AMD – The Blueprint

AMD emulates Nvidia but has more to do.

• AMD is building a vertically integrated AI toolkit based on its silicon but there are two elements still missing that will still prevent it from really taking the fight to Nvidia.
• AMD will acquire Silo AI, a Finnish AI body shop that creates AI on behalf of its clients for $665m in cash giving AMD further vertical integration in terms of what it can offer on top of its silicon to its customers.
• This is in addition to Mipsology which AMD acquired a year ago and Nod.ai which was acquired 9 months ago.
• Both Misology and Nod.ai were long-standing partners of AMD and were developing tools, services and platforms that make it easier to deploy AI on top of AMD’s silicon.
• Both Mipsology and Nod.ai were small enough acquisitions that AMD did not have to disclose the price which combined with AMD’s low market share at the moment, leads me to suspect that neither of them was in very good financial shape when they were acquired.
• Becoming part of AMD gives them the financial runway to continue and also greatly helps AMD build a competing solution in this space to try and put a dent in Nvidia.
• Silo AI is another kettle of fish as it is a much larger business with a track record, a list of clients and 300 AI experts, 125 of whom are PhD scientists.
• These three businesses will all become part of AMD’s AI group whose task is to provide the development platform, tools and services to make it easy to use AMD silicon to create AI services.
• The remit is much wider than generative AI but expands to all branches of AI including deep learning which is already widely used across many industries.
• However, to challenge Nvidia, AMD needs to achieve two things.
• First, it needs a ubiquitous development platform like CUDA and second, it needs to up its product cadence such that it is in the market with the same generation silicon as Nvidia at the same time.
• None of these acquisitions will achieve these goals but crucially they lay the groundwork for AMD to build platforms and tools to compete with CUDA, Omniverse and so on.
• This is going to take some time and even though AMD announced that it is expecting $3.5bn in AI chip revenues this year, this market is now so large that this is a drop in the ocean.
• Nvidia’s run rate is now well over $100bn annually meaning that AMD will manage just 2.9% market share even if it hits this target.
• This is why I think that AMD and everyone else for that matter has a massive hill to climb while the control point in the AI ecosystem remains at the level of the silicon development platform.
• CUDA is so mature and so universal that I think that all of its competitors are going to have a very hard time making a dent in its position.
• However, this will not always be the case as there are signs that developers may shift from developing to silicon and switch over to using foundation models over time.
• OpenAI, Google, Microsoft and others already offer tools that allow developers to take their models and modify them without having to bother with silicon at all.
• The owners of these models have a vested interest in using silicon other than Nvidia and will ensure that their models are optimized for other offerings and in many cases will use their own.
• This is how AMD and others may finally put a dent in Nvidia, but it will be a long time before the market moves fully in this direction.
• Hence, for the next year or two, I think that Nvidia looks pretty secure with its biggest problem being how to ride out the correction in demand that must inevitably come when the AI bubble bursts.
• Nvidia looks expensive but it remains far more reasonably priced than almost all of the rest of the entire AI industry, and so if I was forced to take a position in AI, I would own it.
• However, I continue to prefer the lateral themes of inference at the edge and nuclear power (to run all the data centres) and am tempted by the largest AI ecosystem in the world that no one wants, Baidu.