$GOOG eats $NVDA

The whole Nvidia-backed AI surge has been tied to the release of ChatGPT and its explosive growth. OpenAI has observed the fastest growing monthly active users (MAUs) of any company ever. ChatGPT is what people think of when they talk about AI.

OpenAI with a first mover advantage has taken a significant market share in LLM API calls and chat bot use. With such a large market share there is only one direction that can go (down). However, if the AI market expands (I don’t think this much is controversial) and OpenAI can maintain a strong market share then it will set up to do very well in the coming decade.

The AI market has been defined by these scaling laws, if you merely increase compute, you will get better model quality. Probably the best example of this is Grok which very quickly surpassed OpenAI on most benchmarks by being bolder on compute.

If throwing more money at the problem is all you need to be better, then $NVDA which is providing the compute at scale is able to: - Increase its operating margin from ~30% to ~60% in just 2-3 years - 10x its revenues in a similar time period

GPUs are highly programmable and general. This is actually how Nvidia was able to quickly change from being a graphics company to an AI infrastructure company.

One argument is that, in something as quickly changing as AI, you need general purpose compute since you don’t want to lock in a hardware design and manufacture it only for it to become irrelevant very quickly.

But, what exactly is quickly changing? Every company is using similar approaches (transformers +. RL) and throwing more money at the problem (scaling laws).

I think what we are talking about when we speak of the rapid change in AI is the application layer. There is a new company being started (and going viral) every day with new ways to apply the current approaches.

TPUs are application specific integrated circuits (ASICs) and are probably a big deal in a world where the kinds of operations being done at scale is unchanging (I.e matrix operations). In addition they are:

- More scalable, can run workloads on more chips connected with higher bandwidth/throughput - Cheaper (better performance per watt)

Every single hyper scaler will build their own ASICs since they need to diversify away from Nvidia in the long run.

In the best case for Nvidia, these ASICs only slightly reduce their margins and market share in an ever growing AI pie.

In the worst case, Nvidia’s business is disrupted completely.

Even if Nvidia were to pivot and build their own ASICs they would be entering a highly competitive space (I.e. lower margins).