Nvidia prepares to exponentially increase AI inference

Chipmaker Nvidia reported revenue of $57bn in its third-quarter filing for 2026. Its datacentre business contributed the most to its bottom line with revenue of $51bn, a 66% increase year-over-year compared to last year’s results.

According to CEO Jensen Huang, the company continues to see growth in AI workloads that require high performance graphics processor unit (GPUs), which Nvidia is specialized in.

According Huang, AI inferences are scaling exponentially because of advances in pre-training and post-training capabilities, as well as reasoning abilities. He said that AI inferences are becoming more complex as AI systems “read, think, and reason” to generate answers. Huang claimed that this exponential increase in computation requirements is driving the demand for Nvidia platforms.

The company’s NVLink AI network infrastructure business grew 162% with revenue of $8.2bn.

Huang stated: “Customer demand for NVLink Fusion is growing.” In October, we announced a strategic partnership with Fujitsu whereby we will integrate Fujitsu’s CPUs and Nvidia’s GPUs via NVL fusion, connecting our large eco-systems. We also announced a partnership with Intel to develop custom datacentre products and PCs connecting Nvidia’s and Intel’s ecosystems via NVLink.” He said, “Our part of the datacentre has increased with each generation from Ampere, to Hopper, to Blackwell, to Rubin.” He said that a datacenter with a power of one gigawatt still has a limited amount of power. “Your performance per Watt translates directly into your revenues. This is why choosing the right architectural approach is so important now.”

Huang responded to a question about the biggest bottleneck for Nvidia’s future growth by saying: “What Nvidia does obviously hasn’t been done before.” We’ve created an entirely new industry. On the one hand we are transitioning from general-purpose computing and classical or traditional computer to accelerated computing, and AI. A new industry was created called AI factories. The idea that in order for software to work, you need these factories, which generate each token, rather than retrieving pre-created information.” We plan the supply chain up and down. We have established a number of partners and therefore we have many routes to market.”

But one market is now effectively closed, which is China. Huang said that geopolitical concerns and the highly competitive market in China prevented large purchase orders from materialising in the first quarter. We are disappointed that we cannot ship more competitive datacentre computing products to China. However, we will continue to advocate America’s ability compete around the globe. Huang wants Nvidia to deliver the best value to customers, saying: “At this point, I’m very confident that Nvidia’s architecture is the best performance for TCO [total cost of ownership]”.