The Tangled Web Nvidia Coreweaves

People like to say what is happening in tech today is nothing like the 2000 bubble. Companies are far more profitable and are less speculative, they say.

While this is true for the most part, today’s note will point out some striking similarities in behavior today vs. 25 years ago.

When Cisco was at its prime, it was widely known (though usually casually dismissed) that they sold a lot of equipment to doomed startups…who they also financed.

So a harebrained web site would come to Cisco, say they wanted to buy some equipment, but didn’t have the money. Cisco would say “great, we’ll lend you the money for the equipment and take a stake in your company”.

If you were Cisco, this was great! You made more sales, you got a venture portfolio that let you ride the bubble up, and, best of all, you got to book income from the loans.

That’s right, Cisco would book revenue for the product sale and then financial income from the money they gave you to buy it!

While they had to disclose the source of this income, investors valued it at the same multiple as regular earnings.

What could go wrong??? Cisco was the most valuable company in the world!

Of course, what went wrong is those startups went bust, couldn’t pay back their loans, and wouldn’t buy any future products from Cisco.

Cisco stock is still below where it was in 2000.

Nvidia

That brings us to the present moment and the Coreweave IPO that took place Friday.

Coreweave is (or maybe was?) one of the hottest AI startups. It’s business model is essentially buying AI chips from Nvidia and then leasing them out to others who can’t get enough chips directly from Nvidia.

If you’re scratching your head over this approach, you’re not alone. Why would Nvidia sell chips to Coreweave instead of directly to their other customers?

Because these aren’t arm’s length transactions.

Reminiscent of Cisco, Nvidia is double dipping in Coreweave’s business. They don’t just sell chips to Coreweave. They also are a major customer of Coreweave’s data centers!

Nvidia is the second largest customer of Coreweave at ~15% of revenue.

So just to make sure this is perfectly clear, here’s how it works:

• Nvidia sells lots of chips to Coreweave and books revenue
• Coreweave leases some chips back to Nvidia allowing Coreweave to also book revenue
• Coreweave uses that “revenue” to raise more capital to allow it to buy more chips…allowing Nvidia to book more revenue

If Nvidia needs access to its chips for its own AI research, why does it need to sell them to Coreweave first???

Maybe there is a good reason I’m not privy to, but on the surface, it appears the reason is to allow both firms to overstate their revenue.

But that’s not all…Nvidia holds nearly $1B in Coreweave equity, so they have an additional vested interest in seeing Coreweave succeed.

Note, at the original IPO terms, Nvidia had 18M shares worth ~$1B. However, once it appeared the IPO was going to fail, Nvidia agreed to buy an additional $250M at $40/share.

Notice how that gets them right back to $1B…but only after writing a $250M check.

Why would they do this? Because if they hadn’t, the IPO would have failed and Coreweave may well have gone to $0.

Nvidia essentially spent $250M on a bailout of Coreweave to ensure Nvidia can continue to sell them chips.

How can I be sure it would have gone to $0 if Nvidia didn’t step in? For that, let’s explore Coreweave itself in more depth.

Coreweave

One thing you may be wondering is why would Nvidia agree to sell all these AI chips to Coreweave in the first place rather than to their more established customers like Google or Amazon…or Microsoft, Coreweave’s largest customer.

There is an interesting answer to that. Before AI, Coreweave was a crypto miner. If you know anything about crypto mining, you understand it requires a lot of Nvidia chips.

So Coreweave was a big Nvidia customer with a large inventory of chips before AI took off. After AI, they realized they could repurpose their inventory as a reseller to AI developers given there was a shortage of Nvidia AI GPUs.

Thus, they pivoted to the business model I mentioned earlier of buying chips from Nvidia to lease to others who couldn’t get as many Nvidia chips as they needed.

Finally, a use case for crypto!

However, there are a number of flaws in this model.

Unsustainable Demand

First, you are marginal capacity. As long as there are shortages, others want access to your chips. As soon as the shortages end, all your demand goes away.

Think of it like going from being a ticket broker while Taylor Swift was on tour to selling tickets today when I’m sure someone is on tour, but there are plenty of seats available.

Ticket brokers have what’s referred to as perishable inventory, meaning you can’t put it in a warehouse until prices improve.

Nvidia chips are similar to perishable inventory. Their shelf life isn’t as short as concert tickers or airline seats, but Nvidia is constantly rolling out better chips making the prior ones less in demand.

Yet, Coreweave assumes the useful life of its chips is around six years!

In a way, Coreweave is an inverse bet on Nvidia innovation. The more Nvidia improves future chips, the quicker Coreweave’s inventory loses value.

You might think this isn’t the end of the world. If the inventory loses value, Coreweave will buy the new chips. Sure, they’ll have some inventory writedowns, but they’ll have plenty of future profits on the new chips.

Unsustainable Leverage

There’s one problem with that thought process – debt.

Coreweave has close to $10B of debt (their revenue last year was only $2B). Most of that debt is collateralized by its inventory.

That means if Nvidia innovates and Coreweave’s inventory value collapses, Coreweave can’t pay back it’s debt.

The other big risk is customer concentration. Microsoft is over half of Coreweave’s revenue.

If Microsoft is able to get more chips directly from Nvidia, they will buy less from Coreweave who may be stuck with inventory they can’t place elsewhere. This again means they can’t service their debt.

The Nvidia Confidence Game

So the bankruptcy risk here is significant. Given that, you can see why Nvidia felt compelled to pony up $250M to keep the Ponzi scheme going.

After all, if Coreweave goes under, Nvidia loses a lot of revenue, which hurts Nvidia stock, which hurts all tech stocks (their other customers), hurts the entire market, and potentially leads to a March 2000 like end of the AI bubble.

$250M was a small price to pay to take that risk off the table.

But if someone tries to tell you this whole thing is no big deal, you should be very skeptical of their motivation.

Coreweave is a huge red flag that there is an AI bubble and it’s closer to unraveling than we all realize.