AI Ecosystem – Capital Bloom.

Nvidia can sleep well for now.

• Competing with Nvidia is currently focused on building better chips but this is ignoring Nvidia’s two key differentiators which mean that for now, no one is going to lay a glove on Nvidia.
• There is almost unlimited capital for investing in AI startups at the moment and outside of large language models (LLMs), the biggest area of focus is on the silicon that is used to train them.
• All of the large cloud companies who have been forced by their clients to allow Nvidia to reside in their data centres are working on or have already created their own chips and there is a rapidly growing number of start-ups all vying to enter this space.
• One of the oldest start-ups is Graphcore which was initially used with Google’s TensorFlow to train GPT-3 but has since lost out to Nvidia at both Microsoft and Google.
• This should serve as a cautionary tale of just how hard dislodging Nvidia is going to be, but everyone seems to be rushing in regardless of the warning signs.
• The latest is South Korean start-up Rebellions which has secured a Series B round of $124m at a valuation of around $650m from a range of Asia (ex-China) investors.
• Others include Groq, Graphcore, Cerebras, and Tenstorrent to name but a few and in this climate, the list is likely to grow as there is plenty of money willing to bet on anything that is AI-related.
• The problem is that these companies appear to be focused on hardware performance and outside of very specific use cases, this is unlikely to be enough.
• Cerebras is a good example of a specific use case as its designs have massive dies to deal with the problems of inter-chip communication and as such, it can offer very high-end performance.
• The problem is that it does not support the industry-leading (and now standard) development platform, CUDA which is owned by Nvidia and only runs on Nvidia silicon.
• This means that Cerebras’ appeal is limited to very specific and high-end applications where performance is so much more important that developers are willing to forgo the advantages of CUDA to access better performance.
• Outside of this, the newcomers need to develop a software development platform as good as CUDA which is no mean feat given its 20-year head start.
• Furthermore, competitors also need to deal with Nvidia’s product cadence where it regularly releases new products or updates every 6 months.
• This is a pace that is almost unheard of in the semiconductor industry and means that Nvidia is at least 1 or 2 generations ahead of its nearest competitors pretty much all of the time.
• This is something that even the best-funded startups are going to struggle to match given that their resources are far more limited making them unable to throw the kind of resources at this that are required to iterate products this quickly.
• The only thing that I can see on the horizon to challenge Nvidia’s dominance is a change in the control point for how generative AI services are developed.
• This is something that Open AI has already launched, and I am certain that Google and Meta will soon follow.
• The idea is to use their foundation models combined with an SDK to make it much easier to develop generative AI services.
• If this is successful, then developers will care less about which silicon is being used to train their services and more about which foundation model and SDK they are using.
• The owners of foundation models are far less wedded to Nvidia’s silicon and in most cases are actively developing their own or broadening the support of their foundation models to the chips from other silicon vendors.
• This is likely to some time and so in the meantime, there is no viable challenge to Nvidia that I can see.
• However, for those looking beyond 2025, this is something to be aware of as it is the one thing that could break Nvidia’s grip on the market.