Nvidia’s grand plan to dominate the global AI brain from the ground up

Nvidia, the firm whose chips currently power a large portion of the world’s artificial intelligence infrastructure, is moving farther into the development of AI software and models, indicating that it aims to be much more than just a manufacturer of hardware.

A portion of the plot is revealed by the numbers. Analysts predict that Nvidia’s annual revenue will surpass $358.7 billion in 2026 after rising from $26.9 billion in 2022 to $215.9 billion in 2025.

The shares of ChatGPT have increased by around 990% from the company’s November 2022 start.

Nvidia announced intentions to invest $26 billion over the next five years to assist the development of open-source AI big models in filings submitted to the U.S. Securities and Exchange Commission.

The company’s appeal has never been just about chips. Its CUDA software platform, which allows customers to get the most out of its graphics processing units, has been central to its rise.

Justin Boitano, Nvidia’s vice president of enterprise platforms, pointed out that most of their staff is software engineers, a fact that often gets overlooked.

To build on that software side, Nvidia recently released a new open-source AI language model called Nemotron 3 Super. The model is built for enterprise-grade AI systems that involve multiple AI agents working together.

It carries 120 billion parameters and uses a design called Mixture-of-Experts.

One of its key features is a context window of up to one million tokens, meaning it can process an entire book or thousands of pages of financial records in a single run.

Nvidia has taken what might be called a middle road with this model.

Unlike OpenAI, which keeps its models closed, or Meta, which fully opens its Llama models, Nvidia will release the model’s key parameters publicly. Businesses and developers can download and run them for free, or adjust them to fit their own needs.

If Nvidia can hold onto its lead in hardware and grab 10% of the foundational model market, financial analysts say the move could bring in an extra $50 billion in yearly revenue within three years.

Partners build out the hardware side

On the hardware deployment side, Nvidia does not build data centers itself. Partners, including Dell, Hewlett Packard Enterprise, and Foxconn do that work.

Arthur Lewis, who heads infrastructure at Dell, said his company assisted one customer set up 100,000 GPUs in just six weeks. 

Concurrently, NTT DATA revealed a plan to implement what it refers to as “AI factories powered by Nvidia hardware.” These are complete configurations that integrate governance systems, software tools, infrastructure, and data.

The program makes use of Nvidia’s NeMo and NIM software tools in addition to its hardware.

A cancer research hospital that uses Nvidia platforms for radiology and diagnostics, a car parts supplier that uses Nvidia-powered cloud services to reduce production setup time from months to days, and a U.S. manufacturer currently testing battery production lines using Nvidia-accelerated simulation are just a few examples of early customer results.

Kari Briski, Nvidia’s vice president of enterprise generative AI software, noted that building these cutting-edge models puts enormous pressure on storage, networking, and computing systems, and that pressure helps shape the direction of future hardware.

CEO Jensen Huang described AI as more akin to fundamental infrastructure rather than a software fad.

“AI is one of the most powerful forces shaping the world today,” Huang stated. “It is not a clever app or a single model; it is essential infrastructure, like electricity and the internet.”

Huang described the AI stack in five layers: energy at the base, then chips, then physical infrastructure such as land and cooling systems, then AI models, and finally applications at the top, where he said actual economic value gets created, through things like drug discovery, industrial robots, and self-driving vehicles.

He acknowledged the build-out is still in early stages.

A few hundred billion dollars have been spent so far, but Huang said the total will require trillions, calling it potentially the largest infrastructure build-out in human history.

He added that AI models have recently crossed a key line, becoming reliable enough to be widely useful, and that open-source models are helping speed up adoption across the board.