equipment (servers, GPUs) and long-lived facilities (buildings, power). Industry data puts the split at roughly 70% IT / 30% facilities A1Industry average from Dell'Oro Group and IoT Analytics (2024) for hyperscaler capex. In practice varies by company and by how many greenfield sites vs. equipment refreshes are in the mix. A more building-heavy split (65/35) would shift ~$25B from servers to buildings, modestly reducing annual depreciation.Dell'Oro Group, IoT Analytics 2024:

CategoryShareAmount
Servers & GPUs70%$350B
Buildings & infra30%$150B

Step 2: Depreciation

Each asset depreciates on a schedule sourced from the companies' 10-K filings:

CompanyServersBuildings
Microsoft6 years5–15 years
Amazon5 yearsUp to 40 years
Alphabet6 years25–30 years

Amazon is the outlier — it shortened server lives in January 2025, explicitly citing AI hardware evolution. Since Nvidia ships new GPU architectures every 12–18 months, we use 5 years for all servers A2Conservative choice aligned with Amazon's Jan 2025 decision to reverse its earlier extension (from 6yr back to 5yr). Microsoft and Alphabet still use 6 years. Using 5 years increases annual depreciation and makes GAAP returns look worse — this is deliberate conservatism.Amazon 10-K (Jan 2025), Microsoft 10-K, Alphabet 10-K and 20 years for buildingsMidpoint of the 15–30 year range reported across the three companies' 10-K filings. Amazon reports "up to 40 years" but that includes land improvements; core structures cluster around 15–25 years.. Annual depreciation:

$77.5B/yearServers: $350B / 5yr = $70.0B/yr Buildings: $150B / 20yr = $7.5B/yr ────────── Total: $77.5B/yrDerived directly from capex (A1) and 10-K depreciation schedules (A2). This is the hardest number in the analysis. total depreciation — $70B$350B in servers ÷ 5-year useful life = $70B/yr from servers, $7.5B$150B in buildings ÷ 20-year useful life = $7.5B/yr from buildings. This is the most reliable figure in the entire analysis.

Step 3: A simple P&L model for the 2026 capex cohort

Now we model revenue, costs, and profit over the 5-year server life. This requires assumptions that cannot be sourced — they are modelling choices.

A3 — Cash operating costs: $35B/yr escalating to $45B

Cash costs are the real dollars spent to keep data centres running (electricity, cooling, maintenance, staff, bandwidth) — distinct from depreciation. None of the three companies disclose this directly. We derive top-down from AWS's 2024 segment data:

AWS 2024 (from Amazon 10-K): Revenue: $107.6B Operating income: $39.8B Total operating expenses: $67.7B Depreciation & amortisation: −$13.3B ────── Non-depreciation operating costs: $54.4B

That $54.4B includes everything: infrastructure costs and personnel, R&D, SBC, sales. AWS does not disclose the split. We estimate roughly half is people/software, leaving infrastructure at ~$25–30B, or about 25–28% of revenue. The "roughly half" is a judgement call — it could be 40/60 or 60/40, shifting the estimate ±$5–10B.

Scaling to our $250B cohort revenue: $250B × 12–14% = $30–35B/year. We use 12–14% (not 25–28%) because A6 already captures personnel/R&D costs at 20% of revenue — using both would double-count.

We start at $35B in year 1 and model escalation to $45B by year 5, reflecting 30–40% energy cost increases projected by IEA and Fabricated Knowledge as data centre power demand surges.

A4 — Steady-state revenue: $250B/yr (the biggest assumption)

How much annual revenue does $1 of data centre capex generate? We derive this from AWS, the only hyperscaler that discloses cloud-specific capex.

AWS cumulative capex 2020–2023: ~$104B AWS 2023 revenue: $90.8B Revenue per $1 of cumulative capex: $0.87/yr

Caveats: This likely overstates revenue efficiency — AWS spent significant capex before 2020 that is still generating revenue (servers from 2019 were still within useful life in 2023). The true installed base is larger than $104B. The 2024 data ($107.6B revenue on ~$157B cumulative) gives $0.69, but that's depressed because the 2024 capex wave ($53B) hadn't yet ramped.

At $0.87/yr (historical): $500B × 0.87 = $435B/yr At $0.69/yr (2024): $500B × 0.69 = $345B/yr

Both overstate the yield on new capex (historical ratio benefits from fully depreciated assets still earning revenue). We haircut to $0.50/yr — well below the range — giving $250B/year.

This is the single largest assumption in the model. The $0.50 haircut from $0.87 is a judgement call. The gap between $0.50 and $0.87 spans $185B in annual revenue. A more optimistic analyst could justify $0.60 ($300B); a pessimistic one could argue $0.40 ($200B). Additionally, cloud compute prices tend to fall over time as competition and newer hardware increase supply. Our flat $250B assumes volume growth exactly offsets price declines.

A5 — Revenue ramp: linear from $0 to $250B over 3 years

Data centres take 1–2 years to build and fill. Even after physical completion, enterprise contracts ramp gradually — sales cycles are long, migrations take time, and workloads scale up incrementally. A 3-year ramp to full utilisation is standard industry practice.

Year 1: $250B × 1/3 = $83B Year 2: $250B × 2/3 = $167B Year 3: $250B × 3/3 = $250B (steady state)

A faster 2-year ramp would improve returns; a slower 4-year ramp would reduce annualised GAAP return by ~1.8 percentage points.

A6 — Operating expenses (R&D, sales, G&A): 20% of revenue

This covers everything that isn't COGS (depreciation + cash infrastructure costs). AWS's non-infrastructure, non-depreciation costs run ~23–25% of segment revenue. We use 20% because this is incremental capacity on existing platforms: the core R&D, sales teams, and corporate overhead already exist. New capacity doesn't require proportionally new headcount.

A higher ratio (e.g. 25%) would reduce steady-state operating income by ~$12.5B/year and cut the GAAP return by ~1 percentage point.

A7 — Tax rate: 15%

Effective tax rates for MSFT, AMZN, and GOOG range 10–18% in recent filings. We use 15% as a round midpoint. Two simplifications: (1) no tax on Year 1's loss, and (2) we don't carry the Year 1 loss forward as a tax shield in later years. Both slightly understate actual returns — making this conservative.

With assumptions A3A7 in place, here's the year-by-year P&L:

Yr 1Yr 2Yr 3Yr 4Yr 5Total
Revenue831672502502501,000
COGSDepreciation ($77.5B, fixed) + cash costs (A3, escalating). Depreciation starts when assets are placed in service regardless of utilisation — Year 1 electricity would actually be lower at ~33% load, making this conservative. See A3 dropdown for full derivation.-113-115-118-120-123-589
Gross profit-3052132130127411
OpexA6 applied to each year's revenue. See A6 dropdown for rationale.-17-33-50-50-50-200
Operating income-4719828077211
TaxA7 applied to positive operating income only. No tax on Year 1 loss; loss not carried forward. See A7 dropdown for detail.0-3-12-12-12-39
Net income-4716706865172

All figures in $B

~6% annualised GAAP return. $172B cumulative net income on $500B invested = 34.4% total, or (1.344)1/5 − 1 = 6.1%$172B / $500B = 34.4% total (1.344)^(1/5) − 1 = 6.1% annualised annualised. Mediocre. This is the number that drives sell-side downgrades and panicky headlines.
Why Year 1 looks so bad

Year 1 is deeply negative (-$47B). You're eating $77.5B in depreciation plus $35B in cash costs on buildings that are mostly empty. GAAP requires depreciation to start when assets are placed in service, regardless of utilisation. That loss hits the income statement in 2027 and fuels the panicky headlines.

Sensitivity analysis
If you change…Impact
Steady-state revenue ±$25B±1.3 pp
Server life 5→4 years-2.3 pp
Ramp 3→4 years-1.8 pp
Revenue ratio $0.50→$0.40-2.5 pp

Optimistic ($275B revenue, fast ramp) → ~9%. Pessimistic ($200B revenue, 4-yr ramp) → ~0%. The GAAP lens puts this investment between 0% and 9% annualised.


Part 3: The Bull Case — What Cash Flow Says

Forget the income statement. Think about cash.

You hand $500B to Nvidia and various construction companies in 2026. That cash is gone. In return you get buildings full of GPUs. Those buildings generate revenue. The question is: how much cash comes back, and when?

A note on the cohort approach: We model the 2026 capex as a standalone investment generating its own revenue stream. In reality, customers run workloads across a mix of old and new infrastructure — you can't cleanly attribute specific dollars of revenue to a specific year's capex. This is a standard simplification in infrastructure modelling (it's how REITs model new developments too), but it means our per-cohort returns are illustrative, not precise. Part 4's fleet-wide analysis partially addresses this by modelling all cohorts together.

Step 1: From operating income to cash

In steady state (years 3–5), operating income averages ~$80B/yearAverage of Years 3–5 from the P&L table: ($82 + $80 + $77) / 3 = $79.7B ≈ $80B. Declining slightly as cash costs escalate from $40B to $45B.. That number was computed after subtracting $77.5B in depreciation. But depreciation is not a cash outflow — no cheque leaves the building each year as your servers age. The cash left in 2026 when you bought them.

However, taxes are a cash outflow. At 15% (A7)Applied to positive operating income only. Steady-state tax: ~$80B × 15% = ~$12B/year. See A7 dropdown for detail., the government takes ~$12B/year off the top. So the actual cash generated:

Cash from operations: ~$145.5B/yearOperating income (avg yr 3–5): $80B − Tax at 15% (A7): $12B = After-tax income: $68B + Depreciation (non-cash): $77.5B ────── = Cash from operations: $145.5BDepreciation is a real economic concept — servers are wearing out — but it does not cost cash in the year it's recorded. The cash consequence shows up later when you replace the worn-out equipment. Taxes, however, are a real cash outflow and must be subtracted. at steady state

Step 2: Subtract maintenance capex

Some cash must be reinvested — replacing failed drives, swapping dead nodes, patching power systems. This is ~$25B/year A8Server failure/replacement: $350B × 6.5% = ~$23B Building maintenance: $150B × 1% = ~$1.5B ────── Total: ~$25B/yrThe 5–8% server rate is an estimate — hyperscalers don't publish component failure rates. This only covers piecemeal repairs (dead drives, failed nodes), not wholesale fleet replacement — that scenario is modelled separately in Risk 1. in maintenance capex.

Step 3: Free cash flow — year by year

Net Inc.+ Dep.= Cash Ops− Maint.= FCF
Yr 1-47+77.530.5-255.5
Yr 216+77.593.5-2568.5
Yr 370+77.5147.5-25122.5
Yr 468+77.5145.5-25120.5
Yr 565+77.5142.5-25117.5
Total$434.5B

All figures in $B

~24% steady-state cash yield.Avg steady-state FCF (yr 3–5): ($122.5 + $120.5 + $117.5) / 3 = ~$120B/yr FCF yield = $120B / $500B = 24.0% Payback in ~5.6 yearsAfter Yr 3: $5.5 + $68.5 + $122.5 = $196.5B After Yr 4: $196.5 + $120.5 = $317.0B After Yr 5: $317.0 + $117.5 = $434.5B Still short by $500 − $434.5 = $65.5B At yr-5 run rate ($117.5B/yr): $65.5 / $117.5 ≈ 0.56 years → Payback at ~5.6 yearsUndiscounted payback. A 10% discount rate stretches this further.. 5-year cumulative return: 87%$434.5B cumulative FCF / $500B invested = 86.9% ≈ 87%. Same $500B, same assets, same customers — completely different conclusion.
FCF sensitivity to revenue assumption (A4)

A4 ($250B steady-state revenue) is the biggest assumption. Here's how the FCF picture changes if revenue is materially higher or lower:

Steady-state revenueSS FCFFCF yield5-yr cumul.5-yr return
$200B (pessimistic)~$86B/yr17%$296B59%
$250B (base case)~$120B/yr24%$435B87%
$300B (optimistic)~$154B/yr31%$573B115%

At $200B revenue (i.e. $0.40 per dollar of capex instead of $0.50), the investment still generates positive cash flow but takes well over 6 years to pay back. At $300B ($0.60 per dollar), it pays back within 4 years. The $100B swing in revenue assumption moves the 5-year return by ~28 percentage points in either direction.

Same investment, two lenses

MetricGAAPCash flow
Annual return (steady-state)~6%~24%
Cumulative 5-year34%87%
PaybackNever really pays back~5.6 years

Why the gap exists

It's not a trick. $112.5BBuilding investment: $150B Depreciation over 5 years: $7.5B × 5 = $37.5B Remaining on balance sheet: $150B − $37.5B = $112.5BThis is data centre buildings that will generate revenue for another ~15 years. GAAP can't show their future value within a 5-year window. Caveat: Data centre buildings are highly specialised — their resale value may be below book value if AI demand shifts or facilities become technologically obsolete. A generic office building has broad resale appeal; a 50MW data centre does not. The $112.5B is an upper bound. of the $500B went into buildings that depreciate over 20 years. Our analysis window is only 5 years. That remaining book value isn't lost — it's buildings that keep earning for another 15 years. GAAP penalises you for expensing a 20-year asset over a 5-year window. Cash flow correctly recognises the buildings were paid for once, in 2026, and are still there.

The REIT analogy

This is identical to how rental property works. Buy a house for $500K. Rent it for $30K/year net. Your accountant depreciates the $400K building over 27.5 years ($14.5K/yr).

GAAP annual return: ($30K − $14.5K) / $500K = 3.1%
Cash annual return: $30K / $500K = 6.0%

Nobody evaluates rental property on net income after depreciation. REITs report Funds From Operations (FFO) — literally "net income plus depreciation" — because the industry learned that GAAP earnings are misleading for long-lived physical assets. Data centres are the same physical object. The accounting should be evaluated the same way.


Part 4: Risk Scenarios

The ~24% steady-state FCF yield is the bull case. Here's what breaks it.

Risk 1: Faster hardware obsolescence

Our model assumes $25B/year in maintenance. But if GPUs don't just wear out but become obsolete — Nvidia ships new architectures every 12–18 months (Hopper → Blackwell → Rubin) — then "maintenance" becomes "full replacement" at $70B/year$350B server base × 20% annual replacement rate (= 1/5yr, matching our A2 server life) = $70B. This is a full refresh on our assumed 5-year cycle. If full refresh every 3 years, replacement jumps to ~$117B/yr, which would collapse steady-state FCF yield to ~6%.:

Base caseObsolescence
Replacement capex$25B/yr$70B/yr
Steady-state FCF$120B/yr$75B/yr
FCF yield24%15%
Payback~5.6 yr~9 yr
15% FCF yield — still positive, but significantly worse. Manageable on its own.

Risk 2: Obsolescence + weak demand

The scenario that genuinely breaks things: fast hardware turnover combined with disappointing demand — data centres half empty because enterprises decided they didn't need that much AI compute.

At ~50% utilisationRevenue at ~50%: $125B COGS: $100B (dep. $77.5B fixed + $22B cash costs scaled down with load) Gross profit: $25B Opex at 20%: −$25B Operating income: $0B Tax (15% of $0): $0B + Depreciation: $77.5B = Cash from ops: $77.5B − Replacement ($70B):−$70B = FCF: $7.5B with accelerated replacement: FCF yield collapses to 1.5%. Payback ~67 years. This is the scenario that should worry investors — not "they're spending too much" but "they're spending on assets useful to fewer customers for less time than expected."

Risk 3: Financing costs

Our model implicitly assumes the $500B is funded from operating cash flow. Combined OCF is ~$440B — but these companies also have existing capex commitments, dividends, buybacks, and debt service. To the extent the incremental spend is debt-financed, interest costs are absent from our model.

At current credit ratings (AA+/Aaa), hyperscaler debt costs ~4–5%. If half the $500B were debt-financed, that's ~$10–12B/year in interest — reducing steady-state FCF from ~$120B to ~$108–110B and the yield from ~24% to ~22%. Material but not thesis-changing. The bigger risk is if debt loads grow large enough to constrain future investment flexibility.

Risk 4: Customer concentration and circular revenue

A non-trivial share of current AI cloud revenue comes from a small number of large customers — including startups that the hyperscalers themselves have funded. Microsoft has invested ~$13B in OpenAI, which is one of Azure's largest AI customers. Google has backed Anthropic. Amazon has invested in Anthropic too. This circular funding dynamic — where the hyperscaler invests in a startup that then spends at the hyperscaler's cloud — flatters current growth rates. If these startups fail to reach profitability and the funding dries up, a meaningful chunk of AI cloud demand could evaporate simultaneously.

Our revenue assumption (A4) does not explicitly account for this risk, though the haircut from $0.87 to $0.50 per capex dollar provides some implicit buffer.

What to watch: the growth-rate threshold

By 2027, four overlapping capex cohorts will generate ~$178B/yearEach cohort's capex is split 70/30 (A1), then depreciated at 5yr servers / 20yr buildings (A2): Capex Servers(70%) Buildings(30%) ÷ 5yr ÷ 20yr Pre-24: $180B → $126B/5=$25B + $54B/20=$3B = $28B 2024: $180B → $126B/5=$25B + $54B/20=$3B = $28B 2025: $288B → $202B/5=$40B + $86B/20=$4B = $44B 2026: $500B → $350B/5=$70B + $150B/20=$8B= $78B ───── Total depreciation hitting by 2027: ~$178BCapex per year sourced from annual filings. Pre-2024 = only the portion with servers still within 5-year life by 2027. in total depreciation. Add ~$75BWe know A3 for our $500B cohort: $35–45B/yr, midpoint $40B. But by 2027, the entire fleet is ~$1.15T of cumulative capex (all four cohorts). If cash costs scaled linearly with capex:$40B × ($1,150B / $500B) = ~$92BWe discount to ~$75B because older cohorts (pre-2024, 2024) use less power-hungry hardware — fewer GPUs, more traditional compute — so their per-dollar electricity costs are lower. in fleet-wide cash costs → total COGS of ~$253B. What cloud revenue growth absorbs this?

Current combined cloud revenue: ~$308BAWS: ~$142B ($35.6B × 4) Azure (est.): ~$95B (MSFT doesn't disclose exact figures) Google Cloud: ~$71B ($17.7B × 4) ────── Combined: ~$308B. Current blended growth: 34%Revenue-weighted blend: AWS 24% × 0.46 + Azure 39% × 0.31 + GCP 48% × 0.23 = 34%All growth rates from Q4 2025 / Q2 FY2026 earnings.

The key question: can revenue grow fast enough to keep gross margins near ~60%~60% is the current cloud gross margin for all three companies (AWS ~62%, Azure ~59%, GCP ~56% in recent quarters). It's the level Wall Street treats as "healthy" for cloud infrastructure — below it, analysts start questioning whether the business model works at scale. — the level all three cloud businesses currently run at? Each row below takes a growth rate, compounds it for 2 years to get 2027 revenue, then calculates gross margin as (revenue − $253B COGS) / revenue:

Blended growth2027 Rev.Gross marginVerdict
43%$633B$308B × 1.43² = $633BThe growth needed to maintain exactly 60%: Revenue × 0.40 = $253B → Revenue = $633B60%Target maintained
34% (current)$553B$308B × 1.34² = $553B Margin: ($553B − $253B) / $553B = 54%54%Modest compression
20%$443B$308B × 1.20² = $443B Margin: ($443B − $253B) / $443B = 43%43%Concerning
10%$373B$308B × 1.10² = $373B Margin: ($373B − $253B) / $373B = 32%32%Severe
0%$308B$308B × 1.00² = $308B Margin: ($308B − $253B) / $308B = 18%18%Disaster

The practical rule: At current 34% blended growth, margins compress modestly — manageable. Below ~20%, compression becomes serious. Below ~10%, the investment thesis is in real trouble. None of these companies need to accelerate — but they can't afford a sharp deceleration.

This is a rougher estimate than the rest of the model — it depends on how cash costs scale across the fleet and how much pre-2024 depreciation is still running. Treat thresholds as approximate.


Part 5: Conclusion

Three of the most dominant businesses in history, with combined operating cash flow of ~$440B/yearMSFT $136.2B + AMZN $139.5B + GOOG $164.7B = $440.4B2025 annual filings, are investing in assets that under our model yield ~24% in annual cash returns at steady state — but only ~6% on a GAAP basis. The stocks sold off 6–10% because the market is pricing the latter.

The gap between those two numbers is not a trick. It is the mechanical consequence of depreciating 20-year buildings over a 5-year analysis window. Whether the GAAP view or the cash flow view better reflects reality depends on one thing: whether the revenue actually shows up. At current cloud growth rates (34% blended), the model says it does. If growth decelerates sharply below ~20%, it doesn't.

The depreciation will hit the P&L exactly as the bears predict. Quarterly EPS will disappoint. There will be confident people on television explaining why the AI bubble has burst. The number to watch is not EPS — it's blended cloud revenue growth.

All assumptions

LabelAssumptionValue
A1Capex split (servers / buildings)70% / 30%
A2Server life / Building life5 yr / 20 yr
A3Cash operating costs$35B → $45B (escalating)
A4Steady-state revenue$250B/yr (biggest assumption)
A5Revenue ramp0 → $250B over 3 yr
A6Opex (R&D, sales, G&A)20% of revenue
A7Tax rate15%
A8Maintenance capex$25B/yr

Sourced facts

FactValueSource
Amazon 2026 capex$200BQ4 2025 earnings
Alphabet 2026 capex$175–185BQ4 2025 earnings
MSFT Q2 FY2026 capex$37.5BQ2 FY2026 earnings
MSFT server dep.6 yr10-K (from 4yr, Jul 2022)
AMZN server dep.5 yr10-K (from 6yr, Jan 2025)
GOOG server dep.6 yr10-K (from 4yr, Jan 2023)
AWS 2024 revenue$107.6BAmazon 2024 10-K
AWS 2024 op. income$39.8BAmazon 2024 10-K
AWS Q4 2025 growth24% YoYQ4 2025 earnings
Azure Q4 2025 growth39% YoYQ2 FY2026 earnings
GCP Q4 2025 growth48% YoYQ4 2025 earnings
Combined op. cash flow~$440B2025 annual filings
MSFT mkt cap loss, Jan 29-$357BCNBC
Capex split (IT/facilities)~70/30Dell'Oro, IoT Analytics