The AI build-out is easiest to understand as a stack. At the top are the companies paying for it. Below them are the chips, then the memory and networking that tie the chips together, then the GPU clouds that rent them out, and at the bottom, the buildings and power that hold the whole thing up. Each layer waits on the one beneath it.

What’s changed in the last two years is where the waiting happens.

Who’s paying

Almost all of it is paid for by a handful of companies. At the end of June, Microsoft, Oracle, Google and Amazon had about $2.3 trillion of contracted business between them that they haven’t yet delivered. Amazon expects to spend about $220 billion on capital this year, and says it still won’t have enough capacity in 2026 or 2027. Microsoft has reportedly turned customers away for lack of space.

Every company further down the stack sells to that small group. When they spend, everyone grows. When they pause, everyone feels it.

The bottleneck keeps moving down

• In 2023 and 2024 it was chips. Nvidia couldn’t make enough.

• In 2024 and 2025 it was packaging, the step that assembles finished chips. That’s why Nvidia signed a $1.5 billion deal this summer to expand Amkor’s packaging plant in Arizona.

• This year it’s memory. On SpaceX’s earnings call, Musk called it “the binding constraint,” with memory supply growing about 20% a year against demand growing much faster. Amazon raised its spending forecast partly because memory costs jumped.

• Underneath all of it is power and buildings. A chip in a warehouse earns nothing. It needs a powered, cooled building with a grid connection, and those take years.

You can see it in the contracts:

• Meta signed 20-year agreements with Vistra for 2.6 gigawatts of nuclear output, some from capacity that won’t exist until the 2030s.

• Amazon signed for up to 1.2 gigawatts more.

• IREN and several other bitcoin miners became AI companies almost overnight, for one reason: they already had power.

The big buyers are now buying the bottom of the stack

The clearest sign that power and buildings are the scarce layer is that the largest buyers have started taking it directly:

• Microsoft agreed in July to buy Aligned Data Centers for about $40 billion.

• Amazon bought the GW Ranch powered-land site in August.

• Google is renting third-party capacity as a bridge while it builds its own.

• Oracle has lined up more than 10 gigawatts through partners, who fund over 90% of it.

When the richest companies in the world start buying powered land rather than waiting for someone to lease it to them, the market is telling you where the shortage is.

Who is in the middle

Between the buyers and the buildings sit the GPU clouds: CoreWeave, Nebius, Lambda, Fluidstack and about 200 smaller ones. They buy chips, rent buildings, and sell computing time, mostly to the same few AI labs and hyperscalers. They are growing faster than anyone, and they carry the most risk. They spend billions now against contracts of four to five years, often with one or two customers.

Nvidia has been funding parts of its own chain:

• $2 billion into Nebius;

• $1 billion into Nokia;

• guarantees and backstops behind several GPU clouds’ leases.

That speeds the build-out. It also means some of the demand in the system is money the supplier put there.

What it means if you own the bottom layer

If the bottleneck is now the powered building, the developer’s position has never been stronger, and never more exposed to the layer above.

Stronger, because the asset is scarce. Signed AI leases are running about $1.65 to $1.9 million per megawatt of IT capacity a year, on 15- to 20-year terms. The buyers have shown they will pay up for sites that can energize soon.

More exposed, because the tenant may be the weak link. If the most likely tenant for your site is a GPU cloud that is itself waiting on chips, memory and financing, your lease is only as good as its next funding round. That’s the subject of the next two pieces: who will sign for your megawatts and who stands behind them, and who funds a powered site before there’s a tenant at all.

The short version: the question has moved from “can I get the chips” to “can I get the building.” If you have the building, the next question is who you let in, and who’s paying their rent.

— Jeff Laughlin, RCG International

Part 1 of 3. Next: Who Buys Your Megawatts, and Who Stands Behind the Lease.

Sources: PitchBook, Hyperscalers and Neoclouds: Executive Commentary From Nine AI Infrastructure Players (Oct 8, 2026); company releases and filings (Nvidia, Amkor, Vistra, Nebius, Nokia, IREN); DatacenterDynamics; Bloomberg as cited by PitchBook. Lease rents are RCG’s arithmetic from disclosed contract values. Figures as of October 2026.