What the number actually is

Capex is what a company spends each quarter.

These are different: purchase commitments and leases that have not started, spread across several years. They stay off the balance sheet until products arrive or leases begin.

The Wall Street Journal went through the footnotes of nine tech companies' filings and found ~$3 trillion.

For the three clouds we track:

• Alphabet: $811B in purchase commitments, $91B in leases not yet started

• Microsoft: $329B in leases not yet started, $228B in purchase commitments

• Amazon: $137B and $130B, split almost evenly

​That is $1.727 trillion — roughly 3X the ~$600B of capex big 3 expecting to to invest this year.

The pace of the AI buildout is already written into contracts.

What it says about their mental model

Nobody signs largely non-cancellable contracts for capacity they think they might not need.

Alphabet's commitments went from $332B to $811B in one quarter — roughly $479 billion added in 90 days. Alphabet describes them as technical infrastructure, inventory and data centre energy. Some of the energy agreements run to 2054.

​All three are reserving data centers, chips and power on a belief that AI usage will grow enough to absorb it. Agentic AI looks central to that model: applications that act continuously could consume far more compute than today's prompt-based tools.

Binding contracts make that direction hard to reverse.​

Three clouds, three different bets

The split between leases and purchase commitments points to three strategies.

• Alphabet is mostly building and buying the inputs: chips/TPU, infrastructure and power

• Microsoft is mostly renting, with more in leases that have not started than in purchase commits

• Amazon is the most balanced, and the one that keeps most of it on its own books

​Capacity was the constraint. Utilization is next.

Infrastructure needs production workloads. AI models need enterprise data, governance and integration. The wave of forward-deployed engineers fits the same logic: hyperscalers are spending downstream to speed up deployment.​

This changes the value of software and channel partners:

• Partners that shorten the path from pilot to production matter more

• Measurable consumption becomes the metric that counts

• Marketplace and co-sell connect customers' cloud commitments to workloads that consume the new infrastructure

• Implementation capacity will matter as much as compute

​Hyperscalers have contracted the supply. They will rely on their ecosystems to activate the demand.

Does your 2027 partner plan have a consumption number in it?

Source: WSJ