I work in the industry Explain the basics

Where the $753B is going: hyperscaler 2026 capex in 12 line items

Raymond Xu

May 10, 2026 · 5 min read

Big-4 hyperscaler capex guidance for 2026 lands at roughly $725B; pulling Oracle in pushes the top-5 to ~$753B, and TrendForce's top-9 CSP aggregate clears $830B. The number is now bigger than the entire 2025 US federal discretionary defense budget. The interesting question is no longer how big it is, but where the dollars actually land.

Five companies, two trillion dollars over three years

Amazon leads with $200B and is the only hyperscaler to break the $200B ceiling for a single year. Microsoft and Alphabet are roughly tied around $190B, with Microsoft's CFO attributing $25B of the increase to memory and component price inflation alone.

Meta lands in the $115–135B range, tracking the Hyperion campus build-out in Richland Parish, LA. Oracle is the smallest of the five at ~$50B but is growing fastest off a small base, almost entirely on the back of OpenAI Stargate commitments.

Where the $753B actually goes

Hyperscalers do not disclose capex at this granularity, so the line items below are estimates derived from analyst aggregates (Tom's Hardware, CreditSights, Introl, TrendForce) plus industry unit-cost data ($20M+/MW for AI shells, $4.5–5.2M/MW for liquid-cooled halls). The IT vs physical split anchors on Microsoft Q2 FY26 (~67% short-lived assets) and Alphabet's reported ~60/40 server/facility mix.

NVIDIA AI systems are the single biggest line at ~$260B — and that's the system-level capture, not just discrete GPU silicon. A single GB200 NVL72 rack runs $3–4M; the dollar concentration into one supplier is unprecedented at this scale.

Custom AI silicon (TPU, Trainium, MTIA, Maia) lands at ~$75B and is the line item most likely to be revised up over the next two years as in-house ASICs displace NVIDIA share at the inference tier. General-purpose servers, networking, and capitalized software round out the IT bucket at ~$485B total — roughly two-thirds of all spend.

The $148B that sits behind a regulator

Of the ~$240B physical-infrastructure slice, roughly $148B is functionally gated by site-readiness clearance: land acquisition, interconnection upgrades, power-generation contracts, electrical equipment with long-lead substation dependencies, and the data center shell itself.

Microsoft is already sitting on $80B in unfulfilled Azure orders because GPUs are arriving faster than the substations to power them. The shell construction can finish on schedule and the racks can land on the floor, but the meter doesn't spin until the interconnection clears — and the dollars stranded in that gap are now bigger than most countries' annual energy infrastructure budgets.

ERCOT Batch Zero (April–July 2026 deadlines) and PJM's flex-interconnection year are where this $148B converges in real time. The capex commitments are already booked; the regulatory clock is what determines whether they earn revenue in 2026 or 2028.

Three implications for site-readiness

First, the asymmetry between IT and physical-infra timelines is widening: chips ship in quarters, substations take 36–60 months. Any capacity that compresses the regulatory tail (PCLR, NFCD, BYONG, RM26-4-000) is worth a multiple of its sticker discount because it converts stranded capex into revenue earlier.

Second, the $148B gated number is a proxy for the addressable market for site-readiness software. If even 1% of stranded-capex value can be unlocked through better interconnect documentation, faster air-permit math, and live docket intelligence, that's a $1.5B annual willingness-to-pay across the cohort.

Primary sources

Get started

Type your site in. See the de-rate.

The calculator returns an effective MW number, the binding rule, and a $/MW-yr net value as you type.

Book a call How it works

Five companies, two trillion dollars over three years

Where the $753B actually goes

The $148B that sits behind a regulator

Three implications for site-readiness