We have been thinking a lot about strategic positioning in the last three quarters amid the “hype” that followed Microsoft’s sizable investment in Open AI. Leaving aside the competitors for the moment (which would require a dedicated post) we have studied three “AI market models”: two come from VCs and one from a top-tier consulting company.
In May 2023, Andressen Horowitz (A16z) published its Preliminary generative AI tech stack to help visualize where Generative AI will accrue value in the market. Of the three we studied, the A16z market model is the simplest and the most “foundational”, as it were. Its simplicity matches the tone of the mentioned post, i.e it is too early to know what’s going to happen, but “there is a tremendous amount of value that will be unlocked”. A16z highlights the fact that companies training the Gen AI models are not – for now – the ones capturing most of the value. Instead, Nvidia share price is up by almost 200% so far this year (as of August 2023).
The McKinsey market model is the most sophisticated – as you would expect from the top strategy consultancy out there. There are two main differences if we compare it with the A16z one. First, it places more emphasis on the “tooling” elements. Model Hub, Prompt Library, MLOps platform, Policy Management etc. are “key elements that need to be incorporated into the technology architecture to integrate generative AI effectively”. Second, it introduces the “data” elements and the importance of developing a data architecture to enable access to quality data.
It is fair to say that in the McKinsey market model, the foundational models are still at the center but resized, compared to A16z. The McKinsey market model is targeted to CTOs and CIOs of big companies, which have tons of proprietary data that can be used to train their Gen AI models. The bottomline is “make your model unique (and fit-for-your-purpose) to gain a competitive advantage”.
The third market model comes from an All-in podcast my dear cofounder posted in our group chat! Craft Ventures put it together as part of the investment memo for MosaicML – the venture capital firm was about to chip in but did not go through due to the acquisition of MosaicML from Databricks for $1.3B.
In case you don’t know, MosaicML sits in the MLOp / Open Source Foundational Model elements of the A16z’s and McKinsey’s market models. The slide is very interesting because it shows the end-to-end tool chain needed to build “company-intranet” chatbots (aka copilots). David Sacks, Craft Ventures founder, says that “every enterprise would like to roll out their own ChatGPT, where their employees could ask questions and get answers. But they don’t want to share their data with the likes of Open AI”. The Craft Ventures-MosaicML market model reinforces McKinsey’s one by zooming in on the importance of 1) “Data Capture” (that is in MosaicML Roadmap), 2) “Data Labeling” and 3) “Safety Guardrails & Constraints”.
We think “Data Capture” can refer to both proprietary and Third Party data (it is worth noting that none of the three market models were putting much emphasis on the Third Party data topic). Regardless of whether open source AI models will prevail against the closed ones, ultimately models will differentiate only by the uniqueness and quality of the data they are trained with. Models trained with external data will have more “insights” than those trained with proprietary data only. Ultimately, your custom-trained models will be scrutinized by the AI regulator(s), which will search for potential bias, so better you have your compliance ready and battle-tested”.