The "Ponzi Game" of GPUs?

Can borrowing money to buy NVIDIA chips and using these chips as collateral be profitable?

Crusoe Energy, formerly a cryptocurrency miner, did just that.

Crusoe has also joined the ranks of cloud service providers, leasing servers loaded with H100 GPUs to startups. It is reported that Crusoe borrowed $200 million in order to purchase a large quantity of scarce NVIDIA H100 chips.

What's special is that Crusoe used these chips as "collateral assets" for the loan. It will take Crusoe three and a half years to repay the loan plus interest. However, Crusoe believes that these chips will generate returns higher than their $32,000 price tag and continue to generate profits for the company over a period of seven years.

This means that the short-term returns on the H100 chips can completely cover this huge loan, allowing Crusoe not only to repay the money but also to continue making money.

Upper90, a hybrid credit fund company, provides loan services to Crusoe and specializes in providing financing solutions for startup tech companies.

This financing model reflects a clear trend: whether it's startups or internet giants like Amazon, they all want to enter the AI cloud services race.

Chase Lochmiller, CEO of Crusoe, said that Crusoe plans to provide approximately 20,000 H100 servers to customers in the coming months. Previously, Amazon had also stated that its H100 clusters could be expanded to 20,000 GPUs for a single customer.

According to Bill Libby, CEO of Upper90, NVIDIA GPUs are "a new category of debt assets." Before signing the debt agreement, he must consider how to value the chips and how to adjust the valuation if the chip version is updated.

This means that the fate of these cloud service providers is tied to NVIDIA. As long as NVIDIA's chips continue to dominate the market, cloud service providers will be profitable.

How long can this "Ponzi scheme" last?

In fact, this "robbing Peter to pay Paul" operation of borrowing money to buy NVIDIA chips and using the profits to repay the loan is not the first time.

CoreWeave, one of the largest AI cloud service providers, raised a staggering $2.3 billion in loans earlier this year, some of which were used to purchase H100 chips.

The company claimed at the time that the value of each H100 chip would be retained for 6 years, with a depreciation of no more than 50% within 2 years. They also stated that the company could not only recover the initial cost of the chips during this period but also generate a gross profit of $100,000 per chip.

According to cloud computing executives, larger H100 orders can be delivered faster, allowing companies to profit more quickly. Given NVIDIA's current market position, these small cloud providers are highly motivated to achieve ideal returns. However, this is obviously based on the premise of the continuous growth in the field of artificial intelligence and the strong position of the H100 chip.

Currently, there have been rumors about the next generation GPU B100 from NVIDIA; in August of this year, Huang Renxun announced that the latest version of the GH200 super chip will be launched in the second quarter of 2024.

As a mortgage asset, there is still a significant risk of depreciation for GPUs. According to data from eBay, the early Nvidia GPU V100 was priced at around $14,000 when it was first launched five years ago, and now it is selling for less than $1,000.

Charles Fitzgerald, a Seattle angel investor and former Microsoft executive, pointed out:

"We know that these GPUs will depreciate. If you use a depreciating item as collateral for a loan, it is best to convert it into a large amount of cash. The margin for error in these businesses is very, very small."