The $740 Billion Data Center Buildout Is a Completion Trade Now. Boockvar Says One Shock Could End It.

Aggregate 2026 capital expenditure at the four largest hyperscalers now sits between $720 billion and $745 billion — up roughly 60% from 2025 — and is on track to exceed their combined operating cash flow sometime in the second half of 2026. Microsoft guided to ~$190 billion in calendar 2026 capex; Amazon raised its outlook twice, to ~$220 billion; Alphabet sits at $195–205 billion; Meta Platforms narrowed to $130–145 billion. Per Bank of America, the five largest hyperscalers (including Oracle) may spend nearly 90% of operating cash flow on capex in 2026, versus 65% in 2025.

What changes with that ratio is the cycle's resilience. When capex is comfortably below cash flow, a project delay is an inconvenience. When it is at or above cash flow, any delay forces a balance-sheet choice — and the balance sheet is now the cycle. That is exactly the wiring Boockvar is flagging.

Vertiv is the cleanest expression of the buildout trade. Q2 2026 net sales rose 24% YoY to $3.27 billion; order backlog reached $15.0 billion, up 109% YoY, with a book-to-bill of roughly 2.9x; the company raised full-year guidance across every key metric. Equinix, trading at ~52x forward earnings with $5.0x net debt to EBITDA, is priced for the buildout to compound — not for any of it to slip. Eaton and Schneider Electric print the same shape: backlog up, guidance up, valuations well above industrial peers.

The contrast with NVIDIA is informative. The chip vendor sits at ~24.9x forward with only5.5% of operating cash flow going to capex, no leverage, and — critically — a customer base whose bookings convert to revenue in quarters, not years. NVIDIA's earnings absorb a capex cut faster than the rest of the stack, which is why the hyperscalers' own multiples (not NVIDIA's) are the cycle's canary.

These three triggers are not mutually exclusive. A 50bp move in 10-year yields tightens the financing gap; that same rate move inflates electricity costs and slows data-center payback, which then compounds the ROI gap. The triggers cascade through the same balance sheet, which is why Boockvar's \"it works until it doesn't\" framing is structural rather than directional.

Telecom providers spent roughly $120 billion on capex at the 2000 peak, then watched it fall ~58% to about $50 billion by 2002 as the dot-com unwind hit, per a 2003 Federal Reserve retrospective. The carriers' suppliers — Lucent, Nortel, the optical-component makers — fell 80–95% from peak even though \"demand for connectivity\" turned out to be permanently real. The wiring — long-cycle capex tied to a short-cycle thesis — is identical to today's. The difference is scale: hyperscaler capex is roughly 6× the 2000 telecom peak, and the AI supply chain is 5–10× more concentrated in a handful of listed names.

The bull case for the buildout has been the \"power wall\": grid constraints force discipline, the argument goes, so supply cannot overshoot demand. That framing inverts when paired with local opposition. The same Heatmap Pro data shows roughly $1 trillion of announced data-center investment now faces sustained or meaningful local resistance, and about half of projects that faced backlash in 2026 were ultimately canceled. Disciplined supply still requires delivered megawatts, and the same transmission constraints that cap supply cap deployment.

That is the transmission link from power to equities: Constellation Energy, Vistra and GE Vernova trade on the assumption that every megawatt announced eventually gets built and financed. Local cancellations undermine that assumption at the megawatt layer first — and the listed utility trade has the longest duration in the chain.

Over a 1–3 year horizon, the structural question is whether AI revenue compounds fast enough to absorb the depreciation that the2026 capex cohort is laying down. On a 6–7 year useful life, the capex Microsoft and Amazon are committing this year will hit operating income in2032–2033. If AI services revenue grows at the rate hyperscalers' current multiples imply, the math works; if it grows at a more typical software-decade rate, it doesn't. That is the question Boockvar's clock is counting down to.