Every time I look at the balance sheets of the generative AI boom, I get a distinct sensation of vertigo. We are told we are living through a technological revolution akin to the industrialization of electricity, yet the financial plumbing beneath it looks less like the power grid and more like a beautifully constructed M.C. Escher staircase. Money goes up, money goes around, and somehow, the same dollar seems to be doing the work of five.
I want to understand how a company like Nvidia can add trillions of dollars in market cap while its primary buyers are often startups funded by the very venture capitalists who are backed by the tech giants who are buying the chips. It is a dizzying ecosystem. If you trace the path of a single dollar in this economy, you quickly realize it doesn't just buy a service and stop; it travels in a loop that makes you wonder where the organic demand ends and the financial engineering begins.
The Three-Body Problem of Modern Silicon
To make sense of this, we have to look at three specific players: Nvidia, the undisputed king of silicon; CoreWeave, a former Ethereum mining company turned specialized GPU cloud provider; and Nebius, a newly emerged European cloud contender. These are not just businesses buying and selling parts. They are nodes in a highly specialized, interdependent network that seems to have bypassed traditional market dynamics entirely.
Take CoreWeave. In 2023, the company managed to secure a massive $2.3 billion debt facility. What makes this fascinating is the collateral used to secure that debt: Nvidia's H100 microchips. This is a level of faith in a single piece of hardware that we rarely see in traditional finance. A bank lent billions of dollars not based on cash flow or real estate, but on the assumption that these specific black rectangles of silicon will retain their astronomical value long enough to pay off the note.
Photo by Marta Branco on Pexels
Then we have Nebius, which is planning to spend upwards of $4 billion to build out AI infrastructure. Where does the money come from, and where does it go? It goes right back to Nvidia to buy more chips to lease to AI startups, many of whom are funded by venture funds that are themselves deeply invested in keeping the Nvidia ecosystem dominant. It feels less like a market and more like a closed-loop life support system.
The Alchemy of Synthetic Demand
What happens when the manufacturer of the most sought-after asset on earth also acts as the kingmaker for the companies buying that asset? Nvidia doesn't just sell chips to whoever shows up with a briefcase of cash. They allocate them. In a world of extreme scarcity, getting an allocation of H100s is the equivalent of being handed a license to print money.
This creates a curious incentive structure. If you are a specialized cloud provider, your entire valuation is based on how many GPUs you have secured. If Nvidia favors you with allocation, your valuation skyrockets. You can then use that valuation to raise more capital, which you immediately use to buy... more chips from Nvidia.
- Nvidia invests in a startup or specialized cloud provider.
- That provider uses the cash (and the prestige of the partnership) to secure loans.
- Those loans are used to place massive orders for Nvidia chips.
- Nvidia reports record-breaking revenue, boosting its stock price.
This is what skeptics call "synthetic demand." It isn't necessarily fraudulent, but it is highly recursive. The demand is real in the sense that purchase orders are signed and chips are delivered, but is it reflective of end-user utility? If the end-users are mostly other startups who are burning through venture capital to train models that do not yet have a profitable business model, then the entire tower is built on the assumption that someone, somewhere, will eventually figure out how to make these models pay for themselves.
Searching for the End User
This is the question that keeps me up: who is the actual customer at the very end of this chain? In a healthy economy, a business buys a tool to make a product, which it sells to a consumer for more than it cost to make. Right now, the AI economy looks like a group of people selling each other increasingly expensive shovels while the gold mine itself remains largely theoretical.
We know that training a frontier model now costs upwards of $100 million, and the next generation is projected to cost ten times that. To break even on that kind of capital expenditure, the software built on top of these chips needs to generate hundreds of billions of dollars in recurring revenue. Currently, the entire software-as-a-service (SaaS) market globally is worth about $350 billion. For the current GPU investment to make sense, AI software needs to essentially double the size of the entire global software industry in the next few years.
Maybe it will. I am open to the idea that we are on the cusp of an economic expansion so vast that these numbers will look quaint in a decade. But if we aren't, if the utility of these models plateaus or if businesses realize they can get 90% of the value using much smaller, cheaper, open-source models, then the collateral backing those multi-billion-dollar loans begins to look incredibly fragile. Silicon depreciates fast. A GPU is not like gold or real estate; its value halving every two years is a feature of physics, not a bug of the market.
What This Actually Means
We are participating in a massive, high-stakes experiment in capital deployment. The circular funding loops between Nvidia, CoreWeave, and the venture ecosystem are not necessarily a conspiracy, but they are a symptom of a market that has run far ahead of its own reality. It is an acceleration mechanism designed to build the future at any cost, using financial engineering to bypass the slow, boring process of waiting for organic customer demand to catch up.
If the gamble pays off, we will look back at this era of circular financing as a stroke of genius—a synthetic scaffolding that held up the cathedral of artificial intelligence until the keystones could be put in place. The scaffolding was necessary because the physical reality of building gigawatt-scale data centers is too slow for the speed of venture capital.
But if the software revenue fails to materialize, the unwind will be spectacular. When the scaffolding falls, it won't just hurt the startups. It will ripple through the banks that accepted silicon as collateral, the venture funds that inflated their portfolios with paper gains, and the retail investors who bought into the narrative of infinite hardware growth. We are riding a carousel, and the music is still playing, but it's worth keeping one eye on the exit.
Quick Answers
Is this circular financing scheme illegal?
No, it is generally legal corporate finance and strategic investment. It only becomes a regulatory issue if disclosures are misleading or if the transactions are shown to be purely sham transactions with no economic substance, which is not currently the case.
What happens if GPU prices start to drop?
If the market becomes saturated or if demand softens, the collateral value of the GPUs holding up billions in debt will collapse. This would force specialized cloud providers to either post more cash collateral or face default, potentially triggering a sell-off of hardware.
How does this compare to the Dot-Com bubble?
There are strong parallels to the telecom buildout of the late 1990s, where companies laid millions of miles of fiber-optic cable based on projected demand that took another decade to actually arrive. The infrastructure was eventually useful, but the companies that built it went bankrupt first.
