It took 30 years for Nvidia, a company whose chips power much of the world's AI (artificial intelligence), to be valued at US$1 trillion (R$5.1 trillion). Reaching US$2 trillion (R$10.2 trillion) took just another nine months. The US$5 trillion (R$25.5 trillion) mark was surpassed less than two years later.
Now, Nvidia is the most valuable company in the world, valued at around US$5.4 trillion (R$27.5 trillion). Next year, its sales are expected to almost double. Some analysts predict that annual revenue will be US$1 trillion (R$5.1 trillion) by 2029.
In addition to chips, CEO Jensen Huang turned to financial engineering to stimulate demand. In August, Nvidia offered a guarantee of up to US$105 billion (R$535.5 billion) for a data center in Ohio.
Previously, he announced a US$500 billion (R$2.55 trillion) package for AI with Wall Street. In July, it promised to supplement revenues from data center customers below targets.
Companies often lend money to customers — just think of the financial arms of automakers. Huang argues that Nvidia is simply helping to facilitate investments in viable projects that would otherwise have difficulty obtaining loans at an adequate cost.
For critics, these deals are reminiscent of the dot-com bubble of the late 1990s, when telecommunications equipment manufacturers lent billions of dollars to companies in the sector that bought their equipment.
When demand for these companies' services fell below expectations, some went bankrupt, leaving Cisco and Lucent with large losses.
For Jay Goldberg, from the analysis firm Seaport Research Partners, Nvidia is walking a fine line between enabling and creating demand. He still doesn't consider, however, that the company has crossed the line, although it is "getting very close."
The question is relevant because Nvidia's financial commitments are enormous and growing fast. Over the past three years, the company has committed to more than US$70 billion (R$357 billion) in investments in startups and offered US$300 billion (R$1.53 trillion) in support to customers.
Some have come to call Nvidia the "central bank of AI", given its importance in financing the sector. This will indeed help the industry grow, as Huang argues, but it also involves risks.
Nvidia's financial engineering is, in part, a response to the transformation of its biggest customers into competitors. The so-called "hyperscalers" — big techs like Amazon, Google, Meta and Microsoft — account for about half of Nvidia's revenue.
This year, they are expected to invest around US$800 billion (R$4.08 trillion), mainly in AI infrastructure. But most have started developing their own chips, which puts future purchases from Nvidia in doubt.
For hyperscalers, proprietary chips are cheaper and more adaptable. Some also sell them: Google sold specialized processors to Anthropic.
Amazon also expects its chip business to become a significant source of revenue. Bloomberg Intelligence predicts these chips will gradually chip away at Nvidia's sales, accounting for about 50% of the market for processors used in AI by the end of the decade, up from about 40% this year.
Hyperscalers have investment-grade credit ratings, which keeps financing costs low. Neoclouds — newer cloud computing companies — need similar investment but have little revenue. Your loans are naturally much more expensive.
Alphabet, owner of Google, sold US$2.7 billion (R$14 billion) in bonds maturing in 50 years in November, at an annual interest rate of 5.7%. The rate paid by CoreWeave, the largest neocloud, to borrow US$2.6 billion (R$13.3 billion) in July was almost double.
It is this difference between financing costs that Nvidia bank intends to reduce. One of the ways to do this is to acquire equity stakes in startups that will be customers or that will indirectly help drive demand for Nvidia chips.
Part of the investments seeks to disseminate open, free and adaptable weighted AI models, in contrast to proprietary offerings from Anthropic, OpenAI and Google.
In August, Nvidia agreed US$6 billion (R$30.6 billion) to license Poolside's software and US$1 billion (R$5.1 billion) for a stake in the startup. It also agreed to purchase Hugging Face for US$12.9 billion (R$65.8 billion).
The investments seek to create AI companies independent of big tech and increase demand for Nvidia chips.
Increasingly, however, Nvidia is not only investing in promising companies, but also directly helping them finance major projects. At the beginning of July, it announced a new strategy by which it committed to supplementing neocloud revenue with new data centers up to a previously agreed floor.
These commitments usually last six years. During this period, Nvidia promises to pay a fixed price for the "compute", as the industry jargon says. If neocloud sells the capacity in question for a higher price, Nvidia receives part of the difference.
The guarantee makes revenue more predictable, lowers debt for new data centers and stimulates demand for Nvidia processors.
Nvidia is also taking another approach with large AI labs, which need enormous amounts of computing power but burn money. The data center the company is supporting in Ohio belongs to SB Energy, owned by Japanese conglomerate SoftBank. OpenAI will be the tenant.
Nvidia will guarantee rental and purchase of energy. In return, the plant will use 1.5 million of its processors.
Even Anthropic, which buys chips from Amazon and Google, has not escaped this net. Through several agreements, the company would have signed a US$35 billion (R$178.5 billion) contract to rent cloud computing capacity from Lambda, a neocloud in which Nvidia has a stake.
Nvidia is also exploring how to attract more outside capital to these projects. Its US$500 billion (R$2.55 trillion) partnership with major Wall Street figures will seek institutional investors, such as sovereign wealth funds, insurance companies and pension funds.
They will finance vehicles that will buy equipment, build infrastructure and sell capacity. Nvidia will not contribute resources or take on debt, but may offer "residual value support" of up to a quarter of the business.
It will also offer technical support, giving lenders more confidence that the financed equipment will be put to good use. Huang presents this as a way to turn computing power into "an investable asset class."
Behind this network of obligations are two fundamental assumptions: that Nvidia's chips will maintain their value over time and that demand will continue to grow rapidly — but neither is guaranteed.
Let's start with the chips. Huang claims they are durable, with software updates increasing their usefulness for years after sale. According to him, they should be seen as financeable assets, which can serve as collateral for loans.
So far, Huang seems to have a point. Older chips are still in demand. In August, CoreWeave announced a contract involving Nvidia's A100 chips that extends until 2029. The A100 was launched in 2020.
SemiAnalysis estimates that renting an Nvidia H100, one of the leading equipment in the AI industry, costs around US$2.80 (R$14.28) per hour on a one-year contract — around 10% less than the launch price in early 2023.
Still, both the longevity and value of these chips may simply be a result of scarcity. When computing capacity is limited, companies have no choice but to keep old chips running.
Nvidia itself launches new chips every year, with the explicit intention of replacing previous ones. This puts it in a curious position: at the same time as it argues that old chips should maintain value, it says that its new chips should replace them.
Nvidia enjoys gross margins of around 75%, versus roughly 55% for smaller rival AMD. Tim Davis, founder of an AI hardware company acquired by chipmaker Qualcomm, believes that as options multiply, the economics of chipmaking will begin to "compress."
The biggest risk to prices is demand itself. As long as AI spending continues to rise, Nvidia will sell its chips, its customers will fill data centers, and its various warranties will rarely be triggered.
But if demand isn't as widespread as expected, neoclouds and other AI providers could struggle to sell capacity or make a profit. This could trigger Nvidia's warranties, forcing it to buy unused capacity, cover price differences, or pay for power it doesn't want.
The same slowdown would also reduce Nvidia's sales and, consequently, the cash flow available to meet these obligations. This scenario does not depend on a collapse in demand; An expansion that is simply below expectations could, in theory, cause problems.
How much could Nvidia have to shell out? The company has committed to around US$25 billion (R$127.5 billion) in future equity investments and has around US$33 billion (R$168.3 billion) in debt.
Potential liabilities to customers total US$300 billion (R$1.53 trillion), off-balance sheet as they are only activated during a slowdown.
It's likely that Nvidia will end up not paying anything close to the full amount of its guarantees. Its commitments are spread over many years and cannot be achieved all at once. The proposed warranty for OpenAI's data center, for example, lasts 20 years from 2028. If OpenAI were to stop paying rent, Nvidia could find another tenant, drastically reducing its exposure.
Furthermore, Nvidia's own finances are strong enough to absorb these liabilities. Bank Morgan Stanley estimates that Nvidia's "total" debt will increase from US$53 billion (R$270.3 billion) at the beginning of next year to US$200 billion (R$1.02 trillion) at the beginning of 2029, as guarantees come into force.
The risk is cushioned by US$99 billion (R$504.9 billion) in cash and liquid securities and by the expected generation of US$200 billion (R$1.02 trillion) in cash this year; only a catastrophic slowdown would threaten the company.
The scenario could change, however, if Nvidia's commitments continue to grow. Andy Li of financial research firm CreditSights fears the company will continue to "accelerate" until "something breaks."
SemiAnalysis estimates US$5.9 billion (R$30.1 billion) in guarantees for every 100 MW of data centers supported; exposure could reach US$175 billion (R$892.5 billion) in 2028.
Amid the rush to secure sales and loans, there is a risk that speculative projects, which would otherwise have difficulty obtaining financing, end up being built. If industry returns fall short or take longer to materialize than these agreements assume, the result will be a glut of expensive, idle chips.
