The big picture: Starting tomorrow, Nvidia will be hosting a GTC developer conference. Once a semi-final sideshow, this event has become the focus of much industry attention. Nvidia’s rise has led many to question the extent to which its software provides a durable competitive moat against its hardware. We’ve gotten a lot of questions about it, so we’d like to give our thoughts here.
Aside from the possible announcement of the next-generation B100 GPU, GTC isn’t really a chip event, it’s more of a developer show. This is his Nvidia flagship event for building a software ecosystem around CUDA and other parts of its software stack.
It’s important to note that when talking about Nvidia, many people, ourselves included, tend to use “CUDA” as an abbreviation for all software provided by Nvidia. This is misleading because Nvidia’s software moat is more than just his CUDA development layer, which is very important for his Nvidia to defend its position.
Editor’s note:
Guest author Jonathan Goldberg is the founder of D2D Advisory, a multi-functional consulting firm. Jonathan has developed growth strategies and partnerships for companies in the mobile, networking, gaming, and software industries.
At last year’s GTC, the company issued 37 press releases highlighting a dizzying array of partners, software libraries, and models. Expect more next week as Nvidia tightens its defenses.
These partners are important because today, hundreds of companies and millions of developers are building tools based on Nvidia’s products. Once these people have built their models and applications, they are unlikely to rebuild them to run on other companies’ chips, at least for the time being. Nvidia’s partners and customers span dozens of industries, and while not all of them are fully on board with his Nvidia, he still shows immense momentum in Nvidia’s favor. is noteworthy.
Simply put, the defensibility of Nvidia’s position at this point hinges on the inherent inertia of its software ecosystem. Businesses invest in software, writing the code, testing it, optimizing it, and educating employees on how to use it, but once that investment is made, they are very reluctant to switch software.
We’ve seen this with the Arm ecosystem’s datacenter migration efforts over the past decade. Even as Arm-based chips begin to demonstrate true power and performance advantages over x86, software companies and their customers are still years away from making the transition, and the transition is still underway. Nvidia appears to be in the early stages of building just such a form of software dominance. And if you can achieve that across a wide range of companies, you could potentially sustain it for years. More than anything, this means Nvidia is best positioned for the future.
There are formidable barriers to entry into Nvidia’s software. CUDA is a big part of that, but even if his CUDA alternatives come along, the way Nvidia provides its software and libraries to so many people is a very defensible ecosystem. It shows that you are building.
We point this out because alternatives to CUDA are starting to emerge. AMD has made great strides with his answer to ROCm which is CUDA. However, progress means that we now have a good viable platform, but it will take years for it to capture any share of CUDA adoption. ROCm is currently only available on a few AMD products, while CUDA has been working on all his Nvidia GPUs for many years.
Other alternatives are equally interesting, such as UXL and various combinations of PyTorch and Triton, but are still in their early stages. Especially since he UXL is backed by the biggest group in the industry, it looks promising. Of course, these members have very diverse interests, which is also their biggest weakness.
We would argue that this matters little if Nvidia can stick around. Here we need to differentiate between CUDA and the Nvidia software ecosystem. The industry will come up with alternatives to CUDA, but that doesn’t mean Nvidia’s software barrier to entry can be completely eliminated.
Also read: Farewell to graphics: How GPUs came to dominate AI and computing – it’s no longer “just” a graphics card
That said, the biggest threat to Nvidia’s software moat is its biggest customers. The hyperscaler has no interest in being locked into Nvidia in any way and has the resources to build alternatives. To be fair, they aren’t immune to staying close to his Nvidia, and Nvidia remains the default solution and still has a lot of advantages, but someone might choose his Nvidia If you’re going to have any influence on your software ambitions, it’s probably going to come from this corner.
And, of course, the question arises: What exactly are Nvidia’s software ambitions?
Over the past few years, Nvidia has signaled its ambitions to create new strands of revenue streams with software products ranging from its Omniverse cloud service. In their latest financial report, they noted that they had $1 billion in software revenue. But these days, software is being positioned more as a service to chip customers rather than a full-fledged revenue segment in its own right, giving us the sense that those ambitions may be changing or scaling back a bit. is obtained.
After all, selling software risks putting Nvidia in direct competition with all of its biggest customers.