Nvidia & Reflection – Health Insurance pt. II

Another move to secure its future.

• Whether it is through funding or acquisition, Nvidia’s strategy to ensure that its leading position in AI hardware and software is maintained continues to evolve as CUDA becomes less relevant.
• Reflection AI is a half-open AI model company that provides the weights of the models that it creates but stops short of making them all available.
• This is the equivalent of offering a cake for sale and letting clients put whatever icing or decorations that they want, but does not let them change the recipe of the cake itself.
• A fully open-source model provides everything and allows users to change the cake itself to whatever requirements that they have.
• By providing the model and the weights, clients can download the model and run it in their own infrastructure, which is the ideal use case for enterprise and sovereign clients, most of whom will want to keep their data within their own infrastructure.
• This market is currently dominated by China, where almost all of the models are made freely available to anyone who wants them, which, combined with their good performance, has led to a lot of uptake outside of China.
• Reflection makes money by selling the cake (which it just upgraded to a 501bn parameter mixture-of-experts model called Beam) as well as the tools that go with it to help clients customise the model to their specific datasets and requirements.
• Nvidia has long been a major investor in Reflection and is now in talks to either put more money in or buy the company outright.
• An acquihire makes much less sense here because the value of the company to the market and to Nvidia is through the product rather than the engineers.
• Reflection’s appeal is that it offers models that are competitive with the best of what China is producing but coming from a Western company, which will make some clients feel more comfortable.
• Beam is significantly better than Nvidia’s Nemotron and competes very effectively with Chinese models that are much larger, making it a better model for Nvidia to champion.
• Nvidia has risen to dominate the AI industry thanks to being first, releasing more advanced products much faster than anyone else and the CUDA development platform that everyone knows how to use.
• However, CUDA is far more relevant for training than it is for inference and with the vast majority of the market going to inference, CUDA’s lock-in is fading fast.
• Nvidia has been well aware of this for a long time, and products like Dynamo, Nvidia Micro-inference services (NIMs), AI Foundry, and so on are all directed at mitigating the impact of CUDA’s fade.
• The AI ecosystem is still very immature, and looking at history, the biggest winners are those that own the ecosystem with the touch point with the user such as Apple, Google, Tencent and Microsoft in PCs.
• Nvidia is employing a two-pronged strategy, which is to develop products and services that are optimised to run on Nvidia silicon and, at the same time, ensure that no one else becomes dominant.
• Apple, through its dominance of its Apple ecosystem and its skill in chip design, has been able to become fully vertically integrated from top to bottom.
• If the same thing happens in AI where a model maker becomes so dominant that everyone starts using its in-house silicon, then Nvidia’s business model is in real trouble.
• This is what I think Nvidia’s open-source efforts are all about and lie in ensuring that open platforms like Hugging Face and now Reflection AI are successful enough to prevent anyone from becoming truly dominant at the top of the technology stack.
• That way it can continue to compete with its AI products as well as its product cadence to maintain its dominant position.
• I have no doubt that Nvidia is going to lose market share over time, but its strategy and Jensen’s skill in predicting the long term are going to ensure that it remains dominant.
• Hence, I am confident it will meet its targets for next year, meaning that the shares remain very attractively priced.
• I continue to consider Nvidia.