Nvidia chief says AI compute needs exploded 100x

When DeepSeek dropped its R1 reasoning model a few weeks ago, Wall Street panicked. Nvidia suffered a brutal 17% single-day stock plunge—its worst since 2020—as investors feared cheap AI training would destroy demand for high-end GPUs.

Now, Jensen Huang is setting the record straight: he thinks that logic is completely backward.

The 100x reasoning tax

Speaking with CNBC following Nvidia's blockbuster earnings, Huang argued that the next generation of AI actually requires 100 times more compute than older models like the original ChatGPT.

The culprit? Reasoning. When models take time to think through problems step by step, inference compute explodes.

Huang explicitly praised DeepSeek, calling R1 fantastic and world-class for open-sourcing high-end reasoning. But instead of killing chip demand, he argued DeepSeek just popularized a computational paradigm that devours hardware.

He lumped R1 in with OpenAI's GPT-4 and xAI's Grok 3 as proof that the entire industry is shifting toward compute-heavy test-time reasoning.

Nvidia's fourth-quarter numbers back up his confidence. Revenue rose 78% year-over-year to $39.33 billion, driven by a 93% surge in data center revenue to $35.6 billion, which now makes up over 90% of total sales.

Export curbs and a 60x gap

While the demand picture looks strong, Nvidia's geopolitical reality is far messier.

Huang acknowledged that export restrictions tightened under the Biden administration have cut Nvidia's revenue share from China roughly in half. Meanwhile, domestic rivals like Huawei are ramping up pressure.

The hardware performance gap is stark. Huang noted that Nvidia's top-tier GB200 system sold in the US generates AI content 60 times faster than the export-compliant chips it is allowed to ship to China.

Yet Huang remains surprisingly philosophical about Chinese developers finding workarounds. Software ultimately finds a way, he told CNBC, suggesting engineers will always optimize code to squeeze performance out of whatever silicon they have.

Why it matters

Jensen Huang just pulled off a classic judo move: turning a perceived existential threat into his biggest sales pitch.

For builders and enterprises, this confirms where the puck is going. Pre-training efficiency may be improving, but test-time reasoning burns massive compute every time a user asks a hard question. If you are building agentic workflows or complex problem-solving tools, compute costs are not going away anytime soon.

For the industry, Huang's argument repositions DeepSeek from an Nvidia-killer to an Nvidia-accelerator. By democratizing reasoning models, open source just created millions of new workloads that ultimately need to run on GPUs.

That said, Nvidia faces real exposure. A 60x performance penalty in China is an enormous wedge that gives competitors like Huawei an open runway. We think claiming software finds a way is partly Huang acknowledging that the Chinese hardware market is slipping out of Nvidia's hands, which could accelerate a permanent split in global AI hardware standards.