The LLMops Infrastructure Landscape

Over the last few weeks, as I've worked on Embedding Sync and chatted with lots of folks in the space, I've been changing my mind on the market opportunity. I've come to the conclusion that at this point in the cycle LLM wrappers or applications may actually be a better opportunity than LLM ops, infra, and tools:

  • Things are changing so quickly that building your product around an architecture or pattern that might not be relevant in 3 months is rough.
  • The engineer's tendency to prematurely optimize and build the general abstraction is leading to everyone in the space going for the infra or framework opportunity. There might be more libraries than applications right now. At some point you have to go and do something with all the picks and shovels.
  • Most of the money in VC (and YC) is going to these sort of companies since the few that work will capture a lot. But that means right now it's extremely saturated and anything you build can be copied quickly by a VC-backed company.
  • The few companies that are building proper AI applications right now are already extremely technical and are less likely to buy a bunch of point solutions. They're more likely to build custom solutions in house, and even discover new patterns before infra or tool companies.

I'm glad these companies exist, and I'm working on a service in this space in my spare time (Embedding Sync) - so this is more of my reflections from investing in the space and hearing from others.

Long term, there will definitely be huge success stories in the space (I really like what LangChain is doing with LangSmith), but the market opportunity for building the marginal LLMops company seems much worse than working a layer above it on consumer or industry application to problems.