You've been thinking about AI as software. It's not. It's real estate. And right now, every AI tool you use is built on a tower of landlords, each one renting from the floor below.
Understanding this doesn't require a computer science degree.
It requires the same instinct that tells a good business owner to ask: who actually profits when I spend money here?
I'm an AI consultant. I help businesses and nonprofits adopt AI with intention. And one of the most important things I've learned is that most leaders are making adoption decisions based on the product they can see… The chatbot, the tool, the platform, without any understanding of the infrastructure underneath it.
That infrastructure is stranger, more fragile, and more consequential than almost anyone is telling you.
Every Floor Rents from the One Below It
Let me walk you through the stack.
At the very top, you have the AI tools your team actually uses. ChatGPT. Claude. Gemini. The AI features built into your CRM, your scheduling software, your HR platform. This is the floor most people live on. It's the one with the nice interface.
One floor down: the companies building those tools. OpenAI. Anthropic. Google DeepMind. Meta AI. These are the names in the headlines, the companies raising billions, the ones announcing the next breakthrough every six weeks. They write the code that makes the AI work.
But they don't own the computers it runs on.
For that, they rent.
OpenAI runs much of its infrastructure through Microsoft's cloud and a company called CoreWeave — a GPU cloud provider that grew nearly 6,000 percent in three years by renting out access to the specialized chips AI needs to function. CoreWeave essentially acts as a landlord: they own the computing hardware, and AI companies pay to use it. As of late 2024, Microsoft alone accounted for 62 percent of CoreWeave's revenue. One tenant, paying rent to one landlord, powering tools used by hundreds of millions of people.
Go one floor further down, and you hit the chips themselves.
The specialized processors that make AI possible — called GPUs — are dominated by a single company: Nvidia. As of 2025, Nvidia controlled more than 80 percent of the market for the chips used in training and deploying AI models. Every major AI lab, every cloud provider, every data center operator is fundamentally dependent on Nvidia's hardware. When Nvidia's CFO described the state of GPU availability in a recent earnings call, she didn't say supply was tight. She said: "The clouds are sold out."
Go one floor further, and you reach the foundry.
Nvidia designs its chips, but it doesn't manufacture them. No American company manufactures the world's most advanced AI chips. That work is done almost exclusively by a single company: Taiwan Semiconductor Manufacturing Company, known as TSMC. TSMC produces more than 90 percent of the world's most advanced semiconductors. Every Nvidia GPU. Every Google TPU. Every Amazon AI chip. All of it flows through one company, on one island.
One floor below that: the machines that make the chips possible. A Dutch company called ASML makes the only equipment in the world capable of producing the most advanced semiconductors. There is no competitor. There is no backup. ASML is a monopoly within a monopoly, at the base of a stack that the entire AI economy rests on.
This Is Not How It Gets Described
When you read about AI, you read about breakthroughs. Capabilities. Benchmarks. The race between American and Chinese AI. The promise of what's coming next.
You do not read about the fact that every single one of those breakthroughs ran on rented hardware, manufactured by a company with no competitors, on an island 100 miles from mainland China.
That's not an accident. It's not a conspiracy, either. It's just the nature of how infrastructure stories get told — or don't.
The software is visible. The stack underneath it is not.
What you use is the product. What makes it possible is a supply chain.
And supply chains, as the last five years have reminded us, are only invisible until they aren't.
The Tenants Are Trying to Become Landlords
Here's where it gets interesting.
The AI companies at the top of the stack are not comfortable being tenants. They're all, to varying degrees, trying to build their own floors.
Google has been developing its own AI chips — called TPUs — for years, and in 2024 began selling access to them externally for the first time, moving it from tenant to competitor with Nvidia. Amazon has Trainium. Meta is acquiring chip startups. OpenAI is designing custom chips through a partnership with Broadcom.
The reason is straightforward. One analyst put it plainly: "They don't want to be stuck behind an Nvidia monopoly."
And Nvidia is watching. The company has invested heavily in the AI companies that depend on its chips — including a reported $100 billion investment in OpenAI. Which means Nvidia is simultaneously the landlord collecting rent and an investor in the tenant who is trying to build a competing building.
This is not normal market behavior. It is a dense web of mutual dependency, where the lines between customer, competitor, investor, and infrastructure provider are almost impossible to separate.
CoreWeave — that GPU cloud landlord renting to OpenAI and Microsoft — is $8 billion in debt, growing explosively, and dependent on a handful of massive clients who are all actively building the infrastructure that would make CoreWeave unnecessary. The tenants are building their own buildings. The landlord is borrowing money to build faster than them.
Every dollar flowing through this ecosystem is simultaneously a bet and a hedge.
What This Means If You're a Leader Making Adoption Decisions
I want to be clear: none of this means you shouldn't adopt AI. You should. The tools are genuinely useful, and the organizations that learn to use them well will have a real advantage.
But it does mean a few things worth understanding.
The tools you're adopting are not standalone products. They are the visible surface of a complex supply chain. When something changes at any layer of that stack — chip shortages, geopolitical tension, pricing decisions by a company you've never heard of — it ripples upward to the tools your team uses every day.
"Reliability" in AI is a different question than reliability in most software. When your accounting software goes down, it's probably a server issue at one company. When an AI capability shifts or disappears, it could be a policy change, a chip shortage, a cloud provider renegotiating a contract, or a foundry in Taiwan running at capacity. The dependency chain is long, and most of it is invisible to you.
The companies building AI tools are, in many cases, not yet profitable. They are spending enormous amounts of money on rented infrastructure in a race to build capabilities they hope will eventually justify the cost. That's not a reason to avoid their tools — but it is a reason to avoid building your entire operation around a single vendor with no contingency.
Understanding the stack doesn't change what you do tomorrow morning. But it changes how you think about what you're adopting, and that matters.
The Floor You're Standing On
The next time someone tells you that AI is just software, or that the AI revolution is happening in the cloud, or that this technology is endlessly scalable and infinitely available — remember the stack.
There's a chip. Made by one company. Manufactured by another. On an island the size of Maryland.
There's a data center. Powered by electricity your neighbors are increasingly paying for. Cooled by water. Staffed by almost no one.
There's a landlord collecting rent from a tenant building tools that your team is learning to use.
And at the very top of all of it: a prompt you typed into a box.
The floor you're standing on goes deeper than you think.
