Nvidia Makes $105 Billion Guarantee Gamble on OpenAI Data Center

Photo of author

By Global Team

U.S. Nvidia has agreed to provide up to $105 billion in support for a dedicated super-sized data center for OpenAI. The competition in AI semiconductors is moving beyond simply selling chips and into the stage of backing customers’ debts.

Source: Nvidia
Source: Nvidia

The New York Times reported on the 17th local time that Nvidia had signed a financing agreement for a huge data center being built in Pike County, Ohio, in the United States. It is said to be one of the largest facilities in the world, and once completed, it will be leased entirely by OpenAI, the company behind ChatGPT.

A data center is essentially a factory that runs AI services. Services like ChatGPT require tens of thousands of high-performance chips to perform calculations each time a question is processed, consuming electricity in the process. As the number of users grows, demand for chips and power increases. Nvidia is effectively the dominant supplier of high-performance AI semiconductors.

The support is closer to a payment guarantee than a cash investment. Nvidia CEO Jensen Huang said in a company blog that the $105 billion would be used to guarantee the rent and electricity bills that OpenAI must pay. If OpenAI cannot pay, Nvidia will cover the cost instead.

It is similar to having a guarantor for a bank loan. The company building the data center reduces the risk of losing 20 years’ worth of rent, while OpenAI secures access to the facility even if its credit is insufficient. Huang explained the background by saying, “AI companies are growing faster than their financial strength and credit can support.”

How large is the data center?

The Ohio data center is estimated to cost up to $500 billion in total, or about 700 trillion won. One single complex will consume an amount of money comparable to South Korea’s annual government budget.

Its power consumption will be 8 gigawatts. That is equivalent to the output of eight large nuclear power plants running nonstop. By U.S. standards, it is roughly the amount used by about 6 million households.

OpenAI’s lease term is 20 years. Rent will begin after operations start in 2028, and only for the completed portions. Until then, it may not have to pay a single dollar. Nvidia’s guarantee initially applies to 4.25 gigawatts of capacity and may be expanded to the remaining sections.

The facility will be built, owned, and operated by SB Energy, a subsidiary of Japan’s SoftBank. Nvidia also plans to invest $1.5 billion, or about 2 trillion won, in SB Energy.

The construction of super-sized complexes is not limited to Ohio. OpenAI and its competitors are pursuing similar projects in several U.S. states, including Texas. The pace of investment is so fast that issues such as grid capacity, land, and water supply have become local concerns.

The circular-trade controversy over money going round and round

As news of the deal spread, debate over circular transactions resurfaced in the United States. This is a structure in which a major tech company invests in an AI startup, and the money received is then returned to the investing company through purchases of chips and cloud services.

The relationship between Nvidia and OpenAI is cited as a representative example. Nvidia had previously invested $30 billion, or about 42 trillion won, in OpenAI. Only Nvidia chips will be used in the Ohio data center. Critics say it is hard to tell whether chip sales reflect genuine demand or simply the return of money Nvidia itself provided.

It resembles a shop owner lending money to a customer, who then uses that borrowed money to buy products from the same shop. Sales accumulate on the books, but no new outside money is coming in. Concerns follow that if similar transactions pile up, the boom may appear larger than it really is.

OpenAI is still operating at a loss. Although the number of paid ChatGPT subscribers and enterprise customers is increasing, many say its earnings fall far short of the scale of data center investments it has promised around the world. On credit alone, it would be difficult for the company to sign a 20-year lease.

Huang said, “OpenAI will pay the rent,” rejecting the claim that this is a circular transaction. The two companies were previously reported to have discussed a support plan worth $250 billion.

More and more cases are emerging in which Nvidia acts as a funding source for its customers. Last week, it announced a plan to create $500 billion in financing with six partners, including major asset managers and banks, so that customers can pay for chips and data center costs.

The practice of sellers financing buyers’ purchases also existed during the telecom bubble of the late 1990s. At the time, telecom equipment companies lent money to customers to sell their equipment, only to be left with bad debt when the bubble burst. Some argue that today is different because there is real demand for data centers.

What impact will this have on South Korea’s semiconductor industry?

Investment in U.S. data centers is directly linked to South Korea’s semiconductor industry. Nvidia’s AI chips are paired with high bandwidth memory (HBM), which is mainly produced by Korean companies. Several HBM chips are attached to each chip. As super-sized complexes increase, memory orders are expected to rise as well.

SK hynix and Samsung Electronics are the main suppliers. In South Korea, expectations for the semiconductor market have tended to rise whenever news breaks of new data center construction in the United States.

There are also reasons for caution. If part of the data center boom is inflated through circular transactions, memory demand could also weaken when the bubble bursts. Some say Korean companies need to reduce dependence on specific customers and diversify demand sources.

Power supply is another key point. If complexes consuming the equivalent of eight nuclear plants keep appearing, investment in power generation and transmission networks will have to follow. Domestic companies making transformers, cables, and cooling equipment have already been increasing orders from the U.S. There are also forecasts that the market for power infrastructure such as small modular reactors and energy storage systems could expand alongside it.

The Ohio data center is scheduled to begin operations in 2028. Until then, whether OpenAI’s revenue grows enough to cover the rent will determine the success of the guarantee agreement. Whether the 146 trillion won guarantee becomes a safety net or a burden will depend on AI demand.