Blog

Sam Altman, the chief executive of OpenAI, says that technological revolutions are driven by more than just technology. They are also driven, he argues, by new ways of paying for them.

“There is always a lot of focus on technological innovation. What really drives a lot of progress is when people also figure out how to innovate on the financial model,” he recently said at the site of a data center that OpenAI is building in Abilene, Texas.

Over the last several years, Mr. Altman’s company has found unusual and creative ways of paying for the computing power needed to fuel its ambitions.

Many of the deals OpenAI has struck — with chipmakers, cloud computing companies and others — are strangely circular. OpenAI receives billions from tech companies before sending those billions back to the same companies to pay for computing power and other services.

Industry experts and financial analysts have welcomed the start-up’s creativity. But these unorthodox arrangements have also fueled concerns that OpenAI is helping to inflate a potential financial bubble as it builds what is still a highly speculative technology.

Here are unusual financial agreements helping to drive the ambitions of OpenAI, the poster child of the artificial intelligence revolution.

From 2019 through 2023, Microsoft was OpenAI’s primary investor. The tech giant pumped more than $13 billion into the start-up. Then OpenAI funneled most of those billions back into Microsoft, buying cloud computing power needed to fuel the development of new A.I. technologies.

(The New York Times has sued OpenAI and Microsoft, claiming copyright infringement of news content related to A.I. systems. The two companies have denied the suit’s claims.)

By the summer of last year, OpenAI could not get all the computing power it wanted from Microsoft. So it started signing cloud computing contracts with other companies, including Oracle and little-known start-ups with names like CoreWeave.

Across three different deals signed this year, OpenAI agreed to pay CoreWeave, a company that builds A.I. data centers, more than $22 billion for computing power. As part of these agreements, OpenAI received $350 million in CoreWeave stock, which could ultimately help pay for this computing power.

OpenAI also struggled to get the additional investment dollars it wanted from Microsoft. So, it turned to other investors. Earlier this year, the Japanese conglomerate SoftBank led a $40 billion investment in OpenAI.

At the same time, OpenAI has been working with various companies to build its own computing data centers, rather than rely on cloud computing deals. This also includes SoftBank, which is known for highly speculative technological bets that don’t always pay off. The company is raising $100 billion to help OpenAI build data centers in Texas and Ohio.

Similarly, Oracle, a software and cloud computing giant, has agreed to spend $300 billion building new data centers for OpenAI in Texas, New Mexico, Michigan and Wisconsin. OpenAI will then pay Oracle roughly the same amount to use these computing facilities over the next several years.

The United Arab Emirates was part of an OpenAI’s fund-raising round in October 2024. Now, G42, a firm with close ties to the Emirati government, is building a roughly $20 billion data center complex for OpenAI in the Emirates.

Last month, Nvidia announced that it intended to invest $100 billion in OpenAI over the next several years. This could help OpenAI pay for its new data centers. As OpenAI buys or leases specialized chips from Nvidia, Nvidia will pump billions back into OpenAI.

Two weeks later, OpenAI signed an agreement with AMD that allows OpenAI to buy up to 160 million shares in the chipmaker at a penny per share. That translates to roughly a 10 percent stake in the company. This stock could supply OpenAI with additional capital as it works to build new data centers.

OpenAI pulls in billions of dollars in revenue each year from customers who pay for ChatGPT, computer programming tools and other technologies. But it still loses more money than it makes, according to a person familiar with the company’s finances.

If the company can use its new data centers to significantly improve A.I. technologies and expand its revenue over the next several years, it can become a viable business, as Mr. Altman believes it will. If technology progress stalls, OpenAI – and its many partners – could lose enormous amounts of money. Smaller companies like CoreWeave, which are taking on enormous amounts of debt to build new data centers, could go bankrupt.

In some cases, companies are hedging their bets. Nvidia and AMD, for instance, have the option of reducing the cash and stock they send to OpenAI if the A.I. market does not expand as quickly as expected. But others would be left with enormous debt, which could send ripples across the larger economy.

{The electrical} system warning mild had gone on in my Kona EV over the weekend, and all of the guide stated was to take it to the supplier for analysis. I first tried scheduling an appointment through the web site, and it jogged my memory how the net, as soon as a marvel, is wanting awfully clunky lately. There have been a number of choices for providers to schedule, however it wasn’t in any respect clear which ones I’d need.

Not solely that, I’d solely reached this web page after clicking by means of numerous promotions and testimonials about how nice the dealership is—in brief, content material designed to serve the pursuits of the supplier quite than the pursuits of the shopper. Finally, I did discover a free-form textual content subject the place I may describe the issue I truly wished the appointment for. However then it pushed me to a scheduling web page on which the primary accessible appointment was six weeks away.

So I attempted calling the service division immediately, to see if I may get some indication of how pressing the issue may be. The telephone was busy, and a pleasing chatbot got here on providing to see if it’d assist. It was fairly an exquisite expertise. First, it had already recognized my automobile by its affiliation with my telephone quantity, after which requested what the issue was. I briefly defined, and it stated, “Received it. Your EV service mild is on, and you should have it checked out.” Bingo! Then it requested me after I wished to schedule the service, and I stated, “I’m unsure. I don’t understand how pressing the issue is.” As soon as once more. “Received it. You don’t understand how pressing the issue is. I’ll have a service advisor name you again.”

That was practically an ideal customer support interplay! I used to be very happy. And somebody did certainly name me again shortly. Sadly, it wasn’t a service advisor; it was a poorly skilled receptionist, who apparently hadn’t acquired the knowledge collected by the chatbot, since she gathered all the identical info, solely far much less effectively. She needed to ask for my telephone quantity to search for the automobile. Half the time she didn’t perceive what I stated and I needed to repeat it, or I didn’t perceive what she stated, and needed to ask her to repeat it. However ultimately, we did get by means of to the purpose the place I used to be provided an appointment this week.

This was not the one difficult customer support expertise I’ve had not too long ago. I’ve had an issue for months with my gasoline invoice. I moved, and by some means they arrange my new account fallacious. My on-line account would solely present my former tackle and gasoline invoice. So I deleted the present on-line account and tried to arrange a brand new one, solely to be instructed by the net interface that both the account quantity or the related telephone quantity didn’t exist.

Calling customer support was no assist. They might search for the account quantity and confirm each it and the telephone quantity, and inform me that it ought to all be OK. However after I tried once more, and it nonetheless didn’t work, they’d inform me that somebody would look into it, repair the issue, and name me again when it was completed. Nobody ever referred to as. Not solely that, I even acquired a plaintive letter from the gasoline firm addressed to “Resident” asking that I contact them, as a result of somebody was clearly utilizing gasoline at this tackle, however there was no account related to it. However after I referred to as again but once more and instructed them this, they might discover no document of any such letter.

Lastly, after calling a number of occasions, every time having to repeat the entire story (with no document apparently ever being saved of the a number of interactions on the gasoline firm finish), I wrote an e-mail that stated, primarily, “I’m going to cease attempting to resolve this downside. The ball is in your courtroom. Within the meantime, I’ll simply assume that you’re planning to supply me gasoline providers totally free.” At that time somebody did name me again, and this time assured me that that they had discovered and stuck the issue. We’ll see.

Each of those tales emphasize what an enormous alternative there may be in customer support brokers. However in addition they illustrate why, in the long run, AI is a “regular know-how.” Regardless of how clever the AI powering the chatbot may be, it needs to be built-in with the programs and the workflow of the group that deploys it. And if that system or workflow is unhealthy, it must be reengineered to utilize the brand new AI capabilities. You’ll be able to’t construct a brand new skyscraper on a crumbling basis.

There was no chatbot on the gasoline firm. I want there had been. However it could solely have made a distinction if the knowledge it collected was saved into information that have been accessible to different AIs or people engaged on the issue, if these assigned to the issue had the experience to debug it, and if there have been workflows in place to comply with up. It’s attainable to think about a future the place an AI customer support assistant may have truly fastened the issue, however I believe that it will likely be a very long time earlier than edge circumstances like corrupted information are solved mechanically.

And even with the nice chatbot on the Hyundai supplier, it didn’t do a lot to vary my general buyer expertise, as a result of it wasn’t correctly built-in with the workflow on the dealership. The data the chatbot had collected wasn’t handed on to the suitable human, so a lot of the worth was misplaced.

That implies that the issues that face us in advancing AI should not simply making the machines smarter however determining how one can combine them with present programs. We could ultimately get to the purpose the place AI-enabled workflows are the norm, and firms have discovered how one can retool themselves, however it’s not going to be a straightforward course of or a fast one.

And that leads me to the title of this piece. What’s the aggressive moat if intelligence turns into a commodity? There are numerous moats ready to be found, however I’m positive that certainly one of them will likely be integration into human programs and workflows. The corporate that will get this proper for a given business may have a bonus for a surprisingly very long time to come back.