Nvidia thinks it has a way to grind plateaus into vertical scale. Trillions of dollars are riding on it

In the past few weeks, there’s been a lot of chatter in the AI world that the current method of building increasingly powerful models may be reaching its limits.

The current “scaling law” of today’s large language models has essentially boiled down to more data plus more GPUs equals more capable models. This simple equation has helped push GPU behemoth Nvidia rise to the most valuable company in the US.

Now evidence is starting to appear that indicates these consistent gains may be starting to plateau. That would be very bad news for Nvidia, which Wall Street expects extremely high growth from.

But in yesterday’s third-quarter earnings call, Nvidia CEO Jensen Huang did not seem alarmed. When asked about this on the call, Huang said:

“As you know, this is an empirical law, not a fundamental physical law. But the evidence is that it continues to scale. What we’re learning, however, is that it’s not enough, that we’ve now discovered two other ways to scale.”

Huang pointed to OpenAI’s latest model, OpenAI o1, as “one of the most exciting developments” in the effort to keep scaling AI gains, as it uses a multistep “reasoning” process to break queries down into steps. “The longer it thinks, the better and higher-quality answer it produces,” Huang said.

Nvidia’s GPU business is booming, and the company pulled in over $30 billion in Q3, up 112% year over year. The company is in the midst of a transition to a new class of “Blackwell” chips after supplying pretty much every company with its “Hopper” H100 GPUs, which helped train many of the foundational models in use today.

Intense competition for Nvidia’s GPUs have led to supply constraints, raising questions about the company’s ability to ramp up enough chips to meet demand, though Huang expects a smooth transition from the Hoppers to the Blackwells.

“Hopper demand will continue through next year, surely the first several quarters of the next year. And meanwhile, we will ship more Blackwells next quarter than this, and we’ll ship more Blackwells the quarter after that than our first quarter,” Huang said.

Huang also noted an opportunity for Nvidia to update existing data centers to more modern computing clusters that are built to process AI.

“If you just look at the world’s data centers, the vast majority of it is built for a time when we wrote applications by hand and we ran them on CPUs. It’s just not a sensible thing to do anymore,” Huang said.

Huang said that any company looking to build a data center tomorrow “ought to build it for a future of machine learning and generative AI because they have plenty of old data centers.”

Colette Kress, Nvidia’s CFO, said the company’s focus on helping countries build “sovereign AI” is “such an important part of growth” and that the company continues to help countries that are “working to build these foundational models in their own language, in their own culture, and working in terms of the enterprises within those countries.”