Even Nvidia’s Rivals Are Buying Jensen Huang’s Rubin Chips

Observer ·

Even Nvidia’s Rivals Are Buying Jensen Huang’s Rubin Chips

Huang introduced Rubin as a complete computing system. In his last full quarter as CEO, Nvidia was also guaranteeing some of the data centers where its chips would run.

Nvidia co-founder Jensen Huang took the stage in Las Vegas in January with six chips designed to work together . He introduced them as Rubin, a system that lets cloud companies buy the processors that run A.I. models and the networking that connects them in a single rack. In Huang’s last full quarter as CEO, the first production shipments of Rubin brought in nearly $20 billion, according to the company.

Nvidia made its pitch to another set of buyers at the SC26 supercomputing conference , which closes Nov. 20 in Chicago. Ian Buck , who runs high-performance computing at Nvidia, introduced a separate Rubin chip for scientific simulations on Nov. 16. Nvidia expects that version to be available in 2028.

Rubin racks are already running at Google Cloud, Microsoft Azure and Oracle Cloud. Google sells cloud customers time on its own chips. Microsoft designed its own A.I. processor , too. That gives Colette Kress, Nvidia’s interim CEO , an unusual set of negotiating partners. Some of her biggest buyers have an alternative to Nvidia. Others need Nvidia’s financial backing to buy its equipment in the first place.

Huang was still in charge when the quarter closed on Oct. 25 . Nvidia’s revenue had jumped 103 percent from a year earlier to $115.8 billion. Data centers accounted for $107.6 billion, up 21 percent from the previous quarter and 110 percent from a year earlier. By the time Kress delivered those results on Nov. 17, the board had accepted Huang’s resignation . Board chair Harvey Jones told investors the directors would choose his permanent successor and had ruled out reinstating him.

Kress, who is still chief financial officer, called Rubin the fastest rollout in Nvidia’s history. Nvidia said the early stages of the rollout showed up in its gross margin, which edged down to 74.8 percent from 75 percent the previous quarter. Nvidia expects 71 to 72 percent in the next quarter as Rubin shipments grow.

For a company that rents out computing power, the arithmetic looks different. Prices for tokens, the units in which A.I. services meter usage, have fallen nearly 60 percent since May, Kress said on the earnings call . As usage climbed, customers renewing multiyear rentals of Nvidia’s older Hopper processors agreed to pay up to twice their original rate, she said. Rental rates for the newer Blackwell Ultra B300 were up more than 60 percent since its introduction.

“Nobody knows how long that pricing power is sustained, but at least for the moment, there’s a lot more,” Tomasz Tunguz , founder and general partner of the venture firm Theory Ventures , tells Observer. “The neoclouds and the hyperscalers are able to monetize those Nvidia chips better than I think most people would have expected.”

Still, “the TPU business for Google is monstrous,” says Tunguz. Google’s tensor processing units give customers another way to train and run their models. Sales of those systems contributed to Google Cloud’s 103 percent revenue growth in the third quarter, its parent company, Alphabet , reported on Oct. 28.

“As tokens become cheaper or the usage increases, the same buyers are buying more, or more buyers enter the marketplace and are consuming tokens,” Lior Prosor , co-founder and managing partner of the venture fund Deep33 , tells Observer.

Prosor thinks Nvidia may lose some share of the market for inference chips, the processors that put a trained model to work answering a query. Google Cloud can offer its customers a choice of its own chips or Nvidia’s.

Nvidia expects about $130 billion in revenue in its fourth quarter, give or take 3 percent. That quarter has an extra week. The forecast assumes no data center compute revenue from China, where Nvidia shipped $2.4 billion of Hopper data center products in the quarter just ended.

For Prosor, what changed on Kress’s first call was how plainly Nvidia described its role in financing the companies that buy its hardware. “It was implicit when Jensen was in the seat,” he says.

Nvidia has bills of its own. At quarter’s end, it had $102 billion in cloud service and data center lease commitments , up from $74 billion three months earlier. The new total includes $74 billion for leases that have yet to start and $28 billion for cloud services. Nvidia uses that capacity to design and test chips and develop its own models.

Then there are the specialist A.I. cloud providers, or neoclouds. Nvidia separately listed $32 billion in agreements that depend on those companies securing financing and getting their equipment up and running. Nvidia can earn money selling them the equipment and, in some deals, collect a share of what they earn renting it out.

In Ohio, Nvidia agreed under Huang to back an SB Energy campus where OpenAI plans to lease space for Nvidia hardware. That one project accounts for up to $105 billion of Nvidia’s $118 billion in maximum exposure across its guarantees. The Ohio guarantees take effect one building at a time. Nvidia said it had made no payments on its guarantees as of Oct. 25.

Kress told investors that with other sources of capital becoming available, Nvidia was “rigorously evaluating future investments, commitments and guarantees.”

Some customers need help securing a site. Others can afford to design their own chips, even as they order racks of Rubin. Tunguz has a question about that second group. “Is that ultimately healthy for Nvidia? Because they’re selling to people who are developing chips that will ultimately compete.”

Источник: Observer