Nvidia CEO claims reasoning models will boost GPU demand

Nvidia continues to dominate artificial intelligence (AI), with its latest quarterly revenue results showing a 16% increase – a 93% rise from the same time last year.

Datacentre revenue for the company was $35.6bn in the quarter and $115bn over the entire year, a 142% rise compared to last.

Jensen Huang, Nvidia’s CEO and founder, said in his prepared remarks: “Demand is increasing for Blackwell adds another scaling rule to reasoning AI – increasing compute makes models smarter, and increasing compute makes the answer smarter.

We’ve successfully ramped-up the massive-scale manufacturing of Blackwell AI Supercomputers, achieving millions of dollars in sales during its first quarter. AI is advancing at a rapid pace as agentic AI, physical AI, and other AI technologies set the stage for a new wave of AI that will revolutionise industries.

During the earnings conference Financial analysts questioned Nvidia about DeepSeekwhich requires less powerful graphics processor units (GPUs), as well as the fact that cloud services providers (CSPs), such Microsoft, are designing their custom chips optimised to AI workloads. According to the transcript of the earnings conference,

Nvidia is developing its own custom chips optimised for AI workloads. CSPs make up about half of Nvidia’s business, according to a post on Seeking Alpha. There is also a growing demand from enterprise clients. Huang said, “We see enterprise growth going forward,” which he thinks represents a greater opportunity to sell Nvidia GPUs on the long term.

Huang discussed in the earnings call why he believes that new AI models will increase demand, even though AI models are becoming more computationally efficient. “The more a model thinks, then the smarter it is,” he said. “Models such as OpenAI, Grok-3, and DeepSeek R1 are reasoning models which apply inference time scale. Reasoning models consume 100 times as much compute. Future reasoning models could consume a lot more compute.”

Huang, when asked about the risk of CSPs developing application-specific integrated chips (ASICs) rather than using GPUs, responded by talking about how complex the technology stack is. “The software stack can be incredibly difficult.” “Building an ASIC is the same as what we do, we build a brand new architecture,” he explained.

Huang claims that the technology ecosystem built on top of Nvidia’s architecture is 10x more complex than it was just two years ago. “That’s pretty obvious,” he said. “The amount of software being built on top of architecture has grown exponentially, and AI is progressing very quickly.” It’s hard to bring that whole ecosystem [together] onto multiple chips.

Forrester senior analyst Alvin Nguyen commented on the Nvidia results: “Having another record performance from Nvidia feels commonplace, despite its enormity.” The record earnings are a reflection of the demand for Nvidia’s AI products. The emphasis on reasoning models driving greater, not lesser, computation is a great verbal counter to worries about DeepSeek affecting their demand.”

Nguyen, however, felt Huang’s answers to questions about custom chips that are an alternative to Nvidia GPUs were “dismissive”.

According to Nguyen, Huang’s response to questions about how custom chips from Amazon and Microsoft could threaten their business was “dismissive”.