$725B
Big Four capex, 2026 (FT) — up 77% from $410B in 2025
Market Dossier · AI Infrastructure Cycle
The $600 Billion
Question
The companies driving the market are real, profitable, and dominant. The spending they are doing is not yet justified by the revenue that spending requires. This is the gap — and the math of what happens if it doesn't close.
01 — The Coverage Ratio
The build is running years ahead of the revenue.
Using David Cahn's framework: every dollar of compute needs a second dollar of energy, buildings and power to run it, so AI infrastructure must throw off roughly 2× the chip spend in end revenue to justify itself. Against that hurdle, here is where the ecosystem actually sits in 2026.
Annual Revenue Required vs. Reality — 2026
All figures annualized, US$ billions. "Required" = revenue needed to justify current infrastructure spend.
Revenue
required to justify
required to justify
~$600B
AI-specific
capex spent
capex spent
~$450B
AI ecosystem
revenue actual
revenue actual
~$100B
COVERAGE RATIO ≈ 0.25× — up from 0.15× in 2025, still roughly one-sixth of the hurdle
Source: Sequoia / D. Cahn "AI's $600B Question"; Futurum & philippdubach.com capex-revenue analysis (Mar 2026); IEEE ComSoc / Introl 2026 capex estimates.
02 — The Escalation
The largest concentrated infrastructure cycle in corporate history.
Big Four hyperscaler capital expenditure has gone vertical. The 2026 figure represents a 77% jump in a single year — and quarter-on-quarter, the spend is now nearly 4× what it was three years ago.
75%
Share of hyperscaler capex tied directly to AI infrastructure
3.7×
Q1 2026 spend vs. Q1 2023 ($130B vs. $35B)
~90%
Of GPU accelerator spend captured by a single vendor (Nvidia)
Big Four Combined Capital Expenditure
US$ billions per year. Amazon · Microsoft · Alphabet · Meta.
2023
2024
2025
2026
(guided)
(guided)
Source: Financial Times capex compilation; CreditSights / Introl 2026 hyperscaler estimates. 2023–24 figures approximate combined reported capex.
03 — The Squeeze
The income statement has cushion. The cash flow has almost none.
This is the genuinely new risk versus 1999. When capex was a fraction of cash flow, a demand miss was survivable. Now the spend consumes nearly all internally generated cash — and the gap is being plugged with a wall of new debt.
How Much Room Is Left
Each bar shows a balance-sheet pressure gauge for the Big Five hyperscalers, 2026.
Source: Bank of America; Morgan Stanley debt-issuance forecast; philippdubach.com FCF analysis; CreditSights capital-intensity data (45–57% of revenue).
04 — Then vs. Now
A different bubble than 1999 — sturdier companies, tighter structure.
On cyclically-adjusted valuation the two eras nearly match. On profitability and forward multiples, today is far healthier. On concentration, today is actually worse. The verdict column reads from the perspective of bubble risk.
| Metric | Dot-Com Peak · 2000 | AI Cycle · 2026 | vs. 1999 |
|---|---|---|---|
| Shiller CAPE ratiocyclically-adjusted P/E | 44.2 | 41.6 | Comparable |
| Nasdaq forward P/Eprice ÷ next-yr earnings | ~60× | ~26× | Healthier |
| Leading "picks & shovels" stockCisco 2000 vs. Nvidia 2026, fwd P/E | ~100× | ~30× | Healthier |
| Profitable tech IPOs at peakshare operating in the black | 14% | Majority | Healthier |
| Investment as share of US GDPAI 2026 vs. dot-com telecom/tech | 1.2% | ~0.4% | Smaller |
| Top-7 share of S&P 500index concentration | ~18%* | ~40% | Worse |
Source: multpl.com / Shiller (CAPE); VisualCapitalist & IntuitionLabs (fwd P/E); Bravos Research & Benzinga (Cisco/Nvidia); Red Lotus Capital (IPO profitability); Richmond Fed (GDP share); MacroMicro (Mag-7 weight). *2000 top-7 S&P weight approximate.
05 — The Multiplier
Why a 10% miss is not a 10% problem.
The chain is leveraged at three separate points. A modest shortfall in end-user demand compounds as it travels down the stack — because capex is gearing, capex is also somebody's revenue, and the customers are concentrated.
1
The trigger
A 10% miss in end-user AI revenue
Roughly $60B of expected demand against a ~$600B target fails to materialize. On its own, manageable.
2
Multiplier 1 · the 2× TCO gearing
→ ~$120B of infrastructure loses its justification
Because chips are only half of total cost of ownership, every dollar of missed end-revenue implies two dollars of infrastructure spend that no longer pencils out. The miss doubles on the way down.
3
Multiplier 2 · capex is someone's revenue
→ A ~$45–70B direct hit to the supplier layer
If hyperscalers cut next year's capex 10%, that is a direct revenue cut to Nvidia, memory makers, builders and power providers — whose valuations are priced on capex compounding, not shrinking.
4
Multiplier 3 · customer concentration
→ Discrete risk to a $1.05T backlog
Half of the $2.1T cloud backlog rests on two cash-burning labs. A demand miss that slows even one major commitment removes a large, lumpy contract, not an even 10% sliver.
∑
The result
A single-digit demand miss → a 20–30% equity revaluation
In the soft-landing case the core thesis survives but multiples compress. The tail risk — the "telecom replay" — sees 85–95% of capacity idle and $500B+ in shareholder value erased.
Source: Breckinridge (customer concentration); businessengineer.ai (backlog $2.1T, $1.05T from OpenAI/Anthropic); Truthbit/Medium (soft-landing & telecom-replay scenarios).
06 — The Valuation Mechanism
Bubbles de-rate violently because both sides move at once.
A forward P/E only holds while the growth story holds. A demand miss doesn't just lower next year's earnings (the E) — it removes the justification for paying 26× (the multiple). The two effects multiply; they do not add. A worked illustration:
Forward earnings
100 → 95
~5% cut from impairments & slower growth
×
The multiple
26× → 19×
"Durable compounding" premium comes out
=
Price level
2600 → 1805
earnings × multiple
Even a modest 5% earnings cut, paired with a de-rate from 26× to 19×, produces a price decline of —
−27%
For comparison: the Nasdaq fell ~78% in 2000–02 — not because revenue fell 78%, but because a ~60× multiple collapsed onto disappointing earnings. The same two-sided mechanism is live today. The difference is that today's earnings are real, so the floor sits far higher.
Illustrative model. Inputs flexible — actual outcome scales with how much of today's ~26× is "premium" vs. justified growth, and with the realized depreciation life of GPU assets (5-yr straight-line vs. the 2–3-yr life some short-sellers assume).
07 — The Fragile Base
Half the backlog rests on two companies that don't yet profit.
The $2.1 trillion revenue backlog underpinning hyperscaler valuations looks diversified. It isn't. The single largest source of demand is a pair of venture-funded labs still burning billions a year.
Cloud Revenue Backlog — Big Four, 2026
Total ≈ $2.1 trillion in contracted future revenue.
$1.05T
OpenAI & Anthropic — two cash-burning AI labs, each losing billions annually despite explosive revenue growth
$1.05T
All other customers — the broad base of enterprises and cloud tenants combined
Source: businessengineer.ai "AI Capex Map" (May 2026), citing FT & The Information disclosures incl. Anthropic's $200B / 5-yr Google Cloud commitment.
08 — The Synthesis
The income statement has plenty of cushion — these are wildly profitable companies. The cash-flow and valuation structure has almost none. AI revenue must roughly sextuple toward the ~$600B hurdle before the first vintage of GPUs fully depreciates. A 10% miss is survivable as a business — but not as a valuation.
↑ The case it holds
Cloud took 7+ years from hype to profit yet created enormous value. Railroads and fiber were both overbuilt — and both excesses funded infrastructure that later ran at capacity. At <0.4% of GDP, the bet may simply be early, not irrational.
↓ The case it breaks
Coverage sits near 0.25×, capex eats ~90% of cash flow, and the backlog leans on two pre-profit labs. The payback math assumes near-term monetization that 95% of enterprise AI pilots have so far failed to deliver.