Source: Semiconductor Industry Watch

Starting tomorrow, Nvidia will host the GTC Developer Conference. The event is now the focus of the industry's attention. With the rise of Nvidia, many people have been asking to what extent Nvidia's software provided a long-lasting competitive moat for its hardware. Since we've received a lot of related questions, we'd like to share our thoughts here.

Other than the potential launch of the next B100 GPU, GTC isn't really a chip event; GTC is an exhibition for developers. This is Nvidia's flagship activity to build a software ecosystem around CUDA and the rest of its software stack.

It's worth noting that when talking about Nvidia, many people, including ourselves, tend to use “CUDA” as a shorthand for all software Nvidia provides. This is misleading because Nvidia's software moat is more than just the CUDA development layer, which is critical for Nvidia to defend its position.

At GTC last year, the company issued 37 press releases covering a dazzling array of partners, software libraries, and models. As Nvidia strengthens its defenses, we expect more of this next week.

These partners are important because hundreds of companies and millions of developers are now building tools on top of Nvidia's products. Once built, these people are unlikely to rebuild their models and apps to run on other companies' chips, at least in the short term. Notably, Nvidia has partners and customers across dozens of vertical industries, and while not all of them fully support Nvidia, it is still showing tremendous momentum in favor of Nvidia.

Simply put, Nvidia's current position's defensive capabilities depend on the inherent inertia of the software ecosystem. Companies invest in software — writing code, testing it, optimizing it, educating employees on how to use the software, etc. — once invested, they are very reluctant to replace it.

We've seen this in the Arm ecosystem's attempts to expand into data centers over the past decade. Even if Arm-based chips begin to show real power and performance advantages over x86, it will still take years for software companies and their customers to act, and this transformation is still ongoing. Nvidia appears to be in the early stages of establishing this software advantage. If they can achieve this goal across a broad range of businesses, they're likely to stick with it for many years. The most important thing is that Nvidia is in the best position for future development.

Nvidia faces a huge barrier to entry in its software space. CUDA is an important part of this, but the way Nvidia provides software and libraries will help them build a very reliable ecosystem even as alternatives to CUDA appear.

We're pointing this out because we're starting to see alternatives to CUDA appear. AMD has come a long way in responding to CUDA, RoCm. However, when we say progress, we mean that they now have a good, working platform, but it will take years to get their share of CUDA adoption. Currently, ROCm only works with a few AMD products, while CUDA has worked with all Nvidia GPUs for years.

Other alternatives, such as UXL or different combinations of PyTorch and Triton, are just as interesting, but are still in their early stages. UXL looks particularly promising because it is supported by some of the biggest names in the industry. Of course, this is also its biggest weakness because the interests of these members vary greatly.

We believe that none of this matters if Nvidia can gain a foothold. That's where we need to differentiate CUDA from the Nvidia software ecosystem. The industry will be proposing alternatives to CUDA, but that doesn't mean they can completely remove Nvidia's software entry barriers.

That being said, the biggest threat facing Nvidia's software moat is its biggest customer. Hyperscale businesses aren't interested in locking down Nvidia in any way, and they have the resources to build alternatives. To be fair, they're also not immune to being closely linked to Nvidia, which is still the default solution and still has many advantages, but if anyone weakens Nvidia's software ambitions, it's likely from this corner.

Of course, this raises the question: What exactly are Nvidia's software ambitions.

Over the past few years, as Nvidia launched its software products, including its cloud service Omniverse, they conveyed a sense that they were ambitious to create a new component of their revenue stream. In their latest earnings call, they indicated that they have generated $1 billion in software revenue. Recently, however, we have a sense that they may be repositioning or scaling back these ambitions slightly, and the software is now positioned as a service they provide to chip customers rather than as a mature revenue unit in its own right.

After all, selling software could put Nvidia in direct competition with all of its biggest customers.

18 years of perseverance, in exchange for a big explosion

According to the “BBC” report, Nvidia's rapid rise is a bold bet on its own technology plus good timing.

Nvidia was founded in 1993, and the current Nvidia CEO Wong In-hoon was one of the founders. At the time, Nvidia was still a company focused on creating images for video games and apps.

In 1999, Nvidia developed the first graphics chip (GPU) GeForce256 to enhance the computer's image display effect. At the time, it was publicly listed at a price of 12 US dollars per share.

In 2006, researchers at Stanford University discovered that GPUs have another use: they can speed up mathematical operations, which ordinary chips cannot do. In the same year, Hwang In-hoon made an important decision in AI development: he invested Nvidia's resources to build a tool that allows GPUs to program, so that GPUs are not just tools for creating images, that is, CUDA (Compute Unified Device Architecture, Unified Computing Architecture).

What is CUDA? This is a set of programming tools provided by Nvidia to developers. It allows engineers to use CUDA to save a lot of time writing low-level syntax, and then directly use higher-level syntax such as C++ or Java to write algorithms applied to general-purpose GPUs to solve complex problems in parallel computation.

For researchers, this is a new approach to high performance computing on consumer hardware.

When Huang Renxun invested in CUDA 18 years ago, he was underestimated by many investors. The 2024 results now prove that this choice will result in a market capitalization of 2 trillion US dollars in the future.

According to Forbes, most AI innovations are based on Nvidia's CUDA platform. Nvidia's strategy is to make the platform spawn a huge software ecosystem, making it difficult for the latter to break through.

In 2012, Alexnet, an artificial intelligence that can classify images, came to the market. It used two of Huida's programmable GPUs for training.

At the same time, scientists discovered that GPUs can greatly speed up the processing speed of neural networks and begin to be used in work.

Nvidia leverages its strengths by investing in GPUs that are more suitable for AI and software that makes this technology easier to use.

However, while Nvidia's current dominance seems solid, it's harder to predict in the long run.

Kevin Krewell (Kevin Krewell), an analyst at consulting firm TIRIAS Research, pointed out that many opponents behind Nvidia are rushing to catch up.

For example, both AMD (AMD) and Intel (Intel) are famous for manufacturing central processing units (CPUs), but they have also recently invested in making AI-specific GPUs, and Google has developed tensor processors (TPUs) with AI computing capabilities, which can be used not only in search engines, but also in some machine learning tasks. Microsoft and Meta are also in the process of developing them.

Nvidia has also had success in cryptocurrency mining. According to the Wall Street Journal, as cryptocurrency prices rose, Nvidia surpassed chip giant Intel in 2020 and reached a record of nearly $330 per share at the end of 2021.

Although the cryptocurrency winter has recently arrived, analysts pointed out that the rise of AI will bring prospects for Nvidia to be more prosperous than cryptocurrencies. UBS analysts estimate that ChatGPT alone will require about 10,000 Nvidia GPUs.

Nvidia designs but does not manufacture its own chips; they outsource production work to chip makers, including “Foshima Kamiyama” TSMC.

In an exclusive interview with The New York Times in 2010, Wong In-hoon pointed out that Nvidia's core values include two elements, the first being “tolerance to take risks” and “the ability to learn from failure.” He mentioned that in this rapidly changing world, “celebrating failure” is an important element for all companies, and the second is “intellectual honesty,” which can bluntly point out mistakes made by companies or individuals, learn lessons from them, and make quick adjustments.

Hwang In-hoon said that Nvidia's corporate character is that if you have a good idea and no one has done it before, then just let go and try. If you fail, learn and adjust from it. Every failure is a little bit of progress.

When Hwang In-hoon was in high school, he also learned quite a bit while playing with electric cars. “When you try to break a level and don't get the results you want, you try again and before you know it, you break the level!”

Hwang In-hoon believes that playing with electricity has cultivated the ability to experiment, which is one of the nutrients for innovating and trying new things in the future.

In an exclusive interview, Hwang In-hoon also revealed the characteristics required to join Nvidia. The first one is “to be able to love something you are interested in”. He emphasized that the ability to fall in love with something is very important for successful people.

The second is the ability to “take risks and make mistakes”. Keep trying, learn from mistakes, and make adjustments to meet the next challenge.

The third one is “see the world like a child” and “what if we do it?” “How do I do that?” All great ideas are based on this.

edit/lambor