The $750 Billion Loop
How AI’s chipmakers started financing their own demand
Welcome to Silicon Sands News—the go-to newsletter for investors, senior executives, and founders navigating the intersection of AI, deep tech, and innovation. Join ~35,000 industry leaders across all 50 U.S. states and 117 countries—including top VCs from Sequoia Capital, Andreessen Horowitz (a16z), Accel, NEA, Bessemer Venture Partners, Khosla Ventures, and Kleiner Perkins. Our readership also includes decision-makers from Apple, Amazon, NVIDIA, and OpenAI.
This week, we explore the quiet accounting trick holding up the AI buildout: the companies selling the chips are now writing the checks that buy them.
Let’s Dive Into It..
Key Takeaways
For VCs and LPs:
The demand signal is contaminated: Nvidia is arranging more than $750 billion in deals where it invests in, lends to, or guarantees the debt of the customers who then buy its chips. When you underwrite a position in this sector, model what the revenue looks like with the vendor financing stripped out. The number you get is the real one.
One customer, no credit rating, a quarter-trillion in guarantees: Nvidia is in talks to backstop up to $250 billion of OpenAI data-center debt and separately finance up to $350 billion in chip purchases. OpenAI carries no investment-grade rating. Nvidia’s balance sheet is the collateral. Price the counterparty risk accordingly, because right now one company sits behind both sides of the trade.
The multiple is priced for a straight line: OpenAI runs at roughly $25 billion in run-rate revenue and is on track to lose about $14 billion this year. The financing being arranged around it is roughly 24 times its annual revenue. That gap is the bet. Decide whether you believe the line stays straight through 2029.
AMD just ran the same play, smaller: AMD will invest up to $5 billion in Anthropic in exchange for a commitment to deploy 2 gigawatts of AMD Instinct MI450 GPUs. The investment releases as Anthropic buys the chips. Watch this structure spread when it becomes the standard way to book a marquee customer; GPU revenue and GPU financing stop being separate lines.
For Senior Executives:
Vendor financing is a negotiating tell, not a gift: If your GPU supplier is offering to fund your buildout, you have more leverage than the sticker suggests, and they have a demand problem they’re not putting in the press release. Use it. Ask for the terms OpenAI and Anthropic are reportedly getting, then benchmark your own.
Concentration risk now runs through your supplier’s customers: If Nvidia’s balance sheet is quietly guaranteeing a chunk of the AI economy, then a stumble at one lab reprices your chip vendor, your cloud contract, and your capacity roadmap at once. Map that exposure before your next procurement cycle, not after.
Multi-vendor compute stopped being a science project: Anthropic committing 2 gigawatts to AMD tells you the largest buyers no longer want a single source. Neither should you. The lock-in premium you’re paying for a one-vendor stack is a line item you can now attack.
For Founders:
The tourist capital is in the loop: Money flowing through supplier-financed structures inflates the comps you’ll be measured against. If a competitor raised at a number that assumed a vendor-backed compute deal, your investors will still expect the number. Name the difference in the room before they do.
Compute independence is a fundraising asset: Anthropic is diversifying off Nvidia months before an October IPO. Being able to say your economics survive a single supplier’s price hike is worth real basis points on your valuation. Build the second source early.
Position where the loop can’t reach: The circular deals cluster around frontier-scale training. If your business runs on inference, small models, or specialized workloads, your unit economics are legible in a way the giants aren’t. That legibility is your pitch. Lead with it.
The Loop Nobody Wants to Name
Start with the number that broke the week. Nvidia is working on a fresh set of AI deals worth more than $750 billion, a web of investments, guarantees, and financing commitments spanning OpenAI, SK Group, CoreWeave, and others. Every one of those partners uses the money, directly or indirectly, to buy Nvidia GPUs.
Read that structure slowly. The supplier invests in the customer. The customer buys the supplier’s product. The purchase shows up as the supplier’s revenue. The revenue supports the supplier’s valuation. The valuation funds the next investment.
A finance professor at Boston College gave it the plain name: circular financing, which is when a company lends a customer the money to buy its own goods. Jim Cramer went further and said it echoes the dot-com bubble, when telecom-gear makers financed the upstarts who bought their equipment, right up until the upstarts folded and took the vendors’ revenue down with them.
The market noticed. Nvidia traded down roughly 15% from its peak on these reports, with its next earnings due August 26. When a company worth trillions wobbles because people started asking where the demand actually comes from, that’s the question worth sitting with.
A Quarter-Trillion Behind One Unrated Borrower
The centerpiece is OpenAI. Nvidia is in talks to guarantee up to $250 billion of debt for a 10-gigawatt data-center campus in Pike County, Ohio, and separately to finance as much as $350 billion in OpenAI’s chip purchases. The full project could run past $500 billion.
Here’s why the guarantee exists at all. OpenAI has no investment-grade credit rating. On its own, it can’t borrow a quarter-trillion dollars at rates that make a data center pencil out. So Nvidia’s balance sheet steps in as the collateral. Lenders get to price the debt against the most valuable company on earth instead of against a startup that lost $14 billion this year on roughly $25 billion of revenue.
Sit with the ratio. The financing being assembled around OpenAI is about 24 times its current annual revenue. In 2025, the company booked around $13 billion in revenue against a leaked $20.9 billion operating loss. This is a company that burns cash by design, and the entire structure assumes it keeps growing into the debt. Nvidia’s own forecast has OpenAI reaching $100 billion in revenue by 2029. Maybe it does. But if it doesn’t, the guarantor is holding the paper.
That is the trade the whole sector is now leaning on: one customer’s revenue curve staying vertical for four more years. The chips are almost beside the point.
The Same Machine, One Size Down
Two weeks before the Nvidia news, AMD showed everyone the compact version.
On July 22, AMD said it would invest up to $5 billion in Anthropic. In return, Anthropic committed to deploy 2 gigawatts of AMD’s Instinct MI450 GPUs on the Helios rack platform, with the first gigawatt landing in the first half of 2027. The investment releases in tranches as Anthropic hits its deployment milestones. The money and the chip orders are bolted to the same track.
Run the arithmetic AMD’s own executives handed us. They peg a single gigawatt of compute at double-digit billions of spend. Two gigawatts, then, is somewhere north of $20 billion in AMD silicon. AMD is putting up as much as $5 billion in equity to help secure roughly $20 billion or more of its own chip sales. The investment is the cost of booking the customer.
I want to be fair to AMD here, because there’s a real strategic logic underneath the financial engineering. AMD’s problem has never been the raw silicon. It’s been the software and the reference customers who prove the silicon at scale. Anthropic gives AMD a lighthouse account that says the MI450 can carry frontier training. That’s worth paying for.
And for Anthropic, the logic is sharper still. This is a hedge.
Anthropic Is Buying Its Way Out of Nvidia
Anthropic is reportedly filing to go public as soon as October, targeting a Nasdaq listing behind Goldman Sachs, JPMorgan, and Morgan Stanley, on the back of a $965 billion valuation and roughly $47 billion in run-rate revenue. When you’re about to hand a prospectus to public investors, “we depend entirely on one supplier who also invests in our biggest competitor” is a risk factor you’d rather not write.
The AMD deal rewrites that sentence. Anthropic can now tell the street it runs a two-vendor compute stack, that its cost curve isn’t hostage to a single monopoly, that it has optionality on price. Each of those lines is worth real basis points on the valuation.
This is the through-line I keep coming back to in this newsletter. The winners in AI infrastructure will be the ones who route across suppliers, specialize their workloads, and refuse to be captive to a single monopoly. Anthropic just paid tuition for that lesson in public, and the tuition was a strategic partnership with the one company positioned to break Nvidia’s grip.
Notice the irony, though. Anthropic’s escape from one circular deal runs straight through another. AMD invests in Anthropic so Anthropic buys AMD chips. The mechanic doesn’t change. Only the logo does.
What the Loop Hides
Circular financing isn’t fraud. It’s not even new. Vendors have financed customers since the railroads. The problem is what the structure does to the numbers everyone downstream relies on.
When Nvidia books revenue from a customer it funded, that revenue looks identical to revenue from a customer paying with its own cash. The income statement can’t tell you which is which. Neither can the demand forecast built on top of it. So the whole sector prices off a signal that has the vendor’s own money mixed into it, and nobody can cleanly separate organic demand from manufactured demand.
Now stack the second-order effect. Analysts have flagged that Nvidia’s guarantees mean a single lab’s stumble no longer stays contained. If OpenAI slips, it doesn’t just hurt OpenAI. It reprices Nvidia, which reprices every neocloud leaning on Nvidia, which reprices the debt those neoclouds issued, which reprices the funds holding that debt. The guarantee that made the borrowing possible is the same wire that transmits the failure.
For scale, remember what’s being financed. JLL projects global data-center capacity roughly doubling from about 103 gigawatts today to near 200 by 2030, needing as much as $3 trillion in new spend. A meaningful slice of that buildout is now being underwritten by the companies selling the hardware that fills it. That’s the exposure hiding inside the growth story.
Let’s Wrap This Up
The AI infrastructure boom is real. The compute demand is real. I run a company that pays for both. None of that is in question.
What’s in question is how much of this demand can stand on its own once the vendors stop financing it, and whether the balance sheets holding the guarantees can absorb a miss. Anthropic’s October IPO is the first place we’ll get a clean read, because public markets will price the compute economics that private circular deals have kept comfortably blurry.
So here’s your positioning question for the next two quarters. When the AI trade gets marked to market, are you holding the companies whose revenue survives the financing being stripped out, or are you holding the ones who only look like customers because someone sold them the money first?
Selected the cover for Dr. Dobrin’s upcoming book, AI iQ for World Models: Physical AI & the Post-LLM Economy.
As a thank-you to our community, all paid subscribers as of August 15th will receive a complimentary digital copy.
Disclaimer: This is educational content, not financial advice. Full source links above. The views and opinions expressed above are current as of the date of this document and are subject to change.