How Exactly NVIDIA Is Pulling This Off

It’s not yet March and I’m already admitting defeat on my prediction that NVIDIA would remain flat in 2024. The company’s stock is up 64% this year, adding $700 billion in market cap since January. Its earnings report this week soundly beat expectations with a 265% revenue increase, sparking a $277 billion stock market pop on Thursday, the largest single-day jump in history. So I’ll say it here, mea culpa.

While NVIDIA’s competition is indeed rising, as I reported in December, competitors like AMD and Intel have struggled to match its software advantage, which is proving to be a key lock-in, one I underestimated last year. NVIDIA’s customers buy its A100 and H100 chips to train and run AI models, but do that work via NVIDIA’s proprietary software. The software is hard to learn and switching off it is a pain, so NVIDIA customers stay and reorder. And reorder again. Hence, the sustained dominance.

“It’s not just the chips. That’s the thing,” Dion Hinchcliffe, VP and principal analyst at Constellation Research, told me. “NVIDIA’s got the lock.”

ServiceNow, for instance, is an NVIDIA customer that uses generative AI to answer customer questions, summarize virtual agent conversations, and build internal workflows. The company bought hundreds of NVIDIA chips to train its AI models — building a super pod of GPUs — and uses a NVIDIA program called Triton to run them. It also uses one of NVIDIA’s off-the-shelf models called Nemo for some customer conversation work.

NVIDIA’s software advantage makes continuing to order from the company a relatively straightforward decision, Pat Casey, ServiceNow’s CTO and executive vice president of DevOps, told me. “When I talk to people I know, NVIDIA is still what people are ordering,” Casey said. “I think a lot of that is based on the software superiority. And it’s a known quantity.”

Even as competitors build chips approaching — or exceeding — NVIDIA’s capabilities, the software differential can be the difference for its customers. “If you want to go to one of the AMD chips, Triton doesn’t work there yet,” Casey said.

ServiceNow put in a big order for NVIDIA chips late last year, which are about to go live. After thinking that would be enough for a while, Casey just put in a new order for more NVIDIA chips this week.

NVIDIA also posses this advantage at the exact moment tech is swept up in generative AI euphoria, helping it capitalize before its competitors arrive in force. So, it’s sat back and reaped the benefits of a GPU land grab among big tech companies, who want to make AI capabilities available to their clients, and train their own models.

“Right now we’re seeing a dramatic amount of growth in the cloud services piece,” Harsh Kumar, senior semiconductor analyst at Piper Sandler, told me. NVIDIA’s data center revenue jumped 409% in the fourth quarter, with more than half of sales going to large cloud providers. Meta, meanwhile, will have 350,000 NVIDIA H100 GPU chips in its arsenal by the end of the year, Mark Zuckerberg said recently.

NVIDIA’s run of explosive earnings growth will eventually top off. “It’s not gonna last forever. We know this,” Insider Intelligence analyst Jacob Bourne told me. “But it’s not an overnight plummet either.” The question, of course, is how that plateau arrives. Competition might be one answer, as AMD, Intel, and other Big Tech companies get their chips and accompanying software online. But the current NVIDIA bull run is mostly predicated on the emergence of clear, profitable, widespread use cases for generative AI, which is still no sure thing.

Right now, 40% of NVIDIA data center revenue comes from inference, or using the AI models in production, meaning training the models is the bulk of spending. The companies training these models will eventually deploy them, and they’ll need to see a return on investment or that spending might vanish. We may still be years away from settling the ROI question, which is good news for NVIDIA in the short term, but some companies are already seeing results. “Is the margin positive? Yeah. Absolutely. It is,” said ServiceNow’s Casey. “They’re expensive to run, but they’re valuable too. The customers will pay for that value.”

The post How Exactly NVIDIA Is Pulling This Off appeared first on TheWrap.