Nvidia: The Most Important Company for the rest of our Life

A Fund Manager’s Perspective

Author: Rohit Sah, PhD

Published: June 6, 2024

Nvidia is set to re-write the computer industry in a way that would have been incomprehensible just a few years back. How they got to this point needs fleshing out in a way I have not seen elsewhere.

Initial Days: A Near Death Experience

Nvidia was started in 1993, one of ninety similar chip companies wanting to create a custom graphic card: a semiconductor chip to accelerate the graphics on a normal PC.

Soon after they started, the company nearly died. To create graphics, you need polygons. Nvidia decided to use quadrilaterals for vertices instead of triangles, which was a perplexing decision; even Nvidia’s headquarter building is dominated by triangles. As Jensen Huang puts it today, “the triangle is the fundamental building block of computer graphics.” Sure enough, Microsoft decided to standardize 3D graphics on triangles. Soon after, Sega canceled Nvidia’s only contract. It was only because of the benevolence of Sega’s then President Soichiro Irimajiri deciding to invest $5 million in Nvidia that Nvidia lived to fight another day. By the time Nvidia’s first successful chip, the RIVA 128, was released in August 1997, Nvidia’s cash was down to only one month of payroll.

This extreme desperation is what created Nvidia’s unofficial company motto: “our company is thirty days from going out of business”. This paranoid ethos stayed with Nvidia. As Huang puts it, “when technology moves this fast, if you’re not reinventing yourself, you’re just slowly dying.” Nothing moves faster than the semiconductor industry.

Eventually, Nvidia started moving away from making commodity graphic cards. Nvidia made graphic cards programable by extending the C programming language (Nvidia called it CG) with libraries and other capabilities. This combination of chips and CG language is branded as GPUs, or graphical processing units. GPUs became more generalized computing devices.

CPUs are the central processing unit, and as the name suggests, it centralizes the processing of the device’s tasks and sends instructions serially through the processor.

GPUs are built on quadrilaterals and allow matrix calculations to be performed independently of one another. This allows the device’s tasks to be broken into smaller independent tasks and performed simultaneously. And this is how pixels get created; each pixel creation can be done independently of others. Parallel computing was extended to many other applications over the last few decades. Parallel computing method is what makes GPUs so much faster than CPUs, and how we overcame Moore’s Law implicit constraints on CPU computing power.

From GPUs to CUDA: Building the Unthinkable in Darkness

In the early 2000s, a Stanford quantum chemistry researcher called up Jensen Huang, and said that when he used Nvidia’s GPUs in running his computing intensive models, his models finished tabulations in 2 hours, compared to the 2 weeks he managed at Stanford’s super computers.

This realization that there was a scientific community that was engaged in extremely computationally intensive activities, and could benefit from GPUs, was the spark for what eventually would create the most important company in the world.

But before I get to that, let us examine how the world looked in 2006, before Nvidia started spending vast resources on their CUDA project.

Whenever a company embarks on any project, it needs to take stock of several issues. What is the size of the addressable market? How much will this project cost? How long will it be before this project starts making money? Are there any supply chains available to enable this project’s success? Will this new project take away attention and people from the current engine of profitability and revenue streams? Will the software stack be open source?

In 2006 the answer to each of these questions was a huge negative. If a startup says that they will change the world by serving a sliver of the computing market demand coming from academicians, they would not be able to raise any capital. This would also have meant that Nvidia’s share price would be vulnerable to horrific swings, as markets naturally prefer predictable profit growth, not wild swings from unprofitable projects. Few CEOs would risk their careers on such moonshots.

But Nvidia went ahead anyway. As Jensen Huang put it, “If we don’t build it, they cannot come”. They saw a future that no one else did: a complete overhaul of the way we use computers. Perhaps they realized that the entire world would eventually be doing what the Stanford researcher was doing: using enormous computing power. In a way, Nvidia was answering a question that stumped IBM’s President in 1943, when he claimed that there was a world market for only five computers. Nvidia rationalized that the world will remain in need of ever more computing power. But in 2006, this was not clear.

Nvidia poured lots of resources into this project over the next several years, with barely any success. The company suffered along the way. In 2008, the share price went down 80 percent. A company whose share price goes down that much automatically becomes a target for a buyout or a CEO change. Nvidia lost all credibility with markets as no one could see the profits from Nvidia’s moonshot project. But Nvidia managed to survive these violent swings. It would be almost ten years before Nvidia’s market capitalization would exceed what it peaked in 2007, which was $20 billion.

First Contact with Artificial Intelligence: Light at the End of the Tunnel for Nvidia

Princeton Professor Fei-Fei Li worked on an online database with 14 million images, all labeled by hand along with 20,000 categories (with a category being balloon or strawberry etc.). For any given category, ImageNet contains several hundred images. They started running an annual contest where software programs competed to correctly classify objects and scenes.

In 2012, AlexNet, a software program from the University of Toronto won this contest by a large margin using neural networks, powered by Nvidia GPUs with code written in Nvidia’s CUDA. This was a breakthrough result: for the first time ever, “deep learning” produced a high accuracy (85% vs humans at 95% but well above the 72% achieved before). Deep learning algorithms are inspired by the human brain’s neural networks. Such algorithms attempt to learn patterns directly from data.

The idea of using artificial neural networks had been around for decades. What caught people’s attention was that deep learning could be used for many domains. Google is using deep learning to boost the quality of web-search results. The beauty of one branch of deep learning, i.e. unsupervised learning, is that there is no need for a human to draw up a list of rules; the system learns directly from labeled data.

But for the first time, it was possible to use massive amounts of computing power and vast quantities of training data. All this was only made possible by Nvidia’s GPUs and CUDA software.

On to Revolutionizing the World

CUDA is a software development framework for doing any kind of computation that you want on GPUs. In simple terms, it has hundreds of software development kits that serve all kinds of industries and applications. There is one for Physics, one for weather, one for data science etc.

There is no fee to use CUDA, unlike say, Microsoft Office. But CUDA works only with Nvidia chips. Anyone can download and use it, but it is a closed source and proprietary exclusive to Nvidia chips. The software is optimized for Nvidia hardware.

By 2016, many companies were applying deep learning in several areas, and all of them were building on Nvidia platforms. What further turbocharged Nvidia’s growth was a landmark research paper referred to as the Transformer model published in 2017.

The “Attention is All You Need” paper, or the Transformer paper, provided a framework for deep learning models. One example is Large Language Models, like ChatGPT. But such models require unfathomable amounts of computing power and data. To take one example, if you are training one single speech recognition machine learning model these days, the number of math operations like adding or multiplying to accomplish it is greater than the number of grains of sand on Earth. Only Nvidia’s GPUs and CUDA enabled such gargantuan calculations.

By 2019, the building blocks of Nvidia’s future success were in place. The last block came when Nvidia bought up Mellanox. Data centers and most of the computer industry are standardized on Ethernet technology. Mellanox specialized in InfiniBand, a computer networking standard that has even higher throughput and very low latency compared to Ethernet. When Nvidia bought Mellanox, no one outside Nvidia could understand what on Earth would require a throughput of 3200 Gigabits a second of bandwidth running down a wire in a data center.

With Mellanox under its fold, Nvidia has now built out its entire portfolio of products. If you are a gamer wanting to speed up your desktop PC, there is a Nvidia gaming chip. If you are a sophisticated data center business (like AWS), Nvidia can sell you specialized chips for you to model your data center around. If you already have a data center, and like the PC gamer, you want to add capacity to your data center, Jensen Huang can sell you a box to add on. And if you need a whole data center on a turnkey basis, Nvidia can do that too, thanks to the InfiniBand networking equipment. All this is helped along with an ever-increasing CUDA library.

Concluding Thoughts

Investors sometimes compare Nvidia to Cisco’s run in the 1990s. That does not do any justice to Nvidia’s future in recreating the computer industry. A better comparison would be the IBM-Microsoft-Intel dominated standard of the last 35 years. Nvidia owns the standards, and with its TSMC partnership, it also has manufacturing capabilities.

My preferred comparison is with General Electric, when the electricity network was being rolled out across the United States. GE sold electrons through electric wires. Nvidia “sells” tokens through its data center-oriented hardware. GE created the electricity generation, distribution and retail of electricity. Nvidia is about to do much of the same for Artificial Intelligence. As we move towards an AI centric industrial revolution, the underlying structure will be powered by Nvidia.

Nvidia’s growth will be built on the back of TSMC’s manufacturing capacity. TSMC has penciled in 50% annual growth in their data center business (i.e. mostly Nvidia) for the next 5 years, implying that Nvidia sales will be seven times larger than current levels in 5 years.

Although Nvidia’s market cap is nearly the largest in history, there is no reason why it cannot keep going much higher. If Nvidia was to parallel the growth trajectory of General Electric, then Nvidia will be investing in industries and companies that directly benefit from AI. General Electric was instrumental in the development of the Electric Locomotive, X-ray machine, Radio Broadcast, electric home appliances, vacuum tubes, home television, and all this was just by 1927. Nvidia is sitting at the heart of the AI industry with a gusher of cash flow coming from its 70% gross margin business. Nvidia reinvesting this cash flow not only enables the industry for exponential growth, but it also sets the company apart to be an investor in new ways of doing things.