For 15 years now, we’ve had a SaaS movement where some of your defensibility was just the maturity of your software product and the fact that as you become more feature-rich it is difficult for a competitor to take business away from you. But as AI makes software easier and easier to create, it will change the build vs buy decisions of large corporations, and increase competition at the application layer as more people can offer solutions that are slightly differentiated. Lower costs mean you can serve smaller more niche markets profitably. The process, along with the benefits and challenges, is well described by my friend Eric Koziol in his post on personalized software in Embracing Enigmas. (View Highlight)
By the “top of the stack” I mean agents. While I admittedly don’t have a strong thesis yet on who wins in agents, and why, I do think the agent market will be an oligopoly of a few platforms. Those platforms will have great economics, much like mobile app stores or other application marketplaces do today. (View Highlight)
By the “bottom of the stack” I mean AI chips and the low level infrastructure software that runs them. There is an explosion of innovation at this level that has received several billion dollars in financing over the past few years, and is still nascent, generating maybe 300M−500M in revenue in total across all the non-GPU AI hardware companies. (View Highlight)
I think the TAM of AI hardware, that people are estimating at 400B−500B by end of the decade will be 5-6x that. It will be several trillion dollars. Why? Because AI is effectively taking human tasks and pushing them down through the stack to become compute workloads. And the workloads are so different for so many AI use cases that I believe several AI chip companies will be wildly successful. (View Highlight)
Now, one word of caution - there are always some specific opportunities at every level of the stack. There will be a few software markets that don’t get squeezed because not enough people know how to define the requirements of the software to effectively instruct AI how to design it. And there will always be some pockets of opportunity with unique and unusual and differentiated data sets. (But I think in the case of data sets we are overestimating how many truly unique and defensible ones are out there.) (View Highlight)
So what I believe is, you want to be at the agentic level close to the customer, or at the hardware/infra level close to the computation. And markets in between will be squeezed on profitability. Thus the “hourglass” figure of AI. (View Highlight)
