|
Thanks for reading The Briefing, our nightly column where we break down the day’s news. If you like what you see, I encourage you to subscribe to our reporting here.
Greetings! Nvidia wants customers to buy as many of its next-generation AI chips as possible. It also wants its customers to know their old AI chips will remain valuable for a long time. Can it manage both? Nvidia has been able to grow so quickly because it has convinced big customers to shift almost completely to its latest and greatest hardware nearly every year. At the Computex conference in Taiwan in 2024, for example, Nvidia even shortened its release cycles for new hardware from two years to one to get clients to upgrade more frequently. And on Monday, Nvidia, OpenAI and SB Energy signed a deal to help OpenAI fill 8 gigawatts of space exclusively with Nvidia’s next-generation chips—which CEO Jensen Huang said could amount to $150 billion to $200 billion in revenue for Nvidia from each new generation of its hardware. Each generation is much more powerful than the last—but also much more expensive, helping Nvidia’s profits skyrocket and keeping its margins fat. Now Nvidia is telling the market that its graphics processing units are an investable asset class—and can thus hold financial value for much longer. These two positions feel somewhat at odds. Taking both sides makes sense in an environment of compute scarcity, however. There seems to be demand for any type of usable GPU a customer can get their hands on, whether it’s the newest Vera Rubin or an Ampere chip from several generations back. A longer life cycle for GPUs could be Nvidia’s hedge against the moment when demand doesn’t outpace supply. Right now, bigger players absolutely want Nvidia’s newest Vera Rubin server racks. Cloud executives told me most of the demand they’re seeing from customers right now is for those racks, which are set to ship later this fall at around $7 million each. But right now, the biggest buyers are just a handful of players: cloud giants, Nvidia-supported neoclouds (like CoreWeave, Nebius, Nscale and Firmus), Nvidia-supported AI labs (like OpenAI) and, maybe, Wall Street quant firms. The next phase of AI adoption, and thus the next big driver of demand for computing power, will likely come from a wider range of businesses than a handful of cloud and AI companies. Those customers will also likely be more cost-conscious—and they may not need the most powerful, most expensive Nvidia server racks. Market dynamics could also change: My colleague Ann reported today that some analysts and data center developers are worried there could be a glut of server chips due to holdups in building and powering up data centers. Some investors are also getting more concerned that big tech’s AI data center spending spree could lead to the accumulation of too much debt before the companies deliver the profits to back it up. The cloud giants could pull back on spending as well. One credit executive said that as time goes on, only a few companies may need to keep switching to Nvidia’s newest hardware every year. Switching entails a lot of cost and risk that a large or decades-old business might not want to take on. Lee Kestler, CEO of data center developer EdgeCore Digital Infrastructure, compared where the GPU market is going to what happened in the cellphone business when most users decided they were fine holding on to their devices for years. “The older models still make calls and text. Value isn’t directly related to ‘latest and greatest’ when it comes to use cases to keep revenue flowing,” he said. If Nvidia amplifies the message that older-generation GPUs can be useful and financially valuable for a long time, that could push fewer people to want its forthcoming Rubin Ultras and make asset-backed debt harder to raise. Older or secondhand chips are also cheaper, and that would allow more customers to afford buying or renting Nvidia hardware. Last week, for example, CEO Jensen Huang said that Nvidia’s Ampere chips, first launched in 2020 and no longer in production, could be “mission-capable” through 2029, thanks to new software that can improve their performance. Still, for now, most big buyers of GPUs aren’t changing how they account for them on balance sheets as a result of the chips’ longer lives. Most companies are still depreciating GPUs linearly over five or six years, data center and credit executives said. (That means GPUs lose a fifth or sixth of their value on a customer’s balance sheet each year.) If you come across a 10-year depreciation schedule for Nvidia hardware, let me know. Over the weekend, Dario Amodei dusted off his X account for only his fourth post of the year to make the case that he is not, in fact, an AI doomsayer. He argued that when Anthropic pushes for regulation, it’s not just trying to create a moat that will guarantee its continued dominance. Instead, Amodei wrote, the AI giant advocates holding the most powerful companies to the highest standards while giving smaller competitors a chance to catch up. Amodei’s frequent nemesis, former White House AI czar David Sacks, didn’t buy the defense. In a response late Sunday night, Sacks argued that regulation often occurs not for the good of the public, but because deep-pocketed industry groups convince the government to write rules as they see fit. It’s hard to have this argument absent the actual existence of any substantial regulation. Buried in Amodei’s thread was a curious line about the White House’s voluntary AI framework, which the Trump administration briefed companies on earlier this month. Amodei wrote that he supports testing frontier models prior to their deployment, but he added: “Of course I have to see the details to be sure.” Even Sacks, who spent more than a year shaping the administration’s AI strategy and still chairs an ostensibly influential advisory board, posts on X as if he’s waiting for policy developments to emerge, just like everyone else. The landscape following Anthropic’s unveiling of its Mythos AI model has been defined by ad hoc government actions, open letters and proposals for pauses. We have yet to see much tangible AI policy, though. When the Trump administration does take action on that front, as with its AI framework, it often shrouds the details in secrecy. For now, the debate over regulatory capture is largely theoretical.—Leo Schwartz
- Tesla is gearing up for a public launch of its Cybercab service in Austin, Texas, as early as this month.
- A trial is set to begin Tuesday in a federal court in Oakland, Calif., in which state attorneys general will argue that Meta Platforms has designed its products in ways that have harmed young people.
- The U.S. Department of Justice has launched an antitrust probe examining whether Andreessen Horowitz’s partners are improperly holding board seats at competing AI companies, Bloomberg reported.
| 
