Is Nvidia Really Paying for Its Own Growth? The Numbers Tell a Different Story

A couple weeks ago, in a Grow or Die piece titled “Where Nvidia (NVDA) goes from here,” I said that while the circular financing you’ve heard about is a real phenomenon and deserves scrutiny, it doesn’t concern me much.

But I only gave you four short paragraphs on why.

That’s nowhere near enough for an issue being taken so seriously by so many investors.

So today I’m revisiting the topic today and giving it the consideration it deserves.

First, let’s look at why many investors are concerned…

The basic argument goes like this: Nvidia hands an AI company money. The AI company spends that money on Nvidia chips. Nvidia books the sales as revenue. Then Nvidia takes some of what it earns from those transactions and puts into the next AI company. Rinse and repeat.

If that’s really what’s happening, and a meaningful chunk of one of history’s greatest growth stories does amount to Nvidia paying itself, the company and the stock could be in serious trouble.

People point to the dot-com era, when telecom equipment makers lent money to startups so the startups could buy their gear. When the startups died, the telecoms/lenders got crushed.

The argument is serious.

So let’s start with some numbers to begin to ground things…

Everyone throws around the term “circular financing” without mentioning how big it is. Let’s fix that.

I went through Nvidia’s guarantee and commitment schedules and added up everything that could fairly be called “Nvidia putting its balance sheet behind a customer.” I found three buckets of such deals:

• About $108.5 billion of guarantees on land, power, and building shells for AI cloud companies

• About $36 billion of what are called take-or-pay agreements with cloud partners

• About $20 billion of data center leases Nvidia signed on behalf of somebody else.

Add it up and you get $164.5 billion—we’ll call it $165 billion. That’s a huge number. For context, the market caps of three iconic American companies:·

• AT&T (T) = $172 billion

• McDonald’s (MCD) = $179 billion

• Walt Disney (DIS) = $180 billion

That $165 billion is a useful total for “Nvidia putting its balance sheet behind customers.” It’s not one kind of risk.

The $108.5 billion of land, power, and shell guarantees only cost Nvidia money if a tenant defaults. The $20 billion of leases it signed and plans to hand off only stick if the handoff doesn’t happen. The $36 billion of take-or-pay agreements is different. Those are minimum-revenue floors. They can cost money even if the cloud partner is healthy, if the machines sit underused.

But a huge number by itself tells you nothing. What matters: huge compared to what?

Over the last 12 months, Nvidia generated about $127 billion in free cash flow—the actual cash left over after running the business and buying equipment. I computed that from the company’s filings: $134 billion of operating cash flow, less about $7 billion of capex.

So Nvidia’s entire web of customer commitments—every guarantee and every backstop—comes to about 1.3X trailing twelve-month (TTM) free cash flow. And only about half of the annual free cash flow I expect the company to be generating two years from now.

That ratio is a big part of the story. Here’s why…

Lucent Technologies was the poster child of the dot-com era’s vendor financing fiasco and had the largest book of these types of deals among telecom equipment makers. In 2000, at the peak, Lucent had about $8.1 billion of customer financing commitments outstanding.

That same year, Lucent’s free cash flow was negative about $2.4 billion according to its SEC filings. Its operating cash flow was a mere $304 million. And its cash and equivalents totaled about $1.7 billion at the end of fiscal 1999 and $1.5 billion at the end of 2000.

In other words, Lucent was burning a couple billion dollars annually on operations and capex so there was nothing to cover the loans it made. Every dollar it lent customers had to be financed by borrowing or selling stock. And the company’s customers were taking 114 days to pay their bills, up from 95 the year before.

Nvidia today: $165 billion of commitments against $127 billion of TTM free cash flow, $99.4 billion in cash equivalents and marketable securities on the balance sheet, and customers paying in about 60 days.

So Nvidia is not a bigger Lucent.

Lucent’s balance sheet and cash flows could not justify financing customers. Nvidia’s can.

What’s more, what Nvidia is financing is much more valuable. If a customer folds, what’s left behind are the Nvidia GPUs and rack-scale AI systems that comprise a major part of every cloud and help power every AI model in existence. And there’s a line of buyers out the door. Compare that to dot-com era fiber optic cable, which sat dark and unused for a decade.

Why these deals exist at all

Here’s something that matters, which most critics miss…

Many investors treat all “circular financing” as inherently bad. That’s wrong. And lazy. These are deals struck among super sophisticated parties in the real world, and every one of them thinks they’re getting more than they’re giving up.

As an investor or analyst, you can’t form a valid opinion on these things unless you dig into the specific companies involved and the specific terms of the deals.

To me, a big part of what’s really going on is that these deals are a way to vertically integrate without buying every company in the supply chain. So companies can have more control over the things they need to grow their businesses.

If you depend on something, you want some control over it. Nvidia’s growth depends on things like more data centers, more power, and a healthy growing customer base. So it’s doing what it can to make those things happen.

In general, these deals are also a way to share risk and provide longer-term demand certainty so big, expensive, slow things actually get built—and innovation happens faster than it otherwise would have.

Think about what’s happening with Nvidia’s neocloud deals. These customers have exploding demand to build new data centers today but basically no credit history. Lenders want 20 year contracts; neoclouds want more capacity now. Nvidia basically stands in the middle as a credit proxy so the buildout happens on the technology’s timeline instead of the paperwork’s.

Nvidia CFO Colette Kress was explicit on the recent earnings call: “We’re not making loans.”

What Nvidia is providing: a floor—”a minimum revenue guarantee that gives lenders the confidence to underwrite the project.”

In exchange Nvidia shares a portion of the neoclouds’ revenue above that floor. As Kress put it: “we get paid twice, once on the hardware sale and again through the share of rental revenue, a highly reoccurring stream layered on top of a onetime equipment purchase.”

That’s not a subsidy. It’s getting paid for providing credit, like a bank does.

The biggest number nobody is discussing

I haven’t seen much coverage of the deal that makes up two-thirds of Nvidia’s $165 billion “circular financing” total so let’s look at that…

In August, Nvidia guaranteed up to $105 billion for a data center campus in Pike County, Ohio, built by SB Energy, with OpenAI as the tenant. Here’s the 8-K filed with the SEC if you’re interested.

The terms are better than the headline:

• OpenAI has agreed to reimburse Nvidia for anything Nvidia does pay

• The guarantee ends automatically if OpenAI “achieves a satisfactory credit rating”

• Nvidia’s obligation doesn’t begin until the buildings are ready, which isn’t expected until 2028

The most interesting detail: this is a guarantee on land, power, and buildings, not on chips.

And it’s not a promise to pick up OpenAI’s rent check. It’s a residual-value guarantee. Nvidia only pays if OpenAI goes insolvent and defaults or stops paying rent. Even then, SB Energy has to try to re-lease the space or sell it first. Nvidia owes the gap between a pre-set minimum value and whatever that re-lease or sale brings in, capped at $105 billion across the initial 4.25 gigawatts.

What that actually means: For Nvidia to owe anything close to $105 billion, a fully permitted, fully powered 4.25 gigawatt campus in the middle of the largest infrastructure buildout in history would have to be worth close to nothing. Not gonna happen.

Nvidia also chooses the next step. It can take over the lease, force a re-letting, start a sale, let the lease end, or wait up to a year while covering specified project costs. It doesn’t automatically inherit the campus.

That still leaves a strange kind of worst case. The thing standing behind the guarantee is not a warehouse of unsold chips, and it’s not the dark fiber a la 2001. It’s scarce, permitted, powered infrastructure in Ohio. Power is the hardest part of AI right now. You can still find land, even with NIMBY getting louder. You can pour concrete. You can’t conjure a gigawatt. That takes turbines, substations, transmission lines, and years of permits.

If OpenAI failed, Nvidia wouldn’t be given the campus. It would be covering a shortfall on an asset the rest of the industry is already fighting over. That’s a bad outcome. But a different sort of bad than Lucent collecting unused cable.

The claim I need to correct

You may have read something along the lines of “Nvidia admits 25% of its revenue depends on circular financing.” I’ve seen it repeated a dozen times.

What Kress actually said: “We expect demand from the AI labs for which we expect to leverage our balance sheet to contribute toward roughly a quarter of our business next year.”

That’s narrower than critics claim, but broader than defenders admit. Leveraging the balance sheet covers guarantees, backstops, and equity stakes.

Guarantees on land, power, and buildings cost nothing unless a tenant defaults. The take-or-pay floors can cost money if utilization comes in light.

But the point is that Nvidia is not writing checks for a quarter of its own sales.

Red-ish flags that do warrant your attention

In the first six months of this fiscal year, Nvidia recorded $23.7 billion of gains on its equity investments. Those gains don’t sit off to the side. They run through the income statement as “other income” and go straight into net income.

That’s about 17% of Nvidia’s first-half pre-tax profit. And it caused net income to slightly exceed operating income in the first half of the year.

About $7.5 billion of those gains are on companies that aren’t publicly traded according to Nvidia’s 10-Q. They get marked up when a later funding round prices the company higher—sometimes a round Nvidia itself joins. So Nvidia invests, the company raises at a higher valuation, Nvidia marks up its stake, and the gain lands in earnings.

That’s an actual circle. We’re only talking about 5% of pre-tax profit, not 50%, and the operating business grew 106% year over year on its own. But the marks have run almost entirely one direction: $7.5 billion of markups against $150 million of impairments and write-downs. Nothing wrong with that in a rising market of course. It just hasn’t been tested by a falling market yet, which is bound to happen at some point.

Something else that gets no coverage: Nvidia committed $279 billion to its suppliers through 2032, with the bulk committed over the next two and a half years. The company reported this number in the most recent quarter and said it was up from $119 billion the prior quarter. That’s an enormous jump in one quarter, and an enormous total. It’s mostly to secure memory supply during this shortage.

Importantly, these are owed no matter what. So is the $36 billion of take-or-pay I flagged earlier. Put them together and Nvidia is carrying $315 billion of firm obligations against $128.5 billion of contingent ones.

If you’re desperate to find something to worry about on Nvidia’s financials, it’s the firm column, not the contingent one everybody is arguing about.

Wrapping up

Hopefully I’ve made it clearer why I’m not concerned about Nvidia’s “circular financing” deals at this point.

The two real issues I found: A small but meaningful slice of reported profit now comes from marking up private companies rather than selling products. And the firm obligations, $279 billion to suppliers plus $36 billion of take-or-pay, dwarf the contingent guarantees everybody is actually arguing about. Both are things to watch, not panic over.

A few other things to watch:

• Days sales outstanding. It’s 60 now, and worth noting it already moved up this quarter. Kress attributed that to extended payment terms for large purchases by investment-grade customers, and the 10-Q says Nvidia will stretch from 90 days to a full year on big data center builds. Not alarming yet. But this is the drift I’d watch, and if it keeps going it means Nvidia is moving closer to lending chips than selling them.

• Write-downs in the investment portfolio. Cumulative losses and impairments on the private book stand at $250 million. The first billion-dollar one tells you the marks were optimistic.

• If and when OpenAI earns “a satisfactory credit rating,” because that $105 billion guarantee goes away the day it does.

• Whether that guarantee ever moves from a footnote to a booked liability. It was signed after the quarter closed, so Q3 is the first real look at how Nvidia’s own accountants treat it.

We’ve been long Nvidia in my paid Disruption Investor advisory since September 2020. We didn’t sell during the fashionable bearishness that characterized most of 2026, and this sort of work is a big reason why.

There will be pullbacks in the stock, sometimes sharp ones. But I still think Nvidia will become the first $10 trillion market cap company. You can see my rough easy math here.