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

The five biggest cloud companies — Amazon, Microsoft, Google, Meta, and Oracle — plan to spend more than $750 billion building AI infrastructure in 2026 alone. That is more than the entire US federal defense budget for non-personnel discretionary spending. Where do those dollars actually go?

Five companies are doing almost all of it

Amazon plans to spend $200 billion this year. Microsoft and Google are each around $190 billion. Meta is $115–135 billion. Oracle is the smallest at $50 billion but is growing fastest because of its OpenAI deal.

These five numbers have nearly doubled in a single year. The previous record — set in 2025 — was $410 billion across the same companies. 2026 will be roughly $753 billion.

Twelve places the money actually goes

These companies don't publish a line-item breakdown, so the numbers below are estimates built from analyst reports plus industry construction-cost data. The biggest single bucket — about two-thirds of total spend — goes to computer hardware: AI chips, servers, memory, and networking gear.

NVIDIA's AI systems alone account for ~$260 billion. A single AI server rack from NVIDIA costs $3–4 million; hyperscalers are buying these by the tens of thousands.

The remaining one-third (~$240 billion) goes to physical infrastructure: the buildings, electrical equipment, cooling systems, land, power contracts, and grid connections. Even at this smaller share, the physical-infrastructure number alone is bigger than the entire global semiconductor capex of 2018.

$148 billion of it is waiting on permits

Of the $240 billion physical-infrastructure spend, about $148 billion is committed to projects that cannot turn on until a regulator says yes. This includes land that needs zoning, transmission upgrades that need utility approval, and power generation that needs interconnection studies.

Microsoft has reported $80 billion in cloud orders it cannot fulfill because there isn't enough power available at the right places. The chips are ready; the grid is not.

Two specific regulatory windows — ERCOT (Texas) Batch Zero in spring 2026, and PJM (the mid-Atlantic grid operator) flex-interconnection rulings — will decide which of these stranded projects come online in 2026 versus 2028.

Why this matters

Computer chips can ship in three months. New electrical substations take three to five years. The gap between those two timelines is where billions of dollars sit idle.

Software that helps developers navigate the regulatory side of getting a site online — the permits, the interconnection paperwork, the air-quality rules — is leveraged against this $148 billion of stranded capital. Even a small efficiency gain unlocks an outsized dollar return.

Glossary

Capex: Capital expenditure — money spent on long-lived physical assets (buildings, equipment, land), as opposed to operating costs like salaries.
Hyperscaler: The largest cloud computing companies — typically Amazon, Microsoft, Google, Meta, and Oracle — that operate data centers at globally significant scale.
Interconnection: The regulatory and engineering process by which a new electrical load (like a data center) is approved to connect to the transmission grid.
Substation: An electrical facility that steps high-voltage transmission power down to the distribution voltages a data center can use. New ones take 3–5 years to permit and build.
Stranded capex: Capital that has been committed and partially deployed but cannot generate revenue because of an external blocker — typically the inability to energize the site.

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 are doing almost all of it

These five numbers have nearly doubled in a single year. The previous record — set in 2025 — was $410 billion across the same companies. 2026 will be roughly $753 billion.

Twelve places the money actually goes

NVIDIA's AI systems alone account for ~$260 billion. A single AI server rack from NVIDIA costs $3–4 million; hyperscalers are buying these by the tens of thousands.

$148 billion of it is waiting on permits

Microsoft has reported $80 billion in cloud orders it cannot fulfill because there isn't enough power available at the right places. The chips are ready; the grid is not.

Why this matters

Computer chips can ship in three months. New electrical substations take three to five years. The gap between those two timelines is where billions of dollars sit idle.

Glossary

Capex

Capital expenditure — money spent on long-lived physical assets (buildings, equipment, land), as opposed to operating costs like salaries.

Hyperscaler

The largest cloud computing companies — typically Amazon, Microsoft, Google, Meta, and Oracle — that operate data centers at globally significant scale.

Interconnection

The regulatory and engineering process by which a new electrical load (like a data center) is approved to connect to the transmission grid.

Substation

An electrical facility that steps high-voltage transmission power down to the distribution voltages a data center can use. New ones take 3–5 years to permit and build.

Stranded capex

Capital that has been committed and partially deployed but cannot generate revenue because of an external blocker — typically the inability to energize the site.