How Open Source Models Will Erase AI Economics
The table below is a bit of an eye test, but what it shows is that EV/FCF multiples are not cheap.
The risk to the chip, cloud and frontier lab companies is open source models, as the economics largely go away under open source.
The open source model companies are primarily Chinese companies. American chip companies do not enjoy the same market share in China - a negative for NVDA, AMD, AVGO and MU.
Similarly, companies can host their own open source models without having to rely so heavily on AWS, Azure and Google Cloud to serve up those models.
When we talk about Anthropic and OpenAI versus open source models, it is important to remember that Anthropic and OpenAI have built harnesses, similar to what we have done with Kilby. Claude and GPT are base layer models plus tools. The open source models lack the tools layer, so depending upon what tools you may need, you would have to build the tools yourself should you choose the open source model route.
Bloomberg, Factset and Refinitiv ought to re-write their codebase to incorporate open source models (essentially build APIs off of every product so that models may ingest the data).
For example, if Kilby had a database of 10,000 EPS call transcripts, Kilby would train an open source model on those transcripts and replace Anthropic for Kilby’s /debrief function, which today costs the user $0.25 per query, and could take that cost down to effectively zero.
In a perfect open source world, Kilby could take AI query costs down to zero, and would only charge users a small fee for file storage, which would effectively make Kilby a low cost provider of hyper-personalized, frontier-level intelligence (today for Investment Management and general queries, in the near future Kilby will house other Industry-specific functionality).
Bloomberg, Factset and Refinitiv will not be able to follow Kilby on our low-cost journey because it would cannibalize their business and destroy each company’s revenue line.
So hang tight, Kilby has a bright future with open source models as do other AI-native companies. Incumbents do not.
By the way, here is a copy of Kilby’s debrief output for META’s most recent EPS call complete with Bull Signals, Bear Signals, Deflections and Follow-Up Questions: