Nvidia – Bear Attack

Nvidia is the subject of pretty baseless market chatter.

• The knives are out for Nvidia, with press stories and social media all getting excited about something that has been well known and priced into the shares for some considerable amount of time.
• Reuters is reporting (see here) that OpenAI has been seeking alternatives to Nvidia (well known) and that Nvidia had to give OpenAI a 30% discount in order to prevent it from going elsewhere (dubious).
• This has lit up social media with all of the bears coming out of the woodwork to call time on Nvidia’s dominance of the AI industry, but reality paints quite another picture entirely.
• Nvidia is almost certain to lose market share, but demand for AI silicon is currently so strong that this does not meaningfully affect the investment case for the shares.
• Nvidia currently has around 85% market share of the market for merchant market silicon for AI data centres, which it has been able to build with 2 key differentiators.
  • First, CUDA: which is the development platform that a creator of an AI model will use to access Nvidia’s chips.
  • Unlike all of its competitors, CUDA has been around for about 25 years and over the last 12, it has become the standard for developing AI algorithms.
  • This means that there is a large and vibrant ecosystem in the industry that knows how to use CUDA, which on its own serves as a strong lock-in and a reason to buy Nvidia.
  • However, CUDA’s appeal is far stronger in training than it is in inference, and as the market is now rapidly moving to inference, there is a threat to Nvidia’s dominance, which is where the second differentiator comes in.
  • Second, product cadence: where Nvidia is routinely at least one generation ahead of everyone else.
  • This is an advantage that is fairly routinely ignored, but as CUDA’s lock-in weakens, it becomes increasingly important.
  • The most advanced AI system available today is, without question Nvidia’s Vera Rubin, which is currently in production but has yet to really make it into the market in real volumes.
  • Vera Rubin is a large step up from Grace Blackwell, which is the current generation, and it offers a 66% increase in performance at FP4 while increasing power consumption by around 40%.
  • Vera Rubin is the new benchmark against which Nvidia’s appeal should be judged, but while it is not in the market in volume, competitors and critics will continue to compare themselves against Blackwell.
  • Vera Rubin improves the economics of compute and while everyone else is running to catch up, Nvidia can make the case of being more economical to run even at 75% gross margins.
• Nvidia has long been aware of these competitive threats and has moved to address them with the introduction of Nvidia Inference Microservices (NIMs), Dynamo and the effective acquisition of Groq, which I think will have some effect, but some market share loss is inevitable.
• However, the market is growing so quickly at the moment that even with its customers diversifying with other suppliers, Nvidia is still struggling to keep up with demand.
• The signals coming from Micron, Samsung, SK Hynix and TSMC are that demand continues to be very robust, and so I very much doubt we are about to witness an implosion of demand for Nvidia’s silicon.
• In the face of this demand, I find it very unlikely that OpenAI has the market power to demand a 30% discount, as there will be plenty of other takers in the queue to buy Nvidia’s products.
• Hence, I think that the noise surrounding Jensen’s comments at the weekend and the Reuters article is just chatter and speculation, possibly amplified by those holding short positions in the shares.
• In this environment and given where Nvidia’s valuation is, I remain fairly indifferent to the shares, but if one has to have a direct investment in AI, this is, along with Google is one of the better ones to choose.
• That being said, I continue to prefer my adjacencies of nuclear power (which is doing very nicely at the moment) and inference at the edge, which requires a bit more patience.