The Full AI Innovation Pipeline

I’m defining the “Frontier AI Innovation Pipeline” as the complete set of activities with a new AI model that moves the industry’s needle.

And that can be roughly broken down into:

R&D.Model development.Training.Post training.Inference.

Indeed, the AI model pipeline is super complex and multi-faceted.

With the number of GPUs needed to do research and tests for the release of a new model.

Indeed, you have to, in general, look at a proportion of 1/5 or 1/10 in terms of GPU needed to train a model for the final release vs. R&D and tests before the model is complete.

For instance, DeepSeek’s case used 2,000 Nvidia H800 GPUs for training. That means the company had to have at least 10x that amount of chips for research, smaller-scale testing,

And then, there is the elephant in the room: inference.

In other words, even when the AI model is ready for deployment unless it’s open source and its weights can be downloaded and repurposed directly on an application or device, in all other cases, it will be served either via APIs or via an app interface.

In that instance, you will need underlying server infrastructure to serve these models.

And especially on the reasoning side, that can be hard to sustain at scale!

Thus, again, when considering moats on the frontiers of AI, all the above are critical factors to consider as each comes into play.

In addition, just like Google started as a single search engine, then vertically integrated to have a specialized infrastructure to serve its whole suite of applications, a player like OpenAI and Anthropic, over time, will go through a similar path to lock in massive moats.

Google today is among the major players in the data center space.

The massive Google facilities to serve internal AI workloads for Google Image Credit: Semyanalysis

The competitive moat isn’t just in the model itself, but instead on the infrastructure you build out of it, combined with the locked-in market shares of an industry that is still all over the place, as winner-take-all effects haven’t consolidated yet but will, at due time!

And so far, we haven’t even accounted for the most significant expense for these frontier AI players: talent!

The post The Full AI Innovation Pipeline appeared first on FourWeekMBA.